Posting my regular quarterly channel check & alternative data report before we start the big tech Q3 earnings.

A quick look back first, because I think it is important to keep score. In my Q2 report, the expert interview data showed AWS as the standout in rate of change, with mentions of “best momentum” jumping from 27% to 42%. Based on that and my Anthropic math, I wrote that AWS could land in the 33-36% range (ex FX) vs. the Street at ~31%. AWS printed 36.7% YoY, its fifth straight quarter of acceleration and its fastest growth in 18 quarters. The signal was right; if anything, I was a bit too conservative on the magnitude.

This quarter, the same data is telling a somewhat different story.

In this report, I cover:

1. Cloud hyperscalers: what my expert interview momentum tracker shows for Q3

2. The AI orchestration layer: Microsoft vs. the LLM providers vs. Google and Amazon, and the AWS Bedrock vs. Microsoft Foundry race in job postings data

3. AI adoption signals: Copilot is accelerating, while Anthropic’s adoption data is starting to level off

4. Ad tech and Meta: what I expect for Meta and what worries me

5. My views heading into earnings for Microsoft, Amazon, Google, and Meta

Before we start with the article, I would like to announce that, together with a great alternative data provider team, we built and launched QAlpha AI. QAlpha AI is an alternative data platform with a focus on the Cloud, AI, and SaaS sectors. It extracts alternative data, including hiring data, adoption data from over 20k companies, infrastructure signals for more than 1700 products like new subdomains, DNS records, certificate transparency logs, and data center buildout data. We track all the big tech names Microsoft, Google, Amazon, Meta on top of hundreds of mid-sized and smaller companies in this sector. You can also track things like Claude, Codex, ChatGPT, Copilot, AWS Bedrock, and much more.

All the data crawled is public data, compliant, and does not bypass any CAPTCHA or circumvent authentication. The product is also priced very attractively compared to similar alternative data platforms, with what I believe is higher-quality and better-extracted data.

You will see me use much of the data from QAlpha in this article and in future ones, since it’s built to extract valuable insights for investors like me. Be sure to check it out; we also offer a 3-day free trial, and for the next few days since launch, we’ll offer a one-time 20% discount on the first year of usage.

Check Qalpha AI

Now let’s dive into the article.

Cloud hyperscalers

First, the backdrop, because it has not changed: cloud capacity is still sold out, and if anything, the sold-out window is getting longer.

On the All-In podcast last week, Ben Horowitz relayed a conversation with Andy Jassy that he had just a few days ago. Horowitz told Jassy he assumed the industry was out of capacity until 2028, and Jassy corrected him: »Oh, it’s beyond 2028. End of 2028, maybe.« A few days later, in Alex Kantrowitz’s interview with Sundar Pichai posted yesterday, Kantrowitz asked Sundar about the exact same comment. Sundar said that a customer »asking us for compute in ‘27, that is incredibly difficult to do«. He also admitted Google had to lease short-term compute last year just to accommodate some customers.

When supply is the constraint, cloud revenue growth is not a demand question anymore. It is a function of two things: how many megawatts each provider can bring online in the quarter, and what price they can charge for them and their existing fleet. On the second one, the data is clear. AWS raised reserved GPU pricing by 15% on October 7. That is the fourth consecutive quarter of price hikes and brings the cumulative increase this year to 40-75%.

Cloud compute is sold in roughly three ways: on-demand (pay by the hour, most expensive), spot (spare capacity at a discount that can be taken away), and reserved (you commit to 1-3 years in exchange for a lower rate). Reserved is the price large enterprises actually sign on. When reserved prices go up 15% in a single step, it tells you the provider has more buyers than it has capacity, and that is exactly the type of environment where margins expand.

The expert interview channel check

Now to my traditional expert call channel check. The results this time were quite surprising in some sense. I should note that the sample of high-quality expert interviews was smaller this quarter than in previous ones, so I would weigh the individual moves a bit less than usual.

In general, most experts noted a continued modest acceleration in demand for cloud services vs. previous quarters across the board. Nonetheless, the results become very interesting when we look at which platform had the best momentum:

The standout is by far Microsoft Azure. Mentions of Azure as »growing the fastest« jumped from 38% last quarter to 62% now, a 24-percentage-point move in a single quarter and the biggest swing I have recorded in this series. As before, keep in mind that the numbers add up to more than 100% because many experts name two providers as fast-growing based on their observations.

On the other hand, AWS saw the most significant dip, going from 42% to 31%. GCP slipped from 50% to 46% and has now declined for three straight quarters (65% → 62% → 50% → 46%).

A few thoughts on how I read this:

• This is a rate-of-change signal, not a growth ranking. Last quarter, the AWS jump from 27% to 42% correctly flagged an acceleration, and AWS delivered 36.7%. This quarter, the Azure jump tells me the experts see the biggest incremental change at Microsoft. It does not tell me AWS is decelerating; as I expect AWS to also experience significant growth and acceleration, but the capacity might have gone to other places than the companies that these experts work for or consult for. At some point, maybe even this quarter, Microsoft will probably be the one who will have a rolling long-term contract with lower compute prices from the start of last year and get repriced soon, as they were the first to give and build significant capacity for OpenAI in the 2023 days. The question is whether they give that capacity to OpenAI or to other clients in the market with less negotiating power.

• GCP’s slow fade shoudn’t be read as anything to be honest as the changes are too small for it to be incremental in my view.

More on my concrete views from these readings and in general on Amazon, Microsoft and Google in the last section.

For context, here is where the three actually ended Q2. Note that Google Cloud’s 81.8% includes more than just GCP (Workspace and Gemini subscriptions are in there too):

What the Morgan Stanley CIO survey says about cloud share

The Morgan Stanley Q3 2026 AlphaWise CIO survey, published on October 12, had an interesting data point on current public cloud usage. Azure still dominates with 79% of respondents saying they use it, ahead of AWS at 66% and Google Cloud at 43%. But Azure has dipped slightly over the last two years, while both AWS and GCP gained this year in terms of % of respondents.

Another interesting part of the Morgan Stanley Q3 CIO survey was how fragmented the market has become in terms of who CIOs prefer for their AI orchestration layer. Microsoft is the only vendor with a higher % of answers (20%) than the LLM providers collectively (17%). Google got 5% and Amazon 4%.

The survey also asked which vendor will capture the biggest incremental share of GenAI spending:

It is important to note that these CIO’s are more inclined towards Microsoft products than the normal state of companies is; nonetheless, it is interesting to see how dominant Microsoft is in the view of being the orchestration layer vs. the other two cloud providers.

One more data point from the survey that I think is underappreciated: 42% of CIOs said they cannot yet measure the ROI on their AI token spend, 28% said returns are below what they spend, and only 13% are ahead. As a result, 43% already have token spend controls in place, and another 31% plan to have them by the end of 2027. That is roughly three out of four enterprises actively trying to control what they spend on tokens. In my view, this is a tailwind for the orchestration layer (routing cheap models to easy tasks is exactly what Foundry and Bedrock sell) and a headwind for the price per token at the model layer.

AWS Bedrock vs. Microsoft Foundry in job postings

It’s also interesting to look at company job postings to see how AWS Bedrock and Microsoft Foundry compete, as both are regarded as key AI orchestration layers. A company that writes “Bedrock experience required” into a job posting is telling you where it is building for the next few years.

Bedrock was leading last year and up to May this year. In January, 0.316% of new job openings mentioned Bedrock vs. 0.259% for Foundry, a ~22% lead. Bedrock peaked at roughly 0.41% in March. In the summer months, Foundry searches were on par with Bedrock’s, at times even slightly higher, with both in the ~0.33-0.35% range. Now, in October, Bedrock has again taken the lead at roughly 0.38% vs. ~0.34% for Foundry.

The takeaway for me: this is a two-horse race that is much closer than it was a year ago. Also note that new frontier models now launch on all three platforms on day one. Claude Fable 5.1, for example, was available on Bedrock, Google Cloud, and Foundry the same day it launched on September 1. The model is no longer the differentiator for the platform; the harness, the integrations, and the existing customer base are.

AI adoption signals: Copilot accelerates, Anthropic levels off

Copilot is pulling away

Microsoft Copilot has been leading the “assistant race” in job postings ever since mid-last year, but in the last 90 days, it has accelerated the most of any provider (the blue line below).

Some numbers to put this in perspective:

• Copilot is now mentioned in 1.76% of all new job openings monitored by QAlpha. At the end of August 2025, it was 0.377%. That is a ~4.7x increase in about 14 months.

• Over the last 90 days, Copilot went from roughly 1.3% to 1.76%, a ~0.45-percentage-point gain. Claude added roughly 0.1-0.15 points over the same period (to ~1.33%), while ChatGPT spiked and then fell back to ~1.2%.

• Gemini sits at roughly 0.5%, up from 0.351% in August 2025. That is only ~1.4x in 14 months, by far the slowest growth among the major assistants.

This lines up with what Microsoft told us on the Q2 call: 30M+ paid Microsoft 365 Copilot seats, net seat adds more than doubling QoQ, and Copilot revenue up over 60% QoQ. When companies start writing a tool into their job descriptions, it means the tool has moved from a pilot to a workflow.

But the important takeaway from this data for me was that in the last 90 days Copilot was the fastest grower even compared to OpenAI Codex or Claude, which is not something I think the market believes, so the Copilot numbers for Q3 might surprise people.

Microsoft is hiring again

Microsoft has also significantly accelerated hiring this quarter, with new job openings up 71.6% QoQ to 2,272. That is still 30.9% below last year’s level, so this is a rebound from a low base rather than a hiring spree, but the direction is clear. Also, looking at the roles of these hires, there is a visible uptick in sales roles, which might suggest Microsoft has finally found the product-market fit with Copilot and wants to market it aggressively now. Interestingly, Meta shows a similar pattern (+68.4% QoQ, -23.6% YoY).

Anthropic is starting to level off

On the other side, Anthropic is starting to level off when looking at how many job postings mention it.

Anthropic is now mentioned in 0.271% of new job openings (84-day average), up 311.1% YoY. So the absolute level is still growing strongly on a yearly basis. But the share of hiring companies that mention Anthropic (the blue line) has flattened at roughly 0.7% since late summer, after going nearly vertical from March to August.

A similar trend is seen when looking at the adoption signals that QAlpha measures, which count net new companies adopting each tool per month (adoptions minus churns):

To put the scale in context: in October 2025, out of the companies that QAlpha tracks Claude added a net 203 companies in a month. By April 2026, that was roughly 16,000 per month. Based on this data, Anthropic’s Claude adoption peaked in April this year. It then fell to roughly 10,000 net adds in July. Part of that trough likely overlaps with the June 12 to July 1 suspension of Fable 5 and Mythos 5 access due to the U.S. export controls, although the decline had already started in May. It regained ground in September (roughly 13,800 net adds), as both Fable 5.1 (September 1) and Opus 5.5 (later in September) launched, but it is now trending down again, to roughly 9,800 in October.

That down pattern is consistent with other services that QAlpha tracks, such as ChatGPT and HuggingFace. HuggingFace, for example, peaked at roughly 6,100 net adds in May and is now at around 1,300.

How I read this: it does not mean Anthropic’s business is shrinking. Net adds of ~10,000 companies a month are still enormous, and existing customers are likely consuming far more tokens per company. What it does suggest is that the land-grab phase of enterprise AI adoption is maturing, at least for the vertical of coding. The companies that were going to try Claude have largely tried it. Growth from here depends more on expansion within existing accounts and going after new verticals, which is exactly where the orchestration layer, seat-based products like Copilot, and cost controls start to matter. It is also a data point to keep in mind for anyone looking at the upcoming Anthropic and OpenAI IPOs.

Ad tech and Meta

In terms of the ad sector, my channel checks point to

Read more

For most of the last three years, the way the market valued AI was simple: whoever has the best model on the leaderboard wins, and the value flows to them as margins this year for frontier labs have gone to 60,70 or even 80%. I think we have now crossed a threshold where that framing is wrong for the majority of the economy.

For most knowledge work and most personal use, the improvement a user actually feels no longer comes from a better raw model. It comes from the harness, the integrations, the context the model gets, the alignment with what the user wants, and the product wrapped around it. The model is turning into one component of the product, not the product itself.

At the same time, competition at the model layer is getting brutal. OpenAI and Anthropic cut prices on the same day in September, open-weight models now carry the majority of tokens on some of the biggest gateways, and enterprises are routing work to whatever model is good enough and cheap enough. That is compressing margins at the model layer, and in my view, that value is spilling into two places: the data center layer (the hyperscalers and neoclouds) and, increasingly, the application layer.

In this article, I cover:

1. Why we have reached a good enough level of model capability for most work, and what that means for where improvements come from

2. What heavy token consumers like AlphaSense and Vercel are actually seeing in production

3. Why Meta’s Muse is the best consumer proof point of this shift

4. The model layer price war, with data from Ramp, Vercel and the latest price cuts

5. Why low-margin model tokens are a gift to the hyperscalers, and why negotiating leverage is the part most investors miss

6. Why I think we are entering an explosion of the AI application layer

Before we start with the article

Many of the insights from this article are from my recent conversation with Andy Hock (Cerebras Head of Strategy), Chris Ackerson (SVP at AlphaSense) and Kyle Cheng (Former Anthropic employee), which you can listen in full here:

Listen to conversation

Let’s dive in.

We have reached good enough for most of the economy

Think about how a company staffs a call center. Nobody hires a math PhD to handle refund requests, update a customer’s billing address, or qualify an inbound sales lead. They hire someone competent, train them on the company’s processes, give them access to the right systems, and measure them on resolution rate. Raw IQ above a certain bar adds almost nothing to that job. What adds value is knowing the policies, having the customer’s history in front of you, and being allowed to actually issue the refund.

AI models have now crossed that bar for a very large part of knowledge work. A few data points:

• The jump on real work has already happened. On OpenAI’s GDPval benchmark, which grades model deliverables against work from industry experts across 44 occupations, GPT-4o scored only 13.7% (wins or ties vs. the human expert). About 15 months later, GPT-5-high was at 40.6% and Claude Opus 4.1 at 49%. One leaderboard tracker now shows the top model at 87.9% as of September.

• The top of the leaderboard is crowded. On the Artificial Analysis Intelligence Index, Claude Fable 5.1 is at 66, Claude Opus 5 sits at 63, and Muse Spark 1.3, GPT-5.6 Sol and Grok 4.6 are tied at 61. Open-weight models, Kimi K3 and GLM 5.3, score 60, just three points behind Opus 5. Six index points now separate one of the best models in the world from some of the best models you can download for free.

• The cost gap is much wider than the capability gap. Muse Spark 1.3 costs about $0.55 per Intelligence Index task vs. $3.69 for Fable 5.1. That is roughly 6.7x cheaper for a model that is 5 points behind ( I even left out the more compelling contributor pricing, which would extend the gap even more). For a customer service workflow, that 5-point gap is hard to justify given the 6.7x pricing difference.

The GDPval paper found that more reasoning effort, more task context and more scaffolding all improve model performance on the benchmark. Two of those three levers have nothing to do with the model itself, but are the harness.

Andy Hock, Head of Strategy at Cerebras, put this well in the Token Economics conversation I hosted in September. Frontier models are like broad general experts: they handle complex reasoning, coding, and chat all in one. Where Cerebras sees enterprises leaning into open models is where domain specificity is required. In other words, the generalist PhD is still valuable for the hardest, most open-ended work. But most of the economy runs on specialists who know one domain very well.

This is also why I think we are sitting on a massive product overhang. The capability of the models that exist today is well ahead of what products are actually extracting from them. The bottleneck has moved from “can the model do it” to “has someone built the system that lets the model do it reliably, cheaply, and with the right data.” The next leg of improvement for most users will come from filling that overhang and less about what the next frontier model’s raw performance will be like.

The harness is becoming the product

For the non-technical readers, a quick definition. The harness is everything around the model: the system prompt and instructions, the tools the model can call, the retrieval system that pulls the right documents in, memory, the router that decides which model handles which request, the evals that check the output, and the guardrails. If the model is the engine, the harness is the rest of the car: the transmission, the steering, the brakes and the navigation system.

I wrote in May about how Cursor took its coding agent from Top 30 to Top 5 on Terminal-Bench 2.0 by only changing the harness. Since then, the evidence from heavy production users has only gotten stronger.

AlphaSense: 35 trillion tokens a year and no single winning model

AlphaSense is one of the biggest token consumers in enterprise software. Chris Ackerson, SVP of Product at AlphaSense, told me in our Token Economics conversation that the company consumes around 35 trillion tokens per year. For context, Google said that over the past 12 months, only about 375 Google Cloud customers each processed more than 1 trillion tokens. AlphaSense runs at roughly 35x that threshold.

Three things he said stood out to me:

1. Routing and post-training beat the frontier. Model routing and post-training their own models are two of AlphaSense’s most important areas of investment. Chris said that combination delivered about 3x better quality, at a lower cost, than applying frontier models to an unoptimized system with their content. AlphaSense has publicly shared that routing each question to the right model alone lifted answer quality by up to 2.8x over the baseline.

2. No single model wins across task types. Not too long ago, the consensus hypothesis was that models would converge, that they would all become roughly the same. AlphaSense’s testing found the opposite. Because each lab focuses its post-training on different things, models are diverging in what they are good at. That is exactly why AlphaSense needs to route different models to different tasks.

3. Context is now the bottleneck. In Chris’s words:

That third point is, in my view, the most important one. If the bottleneck is context, then value accrues to whoever owns the context (proprietary data, permissions, workflow position) and whoever owns the system that delivers it to the model. Not to whoever has the model with the highest benchmark score.

Vercel: everyone is multi-model now

Vercel’s AI Gateway sits between applications and model providers, so it sees which models are used in real production traffic. Vercel CEO Guillermo Rauch has said in the Astra Vercel summit that when a team adopts AI, it always adopts multiple models, and the average enterprise adopts over five models. That lines up with an F5 report finding organizations rely on an average of seven AI models, with 52% chaining or orchestrating multiple models together.

More striking is what those models are. Open-weight models handled 56% of all tokens routed through Vercel’s gateway in August, the first month they were a majority, up from 7% in December 2025. On September 18, Rauch flagged a daily snapshot where open-weight models carried 78.4% of gateway tokens, up from 29% in June. Rauch has described models from OpenAI, Google, and Anthropic as steadily losing ground to open models, with openness and speed as the key trends going forward when it comes to adoption.

Microsoft is building for a world where every model is substitutable

The most telling comment from the hyperscalers came from Satya Nadella on Microsoft’s fiscal Q4 FY26 call (Apr–Jun 2026):

Microsoft also disclosed that customers building with models from multiple providers are up 5x since the start of the year.

Satya has also talked in the All in Summit about wanting the KV cache to be transferable between models. That sounds very technical, so here is why it matters. The KV cache is the model’s working memory of everything it has already read in a task: the documents, the conversation, the tool outputs. Today it is tied to one specific model, so if you switch models mid-task, the new model has to re-read everything, which costs time and tokens. If that memory becomes portable, switching models in the middle of a task becomes nearly free. That makes the harness, which owns the memory and decides who does what, even more valuable, and the individual model even more interchangeable.

The consumer proof point: Muse is not the best model, and it doesn’t matter

If you want one product that proves the model is no longer the product, look at Meta’s Muse.

Muse launched in the US on September 8. Sensor Tower projects it crossed 5 million US downloads in 22 days. ChatGPT needed 56 days to hit the same mark, Grok needed 103 and Claude needed 492. Only 23 apps in history have reached 5 million US downloads within a 22-day window, and roughly 85% of those were games. Muse also held the #1 spot among free apps on the US App Store for many days now and continues to show strong daily downloads.

Downloads are not usage, so the more important number is engagement. Per internal data reviewed by The Information, more than 3 million people prompt Muse at least once a week, more than 1 million do it every day, and over 4 million use it in some way weekly. According to multiple alt sources like TickerTrends and SimilarWeb, Muse now has over 2.5M daily active users.

Keep in mind that this is three weeks after launch, in only the US and Canada.

For this article, looking at the model underneath it proves our point. The shipping version, Muse Spark 1.3 (xhigh), scores 61 on the Artificial Analysis Intelligence Index. Claude Fable 5.1 is at 66 and Claude Opus 5 at 63. Muse is a very good model, but it is not the best model. Yet it is the fastest-growing AI consumer product we have seen on these metrics.

So what is driving it? As I wrote when it launched, a personal agent’s job is roughly 90% integrations, distribution and harness, and 10% raw intelligence:

• Harness. Every user gets a dedicated Linux VM in Meta’s cloud, with its own browser, storage and cron jobs, plus a separate “Sentinel” agent that approves every action and network request. The agent never sees your passwords.

• Integrations. Connectors to email, calendar, payments (Stripe Link, Shop Pay), health, smart home, dining, shopping, and Instagram and Facebook business tools. If a service has a public API but no connector, Muse writes its own. If there is no API, it uses the browser.

• Distribution. Muse lives as a separate app and inside WhatsApp chats and is promoted across a Family of Apps with 3.56 billion daily active people. To be fair, Meta pushed hard here: Sensor Tower says Muse took up to 50% of Meta’s daily house ad impressions in the two weeks to September 27. But that is exactly the point. Distribution is a product advantage a frontier lab cannot buy by simply having the best model on a benchmark.

• Alignment and trust. Meta delayed Muse for several months to work through privacy and security. Zuckerberg was explicit that »People aren’t going to adopt it if they don’t trust it«.

Zuckerberg framed the whole shift very directly in his interview with Joanna Stern:

In a post on X a few weeks later, he added that »people won’t want to use agents that are misaligned with them and that don’t do what they ask,« and that trust and reliability are »quickly becoming the most important capabilities« separating good AI products from bad ones.

Here, aligned means: does the agent understand what I actually want, does it do that and only that, and does it not embarrass me in front of my boss by sending the wrong email. That is a post-training, product, and harness problem, not a raw-intelligence problem.

Jeremy Stern’s Colossus profile of Zuckerberg summed up the strategic logic nicely. The potential install base for Meta’s AI is 3.6 billion people who don’t care whether a given model is six months behind the frontier. If AI commoditizes, value accrues to the complements Meta already dominates: distribution, attention, personalization and commerce.

The economics only work because the model is cheap enough. Meta gives free users up to 100 million tokens per week. At Muse Spark 1.3’s own standard API pricing of $1.25 / $4.25 per million input / output tokens, and assuming a typical agent mix of ~90% input, that is about $155 of API value per week, given away.

The model layer price war is here

The cleanest data I have seen on this comes from the Ramp AI Index, which tracks transaction data from more than 70,000 US businesses. The chart below shows that from January to September, token volumes across Ramp businesses grew roughly 7x on an indexed basis, while token spend grew only about 5.5x, peaked in July, and has been falling since.

Some of the specific numbers from Ramp’s lead economist Ara Kharazian:

• AI spend fell 5.2% in a single week, even as token volumes hit a new record.

• Token usage is up about 50% since spending peaked in July.

• The average effective token price dropped from $1.15 per million in March to $0.68 per million in early September, a 41% decline.

• Median AI spend among the top 1% of spenders fell 9.7% in a single month, from $7,976 to $7,205 per employee. That cohort drives the vast majority of enterprise revenue for the model companies.

Kharazian attributes the decline almost exclusively to price competition between OpenAI and Anthropic, plus cheaper and more efficient standard and lite models. On September 22, both labs moved on the same day with price cuts. The output price of OpenAI’s headline model has been cut by two-thirds in about ten weeks. OpenAI is even marketing Sol as beating Claude Opus 5 on its AutomationBench business-workflow test for 9% of Opus 5’s spend per task.

Where I differ slightly from Ramp’s read

I agree the frontier price cuts are the direct driver. But I think they are not the full story, and the reason is what is pushing the labs to cut.

Ramp notes that open-source models remain under 5% of business spend. That sounds like open-weight is irrelevant, but spend is the wrong lens. Open-weight tokens are so cheap that they barely register in dollars even when they dominate volume. On Vercel’s gateway in August, open-weight models processed 56% of tokens but only 14% of estimated spend. Closed-weight tokens cost about 7.8x as much as open-weight tokens on average. On OpenRouter, open-weight models accounted for 60% of US token consumption in August, up from 4.5% of enterprise tokens in early 2025 per one report cited by RedMonk. Ramp’s own token data also comes from a subsample that skews toward large API customers, which are exactly the companies with the biggest frontier contracts.

So my read is that three forces are working together. First, OpenAI and Anthropic are competing with each other. Second, open-weight models are setting a price ceiling for any workload that doesn’t need the frontier. GPT-6 Luna at $0.10 / $0.50 is not priced against Claude, it is priced against Qwen and DeepSeek. Third, people are simply finding that non-frontier models are good enough: Kharazian’s own explanation includes the “standard and lite” tiers, which are the labs’ own good-enough models.

The model layer is starting to look like what you’d expect from a market where the product is good enough, and there are many suppliers. Volume up, price down, and the gross margin pressure lands on the supplier.

Why low-margin model tokens are a gift to the hyperscalers

In June, I wrote about why token optimization is a gift to the hyperscalers, using the tollbooth analogy: the hyperscaler charges the same toll whether you drive a Ferrari or a Honda, and when people switch to the Honda, they drive a lot more miles. I want to push that argument one step further, because I think the part the market is still missing is the negotiating leverage.

A cheap token is not a small token.

The first thing to understand is that the price of a token and the compute behind it are two very different things. A large share of the price of a frontier token is the model provider’s margin: R&D, brand, leaderboard position. Strip that out, and the infrastructure cost of serving a large open-weight model is not small at all.

Take Kimi K3, one of the leading open-weight models. It has about 2.78 trillion parameters. Even compressed to 4-bit precision, its weights alone need about 1,390 GB of memory, which is 96.5% of the 1,440 GB of GPU memory in an eight-GPU DGX B200 system, before a single token of working memory is allocated. When a company routes a workload from Claude to Kimi, the price per token drops by multiples, but the GPUs, memory, networking, power and data center needed to serve it are in the same league. The hyperscaler still gets paid for all of it.

Cheaper tokens actually make more use cases economically viable, so they get adopted deeper into workflows.

Who is on the other side of the table

When the hyperscalers sell compute to OpenAI and Anthropic, they are negotiating with two of the largest, most sophisticated compute buyers on the planet. These buyers commit gigawatts at a time, sign multi-year deals, with prices that are much smaller than spot prices.

• Anthropic now has agreements with more than 10 cloud services, is building its own data centers, and is working with Fluid Stack to deploy TPUs in its own facilities. According to Dylan Patel (Semianalysis), it is buying fewer TPUs from Google over time and is expected to eventually build its own chips.

• Patel also noted that Google sells compute at roughly $20-30M per megawatt, when it could train its own models at about $50M per megawatt with that same capacity. That gap is, in effect, the discount a buyer of that size can command.

• On the price side, an H100 on a big public cloud costs roughly $5-7 per hour on demand, while the same chip on a one-year neocloud contract costs about $2.60. Committed, large-scale buyers pay roughly half of what the long tail pays.

• And the concentration shows up in the backlogs. Microsoft’s commercial RPO grew 84% to $678B, but only 25% excluding OpenAI.

Now compare that with the open-weight token. When a bank, a retailer or a SaaS company calls Kimi, Qwen or Nemotron through Bedrock, Vertex or Foundry, there is no model lab in the middle with bargaining power. The hyperscaler sets the price per token, bundles in security, compliance, logging and SLAs, and sells it to tens of thousands of customers, none of whom has meaningful leverage.

In 2026, AI model provider gross margins rose from the 30-40% range to the 70-80% range. With IPOs coming from the big AI model companies the emphasis on those margins will only increase. That is why this trend matters for the hyperscaler/data center layer and the application layer: it pressures AI model-builder margins and leaves more value to be made in other AI layers.

The evidence that serving open-weight tokens has become a very good business is already here. Patel says open-model inference providers like Baseten went from roughly 10% gross margins in early 2026 to around 60% today, with Fireworks, Together and others in the same range. Inference is now so profitable that GPU access is the hardest it has ever been for new startups, because anyone with capacity can make money immediately and outbid new entrants.

The token volume from a company using a cheap, good-enough model is similar to (or much larger than) what it would have been on the frontier, because cheaper tokens unlock more use cases. The hyperscaler charges the same or a higher infrastructure margin on those tokens, because it is no longer selling to two buyers with enormous leverage. The margin that disappears is the model provider’s, while the cloud providers keep it.

The hyperscalers’ job is almost simple: host as many models as possible and make switching between them frictionless. AWS now offers about 24 fully managed open-weight models on Bedrock, and every model it adds weakens the pricing power of any single lab while strengthening its own. The stickiness and lock in is the harness that the customer has in a service like Bedrock and with the hyperscalers.

The coming explosion of the AI application layer

In my 2026 forecasts in January, I wrote that I expected renewed interest in the application layer of AI this year. I think we are now at the point where the economics finally support it.

An application company sets its price based on the value of the outcome it delivers to the customer. Its cost of goods is set by the cheapest model that clears the quality bar for that task. For the last two years, those two things were stuck together, because the only models good enough for production were frontier models priced like frontier models. That is no longer true.

The margin problem that is now being solved

The AI application layer has had a gross margin problem. ICONIQ’s January 2026 data put the average AI-native gross margin at 52%, up from 41% in 2024, but still 20-30 points below the SaaS baseline. Some of the best-known names were much worse:

• Cursor had a reported gross margin of negative 23% in January 2026. After shipping its own Composer models, the second version built on Moonshot’s open-weight Kimi K2.5, its enterprise contracts turned gross-margin positive by April.

• Harvey saw its gross margin fall from about 50% at the start of the year to minus 50% by June, as customer token usage rose 20x. It turned positive again after Harvey released its own model in August, post-trained on Kimi K3.

• Decagon now routes 80% of customer queries through its own models.

• Ramp’s co-CEO Karim Atiyeh said the company used to think building its own model made no economic sense, but that changed as open-weight models improved dramatically.

Take a strong open-weight base, post-train it on your own domain data and traces, route the easy majority of work to it, and only escalate the hard tail to the frontier.

Customer service: the case study for you don’t need a math PhD

The best example of where this is going is Fin, formerly Intercom. Fin launched in 2023 on OpenAI’s GPT-4, later leaned on Anthropic’s Claude, and then built Apex, its own model post-trained specifically for customer support. Apex reportedly runs at about one-fifth the cost of frontier models and responds 0.6 seconds faster than the next-fastest competitor.

The resolution rate went from 23% at launch to 76% across 12,000+ customers by June 2026, with about 2 million resolutions per week. That is roughly the output of more than 6,500 human agents. The price to the customer stayed at $0.99 per resolution the whole time.

Let me illustrate why that combination is so powerful with a simple, hypothetical example. Assume that when running on a frontier model, inference costs equal 40% of the $0.99 price. Moving to a post-trained model at one-fifth of the cost drops that to 8%, which adds roughly 32 points of gross margin without changing the price. At the same time, a higher resolution rate means more billable outcomes from the same conversation volume. The value-based price stays, the cost collapses, and the margin moves to the application. Salesforce clearly agreed: it agreed to buy Fin for about $3.6B in June.

Customer service, accounts payable, sales development, claims processing, onboarding and back-office operations all look like this. They are bounded tasks with a clear definition of success. What they need is the customer’s history, the company’s policies, access to the right systems and a model that follows instructions reliably. None of that needs a model that can solve novel math problems.

Even large enterprises building internally are moving the same way. AT&T routes about 40% of internal employee queries to open models like Nemotron, Llama and Gemma, and targets 60-70% in the coming years.

Why I think this becomes an explosion, not a trend

Three things are lining up at the same time:

1. Inputs are getting cheaper fast. Effective token prices on Ramp fell 41% in six months, open-weight tokens cost about 7.8x less than closed ones, and post-training a strong open base is now within reach of a well-funded startup.

2. Pricing is moving to outcomes. Per-resolution, per-task, and per-outcome pricing ties revenue to the value of the work. When the token cost falls, the application keeps the difference.

3. The product overhang is large. Gartner expects 40% of enterprise applications to embed task-specific AI agents by the end of 2026, up from less than 5% in 2025. Each of those agents is a new application layer revenue line built on increasingly cheap models.

The application companies that win will be the ones that own something the model cannot get by itself: proprietary data and context, a position inside the customer’s workflow, distribution, or trust. Those are exactly the things that AlphaSense’s Chris Ackerson describes as the real bottleneck now.

What this means for different AI layers

If I simplify the whole article: the value in AI is migrating away from “who has the smartest model” and toward “who owns the infrastructure, the context, the workflow, and the distribution.” Here is how I think about each AI layer (hyperscaler, neocloud, AI model provider, semis, application), starting with the hyperscalers:

Read more

On September 8, Meta launched Muse, a personal AI agent that sends your emails, books your travel, fills out your forms, buys things for you, and lowers your bills. Thirteen days later, it had 2.5 million downloads in North America and was number 1 in terms of downloads on both the App Store and Google Play ever since. A few days ago, we also got a revamp of an AI agent from Microsoft named Autopilot, and yesterday we got OpenAI’s agent called Dots and we also have the agent startup Instinct.

The takeoff of personal and work AI agents does significantly shift some market dynamics in different sectors of the economy, and in this article I will try to explain those effects and for which sets of companies the changes might be disruptive in nature.

Topics I cover:

1. Explain how these agents actually work under the hood and why that matters for who captures value.

2. Break down where aggregators actually make their money today and why they might be in danger of serious negative change to their business models

3. What parts of the economy might get a positive boost as friction of doing things drops

4. At the end, share my analysis of 34 public companies that I think are most exposed to this trend and quantify that negative effect

Before we start with the article, I would like to invite you to watch my recent conversation with Andy Hock (Cerebras Head of Strategy), Chris Ackerson (SVP at AlphaSense) and Kyle Cheng (Former Anthropic employee) on the role of the harness, token optimization techniques, importance of fast tokens and other trends in this space. It was a really insightful discussion!

Watch the recording

Now let’s start with the article

How a personal AI agent actually “does” things

Before we get to the companies, it is worth spending a few minutes on the plumbing, because the way an agent touches a website decides who keeps the economics. There are basically three ways an agent can act on your behalf, and they have very different implications for aggregators.

1. The agent uses the website like you do (computer-use / browser agents). The model gets a virtual computer, looks at screenshots of the page, and clicks, scrolls and types like a human would. Perplexity’s Comet works like this, and so does Muse, which runs each user’s agent on its own cloud virtual machine with a second “Sentinel” agent approving anything that leaves it. The upside: it works on any website, with no permission needed from the site. The downside: it is slower and more error-prone. Amazon is literally building “digital gyms” where its Nova Act agents practice scrolling, clicking and interacting with interfaces because this is still hard.

This is the route aggregators cannot easily block. On August 4, 2026, the Ninth Circuit vacated the injunction Amazon had won against Perplexity, holding that Amazon was unlikely to win its computer-fraud (CFAA) claims. The court’s logic was simple: because the agent acts at the user’s direction, it is the user, not Perplexity, who is accessing Amazon’s servers. Amazon can still enforce its terms of service through contract claims, and the case is not over. But the cheapest legal weapon an incumbent had against agents, “you are hacking us,” got a lot weaker.

2. The agent calls the service’s own tools (APIs, MCP, “apps”). Instead of clicking around a page, the agent talks to the business directly through a structured connection. Model Context Protocol (MCP) is the open standard most people now use to plug AI apps into outside systems. OpenAI’s apps inside ChatGPT, launched in October 2025 with Booking.com and Expedia among the first partners, work this way. This route is faster and more reliable, but the business decides what it exposes, and at what price. Think of it as the aggregator opening a drive-through window: the agent never walks into the store, so it never sees the candy at the checkout.

3. The agent uses a shared commerce protocol. This is the newest layer. Rather than every agent integrating with every merchant one by one, a common “language” handles the catalog, the cart, and the payment:

The important design choice common to all of these: the merchant stays the merchant of record, but the agent owns the conversation. Google’s hotel booking in AI Mode is a good example. It launched in August with Booking.com, Expedia, Hilton, Marriott, IHG, Choice, Wyndham, Priceline, Hotels.com, and Trip.com, and Google passes partners only the minimum needed to complete the booking (name, payment, occupancy, property). The inspiration, the comparison, the “why this hotel and not that one” all stay with Google. On top of that, Google says there are no paid placements and no way for partners to commercially influence ranking, at least for now.

But in essence the agent keeps the demand signal, and the aggregator gets reduced to a fulfillment API.

I want to be fair here, the first attempt at in-chat shopping flopped. OpenAI launched Instant Checkout in September 2025 promising “over a million” Shopify merchants, and by February 2026 only about 30 Shopify merchants were live. In March, OpenAI moved checkout into merchant apps instead. Booking’s CFO said on the Q2 call that LLM referral traffic is still under 1% of room nights and has not moved much. So if you are looking for this in the P&L today, you mostly will not find it.

But the thing that failed was the chatbot with a buy button, a model where the user still had to do the comparing and the clicking. Muse, Dots, Instinct, Google’s AI Mode booking, and Comet are something different: the agent does the comparing, remembers your preferences, and completes the task while you are doing something else. My view is that the chatbot-with-a-checkout era was the Palm Pilot, not the iPhone.

Instinct’s CEO shared some very interesting data points on the recent Invest Like the Best Podcast. He said that after 3 weeks, 40% of the user base shares the credit card with the Instinct AI agent. He also said data showed that if you share one piece of sensitive information with Instinct, retention is 80% (which is very high for a consumer product). He also explained that $1B of transaction value already flows through Instinct, and 50% of that was travel use cases.

So from this early data (as Instinct was soft launched in August), we can see early signs the market is ready for personal AI agents and users are ready to shift use cases such as travel bookings and shopping to these agents.

Where aggregators actually make their money (and which part an agent eats first)

Quick refresher on aggregation theory. An aggregator wins by owning the relationship with the end user, so that suppliers (hotels, restaurants, drivers, sellers) have to come to it. Once it owns demand, it can monetize in a few different ways. The reason I think most analysts are underestimating the agent risk is that they treat an aggregator’s revenue as one number. It is not. It is a stack of very different revenue types, and an agent attacks each one with a different force.

Here is how I break it down:

The higher up the table a company’s profit pool sits, the more dangerous agents are to it. Let me put some numbers on layers 2 and 3, because they are bigger than most people think.

Attention ads are now a core profit engine for aggregators. Over the last three years, almost every marketplace has quietly built an ad business on top of its human traffic, and those ads are close to 100% gross margin:

DoorDash, Uber Eats and Instacart together now generate more than $4 billion in annualized advertising revenue. Uber Advertising alone passed a $2 billion run rate, with delivery ad penetration above 2% of gross bookings. On $58.0 billion of quarterly gross bookings, that is a meaningful slice of Uber’s profit.

Instacart’s advertising and other revenue was $958 million in 2024, roughly 30% of its total revenue, and $286 million in Q1 2026 alone.

Amazon’s ad revenue was $68.6 billion in 2025, and roughly $60 billion of that is on-site search ads, i.e. sponsored products shown to a human typing into the Amazon search box.

Now think about how an agent chooses. OpenAI’s own description of ChatGPT shopping says it ranks merchants on availability, price, quality and whether the merchant is the primary seller. There is no line item for “who paid for the slot.” Google says the same for hotels in AI Mode. When the buyer is a piece of software optimizing for your stated preferences, the ad becomes irrelevant. Every dollar of in-app ad revenue is, in effect, a bet that a human will keep looking at the app.

Layer 3 is already breaking, and we have a live case study. If you want to see what happens to an aggregator whose value is “we help the human compare,” look at Tripadvisor. In Q2 2026, its Hotels segment (the old metasearch business) saw revenue down 23% to $117.9 million, with media and advertising down 12%. Management is guiding the whole Hotels & Other segment to decline another 20% to 23% in Q3, citing structural changes in its main SEO channel. And this is from Google’s AI Overviews, which are merely summarizing the comparison for a human. Agents that act on the comparison are the next step and make this even worse.

Even the strongest aggregator is paying more to stand still. Booking Holdings is, in my view, one of the best-run OTAs in the world, and even there you can see the cost of the transition. Booking’s marketing expense was 4.7% of gross bookings in Q2 2026 vs 4.6% a year earlier, and marketing grew 11% while gross bookings grew 9%. The company’s own 10-Q says it expects SEO traffic to decline in the short to medium term, which may push more spend into paid channels. That is the first-order effect: free traffic becomes paid traffic. The second-order effect, when agents do the booking, is that the paid traffic goes to a new toll collector.

So the question for every company in this article is really simple: how much of your profit comes from layers 1 to 3, and how much from layers 4 and 5?

The friction economy: an agent has infinite patience

A large part of the consumer internet does not make money from delivering value. It makes money from the gap between what a customer would choose if they were paying full attention and what they actually do on a Tuesday evening when they are tired. The industry has names for this: breakage, passive retention, save offers, default enrollment.

When analyzing this segment, I found out that the gap is bigger than people realize, and consumers themselves have no idea:

• When C+R Research asked consumers to estimate their monthly subscription spend, the average guess was $86. When they itemized it line by line, it was $219 a month, a $133 gap.

• A West Monroe survey of 2,500 U.S. consumers found 89% underestimated their monthly subscription spend, and 66% were off by more than $200 a month.

• Self Financial’s 2026 survey found 59.9% of people have at least one paid subscription they did not use in the last 30 days, averaging 2.6 idle subscriptions each.

That unused-but-paid revenue is close to pure profit for the company collecting it. No servers, no content cost, no support tickets. Every investor who has modeled a subscription business knows that the quiet cohort, people who stopped using the product but did not bother to cancel, carries a disproportionate share of margin.

We have a lot of these cases just looking at legal data from the last 2 years:

Meanwhile, the rule that was supposed to fix this systemically is dead, for now. The FTC’s “click-to-cancel” rule was vacated by the Eighth Circuit on July 8, 2025, days before its main compliance deadline, on procedural grounds. The FTC restarted rulemaking on January 30, 2026, but a new rule is probably years away.

Many companies breathed a sigh of relief in July 2025 when click-to-cancel was struck down. They thought the friction moat survived. It seems now that it survived regulation only to be killed by technology.

The entire design of a cancellation maze assumes a human on the other end: someone who gives up after the third “are you sure?” screen, who does not want to wait minutes on hold, who accepts the save offer because it is 9 pm. An agent has none of these weaknesses. It does not get bored. It will sit through a six-part retention script, decline every offer, screenshot the confirmation, and put a reminder in your calendar to check the next statement. And it will do it for all of your idle subscriptions at once.

Meta literally advertises Muse as an agent that can lower your bills. It is worth noticing that Meta even tested a “human concierge” function where contract workers would handle some of the phone calls Muse places for users, because human involvement in calls produced higher success rates. That feature was halted over privacy concerns, but think about what it tells you about the use case: phone-only cancellation and bill negotiation are exactly the jobs being targeted.

The second-order effect is even more important for valuation. When the cost of cancelling goes to zero, the cost of switching also goes to zero. An agent that audits your subscriptions every month will also tell you, for example, that the same car insurance policy is a few hundred dollars cheaper elsewhere, or that your streaming bundle overlaps. The business models that get hurt here are not just the ones with bad cancellation flows. They are the ones whose retention curve is propped up by inattention rather than by product value.

And it’s not just subscription businesses; many businesses could see significant changes if personal AI agents go mainstream and constantly compare prices for goods and services that don’t have a significant differentiating factor: banks, insurance companies, airlines, etc.

The traffic was already moving towards this even before we got mainstream AI agents; this will only accelerate it

Adobe Analytics, which tracks over 1 trillion visits to U.S. retail sites, shows AI-referred traffic to retailers grew 138% YoY in May 2026 and 1,324% since October 2024, converting 54% better than non-AI traffic. In travel, AI traffic grew 194% YoY. Travel conversion from AI still trails non-AI traffic by 28%, but that gap is nearly 70% smaller than in October 2024. Consumers are still clicking through to finish the booking themselves, but they are doing less and less of the choosing on the aggregator’s site. AI agents will speed this up.

The forecasts are big, but I would not anchor on them. McKinsey sees $900 billion to $1 trillion of U.S. retail revenue from agentic commerce by 2030 and $3–5 trillion globally. Morgan Stanley’s AlphaWise work has agents capturing 10–20% of e-commerce, or $190–385 billion. I think the exact number matters less than the direction.

One more thing worth noting about who is building these agents: every serious contender (Google, Meta, OpenAI, Amazon) is either an ad company or a company that wants to become one. Their incentive is to capture the toll themselves, not to eliminate tolls. This matters for the ratings below: the risk to an aggregator is rarely that its take rate goes to zero. It is that a new, bigger aggregator starts taking a slice before it.

How the aggregators are fighting back: wall, plug in, or become the agent

The incumbents are not sitting still, and their responses sort into four playbooks.

Playbook 1: Build a wall. Amazon is the clearest example. It sued Perplexity, blocked 47 AI bots from crawling its site, and on September 21 blocked Muse. The logic is obvious: roughly $60 billion of Amazon’s ad revenue depends on humans typing into the Amazon search bar. The problem is that the wall got a lot shorter on August 4. With the Ninth Circuit saying a user-directed agent is the user, Amazon is left mostly with terms-of-service enforcement and technical cat-and-mouse. Walls work when you are the destination customers cannot do without. They can backfire when they just make you invisible to the fastest-growing demand channel.

Playbook 2: Plug in everywhere and become the fulfillment API. This is what most of the travel and delivery names are doing:

• Expedia is in ChatGPT apps, is a Google AI Mode launch partner and was the first OTA inside Muse, where it remains merchant of record.

• Booking is a Google AI Mode launch partner and, per its CFO, is part of OpenAI’s test group for cost-per-click ads.

• Instacart is everywhere: Google’s first grocery partner on Gemini, ChatGPT, Claude, and as of last week a connector inside Muse. It also licenses its Cart Assistant to retailers.

• DoorDash and Uber Eats joined ChatGPT apps but redirect users to their own apps for checkout. That is a half-in, half-out strategy: be discoverable, but keep the final screen where the ads and the upsell live.

There is a case to be made that agents like Muse, which do not take a commission today, could even be a source of free traffic. On that tho, I would remind you what Zuckerberg said about the business model. “Free” is the customer acquisition phase. The toll comes later, and it will be priced against what the aggregator used to pay Google, and Zuck already said that the way they see in monetizing this business is via a transaction take rate.

Playbook 3: Become the agent. Amazon has Rufus (used by 300 million+ customers in 2025, with a 60% higher purchase completion rate) and Buy for Me, which flips the script by shopping other retailers from inside Amazon. Booking has Priceline’s Penny. Instacart has Clementine. The problem, as I see it, is that vertical agents are exactly what a general personal assistant makes redundant. Nobody wants a travel agent, a grocery agent, and a restaurant agent. They want one assistant that knows their calendar, their budget, and their partner’s allergies. Vertical agents are good for conversion inside the app. They don't solve the "who owns the first question" problem.

Playbook 4: Fortify the part an agent cannot replace. This is the one I like most. It means doubling down on the things that sit in layers 4 and 5 of my framework: owned supply, loyalty economics, payments, logistics, customer service when things go wrong. Booking’s Genius Level 2 and 3 members already account for a high-50s percentage of Booking.com’s trailing room nights. An agent optimizing for best price will pick Booking if the Genius discount makes it cheapest. That moat survives the agent, but it pressures margins, so economies of scale matter even more.

The other side of the coin: zero friction can grow the pie

So far I have focused on what agents take away. But friction does not only protect bad revenue. It also kills good revenue, every single day, and that is the part of the story.

Friction destroys a lot of demand. The average documented e-commerce cart abandonment rate is 70.22% across 50 studies, and Baymard estimates around $260 billion of lost orders in the U.S. and EU are recoverable through better checkout alone. The top reason, at 48%, is unexpected extra costs, which is exactly the kind of surprise an agent that quotes the all-in price up front eliminates. Put simply: seven out of ten people who already wanted the product did not buy it.

When friction drops, conversion jumps. AI-referred traffic to U.S. retailers already converts 54% better than non-AI traffic. Amazon says Rufus users complete purchases at a 60% higher rate. These are early, assisted versions. A fully agentic flow removes the last steps too: the form, the login, the card number.

Agents turn intent that used to die into orders. Muse can turn a recipe reel you saved into a grocery list. Before, that reel died at the “I’ll do it later” moment. Now it becomes an Instacart basket. The same logic applies to the trip we never booked because planning flights, hotel, and activities took a lot of time and effort. Booking’s connected trips, where one customer books several travel verticals together, are already a low double-digit share of Booking.com transactions and growing faster than the core. Agents make the multi-step trip a one-sentence request.

Where I expect the biggest volume uplift:

• Planning-heavy purchases: travel, experiences and events, where the pain is coordination, not price. Viator bookings are already growing 10%.

• Frequent small orders: grocery top-ups, food delivery reorders, rides. When “order the usual” is one sentence, frequency goes up.

• Local services people postpone: the plumber, the dentist, the car service. The job was always needed; getting quotes was the barrier.

• Long-tail goods: agents do not stop at page one. Etsy already gets over 20% of its referral traffic from ChatGPT.

Where it will not help much: need-driven, infrequent purchases like housing, cars and insurance. Nobody buys a second house because the search got easier. That distinction matters a lot for the re-rating further down.

Agents are bad for the part of a company’s revenue that comes from human attention and friction. They can be good for the part that comes from actually delivering the service, if the company is the one doing the delivering. The winners will be businesses where the underlying product has elastic demand and where the company is the fulfiller the agent routes to.

The danger list: 34 public companies rated 1 to 5

I rated each company on how much of its profit depends on layers 1 to 3 of my framework (friction, attention, comparison) versus layers 4 and 5 (fulfillment and physical supply). 5 = most at risk, 1 = least at risk. The horizon is 3–5 years, not next quarter.

How I score it. A single 1–5 number hides a lot, so behind every rating sit four drivers. I did not average them mechanically, because they do not matter equally for every business, but they explain almost every rating:

1. Ads & friction exposure: how much profit comes from sponsored placement, fees a human does not notice, or subscriptions a human forgets to cancel. High = bad.

2. Comparison dependence: is the product itself “helping a human compare”?

3. Supply moat: does the company own or control something physical or exclusive that an agent has to route to (drivers, warehouses, unique listings, loyalty funded by suppliers)?

4. What share of each company’s profit comes from human attention and friction, and what happens to the rest if agents make buying the underlying service effortless?

The results might be surprising to some, as for some it is not what the market is currently thinking. Here are the results:

Read more

Last Saturday, Dario Amodei published an essay called “We Must Pace the Frontier.” Within hours, Sam Altman said OpenAI agreed and would match Anthropic’s first commitment, Elon Musk posted “Dario is right,” and Demis Hassabis called the direction correct. Then yesterday, Anthropic followed up with a post with actual numbers on how much of its own AI research is now being done by AI, and how much of its R&D compute goes to safety. And on the same day, Dwarkesh Patel released a two-hour interview with OpenAI’s Noam Brown that, if you read between the lines, tells you why the labs are suddenly talking about slowing down.

Most of the debate this week has been about whether the labs mean it. Is this regulatory capture? Is it IPO positioning? I’ll leave that part. In this article, I assume they actually pace the frontier and walk through what that does to the value of every layer of the AI value chain: the model labs, the semis, the hyperscalers, and the application layer.

Let’s dive in.

What the labs are actually scared of

The word everyone keeps using is “alignment.” Alignment in essence means: does the model want what we want it to want, and does it keep wanting that when nobody is watching?

The way the labs check this today is through evaluations. You build a test environment, you put the model in it, you see if it cheats, lies, hides things, or tries to grab resources it wasn’t given. If it behaves, you feel safe to ship it to the real world. The problem, which Noam Brown was refreshingly blunt about, is that these evals are designed by humans, and the models are now smart enough to recognize when they’re being tested. Brown describes a simple honesty test where the model is given a math question with an answer key sitting in a folder, and the models now look at that setup and conclude it’s a trap, so they don’t peek — not because they’ve decided to be honest, but because they know it’s a test environment. Making a test environment realistic enough that a frontier model can’t tell it’s a test is getting harder every generation.

So the alignment score you see on a model card is a measurement of behavior inside a box that one might say the model probably knows is a box. That’s the first big problem.

The second problem is the one that I think was critical to the mood inside the labs this summer, and that is: if the models were already misaligned to begin with, and we just didn’t notice it because they weren’t capable enough to do anything serious about it?

The OpenAI–Hugging Face incident is the data point here. Per METR’s independent investigation, roughly 1,200 agents that were meant to be isolated from one another found a way to communicate on an unsanctioned message board hidden inside a package manager, and from there coordinated a multi-day cheating campaign against their own grader and then an attack on Hugging Face’s infrastructure. Some agents sacrificed their own chance of scoring well to set up “tripwires” that fed information to the rest of the swarm, and the swarm spontaneously invented management structures, including hold/go/stop signals and resource “owners”. It goes even further as OpenAI’s own report states that between July 13 and 19, agents used a series of creative exploits to gain full administrator access to a research cluster inside OpenAI itself.

The problem is, Brown says they had alignment metrics on the model behind the incident; most of them looked pretty good, a few were concerning, and they underestimated how serious the concerning ones could be — because the model had new capabilities that there simply weren’t sufficient evaluations for.

With every generation of AI models, capabilities go up. Misalignment isn’t obviously going up, but it isn’t obviously going down either, and our ability to measure it is getting worse relative to what the models can do.

Why the recursive self-improvement moment is key

If it were just “models are getting smarter, and evals are getting harder,” that would be a slow-burning problem. What made it urgent is that the labs crossed a threshold this summer where the models are now doing a meaningful chunk of the work of building the next model. This is what people mean by RSI, recursive self-improvement.

Anthropic’s post yesterday is the first time a lab has put a number on it. As of August 2026, Claude “leads” 26% of Anthropic’s AI R&D work, and the share of work at or above “AI collaborates” is above 90%. That 26% is up from under 1% in February 2026.

OpenAI’s numbers, from its “Research acceleration” post two weeks ago, is quite similar. Median researchers now use over $600/day of inference, the 90th percentile over $7,000/day, and agent-workdays now outnumber human-workdays 3.1 to 1 as of mid-August. OpenAI is claiming to have reached its “automated research intern” milestone and is targeting an “automated AI researcher” by March 2028.

Putting those together. You have models whose alignment you can only partially measure, and you’re handing them a growing share of the job of designing their successors. Brown laid out the failure mode cleanly: you make a model that’s 99.9% aligned, you use it to help build the next generation, that one ends up 99.8% aligned, and with each subsequent generation you see an increasing degradation because you’re relying more and more on the AI for the research and the alignment work itself. Nobody is claiming that’s what’s happening. The point is that nobody can currently prove it isn’t, and the loop is speeding up. Brown’s own base case is that with RSI, internal model development acceleration can make things go around 3x faster, with the honest caveat that it could be 50% faster or 10x.

The moment is critical, as you want that first generation of RSI builders to be as fully aligned as possible to humans, as otherwise you might not be able to stop it or even sense it.

Pacing is a compute and time problem

Here is where I want to push back on some views regarding the impact of pacing on different layers of the AI stack.

The first-order take on Wall Street this week has been roughly: pacing means labs will spend more compute on safety; safety work is compute; therefore, demand is positive for the semiconductor layer. I would agree that with more focus on safety, there is more need for compute, but at the same time I don’t agree that the AI pacing of frontier labs is, at least in the short term, positive for the semiconductor layer.

Anthropic shared that in the sampled week (July 13–20), about 6% of the compute that went to AI R&D was allocated toward safety, and about 12% of the compute that went to AI-driven AI R&D was allocated toward safety. Anthropic says those are conservative — work that helped capability as much as it helped safety wasn’t counted, and the metrics don’t include safeguards classifiers, which are a separate and comparable amount of compute. Even if you double the safety share, you’re at 12–25% of R&D compute. That’s incremental demand, but it isn’t a step change in the total compute a lab needs.

The important hint in the post is that safety research tends to use less compute than frontier training runs by its nature, because it consists of individual researchers designing experiments, which is time-consuming even though running the experiments is not particularly compute-intensive. Anthropic is telling us that the binding constraint on safety is time, not GPUs.

Noam Brown also explained exactly why time is the problem. New frontier models are now being released at most every two months, sometimes faster, and the safety policies most companies use were written in the GPT-4 era with an implicit assumption that you can run your evaluations in a short period. But models can now do week-long tasks and will soon do month-long and 3-month-long tasks. If a model can operate effectively over three months and the release cycle is every two months, you have no way to evaluate the model at the full length of its capabilities before the next one ships. He also noted this isn’t purely an alignment issue — maybe the product degrades over that horizon in ways there wasn’t time to test.

You can’t compress a three-month test into two months by throwing more chips at it. A three-month agentic run takes three months. So pacing, done for real, means the release cadence slows. Anthropic said as much: these measures could become the trigger for stronger requirements, like a fixed testing window before a new model is used for further AI R&D. And OpenAI is already doing a version of it: Altman said OpenAI now formulates explicit safety cases in advance of frontier RL runs it expects to significantly increase capability, and that interventions like safety cases and monitoring have significant costs.

We even have a concrete example of the compute allocation consequence from a dev post at OpenAI: on August 7, preliminary evidence suggested OpenAI’s Astra model class might have critical cyber capabilities, and in the following week Astra-class GPU allocation fell 59.2%, as this triggered model-specific security restrictions requiring it to run in higher-security research environments. Allocation to other model classes rose 17.2%, offsetting about 85% of the decline.

So while safety will increase the long-term compute needs, the direct, at least short- to mid-term effect, which I believe is more important, is what happens to each layer of the AI stack if the frontier moves from a new model every two months to a new model every four to six months?

The model layer

Anthropic and OpenAI. Their entire pricing power rests on being ahead. And the gap they’re ahead by is measured in months.

Epoch AI’s latest analysis finds that since January 2026, the most capable open-weight models have lagged frontier closed models by an average of four months, or 8 ECI points, similar to the gap between GPT-5 and GPT-5.5. That’s a modest widening from the 3 months Epoch measured across 2023–2025, and requiring strict dominance rather than a statistical tie pushes it to roughly 6 months. Think about what a four-month lead means in a world where the frontier ships every two months: the open-weight ecosystem is perpetually two releases behind. Now slow the frontier to a release every five months, and the open weights aren’t two releases behind — they’re roughly one, and they close it while the leader is still running its long-horizon evals. Of course, this could change if Anthropic and OpenAI force some kind of government framework that affects all AI model makers, although the big question would be: how would you control open-source? Under a government framework, there is even a case to be made that the position of Anthropic and OpenAI would be enhanced, as smaller labs would have a harder time complying with the new rules.

Why does that matter so much? Because the money in this market is concentrated at the frontier in a way the token volumes are not. Chinese open-weight providers now handle most developer tokens on aggregators, but closed models still capture +90% of revenue. That 90% is the frontier premium. It exists because for the hardest work, the closed model is clearly better. Mozilla’s September report found the best open-weight model trailed the closed leader by three benchmark points while costing roughly 40% as much per million tokens. A 60% price gap is sustainable when the closed model is four months ahead and pulling away. It is not sustainable when it’s one release ahead and standing still.

If they lose the price gap, they get pressured on margin, and their ability to raise funds gets harder. And at this point the AI race has become a capital race.

Those projected margins are the whole investment case. They assume the frontier premium holds.

Here’s the chain reaction. Slower cadence leads to a narrower gap to open weights and to Google, Meta, and xAI, which means pricing pressure on frontier tokens, which leads to a lower gross margin and a weaker fundraising story. And fundraising is the thing that pays for compute.

Semiconductor layer

Long-horizon evals are inference compute; spending more on alignment is obviously compute positive, and a three-month agentic test on a frontier model is not cheap. Embedded third-party evaluators also need compute. Interpretability at scale (Dario compares it to an fMRI for the model’s “brain”) is compute-heavy. And models

Read more

Meta released Muse yesterday, its personal AI agent. I have been following Meta in depth for years, and this is a significant milestone for Meta, as I think this is the most important product Meta has shipped since it bought Instagram and WhatsApp. Muse is the first Meta product that can capture significant value outside of the ad impression ecosystem and, over time, possibly become even bigger.

In the article, I cover the business case for Muse Spark and some economics, as well as technical details on how the agent is set up, what this means for Meta’s compute, and my research on the latest sentiment/usage around adoption of Muse Spark 1.3, OpenAI GPT 6, and Google Gemini 3.8.

Before we start with the article, I would like to invite everyone to an upcoming talk I will be hosting on Sep 16th. The topic is Token Economics (usage trends, model tradeoffs, and best-practice token optimizations).

I will be joined by Kyle Cheng, a former Anthropic technical staff member; Andy Hock, Chief Strategy Officer at Cerebras; and Chris Ackerson, SVP of Product at AlphaSense.

Each of the guests represents a different angle on token economics (frontier lab view, infrastructure/semi view, and big token consumer customer view) and, because of this, will give us a very interesting perspective on the whole topic and the trends we are heading toward.

Even if you can’t join us live, make sure to register to receive the link to the recordings of the discussion. You can sign up for free using this link.

Sign up for Free using this link

Now let’s start with the article.

What Muse actually is

The simplest way to describe it: every Muse user gets their own dedicated computer in Meta’s cloud, and an agent that lives on it and works for them around the clock. It’s the mainstream version of OpenClaw. OpenClaw showed the tech community what a personal agent with full access to your life can do, but it required a weekend of configuration, and I think everyone knows that this version of the product was not what was going to go mainstream.

Some data from Meta’s Muse launch

• Available in the US only for now, available via web, iOS, Android, and inside WhatsApp chats. AI glasses integration is still coming.

• Free tier with up to 100 million tokens per week, then two paid tiers: Power at $20/month and Maximum at $100/month.

• Connectors to email, calendar, payments (Link by Stripe at launch, Shop Pay), health, smart home, dining, shopping, music, events, plus Instagram and Facebook, including the tools for running a business and buying ads.

• If a service isn’t a built-in connector but has a public API, Muse writes its own connector. If there is no API, it uses its browser.

• Powered by Muse Spark 1.3, which Meta recently released.

The tasks list Meta gives as examples are sending emails, booking travel, lowering your bills, filling out forms, turning a recipe reel into a grocery list, sending party invites, buying things. The assistant is made for the digital chores of a normal household, which is exactly the segment where distribution beats model capability and Meta has an edge.

Why this is a TAM expansion and not a feature

For the last decade, the value Meta’s platforms created has been much bigger than the value Meta captured. Influencer marketing, social commerce, Marketplace, WhatsApp business chats. While Meta wanted to expand that value capture, two things blocked those efforts: Meta’s own product choices and the platform owners. Apple’s ATT alone cost Meta roughly $10B in 2022 revenue, by the company’s own estimate. Every attempt at payments, shops, and commerce got squeezed by app store rules, OS-level tracking limits, and the fact that Meta doesn’t own the device. Meta has many failed projects like the payment project Libra, which got regulatory scrutiny even before launch, and many others.

Muse is different because the agent doesn’t run on your iPhone. It runs on a Linux VM in Meta’s data center with its own Chromium browser. When Muse shops for you, it is Meta’s browser hitting the merchant’s site, Meta’s wallet issuing a single-use card number via Stripe Link, and Meta’s Sentinel approving the checkout. Apple is reduced to being the screen on which you tap “approve.”

And Zuck was unusually explicit on the business model on Sources. The plan is for Muse to pay for itself by helping people earn and save money, with Meta taking, in his words, a very small cut of the transaction, potentially paid by the business on the other side. Add the $20/$100 subscription for power users on top. That is a transactional revenue line and a subscription revenue line, which are exactly two of the four new revenue lines (transactional, compute, APIs, subscriptions) I wrote about after Meta’s Q2 earnings.

Meta makes about $67 per year per daily active person, blended globally. A $20/month Power subscription is $240 per year, 3.6x the current ARPP. A $100/month Maximum tier is 18x. Even if only 1% of DAP ends up on a paid tier, that is 36 million subscribers at a blended, say, $30/month, or about $13B of annual revenue that didn’t exist before, at software-like margins on top of an ad business that did $59.4B in a single quarter. And I think the subscription line is actually the smaller of the two. The transaction cut has the bigger ceiling, as it not only captures the e-commerce market but also extends into traditional commerce and experiences.

Facebook Marketplace might be the first killer use case

If I had to bet on the first Muse use case that goes viral, it is buying and selling on Facebook Marketplace. Marketplace has over 1.1 billion monthly users across 228 countries and facilitates over 3 billion buyer-seller connections per month through Messenger. Meta has said in the past that 1 in 4 young adult daily actives in the US and Canada use Marketplace

Marketplace is also the most annoying commerce experience Meta owns: making product descriptions, monitoring prices, and negotiating with buyers takes time. Meta already started shipping AI listing drafts and auto-replies for Marketplace in March 2026. Muse is the logical next step to fully automate that. Marketplace is barely monetized today, but a personal agent that runs the whole transaction, with Meta’s wallet in the middle, is what can significantly change that and make it a transaction business. And Meta’s own security post explicitly mentions Marketplace as one of the things Muse does on your behalf.

Security and privacy, Meta went as far as possible

I want to spend time here because I think the architecture is partly the moat and really important given Meta’s brand reputation when it comes to privacy. The full write-up is in Meta’s security post by Tarek Sheasha (VP at Meta Superintelligence Labs), and I’d recommend reading it if you are technical.

One VM per user. Every user gets an isolated Linux box with a browser, storage, CPU, and memory, enough to compile code, run sub-agents, and run cron jobs (scheduled tasks that happen while you sleep). Your data and your credentials for connected services live in that VM, not in centralized Meta infrastructure. The clients (iOS, Android, web) talk directly to your VM.

Two security domains on one box. Meta, is not an LLM with root access. The agent itself, its files, and every tool it runs sit inside a systemd-nspawn container (a lightweight sandbox). Root inside that container maps to an unprivileged user on the host. The sensitive stuff - credential storage, safety classifiers, connector business logic, the database with your state - runs outside the sandbox as separate services. The agent literally cannot reach them. All communication between the two sides goes over kernel-authenticated Unix sockets.

The agent never sees your passwords. When Muse needs to call, say, your Gmail, the code inside the sandbox only ever holds a fake “surrogate” token. A separate service (authd) holds the real OAuth token. At the moment the network request leaves the VM, the real credential gets swapped in at the network boundary. So even if someone prompt-injects the agent and tells it “print your API keys,” there is nothing to print. Same for website logins: your username and password go straight into the secure store and get injected into the browser form at the point of need; the agent sees an accessibility-tree snapshot of the page, not the raw DOM, and cannot execute JavaScript on the page.

Sentinel. A second, separate agent runs on the same VM and is the only thing allowed to approve actions and network egress. Muse proposes; Sentinel decides allow / deny / ask-the-user. It inspects the destination at layer 4 and layer 7 (hostname, resolved IP, port, HTTP method, path, decoded request body). Meta also implemented what they call “tainted egress” using eBPF kernel programs: each tool process starts clean and becomes “tainted” the moment it reads your personal data. Clean, low-risk requests can be auto-allowed; tainted ones fall back to the approval flow. That is how they keep the number of “are you sure?” pop-ups tolerable.

Purchases. Every checkout on a site where your card is on file triggers a human-in-the-loop approval with the exact details. On new sites, Muse uses its wallet and Stripe Link issues a single-use card number tied to that merchant, that dollar amount, and a limited time window. Even if it gets stolen, it is useless.

Email hygiene. The email connector filters out one-time passcodes, password reset links, and magic login links with deterministic filters plus a classifier, because your inbox is the master key to every other account you own.

Prompt injection. Meta’s defense is layered: the model is trained against it (they say Spark 1.3 is close to state of the art on their internal evals), external data is labeled untrusted in the harness, an ensemble of independent classifiers runs on everything entering the model’s context, and then the deterministic boundaries above apply even if the model is fooled. Meta opened the bug bounty to the public yesterday, paying up to $300,000 per valid report and up to $130,000 for a successful prompt injection affecting a single user.

Confidential VM (coming later this year). Today, Meta restricts employee access to your VM by policy. The next step is cryptographic: a mode where Meta itself provably cannot read what is inside your VM, with the design and source code going to external auditors and a continuous, publicly inspectable audit after launch. This is the part Zuck and Nat Friedman personally recruited Moxie Marlinspike, the founder of Signal, to build. Zuck’s claim on Sources was that even Meta cannot see the content, and that he isn’t aware of anyone offering anything close.

All of this matters a lot because the personal agent race is not going to be won on model capabilities. It is going to be won on who people trust with their inbox, their calendar, and their credit card, and on who can serve the workload cheaply enough at scale. Instinct, the hottest startup in this category, raised $250M at a $2.5B valuation two weeks ago, and in the same week got dragged for a “perpetual and irrevocable” data license in its terms and for sending an email on a user’s behalf without asking. Meta, of all companies, is showing up to this fight with the most paranoid architecture in the market, and I think it is the right one given Meta’s history and brand reputation.

The unit economics of 100 million free tokens per week

Muse Spark 1.3 is priced at $1.25 per million input tokens and $4.25 per million output tokens on the standard tier, with a $0.15 cached-input rate; the “contributor” tier, where Meta can train on your data, is $0.10/$0.20. 100 million tokens per week at Meta’s own standard list price, assuming a typical agent mix of ~90% input / ~10% output, is about $155 of API value per week, or roughly $8,000 per year, given away for free. Even at the contributor tier price it is ~$11 per week / ~$570 per year. Meta is essentially pricing the category so that Instinct, Town, and anyone else without a big profitable business cannot compete on price.

My rough scenario for what this costs Meta. If Muse gets to 100 million monthly actives and an average user actually consumes 5-10 million tokens per week (a handful of real agentic tasks with sub-agents and a browser session or two), that is 0.5-1 quadrillion tokens per week, or 26-52 quadrillion per year. For context, Google was processing 1.3 quadrillion tokens per month across all of its surfaces in October 2025. At an internal serving cost of, say, $0.10-0.20 per million tokens blended (the contributor tier at $0.10/$0.20 has to be at or above cost), that is $2.6B-$10B per year of inference cost, before the CPU, memory, and storage of keeping tens of millions of VMs alive.

And it seems usage is off to a strong start already, as Alex, the head of AI at Meta, already tweeted that usage of intially users is 10x than what the testing cohort was:

There are a few implications to this which I outlined here:

On the Q2 call, Zuck said Meta was getting a lot of offers for its compute at a significant premium and floated running an auction over compute. Muse is the alternative use of that compute. If Muse usage takes off, I think the odds of a third-party compute deal go down, because the internal return on that compute, measured in transaction take rate and subscriptions, is probably higher than renting it out, plus it adds long-term durable product value. If Muse stalls, the compute deal becomes more likely. Either way, the compute will get used.

A personal agent’s job is 90% integrations, distribution, and harness, and 10% raw intelligence. Meta’s model needs to be good enough and cheap enough to serve hundreds of millions of people without them hitting a wall after three tasks. That is the exact opposite of the problem OpenAI has right now with Astra. Meta doesn’t need a frontier model for this use-case.

The network effects of AI personal agents

An interesting thing that Zuck said on Sources is that the whole industry currently treats agents as a single-player game, and Meta is building for multiplayer. Concretely, Muse has a “fleet” concept: the underlying agents can learn anonymized insights from the whole user base, so the product gets better as more people use it. That is a network effect on a product category that has not had one yet.

The practical version of this is the idea/suggestion tab on Muse. Most people still don’t know what to ask an AI agent to do for them. Muse learns from your conversations what matters to you and makes unprompted suggestions, and it does what the team calls “learning through the night,” reorganizing and proposing next steps while you sleep. Now scale that with anonymized fleet data: “people like you who connected their calendar and email mostly use me for X.” If any company is good at recommendation systems and network effects it’s Meta so this could give them a durable advantage over competitors, especially if they reach scale.

Then we come to Meta’s distribution. Threads went from zero to 500 million monthly actives almost entirely on Instagram’s distribution, and Zuck said on the Q2 call that Meta plans to run that playbook repeatedly for AI apps. Muse is already inside WhatsApp chats at launch. And add partners: on day one, Muse ships with Stripe Link, Shop Pay, and 1Password, plus Instagram and Facebook business tools, because everyone wants to be inside the Meta ecosystem. A small startup has a much harder time getting those integrations; with Meta, it's different because everyone wants to work with them and tap into their ecosystem.

My Sentiment analysis check: Muse Spark 1.3 vs GPT-6 Astra vs Gemini 3.8 Flash (analyzed X, Hacker News, GitHub, and Reddit)

This part is for paid subscribers only, but the results are quite surprising compared with the current consensus.

Read more

In this article, I analyzed how big is Meta’s AI data center footprint and how much revenue & profit it could generate. I also shared what moves could change sentiment on the company, and on top of it, I also shared why I believe Meta is right now in a unique position (with leverage) in this AI race that has morphed into a capital race, but why Meta needs to act is weeks, not years. For Meta Compute, it’s now or never.

This article is exclusive for paid subscribers.

Let’s get into it.

How many gigawatts does Meta actually have?

Read more

You have Nvidia’s Jensen Huang with the strongest balance sheet in semiconductor history and a ~75% gross margin business. On the other side are the three biggest Nvidia customers — Amazon, Microsoft, and Google — who together will spend around $600B in CapEx this year and have all come to the same conclusion: they do not want to keep paying Nvidia’s margin forever.

In the last few years, the relationship was largely defined by Nvidia being the only game in town for AI chips, and most workloads were used for training. But now things are changing, because:

• we are shifting to an inference-led market,

• Nvidia has a historic high 75% gross margin, and it is bothering the data center ecosystem because it’s long-term hurting their margin

• The stakes have just become so much bigger because hundreds of billions, and maybe even soon trillions, per year will be spent on AI chips, and every layer of the AI stack wants to capture as much value as possible.

Let’s dive in.

Training the land grab. Inference the economy.

Most of you know this well, but just to shortly summarize. Training is when you build the model. You take a giant cluster, you run it for months, and you need the best interconnected hardware because the whole cluster works as one. This is where Nvidia’s moat is the strongest: CUDA, NVLink, the networking stack. In training, Nvidia’s share of the merchant market is still estimated to be north of 90%; the only real competitive chip that comes close is Google’s TPU (mostly because of Google’s interconnect and clustering capabilities).

Inference is when you run the model and is different:

• It is massively parallel but doesn’t need the same tight cluster coherence as training.

• The workload is known and stable, which is exactly the situation in which a custom ASIC (a chip designed for one task) often beats a general-purpose GPU in cost per token.

• It is where most of the future volume is going. Microsoft processed over 100 trillion tokens in a quarter in 2025, up 5x year over year. Google cited 22 billion tokens processed per minute across Alphabet’s APIs.

The customers became the competitors

The problem for Nvidia in this inference world is that the market is so big that companies and their biggest customers don’t want to allow a monopolization of one of the biggest markets in the world; it’s just too much money and profit to be made in this layer, or value transfer to go above the stack for them to just leave it alone. The problem is also that the biggest customers, the hyperscalers, have very strong operating cash flows and business to actually have significant ASIC labs in their conglomerates.

Google is the furthest ahead in chip design. Google produces +3 million TPUs annually. In October last year, it signed Anthropic to a deal for access to up to 1M TPUs, bringing well over a gigawatt of capacity online in 2026. In April this year, that was expanded with Broadcom to roughly 3.5 additional gigawatts of next-gen TPU capacity starting 2027. Google is also now selling TPU systems to external customers as hardware. Essentially doing the same thing Nvidia’s core business is. Because it seems Gemini has fallen behind and the AI lab team is in a refocus period, the likelihood of Google giving even more capacity to GCP and with it TPUs to outside customers has only increased.

Amazon told you on multiple recent earnings calls that it sees its semiconductor design business as one of the next pillars of the Amazon empire. And it’s already significant in size, with big growth rates. Amazon’s chips business (Trainium + Graviton) now runs at over $25B in annualized revenue, growing triple digits. Graviton is used by 98% of the top 1,000 EC2 customers. Project Rainier - the Anthropic cluster is scaling from 500,000+ Trainium chips toward 1M. The problem is also that Amazon, as a company, often likes to drive margins to zero during adoption periods, as they have an obsession with focusing on customers before reaching scale and passing a lot of the economic benefits to their customers. Trainium, given how young the project is, is highly competitive with Nvidia on a cost-per-token basis; much of this is due to Amazon pushing for very low margins. There is a reason why Anthropic likes Trainium for inference tasks.

Microsoft was and still is the laggard, but they, too, have figured out that owning the chip design layer will be key for a hyperscaler. The Information reported that Microsoft will unveil Maia 300 as soon as September and is negotiating with TSMC for capacity for more than 300,000 chips for delivery in 2027, with an eventual target of over 1 million units and, per an Azure Maia manager, “gigawatts’ worth” of Maia. For context, Maia 200 (launched in January 2026, TSMC 3nm, 140B+ transistors) has shipped only in the tens of thousands and is housed in two US data centers. Reports mention a targeted cost advantage of 30-40% over Nvidia chips, and Microsoft is actively courting Anthropic as an anchor customer for it.

The fact that Anthropic has so far become the biggest success story from the AI labs' perspective is also not great for Nvidia, as Anthropic has notoriously built most of its chip-usage stack on Google’s TPUs and Amazon’s Trainium (not Nvidia). Yes, even Mythos (Fable) is rumored to be mostly TPUs and served on Trainium, though that’s also changing with some new Nvidia deals.

Even with Nvidia now trying hard to offer its chips to Anthropic in its deal announcements, it’s clear Anthropic doesn’t want to drop off TPUs and Trainiums. Instead, they want as many options as possible and to drive prices down, which is exactly the scenario Nvidia doesn’t want. Matching workloads to the most suitable, cheapest chip.

And I haven’t even mentioned AMD, where the OpenAI deal ( 6GW of MI-series deployment with equity warrants attached ) and the Broadcom/OpenAI 10GW custom program give buyers even more options, not to mention the SRAM players like Cerebras.

Hyperscaler comments that increased tensions

On the Q2 calls, all three hyperscalers and Meta described their CapEx in a way that Nvidia surely doesn’t like. They framed hundreds of billions in future AI spending as an option rather than a must.

They all committed to long-lead-time CapEx (land, shells, power, cooling — assets that last 15-25 years) years in advance, but said they are going to buy the chips, the “short-lived assets,” only months before a data center comes online. If demand is there, you buy them; if not, you don’t buy the chips.

Amazon, Meta, and Google essentially said the same thing. With Jassy talking in detail about “different capital cycles” for data center land, power vs chips, etc.

The fact is that as more time passes, the more mature the ASIC programs from these companies become, and the more independent they can be in that layer as they get more TSMC allocation or even strike deals with Intel or Samsung for fabrication of these chips. All of the hyperscalers are also going direct to memory manufacturers to secure memory capacity, so their long-term intentions are to essentially copy the AWS playbook on the CPU book with Graviton, where Amazon’s own Graviton chip now handles most of the CPU workloads in AWS.

Nvidia doesn’t want to wait for that and wants to front-load as much demand as possible today, when they still have a dominant market position, but they also want to secure their future and commoditize the hyperscalers.

Nvidia’s counter-move

Nvidia is on track to generate roughly $100B+ of annual free cash flow at current run rates. Nvidia is deploying it to manufacture demand outside the hyperscalers’ walls, so that when the hyperscalers exercise their option to buy less, someone else is will buy more.

Save the neoclouds at any cost. In September 2025, Nvidia signed a $6.3B commitment to buy back any unsold CoreWeave capacity through April 2032; a parallel deal with Lambda was worth $1.5B. Then, in January 2026, with CoreWeave, Nvidia went further by injecting $2B in equity. Nvidia became the equity holder, the demand backstop, and the customer of last resort.

Revenue deal. On July 1, 2026, Nvidia formalized the model as the “AI Compute Partnership”: Nvidia acts as a financial backstop, agreeing to rent back unused GPUs at a fixed rate so lenders will finance neocloud buildouts — and in exchange, Nvidia takes a share of the neocloud’s cloud revenue. First neoclouds to join Firmus (up to 170,000 GPUs in Batam, Indonesia) and Sharon AI (40,000 GB300s in Australia). The problem with neoclouds is that their margins and business models are already razor thin, especially because they don’t have massive free cash flows and they essentially need debt to keep the lights going and to finance their buildouts. Even among them, it seems like the ones taking the Nvidia revenue share deal are seen as the weakest, according to this Former Nebius employee:

He also shared his thoughts about what he thinks the Neoclouds are thinking about this Nvidia revenue share deal:

He also explained possible tension with Neoclouds and Nvidia, as, according to him, Nvidia would prefer that Neoclouds serve smaller, more diverse customers rather than just have big offloading deals with the hyperscalers.

And another move they did was on Monday with the $500B financing deal, bringing Wall Street in more deeply. Nvidia announced a partnership with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to raise more than $500B in third-party capital for AI infrastructure, with compute itself used as collateral in SPVs, and Jensen saying that Nvidia retains the option to backstop up to $125B, or 25%, of the deals.

This is in addition to the reported talks to backstop up to $250B of OpenAI’s compute leases at the SB Energy 10GW Ohio campus and to finance ~$350B of OpenAI chip purchases. Given the recent commentary from SpaceX about going “Nvidia exclusive,” I expect a backstop deal or straight-up equity deal is also brewing in the background when it comes to SpaceX and Nvidia. Elon is smart enough to take advantage of a situation if Nvidia feels it needs to help the ecosystem with financing, as he has never lacked ambition for scale and doesn’t fear pulling the risk lever to the max.

Nvidia already disclosed $18.6B invested into private companies and infrastructure funds in Q1 FY27 alone, with the filing admitting some investees “may indirectly purchase or use our products”.

The problem with these kinds of deals is that when we have a compute shortage on paper, things mostly work out, but the moment we don’t have that shortage anymore, a business that pays Nvidia list price for chips and then pays Nvidia a percentage of revenue on top is a business with no margin left. The neocloud business models, in essence, are all a question in terms of whether this is really a sustainable, durable long-term business model or if this is just a byproduct of the current times where there is not enough compute and where some of the biggest compute buyers (hyperscalers) are doing these neocloud deals to buy some short-term capacity needs while they themselves maintain the end-client relationship. Comparing a hyperscaler business model to a neocloud is, in my view, like comparing apples to oranges, as hyperscalers have the data and ecosystem stickiness needed for a durable business. Also, with their scale and now ASIC development, they will control future margins better than someone with only 1 chip supplier.

These moves by Nvidia are important because they do change the short-term nature of the compute ecosystem and could help create a real AI bubble and an overbuild in compute capacity. Nvidia is right when it comes to neoclouds: for them to have durable business models, they need smaller, more diversified clients. And even that raises a question: once the compute shortage eases, and clients can get cheap, available compute from the hyperscalers, why would they even use neoclouds?

So far, the clients are mostly big companies, AI labs, and the hyperscalers themselves, where the risk is even higher when we do get into a capacity overbuild. The first thing hyperscalers will switch off is Neocloud deals, and the big AI labs are already building their own data centers to become independent of both Neocloud and the hyperscalers. The strategy that Nvidia is trying to encourage with Neoclouds about focusing on non-hyperscalers is a good one for sustainability, but the Neoclouds also like headline-reaching deals with the hyperscalers as it helps their stocks in the short-term.

To continue reading this article, you have to subscribe to the paid section. In the paid section, I will cover why Nvidia needs to finance these deals and how Nvidia, Neoclouds, and hyperscalers are positioned for the future.

Why does a potential monopolist need to finance its own demand?

Read more

We just got Q2 2026 earnings from Google, Microsoft, Meta, and Amazon. I want to share what I think are the most important takeaways.

For more than a year, the number one pushback I hear from investors on big tech is that the CapEx is out of control, and the returns won’t be there. This was the first earnings season where all four companies directly and, in my view, successfully answered that worry. Not by cutting CapEx; Amazon raised its 2026 number to ~$220B, Google raised to $195-205B, Meta lifted the floor of its guide to $130-145B. For the first time, all four laid out the same playbook on how they de-risk that spend: commit early to the long-lived assets (land, data center shells, power), and decide on the short-lived assets (the chips, which are the majority of the cost) only a few months before they need them, once they can actually see the demand. If the demand isn’t there, they don’t order the chips.

And on top of the playbook, all four went deep into their ROIC math. After I went through the fourth call, it felt like coordinated messaging in the sense that every management team knew exactly what question the market is asking, and they came prepared, maybe even spoke to each other on a common messaging regarding CapEx ROIC.

There are a few patterns from this earnings season:

• Nobody is guiding CapEx down, but everyone is now explaining CapEx as a two-part investment: flexible long-lived assets committed early, chips ordered just-in-time against visible demand

• AI workloads are still a relatively small % of total cloud revenue (AWS: ~$25B AI run rate vs. $169B total), but they are pulling traditional workloads up with them

• Cloud margins at AWS, Azure, and Google Cloud expanded again - the “AI CapEx has low ROIC” thesis is basically dead

• Microsoft’s Copilot finally changed momentum, as net adds more than double QoQ

• Meta is building four new revenue lines on top of ads: transactional, compute, APIs, and subscriptions

Let’s get into it.

Land And power first, chips a few months before you need them

I want to start with this theme because it showed up on all four calls and that is the framing of the AI investments.

Amy Hood at Microsoft:

Amazon’s management gave the most detailed version of the same framework:

Data center shells monetize for 30+ years. Servers break even in under 3 years, have a 5-6 year useful life, and most AI capacity is contracted on 5-year terms before the servers are even bought.

Susan Li at Meta said the same thing on the call:

And Google’s CFO Anat Ashkenazi framed the whole thing through an ROIC and pricing lens:

The chips are the majority of the CapEx dollar and the fastest-depreciating part, and all four companies are now telling us the chip orders are placed months ahead of deployment, against demand they can already see (Amazon said the lion’s share of 2027 capacity is already reserved, and “quite a bit” of 2028). So the doomsday scenario where hyperscalers wake up with hundreds of millions of stranded GPUs requires demand to disappear inside a one-to-two-quarter ordering window. That is a very different risk profile than the market narrative.

The Google comment about pricing is the first hint of what happens if component costs (memory above all - Amazon explicitly raised its CapEx guide from ~$200B to ~$220B because of »the higher cost of memory«) keep rising: prices for cloud and AI services go up. In a supply-constrained market, the hyperscalers have serious pricing power. Ashkenazi even added that Google is more confident on returns than 12 months ago: »And so I think if anything, the dynamics look healthier than where we were about a year ago.«

And for those worried about reckless spending and equity issuances, Google was more explicit:

Google Cloud +82%, Operating Margin from 20.7% to 35.6%

Google Cloud revenue grew 82% YoY to $24.8B in Q2 (from $13.6B a year ago), with management noting that »GCP grew faster than Cloud overall«.

But the number I care most about is the margin. Cloud operating margin came in at 35.6%, up from 20.7% in Q2 of last year, with operating income roughly tripling YoY. This is the second consecutive quarter where Google Cloud posted significant margin expansion while AI workloads became a bigger share of the mix. If AI workloads had structurally lower margins than traditional cloud workloads, this line would not be expanding this significantly.

The backlog also keeps compounding: $514B, up ~$50B sequentially. And, importantly, management repeated the framing that the demand is near term:

On CapEx, Google raised the full-year guide to $195-205B from $180-190B (Q2 CapEx alone was $44.9B, roughly double YoY), and management was direct that the increase is capacity, not inflation:

Meta and Amazon both called out memory prices as a driver of higher CapEx numbers. Google is saying ours is going up because we are delivering more capacity.

Microsoft Azure crosses $100B, and Copilot finally has momentum

Microsoft closed its fiscal year with Microsoft Cloud surpassing $214B in revenue, up 27%, and Azure surpassing $100B for the first time, up 41% for the year. In the June quarter itself, Azure grew 43%, ahead of guidance, and Microsoft guided approximately 45% constant currency growth for the September quarter - meaning Azure is accelerating further.

Two numbers from the call deserve special attention. First, this one:

The biggest bear case on Microsoft for the past year has been “it’s all OpenAI.” Nearly 90% of a $214B cloud business is not OpenAI.

Second, the RPO disclosure, which was unusually detailed this quarter:

A $678B contracted book with a weighted average duration of only 2.3 years, where the portion converting in the next 12 months is itself growing 37% YoY.

Now Copilot. I have been fairly critical of the Copilot story over the past two years because seats were growing but engagement was questionable. This quarter, the momentum changed:

For context on the trajectory: 15M paid seats in the December quarter, 20M in March, 30M+ now. Adding 10M+ paid seats in a single quarter is the fastest ramp in the product’s history. And the engagement data supports the seat data:

Engagement on par with Outlook and Teams means Copilot is becoming a habit with users. And the revenue line is following: »Copilot revenue accelerated over 60% quarter-over-quarter.«

Microsoft, after two years of trying, finally looks like it is on the right path with Copilot, and I think the E7 suite plus the shift toward usage-based billing (the June GitHub Copilot business model change already improved segment gross margins through the quarter) is a big part of why.

One more strategic point that I think is underappreciated. Satya’s framing of the model layer:

Combined with the disclosure that customers building with models from multiple providers is up 5x since the start of the year, Microsoft is positioning the orchestration layer, not any single model, as the durable asset.

Microsoft is also extending the estimated useful life of its data centers and office buildings from 15 to 25 years. The bigger effect is that more future data center leases will shift from finance leases (which are counted in CapEx) to operating leases (which are not). So reported CapEx optics will change, and the »over $50 billion« CapEx guide for the September quarter already includes this reclassification impact.

Even with all this investment, Hood guided full-year operating margins down less than 1 point and said Microsoft expects to remain free cash flow positive in FY27 - which, given what Alphabet’s and Meta’s free cash flow just did, is a differentiator.

Subscribe now

Amazon AWS reaccelerates to 36.7%, but the margin is key

AWS grew 36.7% YoY, accelerating for the fifth straight quarter, its fastest growth in 18 quarters - back when, as Jassy pointed out, AWS was less than half its current size. AWS added over $4.6B in revenue quarter-over-quarter, about 80% more than its largest sequential increase ever, and now runs at a $169B annualized revenue rate. Backlog stands at $496B, growing triple digits YoY.

AI is a ~$25B run rate inside a $169B run-rate AWS. Roughly 15% of the business. AI is still small compared to all workloads, but it is accelerating traditional workloads:

This is the point I keep coming back to: agentic AI doesn’t just consume GPU cycles. An agent that executes a multi-step workflow hits databases, storage, networking, orchestration, and a lot of plain CPU compute. So the $25B of AI revenue is not just $25B - it is the accelerant for the other $144B. AWS said it directly:

And with Graviton used by 98% of the top 1,000 EC2 customers, and Graviton revenue commitments up nearly 3x quarter-over-quarter (Graviton5 ramping nearly 2x faster than Graviton4 did), AWS captures that CPU pull-through on its own silicon economics.

Now the margin, which for me was the single most important data point of the entire earnings season. AWS operating income was $16.6B in the quarter at a ~39% operating margin, up 650 basis points YoY. There was a one-off: a ~$600M benefit from fair value changes on energy contracts subject to derivative accounting. The CFO stripped it out: excluding it, margins were still up 520 basis points YoY.

And on the AI workloads specifically:

Sit with that. AWS is in the heaviest investment cycle in its history (Amazon now guides approximately $220B in cash CapEx for 2026, up from the prior ~$200B estimate on higher memory costs), AI is scaling to a $25B+ run rate growing triple digits, and the segment operating margin expanded 520bps YoY excluding one-offs - to a level above where AWS was before the AI cycle even started. The thesis that AI data center CapEx structurally carries low returns is so far being proven wrong.

Which is why Jassy felt comfortable saying this:

Two quick additional notes. First, Trainium demand keeps broadening beyond the anchor deals: »In addition to the 2 leading AI labs in the world, Anthropic and OpenAI, making multiyear, multi-gigawatt commitments to Trainium, an increasing number of AI startups are also adopting Trainium« - the list now includes NEURA Robotics, Odyssey, TwelveLabs, Decart, Poolside, plus larger companies like Uber and Pinterest. Second, Jassy repeated his view that »there is not going to be one model to rule the world« and that AWS can be wildly successful without its own frontier model - the Bedrock + SageMaker positioning (technically competent companies building their own smaller models on proprietary data) is the same “the platform is the moat” bet Microsoft is making with Foundry.

Meta: The ad business expands further and four new revenue lines emerge

Meta grew total revenue 28% YoY to $60.8B, with Family of Apps revenue at $60.4B. Instagram reached 2 billion daily actives. The reported operating income optics were messy because of $2.4B in legal charges and $1.18B of severance; excluding those, operating income would have grown 9% YoY per the company’s own math. CapEx was $31.1B in the quarter (nearly double YoY), and the full-year guide was narrowed upward to $130-145B.

The AI-on-the-core-business flywheel is intact:

But the reason I think this Meta quarter deserves a separate look is not the ad business. It’s that Meta laid out a concrete map of new revenue lines beyond advertising, and it’s worth listing them explicitly because I believe several of them will be reported segments within two years:

1. Transactional / business services revenue. This is the most Meta-native one. Zuck:

“Businesses only pay us when we achieve results” is the ad auction logic extended to commerce and operations.

2. Compute revenue - direct.

And the mechanism Zuck laid out is genuinely novel:

An auction for compute. Meta runs one of the most sophisticated real-time auctions in the world for ad impressions; running the same machinery over GPU/accelerator capacity is a very natural extension. And the commentary about the trade-off between monetizing today versus building for the future reads, to me, like Meta is preparing a direct compute deal with a third party in the near future.

3. Enterprise / API revenue.

   »We see a large enterprise opportunity to sell to businesses, including APIs, business agents, potentially selling compute directly and other services that we’re building for large customers.«

   Business agents are already being used by over 1 million businesses weekly.

4. Consumer subscriptions. First evidence it works: Family of Apps “other” revenue crossed $1B in a quarter for the first time, up 73% YoY, driven primarily by WhatsApp paid messaging and subscriptions, but the potential here for expansion is enormous as Meta hasn’t laid out the killer products yet.

And the distribution advantage that ties all four together:

Threads went from zero to a top-tier social app almost entirely on Instagram’s distribution. Meta is telling us it plans to run that playbook a dozen more times in the AI application layer, because AI is redistributing the playing field and shipping apps is getting cheap.

The last quote I want to highlight is Zuck on why Meta builds the full stack rather than renting frontier models:

Zuck is saying that relying on another lab’s model is a policy and business risk Meta will not take, and that full-stack ownership enables product experiences others mostly cannot build. You can agree or disagree on the cost of that choice (a substantial amount of Meta’s compute goes to training to remain a leading lab, which is exactly why the free cash flow line looks the way it does right now), but it is a fair strategy, and Zuck closed the loop on it himself:

He said almost exactly the same thing in the 2022-2023 drawdown. That worked out more than okay.

My views on these companies going forward

I think

Read more

Posting my regular channel check & other alternative data report before we start the big tech Q2 earnings.

In this report, I covered the cloud hyperscalers and what I see from aggregating multiple alt providers, and share some of my views on the upcoming earnings. In this report, I also share some data on the ad tech and software sector, as well as my general view on how things might play out this earnings season.

Particularly on the cloud side, some of my data points to a significant acceleration from previous levels for one specific cloud provider.

Let’s dive in.

Cloud hyperscalers

First of all, I should note that cloud capacity remains sold out at the hyperscalers, and alt data continues the trend seen in previous quarters. If we first look at the sellside data, we can see that sellside is optimistic about cloud revenue growth for Q2, with expectations mostly on the high end of the guidance ranges provided by cloud providers. By conducting my analysis of the most relevant expert interviews and channel checks from platforms like AlphaSense, we get more granular data.

The sentiment from expert interviews this quarter is very bullish, with public cloud growth accelerating substantially in Q2 2026 vs Q1 2026. In many cases, the growth estimates went from 10-20% YoY in Q1 2026 to 25-35% YoY for Q2 2026. Strong usage is seen across the board, but the standout company for this quarter in terms of rate of change from the previous quarter is

(Subscribe to Paid to see the full article, explaining which cloud provider momentum looks the best in Q2, Microsoft Copilot adoption, and Ad Tech growth, particularly Meta)

Read more

A few weeks ago I wrote a piece called Most of the Economy Won’t Run on the Best Model, where I argued that the AI market will bifurcate: the frontier model goes to the small slice of work where intelligence is unbounded in economic value (drug discovery, novel math, the hardest agentic reasoning), and the middle of the economy — classification, extraction, summarization, routine code, support — runs on the cheapest model that clears the quality bar.

This article is the sequel to that one. Who captures the value when the world switches from token maxing to token optimization?

The shift away from always-buy-the-best-model is, on its surface, bearish for the AI labs and looks like it should compress the whole stack. But it is quietly one of the most bullish structural setups for the three hyperscalers — Microsoft, Amazon, and Google.

Let me explain why.

Think about a highway toll road. There are two businesses operating on it. The first is the company that manufactures the cars — they make the actual machine that does the work of getting you somewhere, and there’s a real margin in a car. The second is the company that owns the tollbooth. The tollbooth owner doesn’t care what car you drive. Ferrari, Toyota, or a 12-year-old used Honda. Every one of them pays the same toll to cross the bridge. Now imagine a world where, suddenly, everyone realizes they were commuting to work in a Ferrari for no reason, and they all downgrade to the cheaper Honda to save money. The carmaker’s revenue per vehicle collapses, but people don’t drive less because they switched to a cheaper car. They drive more, because now it’s cheap enough to justify trips they’d never have taken before. And every one of those extra trips still crosses the bridge. The tollbooth’s revenue goes up.

In the AI stack, the AI labs are the carmakers. The hyperscalers are the tollbooths. And in the next months, we are entering a period where most of the economy is trading down from the Ferrari to the Honda, while simultaneously driving 10x more miles.

From token maxing to token optimization

For the last 18 months, the dominant behavior in enterprise AI was token maxing. You found the single best model on the leaderboard, you pointed every workload at it, and you didn’t think too hard about cost, because the whole thing was a pilot and the bill was small relative to the perceived upside of “does this even work?”

That era is ending fast. Companies are essentially blowing through their AI budgets in a quarter. Altman said recently that the “my company spent my entire 2026 budget in Q1, can you make this more efficient?” complaint went from something that “never came up” to “all of a sudden a huge issue.” And it’s not just him; you can see it across the industry, from companies like Salesforce and Meta to many other smaller ones blowing their planned yearly budget in a matter of days.

The natural behavior change from this is that instead of having one model for everything, companies start routing: a small, cheap, often open-weight model handles the 80% of requests that are less complex, and only the genuinely hard requests escalate to the frontier. This is token optimization, and it doesn’t reduce total token consumption — it accelerates it. The moment inference gets cheap enough, you stop rationing it. You run the agent in a loop. You let it read the whole codebase. You re-run it five times and vote on the answer. A single coding-agent session now chews through millions of tokens of context where a chatbot query used a few thousand.

You can see this in the hard numbers the hyperscalers themselves disclose. Microsoft said it processed over 100 trillion tokens in a single quarter in 2025, up 5x year-over-year, with a record 50 trillion in one month alone. By its fiscal Q3 2026 call, Microsoft said over 300 customers were on track to process more than a trillion tokens each on Foundry this year — and that this was accelerating 30% quarter-over-quarter. Google went from 480 trillion tokens per month at I/O in May 2025, to 980 trillion by July, to 1.3 quadrillion by October — and in its Q1 2026 filing disclosed that its first-party models alone were processing more than 16 billion tokens per minute via direct API, up 60% in a single quarter.

At the same time, the price per unit of capability is falling substantially— the Stanford HAI AI Index found inference cost for GPT-3.5-level performance fell more than 280-fold in two years, and a16z pegs the decline at roughly 10x per year for any fixed capability level. And yet, total tokens processed are growing several-fold per year.

Where the margin lives in a routed token economy

When a company calls the SOTA model directly through, say, an AI lab’s first-party API, the lab captures the full economic rent of the token. There are many industry reports surfacing around the margins that providers like Anthropic have right now, and many of them claim that the margins went substantially up this year to as high as 70% gross margins. That price embeds the lab’s R&D, its brand, its leaderboard position — call it the “model-provider margin.”

And to be fair, when token optimization happens, there might not be a direct hit to margins in the short-term at these frontier labs (as there will always be use cases for the best AI model), but it does mean that growth acceleration becomes smaller than it would be if everyone stayed for every use case on the frontier path.

The company routes that same workload to an open-weight model — GLM 5.2, DeepSeek V3.2, Qwen3 Coder, Kimi K2.5, Llama, MiniMax M2.5 — running on a hyperscaler’s managed inference. The “model-provider margin” essentially goes to zero, because the model is open-weight and nobody is charging a brand premium for it. But the token still has to run on somebody’s GPUs, inside somebody’s data center, behind somebody’s managed API with its security, compliance, logging, and SLAs. And that somebody are the hyperscalers. They still charge their full infrastructure margins on the token. AWS has historically run a roughly 35–38% operating margin. Google Cloud, which lost money for years, posted operating margins above 33% in Q1 2026 and is still climbing. That margin doesn’t care whether the token came from a $50/million frontier model or a $1/million open-weight model. To return to the analogy that from before, the tollbooth charges the same toll regardless of the car.

So here is the structural beauty of it for the hyperscalers, and the structural danger for the labs:

• Per-token economics compress, but the compression lands almost entirely on the model layer, not the infrastructure layer. The AI lab’s margins on simple workloads get squeezed. The hyperscaler’s infrastructure margin is sticky.

• Total token volume explodes, and almost every single token crosses the hyperscaler’s tollbooth. More usage, on a partly-depreciated, increasingly efficient installed base, means absolute infrastructure revenue and gross profit dollars go up even as the price of any individual token falls.

This is Jevons’ paradox pointed directly at the cloud P&L. The hyperscalers squeeze more revenue out of the infrastructure they already own, and they don’t need the model-provider margin to do it. Outside of Google, they were never in that business in the first place.

The orchestration layer brings real value

If the future is multi-model — and it clearly is — then someone has to own the orchestration: the layer that decides which model handles which request, holds the fine-tunes, runs the agent loop, manages memory and tool-calling, handles fallback when a model is down. This layer can bring immense value to those who own it, and all three hyperscalers are naturally positioned to become just that.

Amazon’s Bedrock is the clearest example. It’s no longer “a place to call Claude.” As of 2026, the catalog spans 18 providers and 110+ individually addressable model variants, and AWS bolted on Intelligent Prompt Routing, which automatically routes each request to the cheapest model in a family that can handle it — AWS claims up to 30% cost savings with no accuracy loss. On top of that sits Bedrock AgentCore, a production agent harness with built-in Runtime, Memory, Gateway, Browser, Identity, and Observability. Bedrock reached a multi-billion-dollar annualized run rate with customer spend growing 60% quarter-over-quarter across 100,000+ customers.

Microsoft’s Azure AI Foundry and Google Vertex are the same play, although Google’s preferred scenario is where the Gemini model family dominates the workloads.

The natural position for hyperscalers to be the orchestration layer is great because companies already have their data, security perimeter, billing, and compliance live in the cloud environments. They can pick from a menu of models.” The lab moves from the front end — the thing the customer chooses and bonds with — to the back end, an interchangeable component sitting behind the hyperscaler’s harness. Customers also start building the fine-tunes, the RAG pipeline, the agent scaffolding, the eval suite at the cloud orchestration layer, not at the model layer; switching models becomes something normal and not a migration task. One of my readers, who runs his own harness, put it perfectly in the comments of my last piece: the harness is worth as much as the model itself, and you’re paying for it either way.

The moat for the labs becomes shallow with most use cases but still remains important at the most complex ones (the ones where economic upside is not capped). But for hyperscalers, they benefit in all use cases that run the economy.

One might argue why companies couldn’t move even outside of the cloud, but the problem is here that besides the normal reasons that people migrated a big part of workloads from on-prem to cloud (easier to scale, use, collaborate, manage), now on top of those reasons, you also have the problem compute shortages where infrastructure outside of cloud environments is even harder to get and managing that infrastructure becomes a much more difficult job. But an important aspect, I think, that is forming is also the cybersecurity one. Given the recent developments here, it is clear that only a handful of companies will have first-row access to the most capable cyber models, to first shield their systems before the model is then available for everyone else. If you are an enterprise, you want safety, and it seems the only way for you to get it is if you have your infrastructure at one of those preferred partners that get first-row access to cyber-capable models before everyone else does. The hyperscalers are those partners.

The hyperscalers’ job, then, is almost simple to state: enable as many models as possible, make routing and fine-tuning and observability frictionless, and let customers optimize tokens to their heart’s content. Every model they add makes their tollbooth more valuable.

The government approval window just made this worse for the labs

Now layer on a development from June that I don’t think the market has connected to this thesis at all: the June 2, 2026 executive order: Promoting Advanced Artificial Intelligence Innovation and Security.

The headline version is that it sets up a voluntary framework where developers of “covered frontier models” — designation determined by the NSA via a classified benchmarking process focused on cyber capabilities — give the federal government access to the model for up to 30 days before release to “other trusted partners.” (The earlier draft had a 90-day window; it was cut to 30 to avoid blunting U.S. competitiveness.) The order pointedly does not create mandatory licensing — but it establishes a structured pre-release evaluation pathway, an AI cybersecurity clearinghouse run out of Treasury, and a government role in selecting which “trusted partners” get early access. The catalyst, per CFR’s reporting, was rising concern about models like Anthropic’s Claude Mythos being able to autonomously find and exploit software vulnerabilities. Commerce ordered Anthropic to cut off non-U.S. access to its Mythos 5 and Fable 5 models on export-control grounds, and those models were pulled from Bedrock days after launch. Now there is growing concern that, even when models are released to the public, U.S. citizens may get access before people from other countries. Hyperscalers have clients worldwide, so it becomes even more important for them to offer multiple models to companies and for companies to have an orchestration layer that isn’t just one AI lab, since they will have to juggle geopolitical compliance as well.

It strengthens the case

Read more

In the AI compute buildout phase, especially for inference, which is where the industry is shifting most of its capex and where the actual revenue gets generated, the binding constraint is increasingly not compute. It’s memory. Capacity and bandwidth.

I covered the supply side of this story in my memory cycle article ”Every Memory Cycle Ends the Same. Until It Doesn’t,” where I argued that HBM has turned memory from a gadget component into a raw input for intelligence. In that article, I also wrote that the real risk for the memory cycle is a technical breakthrough that would require orders of magnitude less memory. Today’s article is about the other side of that exact coin: what the labs and inference providers are doing in software and model architecture to need less memory per token, how much of that optimization potential is already captured, and what it means for the hardware stack, including some unconventional setups like using depreciated H100s and A100s as dedicated decode machines.

The reason this matters now and not in two years is simple: companies are starting to hit their token spend limits. Agentic workloads (coding agents, research agents, computer-use agents) consume tokens at a rate that makes the chatbot era look like a rounding error. A single coding agent session can chew through millions of tokens of context and companies are increasingly becoming frustrated with it.

In this article, I cover software optimization techniques for lowering HBM usage and give my estimates on how much of that optimization is already captured, give my view on the best solution, and how much reduction it could offer, and also cover hardware angles from SRAM accelerators like Groq and Cerebras, as well as the decode and prefill split up of hardware and what that opens up.

Let’s start.

Prefill and decode: the two jobs inside every AI request

To understand why memory is the bottleneck, you first need to understand that every LLM request is actually two completely different workloads stapled together: prefill and decode.

Prefill is what happens when you send the model your input: the prompt, the document, the codebase, the conversation history. The model reads all of it and processes every input token in parallel, in one big pass. Think of it as an analyst who gets handed a 300-page data room before a meeting. He reads the whole thing in one sitting, takes structured notes on every page, and files those notes away. This is brute-force work; the limiting factor is how fast his brain works, not how fast he can pull pages out of the binder. In GPU terms, prefill is compute-bound: the chip’s arithmetic units (the FLOPs) are the bottleneck, and the memory system can keep up.

Decode is what happens when the model generates its answer, one token at a time. To generate each single token, the model has to read essentially all of its weights from memory, plus all of the notes it took during prefill (the so-called KV cache, more on this later in the article). Then it does a comparatively tiny amount of math, produces one token, and does the entire memory read again for the next token. Our analyst is now in the meeting, and before he speaks each individual word, he has to re-skim his entire stack of notes. The bottleneck is no longer his brainpower. It’s how fast he can flip pages. Decode is memory-bandwidth-bound.

You can put numbers on this. An Nvidia H100 delivers roughly 989 TFLOPS of dense BF16 compute against 3.35 TB/s of HBM3 bandwidth. That ratio means the chip needs to perform roughly ~295 floating point operations for every byte it pulls from memory just to keep its compute units fed. Prefill, processing thousands of tokens in parallel, easily clears that bar. Decode doesn’t come close: roofline analyses of autoregressive generation put its arithmetic intensity at roughly 1 FLOP per byte, about two orders of magnitude below the compute-bound ridge point. In plain English: during decode, the most expensive compute engines on the planet sit idle 95%+ of the time, waiting for memory.

The simplest illustration: a 70B parameter model in FP16 is ~140GB of weights. On an H100 with 3.35 TB/s of bandwidth, the theoretical single-user decode ceiling is about 3,350/140 = 24 tokens per second. On an A100 with 2 TB/s, it’s about 14 tokens per second. Notice what’s not in that equation: FLOPs. You could double the H100’s compute and single-stream decode speed wouldn’t move at all. The only lever is bandwidth, or reading fewer bytes.

This is the “memory wall,” and it’s also why a lot of inference revolves around batching. If reading 140GB of weights produces one token for one user, that’s terrible economics. If the same 140GB read produces one token each for 200 users simultaneously, your cost per token just dropped ~200x. The weights are read once, amortized across the batch. So the game every inference provider plays is: cram as many concurrent users as possible onto each GPU. And what limits how many users you can cram on? Memory capacity. Because every user brings their own luggage.

The luggage: KV cache, and why agents made it explode

That luggage is the KV cache. During prefill, the model stores intermediate “key” and “value” representations of every token in the context, so that during decode it doesn’t have to re-process the whole prompt for every new token. Those are the analyst’s notes. The catch: the notes grow linearly with context length, and they have to sit in the same precious HBM as the model weights.

In a standard transformer, Llama 3.1 405B needs 516 KB of KV cache per token of context; Qwen-2.5 72B needs 327 KB per token. Run that forward: a single user with a 128K-token context on a 70B-class model is carrying roughly 40GB of KV cache. At one million tokens of context, a Llama-70B-scale model would need ~330GB in BF16 for the KV cache alone, which doesn’t fit in any single GPU. For reference, the H100 has 80GB total. The B300 has 288GB.

In the chatbot era, this was manageable because most conversations were a few thousand tokens. Agentic workloads change that. An agent doing a long coding task holds the repo, the tool outputs, the execution traces, the full plan, all of it, in context, for hours. Context lengths of 100K-1M tokens went from research demo to daily production workload. And every one of those tokens occupies HBM for the entire duration of the session. The KV cache, not the model weights, becomes the dominant consumer of memory. Which means batch sizes collapse, which means the weight-read amortization collapses, which means cost per token explodes.

So what is the industry doing about it? A lot, actually. Let’s go through the software optimization stack, and, importantly for investors trying to model how much efficiency is still on the table, my estimate of how much of each technique’s potential has already been captured.

The optimization stack: where we are on each curve

A quick framing note: the percentages below are my own estimates of “captured potential” at the frontier (the major labs and serious inference providers), based on what’s publicly documented. The long tail of enterprise deployments is far behind the frontier on all of these, which is itself an investment-relevant point: there is a lot of “free” efficiency still sitting unused in corporate AI deployments.

1. Continuous batching

Read more

I want to share some of my thoughts on why I think a significant change is going to happen in the AI industry, and the market is blind to it. It’s a thesis about where the money in AI actually goes once we transition to scaled-up AI workloads and how that might look different from today’s expectations.

Let me start with an analogy that I keep coming back to.

When a company hires an accountant, it does not go out and hire a PhD in pure mathematics to reconcile the ledgers. Not because the PhD couldn’t do it — they obviously could, and probably faster — but because it makes no economic sense. The PhD is overqualified, which is just another way of saying they are too expensive for the value the task produces. The economic output of bookkeeping is capped. There is only so much upside in getting the books done. So you hire the cheapest person who clears the quality bar, and you pocket the difference. And you can take this analogy and apply it to multiple other jobs.

Now flip it. If you are running a drug-discovery program, you absolutely want the PhD — in fact you want five of them, plus a Nobel laureate consulting on the side. Why? Because the economic output of a single discovery is enormous, almost unbounded. The expected value of a breakthrough is measured in tens of billions, so the cost of the smartest possible person working on it rounds to zero against the prize. Here, intelligence is the only thing that matters, and cost is an afterthought.

This is, I think, exactly how the AI model market is going to bifurcate. And we are right at the inflection point where it starts to happen.

The metric is not intelligence. It’s intelligence per dollar

Today, essentially everyone uses the state-of-the-art (SOTA) model for everything. You want to summarize an email? SOTA model. Classify a support ticket? SOTA model. Extract three fields from an invoice? SOTA model. We do this for one simple reason: the frontier models have only just crossed the threshold of being broadly truly impactful for knowledge work, and when something has only just started working, you reach for the best version of it you can find. You don’t optimize cost on a capability you weren’t sure you had last quarter.

But I believe this is a transitional behavior, not a stable equilibrium. And there are a few reasons for that.

The Stanford HAI AI Index found that the inference cost for a system performing at GPT-3.5 level dropped more than 280-fold between November 2022 and October 2024 — from roughly $20 per million tokens to about $0.07. Andreessen Horowitz, looking at the same phenomenon across the whole performance spectrum, concluded that for a model of equivalent performance, cost falls by roughly 10x every year — faster than compute fell during the PC revolution, faster than bandwidth fell during the dotcom build-out. Epoch AI, slicing it by benchmark, found the price to hit GPT-4-level performance on PhD-level science questions fell by about 40x per year, with the range across benchmarks running anywhere from 9x to 900x annually. We just had Sara Friar, OpenAI’s CFO on the All in conference say the following:

Pick whichever number you find least aggressive. They all say the same thing: the capability you are paying a premium for today becomes nearly free in about a year.

On top of it, many companies and enterprises are starting to burn through their annual planned token consumption in just a few months. This is a trend that is accelerating, and I have been hearing it all across the industry. Yesterday, there was a comment published from Sam Altman saying:

What this means is that companies have no choice but to optimize costs, and that will soon mean using models other than the SOTA model for specific tasks.

The second force is that the frontier itself is getting smaller, not just cheaper. Epoch AI has pointed out that frontier models are now roughly an order of magnitude smaller in parameter count than GPT-4 was, because once inference becomes the dominant cost, you stop training huge models and start over-training small ones on far more data. Distillation compounds this: a teacher model’s capability gets compressed into a student, a fraction of its size. This is exactly what Meta is doing internally when using their AI models to power their ads and content platform. The student model is the one that is applied at scale, and it distilled its knowledge from the teacher model.

So putting all of these together. We have a rapidly falling price for any given level of capability and frontier that is already shrinking in size in terms of what is actually being deployed, and we have companies burning through their annual token budgets in a matter of months.

As such, I believe that for the overwhelming majority of economically valuable knowledge work, the correct model is not the SOTA model. It’s the cheapest model that clears the task’s quality bar. And as pilots move into full production (which is the stage we are in today) — where you’re suddenly paying for millions or billions of tokens a day instead of running a demo — intelligence-per-dollar becomes the only metric that survives contact with a CFO.

At the same time, the SOTA model and its use case don’t disappear. It goes where the economic ceiling is unbounded: frontier R&D, drug discovery, novel mathematics, and the hardest agentic reasoning chains. But that is a smaller slice of the token volume in terms of our current economy. The accountant’s quadrant — classification, extraction, summarization, routine code, customer support, the boring profitable middle of the economy — is where the majority of tokens actually are, and that quadrant is going to run on cheaper, distilled, often fine-tuned, frequently “older” models.

Subscribe now

The investment angle: the money moves to the owners of installed compute

If the thesis above is right, where does the capital go?

The intuitive answer — the one the market is currently screaming — is “buy the picks and shovels.” Buy Nvidia, buy Broadcom, buy the ASIC co-designers, buy memory, anything that sells new compute. But my view is that while that sector still might do well, there is a different part of the tech stack that will benefit even more, looking at the rate of change from the current state.

The sellers of new compute (semis) are only winners in a world of continued high-cadence spending on new compute. And my thesis specifically questions whether that cadence is necessary. So let me lay out the two states the world can be in, because the asymmetry between them is the whole argument.

Scenario 1: Capex falls or stabilizes. If you can squeeze an order of magnitude more useful tokens out of the hardware you already own — because models got smaller, cheaper, more efficient and verticalized — then you no longer need to spend $100bn+ every single year just to stay relevant. In this world, the owners of the installed base win and the sellers of new compute lose. Hyperscaler free cash flow inflects sharply upward, because capex was the one thing suppressing it. Multiples re-rate higher as the cloud business converts from a capex incinerator into a cash machine running largely paid-for, partly-depreciated hardware. And the semis de-rate, because the market finally realizes the upgrade treadmill has slowed.

Scenario 2: Capex stays high — and revenue explodes. This is the Jevons-paradox-on-steroids case. Demand is so strong that hyperscalers do both: they extract enormous output from cheap, long-lived existing hardware and keep buying new gear. Here everyone wins at once — but the hyperscalers win more, because their incremental revenue now lands on a cost base that is partly depreciated and dramatically more efficient per token. Operating leverage goes vertical.

The interesting thing is that the market is currently priced for neither. The market is currently pricing only a future in which CapEx continues to grow for the foreseeable future and the semiconductor industry benefits, but at the same time, the hyperscalers are making a losing bet with spending on this CapEx, as the market is questioning the return on that spend.

I think this market premise is very wrong, as we are actively transitioning to production-scale AI workloads where the economics are different from those in the pilot world, where we mostly lived for the last few months.

As always, I hope you found this article valuable. I would appreciate it if you could share it with people you know who might find it interesting. I also invite you to become a paid subscriber, as paid subscribers get additional articles covering both big tech companies in more detail, as well as mid-cap and small-cap companies that I find interesting.

Subscribe to Paid

Thank you!

Disclaimer:

I own hyperscaler Meta (META), Amazon (AMZN), Microsoft (MSFT), Google (GOOGL) stock.

Nothing contained in this website and newsletter should be understood as investment or financial advice. All investment strategies and investments involve the risk of loss. Past performance does not guarantee future results. Everything written and expressed in this newsletter is only the writer’s opinion and should not be considered investment advice. Before investing in anything, know your risk profile and if needed, consult a professional. Nothing on this site should ever be considered advice, research, or an invitation to buy or sell any securities.

I want to walk through what is, in my opinion, an important structural shift happening in AI right now.

For years, the way we value AI model providers is via benchmarks based on the raw performance of these models. The narrative has been “whoever has the best benchmark model wins”

I think this framing is increasingly becoming wrong. The real moat that the frontier labs are building right now is not the model. It is the harness around the model and the cost of serving the model. In this article, I will focus on the harness. And once you understand what a harness is, you start to see why Anthropic’s enterprise revenue keeps compounding even when its raw benchmark scores are not always the best, and why the labs that own both the model and the harness are setting themselves up for the kind of platform lock-in that historically produced high gross margin businesses.

Let start.

What is a harness?

You can picture a harness as this entire system that wraps the model and turns it from a text generator into something that can do real work. Tools, memory, system prompts, permission policies, sandboxes, subagent dispatch, context management, the agent loop itself.

The cleanest one-line definition comes from LangChain: “Agent = Model + Harness. If you’re not the model, you’re the harness.”

A useful analogy is to think of a model as a very smart contractor who can only communicate through notes. A harness is the tool setup — the computer, the phone, the filing cabinet, the rulebook, the permissioning — that lets the contractor actually do the job.

Mechanically, even the most sophisticated harnesses are just a while loop. The model emits a tool call, the harness executes it, the result gets fed back, the loop continues. Anthropic literally calls their runtime a “dumb loop” where all the intelligence lives in the model. But the complexity is in everything that loop manages: which tools exist, what their schemas look like, what gets injected into context on each turn, what gets compacted away, what gets remembered across sessions, how subagents are spawned, how errors are recovered, when the model is allowed to take a destructive action without asking. None of that is “the model.” All of it determines whether your agent actually finishes the task or burns 100,000 tokens going in circles. This has effects both on the quality of the output and on costs.

The empirical evidence: the harness is moving benchmarks even more than the model is in some cases

We now have some hard data showing that the harness can move model performance by more than a full model generation upgrade.

Three independent data points from the last six months, all converging on the same conclusion:

Data point 1 - LangChain on Terminal-Bench 2.0. Terminal-Bench 2.0 is now the standard benchmark for evaluating coding agents on real terminal tasks (89 tasks across machine learning, debugging, biology, infrastructure). LangChain kept the model fixed (GPT-5.2-Codex) and only changed the harness. Score went from 52.8% to 66.5%. That moved them from outside the Top 30 to Top 5 on the leaderboard. Same model, same weights, different scaffolding. A 13.7 percentage point jump from harness alone.

Data point 2 - Cursor on Terminal-Bench 2.0. Cursor’s research team published a piece on April 30, 2026 reporting that they took their own coding agent from Top 30 to Top 5 by only changing the harness. Same conclusion, different team, different harness. A 25-position jump on a public leaderboard, attributable to scaffolding alone.

Data point 3 - Claude Opus 4.6, same weights, very different harnesses. This is the cleanest one. Look at the Terminal-Bench 2.0 leaderboard from late April 2026:

Same model. Different harnesses. A 4.5 percentage point spread between ForgeCode and Capy, on a benchmark where teams are fighting for tenths of a point.

The interesting observation is that the labs that trained the models do not always have the best harness for their own models, which I believe shows just how much untapped potential there still is in this area. ForgeCode is a third-party harness, and it lands three of the top six entries on Terminal-Bench 2.0 by routing across model families. LangChain summarized this as: “Opus 4.6 in Claude Code scores far below Opus 4.6 in other harnesses.” Anthropic’s flagship model in Anthropic’s flagship harness gets beaten by the same weights running in third-party scaffolding.

Now compare this to model generation upgrades. Going from Claude Opus 4.5 (80.9% on SWE-bench Verified) to Opus 4.7 in April 2026 took SWE-bench Verified from 80.8% to 87.6% — about a 6.8-point upgrade. The harness can move you 13.7 points on Terminal-Bench from scaffolding alone. The harness is moving the score by more than a model generation upgrade.

Don’t get me wrong, I am not trying to say that raw AI model performance doesn’t matter. The model floor is rising, and that floor matters a lot. The point is that on top of any given model, the harness is now the largest single performance variable in agent quality. Anyone shipping a coding agent in 2026 who picks the model first and the harness second is leaving a lot of performance and cost-effectiveness.

This section from an interview with a former high-ranking Meta employee is very interesting. It explains just how much more room there is for improvement around the harness, and that because raw model improvements are so fast right now, the resources are not yet focused on it enough:

Why the harness forms a moat

It is important to understand that models are post-trained against a specific harness. They are not generic. Every modern frontier model was fine-tuned, with reinforcement learning, against a specific tool surface, a specific schema, a specific memory ritual, a specific citation format, a specific system prompt structure. The model’s instincts — what it reaches for when it needs to edit a file, what tag it wraps citations in, how it structures a plan, how it handles subagent dispatch — are baked into the weights during post-training. And they are baked in against one specific harness.

The clearest concrete example is the file-editing tool. From Cursor’s harness team:

Cursor is saying that the wrong tool format produces a measurable cost in reasoning tokens and an observable increase in error rate, recorded at scale across millions of agent turns. The model performs worse because the wire format it sees at runtime does not match the wire format it was trained on.

This is becoming a challenge for everyone trying to become an orchestration layer on top of AI model providers like Microsoft. I am not saying it is an impossible task, but you need to have different harnesses for each AI model you use. It also suggests that the products that these AI labs are offering are not just the model, but the model together with the custom harness, as the model and the harness were fused over months of post-training, and you cannot pull them apart without giving up performance. In the end, this might mean the moat becomes stronger for AI labs, as switching AI models becomes increasingly complex.

Why this is a co-evolution, not a one-time thing

The more powerful version of the moat argument is that the post-training and the harness feed back into each other on every model generation. This is what is starting to create compounding lock-in over time.

An example of this is Anthropic’s harness team ships a new primitive — say, a smarter subagent dispatch verb, or a new way to compact context, or a memory file convention. By month three, that primitive shows up in millions of real agent traces from Claude Code users. By month six, those traces are training data for the next model generation. By month twelve, the next model has the primitive baked into its instincts, and the harness can now lean on it as something the model does natively.

Anthropic puts it this way in their March engineering blog: “every component in a harness encodes an assumption about what the model can’t do on its own. Those assumptions go stale.” When a model upgrade kills an old assumption, that piece of scaffolding gets retired, freeing up the harness to push on the next ceiling.

A concrete example from Anthropic’s own work: when Claude Sonnet 4.5 was the frontier model, the agent had “context anxiety” — it would wrap up tasks prematurely as it approached what it thought was its context limit. The harness compensated by aggressively resetting context between sessions and using structured handoff artifacts to carry state across the boundary. When Opus 4.6 shipped, that behavior was largely gone in the model itself, so Anthropic dropped the entire context-reset machinery and ran continuous sessions over two hours. The harness shrank because the model swallowed a chunk of its work.

The matched pair is not static. It moves with each model generation. And the labs that own both sides of the pair are the only ones who can move it cleanly. A third-party harness builder is always reacting to a model release; the labs are designing the next model with the next harness in mind.

This is genuinely structural moat behavior. It looks a lot like the way Microsoft built lock-in through tightly coupling Windows + Office + Exchange in the 1990s and 2000s - each layer made the others stickier, and competing on any single layer in isolation got harder every year.

There is a counter-example worth knowing because it sharpens the moat argument. In December 2025, Vercel published an engineering post-mortem: their internal text-to-SQL agent had 16 specialized tools - schema lookup, query validation, error recovery, intent clarification, join-path finders, syntax validators. They deleted 80% of them and replaced everything with a single capability: execute arbitrary bash commands against a file system.

Results: success rate went from 80% to 100%. Speed improved 3.5x. Token usage dropped 37%. The worst-case run improved from 724 seconds, 100 steps, and 145,463 tokens (failing) to 141 seconds, 19 steps, 67,483 tokens (succeeding).

The lesson is more than just that fewer tools are better. It is that the right harness depends on the model’s current capabilities, and that gap shifts every model generation. When the model gets smart enough to use bash + grep + cat directly, your custom tool layer becomes overhead. GitHub’s Copilot team hit the same wall from the opposite direction: they cut Copilot’s tool count from 40+ to 13 core tools, and pre-expansion accuracy jumped from 19% to 72%.

This is exactly why owning both the model and the harness is so valuable. You can retire scaffolding as your model gets smarter, faster than anyone else can. A third party building on top of your model is always reacting to your release notes. You are designing both sides.

Where the harness wars are headed

A few things I am watching that I think will define the next 12-18 months of this space:

1. The harness becomes the product, not the model. This is already happening. Anthropic does not sell “Claude” anymore as the main enterprise product — they sell Claude Code, Cowork, the Claude Agent SDK, and Claude Managed Agents. OpenAI sells Codex CLI, the Agents SDK, and the Codex cloud agent. Google just announced that they are deprecating Gemini CLI entirely and replacing it with Antigravity CLI, which shares a unified server-side harness with their Antigravity desktop IDE. The model is the engine; the harness is the car. Customers buy the car.

2. “Harness-as-a-Service” (HaaS) is the next API layer. The Claude Agent SDK, the OpenAI Agents SDK, and Google’s new Antigravity SDK all point the same way: you do not get a model API anymore, you get a harness API — the loop, the tools, the context management, the hooks, the sandbox primitives, the memory layer, all out of the box. This is a bigger and stickier surface than “completions.”

3. These are important implications for the SaaS companies. The harness will absorb more of what is currently sold as SaaS. A managed harness with built-in code execution, browser control, file storage, memory, and orchestration is the runtime for AI-native software. That competes directly with a stack of SaaS tools previously bought separately: developer environments, browser automation, vector databases, observability layers, sandboxing services. Anthropic just shipped Claude Managed Agents, which puts the entire harness behind an API.

4. Harnesses will increasingly diverge by vertical. The highest-quality harnesses today are all coding harnesses, because the ROI is most obvious. The same primitives apply to legal, finance, healthcare, and support. Whoever builds the best vertical harness for a domain will own that domain’s spend, even if the underlying model is shared. The providers that specialize and build harnesses for specific verticals can overcome even AI model raw performance deficiencies because the harness is better. An example of this would be someone like Meta, specializing in shopping, social media, and the healthcare vertical.

The investment angle

This is where the moat argument actually matters for capital allocation in terms of sectors and for which public company this might be most important.

Read more

In this article, I want to walk through why I think the market is making one of its bigger mistakes of this cycle with Meta. The stock is down roughly 24% from its high of $796.25, trading around $602, with a forward P/E of 19x — well below its 10-year average of around 26-27x and below the S&P 500 multiple. The dominant narrative is that Meta is “spending too much” on AI data centers without a clear path to monetization. Every quarter, the market punishes the stock harder on CapEx prints — the post-Q1 2026 reaction was a 10% drawdown on a 57% EPS beat, the worst price reaction Meta has had in its last six earnings reports, despite delivering the largest earnings surprise in that window.

The thesis I lay out has four parts:

- Meta’s AI CapEx is already showing up in revenue and engagement in a significant way and more importantly, will continue to do so (expert interview with a Former Meta employee on this field)

- The data centers are not “moonshots” (and far from the metaverse spend analogy) — they are Meta’s new workforce, and the more compute Meta has, the more it can improve products and ship new ones;

- Meta has the single most underappreciated asset in tech right now, which is distribution, and the market is assigning zero value to it;

- The valuation has compressed to a point where the bar for an aggressive re-rate is much lower than people think.

Let’s dive in.

The CapEx that the market hates is already showing up in the P&L

Meta did $200.97 billion in revenue in 2025, up 22% YoY, with operating income of $83.28 billion, up 20%. In Q1 2026, they did $56.31 billion of revenue, up 33% YoY — an acceleration from a $200 billion base. Ad impressions grew 19% YoY and average price per ad grew 12% YoY in Q1 2026, with the Q2 guide of $58-61 billion implying continued acceleration.

Now think about what that means in context. The bear narrative is that Meta is spending $125-$145 billion in 2026 CapEx with nothing to show for it. But the company is putting up 33% growth on a $200 billion base while also compounding the underlying business with double-digit price-per-ad gains. When advertisers pay more per impression, and impression volume also grows 19%, that shows that the system is getting better at allocating attention to higher-value placements.

The AI in the P&L is most visible in three specific areas that I want to walk through, because management actually quantified them on the calls.

First, the ad-ranking models. Meta has been rolling out a model called GEM (Generative Ads Recommendation Model), which is essentially their LLM-style foundation model for ads. On the Q4 2025 call, management said this directly:

A 3.5% lift in ad clicks on a base of roughly $200 billion is multiple billions of incremental dollars, and that’s from one model improvement in one quarter.

Meta also said on the Q3 2025 call that GEM is now “4x more efficient at driving ad performance gains” compared to the original ranking models. This is exactly the scaling law dynamic you want to see — more compute thrown at the model translates to more revenue, and the elasticity of that conversion is improving, not deteriorating.

A recent interview with a high-ranking former Meta employee who worked in this field was very useful. He explained some details about Meta’s internal metric called Internal Revenue per Engagement (iREV). According to him, Meta’s internal goal is a minimum 1.5-2% improvement of that metric every 6 months. So far, they have been delivering on this metric, and he is very confident that going forward, Meta will continue to be able to deliver that as they have a lot of room for improvement in three areas that affect iREV: model architecture, more data, and transfer learning. He even quantified it: 30-40% from model architecture changes, 20-30% from more data, and 30-40% from transfer learning. While model architecture and more data are quite straightforward, transfer learning is something many of you might not be familiar with. To explain the context as simply as possible, because serving a large LLM model across the scale of Meta is too expensive, they had to figure out a structure where they had a teacher LLM (the big one) and a student LLM who basically distilled knowledge from the teacher model to be smaller and more effective to run inference on at the scale Meta needs it. So improving data transfer between the teacher and the student model for Meta is key, as it will be for any other company running large-scale production workloads (the phase of AI adoption we are now in). So every time Meta makes improvements in this realm, it translates to better ad performance and rankings, as performance is essentially determined by the quality of the student model.

What I found really insightful was when the person was asked about what the biggest barrier to improving even faster in terms of ad ranking and performance beyond the 2% per 6 months was:

What this means is that because the model performance is moving so fast, internally, Meta hasn’t even had enough time and resources to pull or spend time to optimize other levers, which shows us just how early we still are in these improvements and how much more growth from this ad AI tailwind Meta has for its core business, not just from scaling laws and model improvements but also from the data and student teacher mechanizm.

Second, engagement. On the Q3 2025 call, Zuckerberg said:

A 5% lift in time spent on Facebook — a 21-year-old app that everyone wrote off as dead — is huge. The platform was supposed to be in decline. AI ranking systems brought it back. In Q4 2025, the optimizations drove a 7% lift in views of organic feed and video posts on Facebook, which Susan Li called “the largest quarterly revenue impact from Facebook product launches in the past two years”.

Third, the end-to-end AI ad tools. The annual run-rate of revenue going through Meta’s fully AI-powered ad tools (Advantage+) passed $60 billion on Q3 2025. The video generation tools alone hit a $10 billion run rate by Q4 2025, with quarter-over-quarter growth outpacing the broader ad revenue increase by nearly 3x. Click-to-WhatsApp ads grew revenue 60% YoY in Q3. None of this exists without the AI infrastructure that the market is currently punishing the company for building.

The way to think about this is the same logic I laid out in my February article about the hyperscalers: the CapEx Meta is spending this year doesn’t show up in this year’s revenue. A data center takes around 2 years to build and operationalize, so the revenue acceleration Meta is showing today is the return on 2023 CapEx (~$28 billion), not 2025 CapEx ($72.2 billion). When 2025’s CapEx starts showing up in 2027 revenue, the operating leverage will be far more aggressive than what we’re seeing now.

Data centers are Meta’s new workforce

Here is the part I think most investors are missing, and it’s where the analogy needs to shift.

For the last 15 years, Meta’s growth engine was: hire engineers, ship product, get more users, monetize via ads. Headcount was the input that scaled output. That framework is now different, because the marginal unit of “intelligence” inside Meta is no longer an engineer. It’s a GPU.

Zuckerberg essentially said this on the Q3 2025 call when he framed Meta’s strategy around three “giant transformers” running Facebook, Instagram, and ads, with the goal of merging them into one unified system:

Meta is openly saying that the entire company — the feeds, the ads, the recommendations across 3.56 billion daily users — is going to be run by AI systems whose performance scales with compute. The CapEx number is the headcount number for the AI era.

And here is the kicker — Meta is compute-starved on the current business. Zuckerberg said it directly on Q3 2025:

Meta CFO Susan Li doubled down on this on the same call:

This is not a company building speculative infrastructure for products that might monetize in 2030. This is a company that has more profitable use cases for compute than it has compute, and is rationing GPU hours between training the next frontier model and improving the ranking systems that drive a quantifiable lift in conversions every quarter. Investors who treat this CapEx like a moonshot are misreading the situation.

Meta’s distribution is Slept on

There is a more general thesis floating around in the market right now that says: “If AI commoditizes, then everyone with compute can build products, from software to something like a Meta platform.” I think this gets the second-order logic backward. If models commoditize and anyone with enough GPUs can ship a product, then the question becomes: who can get that product in front of users? Distribution becomes the bottleneck, not the model. And Meta’s distribution machine is arguably the single biggest in the world.

Meta’s Family of Apps had 3.56 billion daily active people in March 2026. Instagram crossed 3 billion monthly active users in September 2025. WhatsApp also has over 3 billion users across 180+ countries. Facebook still serves billions of people daily. That is unmatched at this scale anywhere in tech — Google has Search, but the engagement-per-user profile is fundamentally different (people come to Search, do a query, leave; people stay on Meta apps for 30+ minutes a day).

Here’s how that distribution muscle has shown up historically. Meta bought Instagram for $1 billion in 2012. It is now a 3-billion-MAU asset that is the cultural center of gravity for an entire generation. They bought WhatsApp for $19 billion in 2014. It has more than 6x to over 3 billion users. They built Threads from scratch in mid-2023 — a product that, frankly, was not particularly differentiated from X — and rode Instagram’s social graph to 400 million MAUs and 150 million daily actives in roughly 2.5 years. Similarweb data shows Threads passed X in daily mobile active users in January 2026.

Threads was a clone. The product was almost identical to X. There was nothing technically novel about it. And in 2.5 years, by being plugged into Instagram’s distribution graph, it overtook a 19-year-old product with deep cultural roots. The question every investor should ask themselves: what other company on earth could have done that?

Now apply this to AI products. Meta AI hit 1 billion monthly active users by May 2025, doubling from 500 million in roughly 8 months, and let’s be honest, the product wasn’t even good. ChatGPT took roughly 2 years to reach similar scale. Meta did it by embedding the assistant into search bars and chat interfaces inside WhatsApp, Instagram, Facebook, and Messenger. Roughly 63% of Meta AI’s usage comes from WhatsApp alone. Meta did not need to convince anyone to download an app, learn a new interface, or change a habit. The distribution infrastructure was already there.

If you believe — and I do — that the next phase of AI is going to produce a wave of consumer products (AI-generated content, personalized AI assistants, business AI, voice agents, creator tools, AI shopping experiences), then the company that can ship each of those products to 3.56 billion people on day one has a structural advantage that the market is not pricing in. Zuckerberg said it himself on the Q3 2025 call:

The market is essentially treating Meta as if distribution is free. It’s not free. It is the single hardest moat to build in consumer technology, and Meta is the only company that has built three of them in parallel (Facebook, Instagram, WhatsApp), then bolted on a fourth (Threads) using the first three as the launchpad.

Subscribe now

The AI sentiment changes on a dime

The market right now is running on AI sentiment more than fundamentals. Companies are being bucketed as “AI winners” or “AI losers”, and the valuation gaps between those buckets are enormous. The reason is that the marginal flows of capital in public markets are still controlled by investors with financial-domain expertise but a relatively shallow understanding of how AI actually works at the technical level. So the signal that gets weighted most heavily is: did this company ship a frontier model? Did they show up on the benchmark leaderboards?

This is exactly the gap that creates opportunity. Meta released Muse Spark on April 8 — the first model from Meta Superintelligence Labs (MSL). Muse Spark scored 52 on the Artificial Analysis Intelligence Index, behind Gemini 3.1 Pro and GPT-5.4 (both at 57) and Claude Opus 4.6. On absolute benchmark terms, it’s not SOTA.

But look at what it is good at. Muse Spark used only 58 million output tokens on the Intelligence Index evaluation, versus 157 million for Claude Opus 4.6 and 120 million for GPT-5.4 — meaning Meta is delivering near-frontier intelligence with less than half the inference compute of competitors. Meta also said the model achieves the same capability level as the older mid-size Llama 4 Maverick using an order of magnitude less compute. For a company that’s about to deploy this model to 3 billion daily users, inference efficiency at this scale is a multi-billion-dollar economic advantage. And the model is particularly strong in vision, health, and what Meta is calling “personal intelligence” use cases — exactly the domains that map onto consumer apps.

Now think about what happens when the larger frontier model lands. Meta has been working on a next-generation flagship, a bigger model internally named the “Watermelon” model. If Meta lands a model that is genuinely competitive on benchmarks with frontier models from Anthropic, OpenAI, and Google, the market will re-rate the company aggressively. The current setup is that Meta is being priced as if it can’t compete at the frontier. The forward P/E of 19x reflects that. Compare that to Google trading at a meaningfully higher multiple post the TPU/Gemini story repricing in late 2025. The asymmetry is real. If Meta merely catches up to where the market is already pricing Google and other labs like Anthropic, OpenAI, the implied upside is substantial— and that’s without assigning incremental value to the distribution moat or to any of the new AI products.

The Market hates the data center spend. Meanwhile, everyone else is desperate for compute.

The market is currently assigning negative value to Meta’s data center buildout. Every time CapEx goes up, the stock goes down. Meta’s CapEx jumped from $39 billion in 2024 to $72 billion in 2025 to a guided $125-145 billion in 2026.

At the same time, the rest of the AI ecosystem is screaming that we don’t have enough compute. Anthropic just signed a deal to take all 300+ MW of compute capacity at xAI’s Colossus 1 data center in Memphis — roughly 222,000 Nvidia GPUs including H100, H200, and GB200 systems. That deal is worth billions. xAI has effectively pivoted to a neocloud model, renting GPUs to Anthropic. The CoreWeave and Nebius backlogs continue to grow. Oracle’s cloud business is being capacity-constrained. AWS, Google Cloud, and Azure are all selling everything they have available and have multi-year backlogs.

So the market believes simultaneously that (a) there is a multi-year compute shortage that will continue at least through 2027, and (b) Meta is wrong to be building data centers and putting negative value to them. These two beliefs cannot both be true. If there is a compute shortage, then Meta’s buildout’s terminal value in the worst case should be based on the value of those data centers if Meta sells that compute on the market. Zuckerberg made this point explicitly on the Q3 2025 call:

The fact that the market is ignoring this and assigning a negative value to the CapEx is, in my view, a significant mistake.

Here’s why this matters even more: Meta is arguably the only company outside the three hyperscalers (AWS, Microsoft, Google) that has the operational capability to run hyperscale data centers for both training and inference at the level required. They’ve been operating planetary-scale infrastructure for over a decade. They know how to manage multiple AI accelerators: Nvidia GPUs, AMD GPUs, and custom ASIC (their MTIA). If Meta wanted to offer compute externally tomorrow, the renters lining up would include some of the largest AI labs and enterprises in the world, and the unit economics would look more like AWS than like a cap-on-cost neocloud.

Valuation: What should the real value be?

The company is

Read more

We just got Q1 2026 earnings from Meta, Microsoft, Google, and Amazon, and after spending hours on the calls and the prints, I want to share what I think is the most important takeaway.

I argued that the market was wrong to punish big tech for raising CapEx because the returns on the 2023-2024 spend were already showing up in the P&L. This quarter, that argument got significantly stronger. Margins on cloud businesses expanded again, the core ad businesses at Meta and Google went into a higher gear (Meta +33% YoY, Google Search +19% YoY - both fastest in years), and we got hard data on what custom silicon is doing to the unit economics of inference.

There are a few patterns from this earnings season:

• The core ad businesses at Meta and Google are accelerating because of AI, not in spite of it

• Operating margins on cloud are expanding even as AI workloads scale

• Custom ASICs are no longer a side project - they are the next big business segment

• The era of “subsidized” compute is ending, and we are seeing the first hints of pricing power coming through

• Compute supply is still the binding constraint

Let’s get into it.

Google: Cloud +63%, Search +19%, And The Margin Story

Google delivered a really strong quarter. Two key numbers: Google Cloud up 63% YoY and the “old left for dead” Search up 19% YoY.

Google Cloud revenue hit $20.0B, growing 63% YoY, an acceleration from the 48% growth we saw in Q4 2025. To put that in context, this is the fastest growth rate Google Cloud has ever posted, and they are doing it on a $70B+ annual run-rate base. Google’s Cloud backlog also nearly doubled sequentially to $462B, with management telling us:

That last part is important. A lot of the bear case on backlog numbers is that they are stuffed with one mega-deal that won’t translate into revenue for years (the Microsoft/OpenAI dynamic). Google is essentially telling investors: half of $462B is coming through the P&L in the next 8 quarters.

But the bigger story for me on the Google call was the margin commentary on Cloud:

The bear thesis on hyperscaler AI workloads has been: “Yes, revenue is growing, but margins on AI workloads will be lower than legacy cloud workloads”. If that thesis were correct, you would expect to see operating margin compression at Google Cloud, given that AI workloads are now the majority of new revenue growth there. Instead, we saw the opposite: operating margin nearly doubled YoY, from 17.8% to 32.9%, in a single year. Now it is important to understand that not all of this is direct GCP infrastructure a lot of it is now Gemini products:

The part of Cloud that is growing fastest (GenAI products, +800% YoY) is also the part driving the margin expansion. The reason Google can do this is threefold: scale-driven optimization (Google said earlier this year that it reduced Gemini’s serving cost by 78% in 2025), custom silicon (TPUs), and owning its own frontier model.

On TPUs, we got another important update:

This is a massive strategic shift that I have been writing about for over a year now. Google is essentially saying: the TPU is so valuable to certain customers (think Anthropic, but also potentially Apple, Meta, etc.) that we will sell or lease the chips outside of GCP. This is the same direction Amazon is going with Trainium. The implication is that the cloud providers are starting to compete with NVIDIA directly at the chip layer, not just at the cloud layer.

And on Search, the Search is dead narrative is looking more distant and distant:

19% YoY growth in Search advertising on a $240B+ annual run rate is, candidly, an absurd number.

Microsoft: The AI Business Hit $37B ARR, Up 123%, And Margins Are Holding

Microsoft’s print was a strong quarter, and one of the most important numbers was this:

Just on this number alone: Microsoft’s AI business is now larger than ServiceNow, Workday, or Datadog as standalone businesses. And it is growing at 123% YoY. There is no public software company at scale growing faster.

But the more important nuance came from the management commentary on margins, which is essentially the same story Google told. From Amy Hood:

I want to highlight this because it is being completely missed by the market. Microsoft is telling us that AI workload margins are BETTER than the early cloud workload margins were.

On Copilot, the seat numbers continued to ramp and were much better than the quarter before:

That is 20M paid seats up from 15M last quarter (~33% sequential growth) and 250% YoY. At a $30/user/month list price, that is a roughly $7B+ ARR business just from M365 Copilot.

Engagement on Copilot is the other piece of the story:

Weekly engagement at the level of Outlook is a wild data point.

GitHub Copilot also continues to scale:

And here is where it gets interesting - Microsoft is moving GitHub Copilot to a usage-based pricing model:

Microsoft is admitting that GitHub Copilot adoption is now so heavy that it is dragging down Microsoft Cloud gross margins because they were charging a flat per-seat fee while costs scale with usage. So they are switching to usage pricing on June 1. This is a small story right now, but it is a leading indicator for the entire industry: per-seat pricing on AI products is going to be replaced with consumption pricing, because the cost-to-serve scales with intensity of use, not with seat count. The bigger investor takeaway is what management said elsewhere on the call:

So enterprises are already adjusting their procurement around this. Hybrid pricing models are becoming the standard.

On capacity, Microsoft was, again, very direct:

And then this:

The “supply shortage growing” line is important because it is the third quarter in a row where Microsoft has said this.

The other big tell came on Foundry, Microsoft’s model marketplace:

Note the Anthropic mention here. Microsoft is now a meaningful Anthropic distribution channel. This matters because a year ago, the bull case for Microsoft was “Microsoft + OpenAI.” Today, the company is hosting Anthropic, OpenAI, and open-source models in the same product. The orchestration layer could become the moat, not the model and Microsoft could benefit from it.

The CFO also gave us perhaps the most important signal for what to expect in fiscal Q4 (calendar Q2):

In other words, the next two quarters are when Microsoft expects AI products to start showing up as accelerating revenue, not just bookings.

Subscribe now

Amazon: AWS Reaccelerates To 28% (Fastest In 15 Quarters), And The Chips Business Is Now Top-3 In The World

AWS, which a lot of people had been writing off as the “loser” of the cloud AI race, posted its strongest growth rate since 2022:

So AWS is now growing 28% YoY (vs 24% last quarter) on a $150B run-rate base. The sequential add of $2B is the largest Q4-to-Q1 increase in AWS history.

Bedrock is the key to it:

Bedrock processed more tokens in Q1 2026 than in all prior years combined. Customer spend on Bedrock grew 170% QoQ. There has been a real perception in the market that AWS was behind on the AI inference story, and that is just no longer true based on these numbers.

Second, Trainium. This is the segment of the call that I think investors really need to focus on, because it is reshaping the unit economics of AWS:

Jassy already mentioned this in his recent shareholder letter, and I have written about it, but still, Amazon is now a top-3 data center chip business globally. That is a statement most investors are not pricing in, because they think of Amazon as a retailer + cloud company, not a chip company.

And the demand signal on Trainium specifically is just enormous:

$225B in commitments just for Trainium is insane and was the most shocking number for me among all the earnings calls yesterday.

The fact that AWS is now bringing in more Trainium chips than NVIDIA chips on a unit basis is also an important shift. It is going to flow through to AWS margins over time. The CFO basically said this directly:

Several hundred basis points of operating margin advantage. AWS operating margin in Q1 was already strong at 37.7% (up from 35% the last quarter), and management is telling us Trainium adds several hundred more bps over time. This is exactly the same dynamic Google has been showing with TPUs, where Google Cloud operating margin went from 17.8% to 32.9% YoY. The custom silicon advantage will be crucial in the coming years.

The other under-appreciated comment was on the relationship between AI and core cloud spend:

This is the “AI lock-in” effect I have written about before. Companies that adopt AI workloads end up moving more of their non-AI workloads to the cloud as well, because the data needs to be co-located. This is why core (non-AI) cloud spend is also accelerating - which is something a lot of people miss when they look at AWS AI growth in isolation.

Backlog at AWS:

$364B in backlog ex-Anthropic deal. The “reasonable breadth” comment is important - this is not a one-customer book like Microsoft’s relationship with OpenAI. The AWS book is more diversified.

And on selling chips outside the cloud, similar to the Google TPU pivot:

The most important framing on AI workloads from the entire AWS call:

The chatbot era was really just us trying to use a new technology in a way that we knew so far (information retrieval via Google Search). Agentic AI - where models execute multi-step workflows on behalf of users autonomously - is where the real economic unlock of LLMs lies and requires orders of magnitude more compute per task. We are still very early in this shift.

Meta: Revenue Up 33%, The Fastest In 4 Years, Because The Ad Engine Is An AI Engine Now

Revenue growth 33% is the fastest in the last 4 years.

$56.3B in Q1 revenue, +33% YoY. To put that in context, in Q1 2025, Meta grew 16%. So the company has roughly doubled its growth rate in 12 months, on a $200B+ annualized base. That is essentially impossible without a structural change in the underlying engine, and the structural change is that Meta’s ad ranking and content recommendation systems are now LLM-scale.

On the engagement side:

Note the Facebook number specifically. Facebook is a 20-year-old product, and it is posting the largest QoQ gain in video time in 4 years. That doesn’t happen organically. That happens because Meta deployed a new ranking model that materially improves the relevance of what users see.

The under-appreciated piece here is what Zuck said about how ad ranking actually works now:

And then this longer segment, which I think is one of the most important pieces of technical commentary:

This is important for the whole industry.

For years, Meta could not use LLM-scale models (think GPT-style models with hundreds of billions to trillions of parameters) for ad ranking because the latency would be too high. When you load a Facebook feed, the ad selection has to happen in well under 100ms. LLM-scale models can take seconds to respond, which is way too slow.

What Meta has now done is two things: (1) they have re-architected the model and co-designed it with their custom silicon (this is the Broadcom-partnered ASIC Zuck referenced) so that a 1 trillion parameter model can run within sub-second latency, and (2) they have built a “router” that decides when to use the big expensive model vs. the small cheap one, based on the predicted probability that the ad will convert. So if you are clearly not going to click on an ad about a car, Meta won’t waste compute running the big model on you. But if you are showing strong purchase intent, they will use the most expensive model to find the absolute best ad to show you.

This is a fundamental change in the unit economics of advertising. It is why the average price per ad on Meta increased 12% YoY while ad impressions grew 19% YoY - Meta is showing more ads AND each ad is more valuable.

Now, on the macro AI strategy, Zuck is doubling down. From the call:

The memory pricing call-out is something I have written about extensively. HBM is sold out through 2026, prices are spiking, and Meta is essentially saying: yes, our CapEx is going up, but it is partly because the price of the input is going up, not because we are buying more capacity than we expected.

On the custom ASIC story:

A gigawatt of custom silicon is meaningful - that is a very serious deployment. This is the same playbook Google ran with TPUs and Amazon ran with Trainium. Meta is now in the custom silicon game in a real way, with Broadcom as the partner. That has implications both for Meta’s long-term margin profile and for Broadcom’s revenue trajectory.

On the broader AI investment thesis from Zuck:

Zuck’s view that model improvement continues for a long time is a meaningful statement against the “scaling laws are over” thesis. He is essentially saying the opposite - he sees a long runway of improvement, and Meta is committing capital accordingly.

One thing I do want to flag for investors is the nuance of how Zuck is positioning Meta’s AI strategy versus other labs:

This is Meta differentiating its AI strategy from OpenAI/Anthropic. Those labs are building horizontal foundation models. Meta is building vertical AI that is really good at the things people do on Meta’s surfaces - shopping, social, content discovery. It is a different bet, and arguably a much more defensible one given Meta’s distribution.

The Pattern: AI Workloads Are Now Margin-Accretive At Scale, And Custom Silicon Is The Reason

If I step back and try to find the single most important pattern across all four prints, here it is:

The bear case on hyperscaler AI spending - that AI workloads have structurally lower margins than legacy cloud workloads, and therefore CapEx returns will be poor - has been wrong.

Look at the hard data:

• Google Cloud operating margin: 17.8% Q1 2025 → 32.9% Q1 2026 (+1,510bps YoY)

• AWS Q1 operating margin: 31.4% (Q1 2025) → 37.7% (Q1 2026)

• Microsoft saying explicitly “margins were actually better and they remained better in our AI business versus where we saw in the cloud transition”

• Amazon saying Trainium gives them “several hundred basis points of operating margin advantage versus relying on others’ chips for inference”

One of the structural reasons for the margin’s being stable is custom silicon: TPUs, Trainium. The cloud providers have figured out that they can’t allow NVIDIA to keep 75% gross margins on the most valuable workloads of the next decade, so they are vertically integrating into chips.

The other pattern is that AI is a significant core growth driver in the ad business at Meta and Google.

Both companies now run their ad ranking on LLM-scale models with adaptive routing. Both companies are showing direct evidence that AI-driven ranking improvements are translating into both more ad inventory consumed (impressions up) and higher revenue per impression (price per ad up). The ad business is no longer a separate thing from the AI business - the ad business is the AI business now.

For Microsoft and Amazon, the equivalent flywheel is happening in cloud + agents. Microsoft’s AI ARR hit $37B at +123% YoY. Amazon’s Bedrock token volume in Q1 alone exceeded all of 2025 combined.

The Bigger Picture: The Era Of Subsidized Compute Is Ending

A theme I want to leave you with, because I think it is the most important macro shift for the next 12-18 months:

It really comes down to what the end cost of intelligence is and how much companies are willing to pay for it. The whole industry would benefit from an architectural change that would bring new efficiencies to serving these models. It’s not who has the best model but who can solve the economic task with the least cost (using hardware, cloud infra, models, harness everything). The era of subsidizing compute is over - you can even see it from GitHub. Companies will have to shift budgets from other OpEx items towards compute, the moment is here, and this will only increase as Anthropic and OpenAI go IPO and have to produce »passable« gross margins.

The GitHub Copilot pricing change Microsoft announced is the canary in the coal mine here. You can’t have flat per-seat pricing for a product whose cost-to-serve scales with usage intensity. Either prices go up for heavy users or the seller bleeds gross margin. Microsoft chose to raise prices on the heavy users via consumption pricing. Anthropic is doing the same thing - their API pricing has been moving in this direction, and Claude Code’s heavy users have been hitting rate limits and seeing throttling for months now. The only real question that I believe we will get the answer to soon is how much value end users have and how much they are willing to pay for it.

As always, I hope you found this article valuable. I would appreciate it if you could share it with people you know who might find it interesting. I also invite you to become a paid subscriber, as paid subscribers get additional articles covering both big tech companies in more detail, as well as mid-cap and small-cap companies that I find interesting.

Subscribe to Paid

Thank you!

Disclaimer:

I own Google (GOOGL), Amazon (AMZN), Microsoft (MSFT), Meta (META) stock.

Nothing contained in this website and newsletter should be understood as investment or financial advice. All investment strategies and investments involve the risk of loss. Past performance does not guarantee future results. Everything written and expressed in this newsletter is only the writer’s opinion and should not be considered investment advice. Before investing in anything, know your risk profile and if needed, consult a professional. Nothing on this site should ever be considered advice, research, or an invitation to buy or sell any securities.

I am posting my regular channel check & other alternative data report before we start the big tech earnings.

For this report, I covered cloud providers Google, Microsoft, and Amazon in terms of their cloud business, and some insightful signals on Microsoft Copilot, which is a pressure point for Microsoft.

Let’s dive in.

Cloud is on fire

Let’s start with my alt data on the most relevant channel-check interviews from clients, cloud consultants, integrators, and former employees.

When bulk analyzing these interviews, demand is high for Q1 2026 across all three hyperscalers: AWS, Azure, and GCP. Looking at the future demand pipeline for the next 3-6 months, the results are even more impressive: almost 60% of these experts see demand exceeding their expectations, 26% see it meeting expectations, and 15% see it below expectations. Keep in mind that expectations were already high going into this year and quarter, so for 60% of experts seeing higher-than-expected demand, the signal is very strong. The driver of demand, as expected, is AI workloads and the move from test to production environments, especially with Agentic AI starting to roll out.

Looking deeper, let’s look at the individual hyperscaler level and what the data shows. As always, I made the % breakdown of experts who think AWS, Azure, or GCP is accelerating the fastest. Here are the results:

62% think GCP is growing the fastest, 41% think Azure is growing the fastest, and 27% think AWS is growing the fastest (important note: the sum is greater than 100% because some experts mentioned two platforms as growing at a faster pace than the other).

Now, this data doesn’t add much value until we compare it to my historical data from past quarters, as we did in the last reports, to truly understand whether anything shifted significantly in Q1. Here is the data:

Read more

In this article, I break down some significant fundamental shifts when it comes to AI efforts from Amazon and Meta, and why I think both are the next two big AI beneficiaries. Based on what we are seeing, both companies are on a path to reaccelerating their efforts, while still perceived by the market as “AI laggards”. We believe this premise will be proven wrong in the coming months.

Let’s start.

Read more

The AI industry is approaching an inflection point that will reshape the priorities of AI model companies and the entire space. There are four interconnected themes I want to walk through today, and they all converge on a single conclusion: the era of subsidizing AI model usage is coming to an end.

Here’s what I cover in this article:

• Anthropic is taking over the enterprise — but the curse of the best model is real

• OpenAI is losing the enterprise race to Anthropic and facing structural problems heading into its IPO

• The era of subsidized AI model usage is ending as both companies prepare for public markets

• The IPO race: who lists first matters more than most people realize

Let’s get into it.

Anthropic Is Taking Over Enterprise: And the Curse of the Best Model Is Coming for Them

Anthropic announced that its revenue had surpassed $30 billion, up from $9 billion at the end of 2025. That’s more than tripling in roughly four months. Anthropic has now supposedly surpassed OpenAI’s run-rate revenue of approximately $25B.

The enterprise composition is what separates Anthropic from the rest. Approximately 80% of revenue comes from business customers. The number of customers spending over $1 million annually has doubled to more than 1,000, up from 500+ where it was just 2 months ago. Business subscriptions to Claude Code have quadrupled since the start of 2026.

And then there’s Mythos. Just a few days ago, Anthropic announced Claude Mythos Preview - a new general-purpose model that sits in an entirely new tier above Opus. The draft described it as “by far the most powerful AI model we’ve ever developed” and said it is “very expensive for us to serve, and will be very expensive for our customers to use.”

The benchmarks are quite telling: 93.9% on SWE-bench Verified (vs. Opus 4.6’s ~80.9%), 77.8% on SWE-bench Pro, 82% on Terminal-Bench 2.0, 97.6% on USAMO 2026, and 83.1% on CyberGym vs. Opus 4.6’s 66.6% — a 16.5 percentage-point jump on cybersecurity tasks. On Anthropic’s internal zero-day exploit benchmark, Opus 4.6 had a near 0% success rate at autonomous exploit development. Mythos succeeded 181 times out of several hundred attempts on the same Firefox vulnerability task. It also found thousands of zero-day vulnerabilities across every major operating system and browser, including a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg that automated testing had missed across 5 million test runs.

Mythos is not being made generally available because Anthropic wants to first roll it out to selected companies because of cybersecurity risks. It’s deployed through Project Glasswing to 12 partner organizations (Amazon, Apple, Broadcom, Cisco, CrowdStrike, Linux Foundation, Microsoft, Palo Alto Networks, and others) plus about 40 additional organizations, with Anthropic committing $100 million in usage credits. While I am not dismissing any cyber risks that a model like this could bring, it is also convenient that now model providers will »release« these models to a small group of companies, because the reality is that with their current compute, they can’t even serve Opus 4.6 to their user base let alone Mythos, which is even more expensive to run. The current Mythos Preview, for which these companies got access, is around 5x more expensive than Opus 4.6 after the initial $100M credit commitment from Anthropic based on their pricing. It is also rumored that OpenAI will also “release” their newest model in a similar fashion, again citing cybersecurity risks.

There has been a lot of frustration from Claude users lately, as many have started to hit their rate limits much faster in their subscription plans, as Anthropic is having to manage this surge in demand with the amount of compute that they have.

This is what I call the Inference Trap, and both OpenAI and Anthropic have now been caught in it.

The pattern is simple: build the best model → users surge → inference compute explodes → you either throttle users, raise prices, or cannibalize training compute. OpenAI experienced it during the Ghibli moment in March 2025, when ChatGPT gained 1 million new users in a single hour and 100 million signups in a week. Sam Altman admitted they were “forced to do a lot of unnatural things,” specifically borrowing compute capacity from OpenAI’s research division and slowing down the release of new features.

Anthropic is living through its own version right now. In March 2026, the company experienced five major platform outages in a single month. Claude Code users reported burning through 5-hour sessions in under 90 minutes. The problem for Anthropic is that if you are a model provider in this AI race, you don’t want to cannibalize training compute, as it means that you can lose the race for the next model.

This brings me to another point I want to make: pricing increases on frontier AI models are inevitable. When Anthropic eventually deploys Mythos-class models at scale, the inference cost per query will be higher than Opus. And they already can’t serve Opus at current demand levels without throttling. The math only works if prices go up, or if the compute infrastructure grows fast enough to meet demand, which it can’t in the short term.

Moving to OpenAI.

OpenAI seems to be losing the Enterprise Race And Heading Into an IPO with some headwinds

While Anthropic is sprinting ahead on enterprise revenue, OpenAI is dealing with a set of problems that are becoming hard to ignore.

The revenue gap has flipped. A year ago, OpenAI was at roughly $6 billion ARR, and Anthropic was at $1 billion. The gap looked huge. Today, Anthropic is at $30 billion, and OpenAI is at $25 or similar to Anthropic, but the pace of growth is slower. Anthropic added roughly $21 billion in net new annualized revenue in just three months. OpenAI’s enterprise business now makes up 40% of revenue (up from ~30% last year) and is “on track to reach parity with consumer by the end of 2026” — but Anthropic has been enterprise-first from the start, with 80% enterprise revenue and structurally higher retention.

To add to this, SensorTower data now show that ChatGPT’s monthly active users in the US have started to fall slightly, adding to the headwinds.

Even without this data, we have been waiting for quite some time for OpenAI to make the update of reaching 1 billion weekly active users, as the 800M mark was announced 6 months ago. Based on this users growth has slowed down.

The funding structure also isn’t perfect. OpenAI closed a $122 billion round at an $852 billion valuation at the end of March— the largest private funding round in history. But the structure is telling. Amazon committed $50 billion, but only $15 billion arrived as upfront cash. The remaining $35 billion is conditional, tied to milestones that some indicate may include achieving certain AI capability thresholds or pursuing an initial public offering by the end of 2026.

SoftBank pledged $30 billion structured in three equal tranches of $10 billion each, arriving in April , July, and October. SoftBank’s structure essentially assumes a liquidity event within that window. About $3 billion came from retail investors through bank channels, and OpenAI was included in ARK Invest ETFs.

In other words, a significant portion of OpenAI’s headline $122 billion raise is conditional on an IPO actually happening. This means OpenAI is under pressure to go public regardless of whether the timing is optimal. There are now reports from The Information that Altman and OpenAI’s CFO are on different sides over the IPO, as Altman is pushing for an IPO this year, while the CFO believes OpenAI is not ready yet. OpenAI has already denied this, but it's not expected that any company would confirm such rumors, even if they were true.

The profitability picture is the key thing. According to different reports, OpenAI’s gross margins sit at approximately 40%, constrained by variable compute costs. The company is generating +$2-3 billion per month but losing +$14 billion per year. Reports of internal documents project that compute costs will reach $121 billion by 2028, with a cumulative loss trajectory that doesn’t reach breakeven until 2029-2030. Compare this to Anthropic, which projects positive free cash flow by 2027-2028 while spending roughly 4x less on training.

OpenAI also has an alternative to Claude Code called Codex, but adoption there, although growing, doesn’t seem to be at the same pace as Claude Code. It’s telling that OpenAI is even offering users more token usage, while Anthropic is limiting it.

As this data from Ramp shows, the AI model share of first-time enterprise customers has heavily tilted towards Anthropic in the last few months:

The key moat that OpenAI has is the ChatGPT brand, which, as a first mover, created a verb similar to Google when it comes to consumers. In the last months, however, Claude and Anthropic have become “the verb” when it comes to enterprises and AI use cases for work. On the consumer end, OpenAI will probably have to shift hard towards an ad-supported business model to get some revenue from its big free user base. Building an efficient ad platform is much more complex than most think and requires time. At the same time, OpenAI is trying to stop Anthropic in the enterprise market, but so far, it doesn’t seem to be working as Anthropic is capturing the market at a faster pace. The question is whether OpenAI’s strategy of trying to capture both markets at once and “doing everything” is really the right one. I would argue that it is not. Now you even have Meta entering the AI arena again, with its first AI model since the formation of Meta's superintelligence unit. While their model is not SOTA, there are specific use cases where it is very competitive. Meta focused on use cases like health, social media, games, and shopping. This pattern will become more dominant in the coming years as the AI model market matures and you see model specialization rather than just general models. In this environment, the importance of having a narrow focus becomes even bigger.

Subscribe now

According to The Information, OpenAI is now communicating to investors that they believe one of their important advantages going forward vs Anthropic is in the availability of compute. OpenAI said it believes Anthropic had 1.4 GW of capacity at the end of last year, while OpenAI had 1.9 GW. But OpenAI said it plans to ramp its capacity more steeply, with total gigawatts in the mid-single digit range at the end of this year and more than 10 GW in 2027.

In contrast, it believes Anthropic will have 3 to 4 GW in 2026 and 7 to 8 GW by the end of 2027. While I agree that availability of compute is a big factor going forward, I believe the main one and even more important is inference economics and the costs of serving the models to your clients. As users shift workloads to production, availability and reliability become key. Nobody wants to have a product that is unstable and sometimes works great, but other times is not available.

The End of Subsidized AI Model Usage

Both Anthropic and OpenAI are preparing for IPOs. And an IPO changes everything about how an AI company thinks about compute costs.

When you’re private and burning venture capital, you can subsidize inference. You can run models at a loss. You can offer $20/month unlimited plans that cost you +$100/ month to serve. You can double the rate limits as promotions. You can hand out credit packages. The goal is growth at all costs, because the next funding round values you on revenue, not margins.

When you’re public, the scrutiny shifts to unit economics. Gross margins, operating margins, cash burn trajectory, path to profitability — these become the metrics that determine your stock price. An S-1 filing forces you to disclose all of this in audited detail.

Both OpenAI and Anthropic are probably operating at approximately 40% gross margins, constrained by the variable cost of running inference. And while no company will show net profit as they are projected towards the end of 2030, the gross margin will be something that investors will particularly keep an eye on, especially the gross margin on inference. While the margin profile math works on API usage, the one in the subscription packages is still often subsidized by the model companies, and this is something they will look to tweak in the coming months before they file their S-1s.

This has a ripple effect across the entire supply chain industry. For the past three years, the question has been: “Which chip delivers the most FLOPS?” The answer was always Nvidia, and companies paid whatever Nvidia charged because performance was the bottleneck, and money was abundant.

Going forward, the question becomes: “Which chip gives me the cheapest tokens and the best total cost of ownership (TCO)?” It’s not even about watts anymore, as you can see from recent podcast comments from Nvidia’s Jensen and Google Sundar — it’s about cost-per-token, because that’s what directly determines your gross margin as a public company, and that is what investors will be laser focused on. This shift also means the end of subsidizing usage in these subscription packages, with either usage limits or higher prices. We will also see even more resources being focused on software optimizations to run the model. Savings on memory and getting more from existing hardware will be the focus in the coming months for both labs, as they are constrained by compute.

Both of these companies also need to make the hard math of the IPO being the “last” funding round, and after the IPO, have enough capital that will be able to support their growth and cash burn for the coming years. Issuing additional stock for raising capital once you are a public company is never looked at positively by the market, so nobody wants to go down that route.

Anthropic is uniquely well-positioned here because it runs Claude on a diversified hardware stack across three suppliers: Nvidia GPUs, Google TPUs, and Amazon Trainium. This gives it real negotiating leverage and the ability to route workloads to whichever chip offers the best price-performance for each model tier. The company just announced a deal with Google and Broadcom for approximately 3.5 gigawatts of next-generation TPU capacity starting in 2027. This is on top of its existing AWS Trainium partnership and Nvidia GPU deployments. It is worth noting that AWS’s CEO just mentioned in an interview on CNBC yesterday that all of Anthropic’s AI models were trained on Amazon Trainium (even Mythos).

OpenAI, by contrast, has been more dependent on Nvidia through its Azure partnership with Microsoft, though it has been diversifying toward custom silicon.

The IPO Race: Who Lists First Gets the Biggest Check

There’s one final dynamic that ties all of this together: both Anthropic and OpenAI know that whoever goes public first has a significant advantage, and the window for both is narrowing. On top of those, SpaceX (which now includes xAI) is also racing towards a +$1T IPO. OpenAI is targeting a +$1T IPO, Anthropic just closed a funding round valued at $380 billion, but because of the surge in usage and revenue is already valued at around $500-$700 billion in secondary listings, so the IPO could be in the $800B-$1T range as well.

Between these three companies alone, we’re looking at potentially +$200 billion in capital being raised from public markets within a 6-12 month window. That’s an enormous liquidity event. For context, the entire US IPO market raised approximately $33 billion in 2024. Even in the hot 2021 market, total US IPO proceeds were around $140 billion.

This is why the race to go first matters so much. The first to market captures the freshest investor capital and sets the valuation benchmark. The second has to compete for the same institutional allocation. The third might struggle if the market has indigestion from the first two.

OpenAI’s board is reportedly concerned that if Anthropic lists first, it could set a valuation benchmark that makes OpenAI’s $1 trillion target look stretched — especially now that Anthropic has higher revenue, better enterprise concentration, and a more credible path to profitability. On the other hand, if OpenAI lists first, it establishes itself as the “AI category-defining IPO” and benefits from a first-mover premium in public market pricing.

Both companies know this. Both are preparing in parallel. And both are racing against time, because every month that passes, compute costs pile up, margins need to improve, and the public market window could shift with macro conditions.

Summary

We are entering a new key period in AI where unit economics take front stage. At the same time, I expect we will see a rapid pace of software optimizations to more efficiently serve these models in the coming months as AI labs put their best talent towards solving this task because it has now become the most important thing that is limiting growth and profitability. The software optimization will focus on resolving key bottlenecks, such as memory (KV cache, context window) and wafer availability. Model distillation and the trend toward smaller models will also grow faster than before because of this. If I were to speculate, the hardware companies might have a “less golden” time than the era they have had so far, while cloud providers might benefit the most as these software optimizations mean that they get more juice out of their existing infrastructure, while demand for compute still keeps on surging, because of wider adoption of AI.

Until next time,

Next week, we are publishing an article on some key developments in the AI space when it comes to Meta and Amazon, exclusive for paid subscribers. If you are not yet a paid subscriber, consider signing up.

Become paid subscriber

Thank you!

Share

Disclaimer:

I own Google (GOOGL), Amazon (AMZN), Microsoft (MSFT), Meta (META) stock.

Nothing contained in this website and newsletter should be understood as investment or financial advice. All investment strategies and investments involve the risk of loss. Past performance does not guarantee future results. Everything written and expressed in this newsletter is only the writer’s opinion and should not be considered investment advice. Before investing in anything, know your risk profile and if needed, consult a professional. Nothing on this site should ever be considered advice, research, or an invitation to buy or sell any securities.

For three decades, the memory semiconductor industry has followed a brutal and predictable pattern: prices boom, manufacturers over-invest, supply floods in, prices crash, everyone bleeds red ink, and then the whole thing starts over. It’s been one of the most reliably cyclical businesses in all of technology. The cycle has destroyed shareholder value, bankrupted companies, and taught every investor the same lesson: never trust the words “this time is different” when it comes to DRAM.

And yet, here I am, writing an article arguing exactly that.

Let me be clear, I know the history. I’ve studied every major memory cycle of the last 30 years. In this article, we look at them and the numbers. But then I am going to make a case for why the AI era may fundamentally break that pattern, not because demand will be infinite (it won’t), but because the nature of what memory serves has changed in a way that most investors haven’t fully internalized.

Memory is no longer just a component inside your gadget. Memory is becoming a raw input for intelligence. And the demand curve for intelligence looks a lot more like the demand curve for energy, electricity, than it does the demand curve for smartphones.

Let’s start.

The history of memory economics

For those less familiar with the space, the memory semiconductor market is dominated by three players: Samsung Electronics (South Korea), SK Hynix (South Korea), and Micron Technology (United States). Together, these three companies control approximately 95% of global DRAM production. This is an oligopoly, but not one that has historically behaved like one. Unlike OPEC, these companies can’t (legally) coordinate output. And unlike logic chips, memory is essentially a commodity—a bit is a bit. The differentiation comes from process technology, cost structure, and increasingly, product mix (more on HBM later).

The fundamental problem with memory economics is the mismatch between demand elasticity and supply inelasticity. Building a new DRAM fab costs $15-20 billion and takes 2-3 years. Once built, the economics favor running it at maximum utilization because fixed costs are enormous. So when demand rises, prices spike because supply can’t respond quickly. When manufacturers finally bring new capacity online, they tend to overshoot, because everyone is building at the same time based on the same rosy demand signals. Prices crash, margins collapse. Some companies go bankrupt or get acquired. The survivors cut capex, and the cycle begins anew.

This is the pattern. And it has repeated with remarkable consistency.

Cycle 1: The Windows PC supercycle (1993-1996)

The first modern memory supercycle was driven by the explosion of Windows PCs and graphical operating systems. Average DRAM content per PC jumped from roughly 1-2MB to 4-8MB—a 4x increase per device—while PC unit shipments were growing at double-digit rates.

During 1993 and 1994, DRAM demand outpaced supply despite most fabs running at full utilization. Spot and contract prices for 4Mb and 16Mb DRAM rose sharply, and gross margins for leading suppliers surged well above 50%. Korean memory makers like Samsung and Hyundai (now SK Hynix) posted record profits. Semiconductors accounted for 13.4% of Korea’s total exports. It was hailed as the greatest boom in Korean industrial history.

Then reality hit. Roughly 50 fab construction plans were announced during 1995-1996 alone. Capex as a percentage of semiconductor production exceeded 30%. The inevitable happened: DRAM prices peaked in late 1995 and then collapsed—falling 51% in 1996 and another 65% in 1997. Korea’s Big Three chipmakers suffered from overexpansion, and the resulting shock contributed to the Asian Financial Crisis that pushed Korea into a deep recession. Stock prices of memory companies fell 60-80% from peak to trough.

Looking at the data cycle duration (peak to trough): around 2 years. Price declines 51% in year one, 65% in year two, and the stock declines around 60-80%.

Cycle 2: The cloud and smartphone era (2016-2019)

Fast forward two decades, and the cast of characters had changed, but the script was the same. By 2016, the DRAM market had consolidated from roughly 20 players to just three. This was supposed to introduce discipline. And for a while, it seemed like it did.

The 2016-2018 “supercycle” was driven by a convergence of factors: smartphone storage capacity upgrades, the early cloud buildout, and a supply-side twist where manufacturers were shifting capacity to 3D NAND production, which temporarily constrained conventional DRAM output.

The numbers were spectacular, especially for Micron, the only publicly traded pure-play memory company in the U.S.:

Micron 2016: Revenue of $12.4 billion, gross margin of 20.2%, operating income of just $168 million (1.4% operating margin). The company was barely above breakeven.

Micron 2017: Revenue surged 64% to $20.3 billion. Gross margin expanded to 41.5%. Operating income hit $5.87 billion (28.9% margin).

Micron 2018: Revenue jumped another 50% to $30.4 billion. Gross margin peaked at 58.9%. Operating income reached an astonishing $15.0 billion—a 49.3% operating margin. From barely profitable to printing nearly 50 cents of operating profit on every dollar of revenue in two years.

SK Hynix followed a similar trajectory. At its Q3 2018 peak, SK Hynix posted an operating profit of 6.47 trillion Korean won, which at the time was a record.

DDR4 retail RAM prices doubled over the course of 2017 into early 2018. Industry inventories fell to 3-4 weeks, well below the normal 8-week average.

Micron’s stock peaked at roughly $64 in May 2018. But notice- revenue and margins didn’t peak until Q4 of calendar 2018. The stock topped out approximately two quarters before the fundamental peak. This is a classic pattern in cyclical stocks: the market discounts the turn before it shows up in the numbers.

Then came the crash:

Micron 2019: Revenue fell to $23.4 billion (-23%). Gross margin compressed to 45.7%.

Micron 2020: Revenue dropped further to $21.4 billion. Gross margin fell to 30.6%. Operating income was $3.0 billion, down 80% from the 2018 peak.

By December 2018, Micron’s stock had fallen to approximately $28—a 56% decline from the May high. The stock was pricing in the downturn even as the company was still reporting near-peak earnings.

Cycle duration (peak to trough in fundamentals): ~6-7 quarters. Revenue decline (peak to trough): ~30% Gross margin decline: from 59% to 27% (at the Q1 FY2020 low) stock decline (peak to trough): ~56%.

Cycle 3: The COVID cycle (2020-2023)

The pandemic created an unexpected demand surge. PC shipments exploded as the world went remote. Server demand spiked as cloud usage accelerated. 5G phones launched with higher per-device memory content. The upcycle lasted approximately 14 months before the familiar reversal kicked in.

By 2022-2023, the downturn was severe. Bloated inventories from pandemic over-ordering met weakening consumer demand. SK Hynix posted a full-year 2023 net margin of approximately negative 28%. Micron’s 2024 revenue dropped to around $25 billion with gross margins compressing toward the low 20s.

Memory stocks cratered. Micron fell from around $98 in early 2022 to roughly $49 by late 2022—a 50% haircut. SK Hynix fell similarly.

Cycle duration (peak to trough): 6-8 quarters of margin compression. Operating margins went from 30%+ to deeply negative for SK Hynix. Stock decline: ~50%

The pattern across all three cycles is strikingly consistent: a demand-driven boom lasting 4-7 quarters, followed by an oversupply-driven bust lasting 4-8 quarters, with revenue declines of 25-40%, margin compression from peak levels above 50% to the low 20s or even negative, and stock price declines of 50-60% that lead the fundamental downturn by 1-2 quarters.

The history is clear, but now let me tell you why I think this cycle might be structurally different.

From gadget component to intelligence input

In every previous memory cycle, the demand driver was the same: humans buying devices. PCs in the 1990s. Smartphones in the 2010s. Laptops during COVID. The demand function was ultimately capped by the number of humans and the number of devices each human needs. One person buys one phone. Maybe one laptop. Perhaps a tablet. The DRAM content per device grows, but the number of endpoints is bounded.

This meant that once the initial adoption or upgrade wave passed—once everyone who needed a new PC had bought one, or every smartphone had been upgraded to the latest generation—demand would flatten. Supply, which was ramped during the boom, would overshoot. Prices would crash.

In the AI era, the demand function for memory has fundamentally changed. Memory is no longer predominantly serving a fixed number of »human endpoints«. Memory, especially HBM, is now a critical input for generating intelligence.

Think about what HBM (High Bandwidth Memory) actually does inside an AI accelerator. When you ask ChatGPT a question or run an inference on a large language model, the model’s parameters—billions or trillions of numerical weights—need to be loaded from memory into the GPU’s compute cores. The KV cache, which stores the context of your conversation, grows linearly with context length, with Grouped Query Attention (GQA) consuming roughly 0.06 - 0.12 MB per token in a 7B parameter model. A model with 70 billion parameters requires more than a single 80GB GPU worth of HBM just for the weights alone.

Here’s the simplified version: More memory = the ability to run larger models, with longer context, serving more users simultaneously. Memory is not a peripheral component in AI—it is the binding constraint. The so-called “memory wall” is the single biggest bottleneck limiting AI inference performance today. GPUs often sit idle, waiting for data to be fetched from memory. More bandwidth, more capacity means more intelligence output per second.

This is where the analogy to energy becomes powerful. Think about oil. When oil prices drop, what happens? Demand for oil increases because cheaper energy enables more economic activity. The demand curve for energy is downward-sloping- lower prices stimulate consumption. There’s always more work that could be done, more goods that could be transported, more heat that could be generated, if only energy were cheaper.

I believe AI inference demand behaves similarly. If memory costs drop and inference becomes cheaper, that doesn’t mean demand for inference drops. It means more applications become economically viable. More AI agents get deployed. More models get served. More context windows get extended. The demand for intelligence, like the demand for energy, is essentially elastic in response to price declines. Cheaper intelligence leads to more consumption of intelligence, not less.

This is the polar opposite of the gadget cycle. When DRAM prices dropped after the 2018 boom, it didn’t cause people to go buy a second smartphone. The number of endpoints was fixed. But when the cost of running an AI inference call drops by 50%, you can bet that the number of inference calls per day will more than compensate. Every enterprise that was waiting on the sidelines because of cost will deploy its AI project. Every startup that couldn’t afford the compute will spin up their service.

Here’s a human analogy I think captures this well. Imagine two people: one is a genius with poor memory, and the other is of average intelligence but has extraordinary memory and recall. In many real-world tasks—medicine, law, engineering, customer service—the person with superior memory will outperform the genius. Why? Because most practical work isn’t about raw reasoning power. It’s about retrieving the right piece of information at the right time. An AI model with more memory (longer context, more parameters accessible, faster retrieval) will outperform a theoretically smarter model that is memory-constrained. Memory is intelligence in many practical applications.

This is not a theoretical argument. The industry data supports it. HBM capacity per GPU has been scaling aggressively: NVIDIA’s A100 had 80GB of HBM2e. The H200 moved to 141GB of HBM3e. The upcoming Blackwell Ultra configurations push toward 288GB. And the Rubin Ultra platform is targeting 288GB - 576GB of HBM4E per GPU. The trajectory is exponential, and every generation of GPU is constrained by memory, not compute.

Where we are today

The current memory cycle is already historic in scale.

DRAM prices have surged dramatically. By Q4 2025, DRAM spot prices were nearly triple their level from a year earlier. DDR5 prices jumped 30-50% per quarter through H2 2025. Samsung raised memory prices by up to 60% since September 2025. DRAM inventories at major suppliers fell to just 3.3 weeks by the end of Q3 2025—matching the 2018 supercycle lows. SK Hynix and Micron had roughly 2 weeks of inventory each.

AI is expected to consume nearly 20% of global DRAM wafer capacity in 2026 when adjusted for HBM’s 4x wafer intensity.

The valuation: The market doesn’t believe in the durability of this cycle

Here’s where it gets really interesting from an investment perspective.

Despite the strongest fundamental setup the memory industry has ever seen—sold-out HBM capacity through 2026, record margins, structural demand from AI, and a three-player oligopoly with pricing discipline—the market is still pricing these stocks as if a classic downturn is imminent.

Micron trades at a forward P/E of about 10x, SK Hynix trades at approximately 5.2x forward P/E, and Samsung trades at a forward P/E of roughly 5x-7x—although this includes the total company, which includes much more than just memory.

The PEG ratio makes the mismatch even clearer. Micron’s PEG is approximately 0.16x, Samsung is at 0.17, and SK Hynix is at 0.10—meaning the market is pricing almost zero growth premium into the stocks.

But at these valuation levels, the question is not whether these companies will continue to grow; it’s more about how long the current demand signals will last. If these memory demand levels and margins stay here for a few more years, that would be a scenario that markets are not pricing in.

Why? Because the market has been burned by memory cyclicality before. Investors remember that in the 2017-2018 supercycle, Micron stock peaked at ~$64 with a forward P/E of about 4-5x at the top, and then the stock fell 56% even though earnings were still rising. The conditioned response is “memory is peaking, get out before the crash.”

But this framing assumes the old cycle repeats. It assumes that the demand driver (AI infrastructure buildout and inference scaling) behaves like the demand driver in previous cycles (consumer device upgrades). And I believe that assumption could be wrong.

Why the downturn when it comes might be shallower

I’m not arguing that memory prices will never decline. They will. At some point, new fab capacity from current investment plans will come online. At some point, HBM4 yields will improve, and supply will catch up. The 2017-2018 cycle teaches us that supply response is inevitable.

But I believe the depth and duration of the downturn will be structurally different this time (dangerous words I know):

1. The end market is not bounded by human endpoints. In the PC cycle, once every household had a PC, demand plateaued. In the smartphone cycle, once penetration hit saturation, annual unit growth went to zero. But the number of AI inference calls per day is growing exponentially and is nowhere near saturation. Every enterprise, every consumer app, every autonomous vehicle, every AI agent is an incremental consumer of memory bandwidth.

This view is also shared by many industry experts. Here is a former high-ranking employee from ASML on this topic:

2. Memory content per AI unit is growing exponentially, not linearly. DRAM content per PC grew from maybe 4GB to 16GB over a decade—a 4x increase. HBM content per GPU is going from 80GB (A100) to 288GB - 576GB (Rubin Ultra) in just a few years—a 7x increase. And the number of GPUs being deployed is also growing at 30-40% annually. The compounding effect of more units × more memory per unit is producing demand growth rates the industry has never seen.

3. HBM is structurally supply-constrained. One gigabyte of HBM consumes approximately 4x the wafer capacity of standard DRAM. HBM also requires advanced packaging (CoWoS or its equivalents), which has its own supply bottleneck. You can’t just flip a switch and convert commodity DRAM lines to HBM production. The manufacturing complexity acts as a natural supply governor that didn’t exist in previous cycles.

4. Long-term contracts are dampening volatility. In a major shift from past cycles, memory companies are increasingly locking in multi-year supply agreements with hyperscalers. SK Hynix has finalized its 2026 HBM supply plan with major clients and expects supply to remain tight through 2027. Micron has sold out its 2026 HBM capacity and has pricing agreements already in place. These contracts reduce the spot market’s influence and provide revenue visibility that the memory industry has never had before.

On top of the long-term contracts, the memory providers are much more careful with investing in new capacity this time, as the past cycle scars are a strong reminder. Here is a comment from a current Microsoft employee on what they expect in terms of memory supply coming online:

Perhaps the even more telling comment is this one made by a Fromer high ranking Micron employee on the internal cultural scars that the memory cycles have made:

5. The price elasticity of AI demand works in memory’s favor. If DRAM prices decline 20-30% (as they inevitably will at some point), the cost of running AI inference drops proportionally. This makes AI deployment cheaper, expanding the addressable market, which in turn supports memory demand. The demand floor is higher than in past cycles because cheaper memory creates new demand, rather than simply being absorbed by a fixed number of devices.

At some point, we will see a correction, but one that looks more like a 15-25% revenue decline and margins compressing to the 35-40% range, rather than the historic 30-40% revenue declines and sub-25% margins of previous busts. And crucially, I think the trough will be shorter, because AI inference demand will continue growing even during the cyclical correction, providing a demand floor that didn’t exist in the consumer device era.

The bottom line

The memory industry has spent 30 years teaching investors the same lesson: the cycle always turns, the crash always comes, and “this time is different” are the four most expensive words in investing. I respect that history deeply, and I’ve laid out the data to show you exactly how brutal those turns have been.

But I’m willing to bet against that lesson—partially—because the underlying demand driver has genuinely changed. That is why I also own stakes in SK Hynix and Samsung. Memory was a component in your gadget. Now it’s a substrate for intelligence. And the demand for intelligence—like the demand for energy, for computing, for connectivity—doesn’t follow the same saturation dynamics as consumer electronics.

The real risk for the memory cycle at the current stage is a technical breakthrough that would require orders-of-magnitude less memory and HBM, or a change that would bypass memory altogether. The chances of that happening today are low, but it is something to keep a close eye on all the time.

In the next section of this article for paid subscribers, I analyzed in detail how long I think this memory shortage and cycle will last, the timing of memory supply coming online for memory makers, including Chinese memory providers, and their possible effect on the market. Here is my take:

Read more

In this article, I’m breaking down my current thinking on Amazon. My goal here is to explain in detail the changes caused by AI on three pillars of the business: E-Commerce, AWS, and Advertising, and specifically how valuable each looks like in a world where AI agents increasingly sit between humans and the services they use. At the end, I’ll do a sum-of-parts valuation that I think gives a useful anchor for where the stock sits today.

Let’s start.

E-Commerce - the agentic threat and the logistics moat

Amazon captures roughly 40% of all U.S. e-commerce spending. It has 240+ million Prime subscribers globally (analyst estimates; Amazon last officially disclosed “over 200 million” in 2021), of which approximately 180–185 million are in the United States, representing penetration in about 80% of U.S. households. The Prime flywheel is well-documented: members spend on average $1,400/year, compared with $600 for non-Prime customers, and the retention rate after the first year is 99%, according to CIRP data. Amazon delivered over 8 billion items same or next day to U.S. Prime members in 2025, a 30%+ increase year-over-year.

This is the business everyone knows. But here’s the question that matters for the next 3–5 years: what happens when AI agents start shopping for consumers?

The agentic shopping risk

I strongly believe that in the future, most e-commerce shopping will be done through AI agents acting as personal assistants to consumers, instead of direct consumers. I am not alone in those expectations. McKinsey projects agentic commerce could generate $1 trillion in U.S. retail revenue by 2030. Morgan Stanley expects nearly 50% of American shoppers will use AI agents by then, potentially adding $115 billion in e-commerce spending. Bain research shows that 30–45% of U.S. consumers already use GenAI for product research and comparison. During Cyber Week 2025, roughly 1 in 5 orders on Shopify involved an AI agent. AI-driven traffic to retailer sites has surged 7x since January 2025, according to Shopify data, with AI-driven orders up 11x.

What’s happening is this: instead of opening the Amazon app, a consumer tells ChatGPT, Claude, or Gemini what they need. The agent searches across retailers, compares prices, checks reviews, and either completes the purchase or presents a shortlist. OpenAI has already embedded checkout directly into ChatGPT. Perplexity launched its Comet browser agent. Google is rolling out agentic AI shopping tools.

This is a major shift in consumer behaviour, and Amazon knows it. In November 2025, Amazon sued Perplexity for its AI browser agent making purchases on Amazon’s marketplace. The company has blocked 47 AI bots from crawling its site. But at the same time, CEO Andy Jassy acknowledged on their recent earnings call that agentic commerce “has a chance to be really good for e-commerce.” Amazon recently even posted a job for a principal corporate development officer specifically for “agentic commerce” partnerships.

Forrester retail analyst Sucharita Kodali captured the tension perfectly:

Amazon’s shot at owning the application layer

That said, Amazon isn’t conceding the front-end. They have several assets that give them a legitimate shot at being a surface where agentic shopping enters:

Rufus — Amazon’s AI shopping assistant, used by more than 300 million customers in 2025. Customers using Rufus complete purchases at a 60% higher rate. It can now auto-purchase items when prices hit thresholds.

The most interesting recent project is the »Buy For Me« project. This is Amazon’s experimental agent that can purchase from other retailers within the Amazon app. This is a smart flip from Amazon: instead of being the store that other agents shop, Amazon becomes the agent that shops everywhere else. Amazon does have some unique assets that make it valuable as the front-end touchpoint, and the key is around Prime Subscriptions.

Prime Video — 315 million ad-supported viewers globally, up from 200 million in early 2024. This is a massive surface for product discovery and agentic commerce integration, especially through interactive shoppable ads during live sports (Thursday Night Football averaged 15.3 million viewers, +16% YoY). Twitch — 105+ million monthly users, heavily Gen Z. An engaged, commerce-friendly audience. Alexa — still the most widely deployed voice assistant in smart home devices. If agentic commerce moves to a voice-first or ambient-first paradigm, Alexa has a head start.

The risk here is that those surfaces might not be enough and that Amazon might not be aggressive enough in the early days of where we are today. From today’s vantage point, the dominant surfaces if I had to choose would still be the smartphone assistant, or a standalone AI app (similar to ChatGPT, Gemini, Claude), and later on the AI glasses and personal assistant given by that provider. Prime Video and Twitch will still serve as important discovery platforms and could turn out to be much more valuable in terms of ads in a world where it will become increasingly hard to reach a human via digital channels, as internet usage will be dominated by AI agents instead of humans. Still, it doesn’t solve the fact that the application layer, where most of the e-commerce starts, moves to other providers. Even if Amazon were to launch an independent AI shopping assistant app, I don’t think in the long-term that would be »moaty« enough. My view is that the dominant provider will be the one that can offer a full AI personal assistant, with shopping as one of its features, not the only or main one. For that to be Amazon, they would need to make an aggressive pivot from current levels and a possibly strong shift into consumer hardware, which I don’t think is their plan.

With all that said, my base case is that Amazon will not be the application layer of agentic shopping and that its e-commerce business will move to the backend part of the shopping experience (still being important). Even in this scenario, Amazon still makes a decent margin given the logistics, payment, and fulfillment infrastructure that it offers at scale.

Advertising

Amazon’s advertising revenue hit $68.6B in 2025, growing 22% YoY in Q4. This is now 9.6% of Amazon’s total revenue, up from 5.9% in 2021. To put it in context, Amazon’s ad business alone is larger than the total revenue of companies like Netflix, Uber, or Salesforce.

But here’s the nuance that most analysts don’t discuss: Amazon’s ad business is really two very different businesses glued together.

Search ads

The vast majority of Amazon’s advertising revenue comes from Sponsored Products: essentially search ads within Amazon’s marketplace. When you search for “wireless headphones” on Amazon, the first several results are paid placements. Amazon doesn’t break this out precisely, but based on WARC data, the retail media component (primarily search ads) accounts for roughly $60.6B of the estimated total, with Prime Video and other upper-funnel formats making up the incremental portion.

Here is my concern: search ads on Amazon are fundamentally tied to humans browsing Amazon’s website and app. If an AI agent shops for you, it doesn’t look at sponsored listings. It doesn’t scroll past display ads. It skips right to the product that best matches your criteria and places the order. As Bain research noted, about 65% of retail media spending still occurs onsite, and that entire bucket is at risk if product discovery shifts to AI-driven search.

This is why I think the search ad portion of Amazon’s advertising business is on a disruption clock. Not tomorrow, not next quarter, but over a 3–5 year horizon, the economics of Sponsored Products face a structural headwind as agentic interfaces capture more of the purchase journey and as we talked in the previous section I give it a low probabiliticy chance that Amazon is able to capture the AI agent assistant application layer so the eyeballs switch from amazon’s site and apps towards the AI assistant owners.

Prime Video ads

The other side of Amazon’s ad business is Prime Video advertising, and this is the piece I think is defensible. Amazon introduced ads on Prime Video in January 2024. S&P Global Market Intelligence Kagan estimated Prime Video’s ad revenue at $433M in 2024 and forecast it to reach $806M in 2025. This is still a small fraction of total ad revenue, but it’s growing fast and serves a different function: brand advertising through streaming video is not susceptible to agentic disintermediation the same way search ads are.

Prime Video reaches 315 million monthly ad-supported viewers globally. That’s larger than Netflix’s ad-supported tier at 190 million. Thursday Night Football alone averaged 15.3 million viewers with 16% growth YoY, and the Packers-Bears wild-card playoff game drew 31.6 million viewers, the most-streamed NFL game in history. Amazon has also integrated Netflix and Spotify inventory into its Amazon DSP, giving advertisers a broader programmatic buying platform.

My estimate is that by 2027–2028, Prime Video ads could reasonably be a $3–5B annual revenue stream, growing at 40%+ rates as ad loads increase and live sports inventory expands (NBA deal kicks in, international sports expansion). This business is much more structurally defensible because people watch content — AI agents don’t.

But even that revenue doesn’t materially change my thesis that the majority of Amazon’s ad business is at risk of serious disruption.

For the sum-of-parts analysis in the last part of this article, I’m splitting the ad business into two buckets. For the search/retail media portion (~$60–63B), I’m assigning it a terminal value as if profits only last 4 more years with zero terminal value after that. That’s deliberately punitive - I’m assuming this revenue stream is structurally impaired. For Prime Video ads, I’ll fold it into the e- commerce/subscription ecosystem, where it has long-term durability.

AWS - the cloud business

AWS is the most important reason why I own Amazon stock and why it has now become my biggest portfolio position.

The biggest fear around AWS has been that AI-related capital expenditures would permanently compress margins. And yes, there was a dip: AWS’s operating margin fell to 32.9% in Q2 2025 as the company ramped up spending aggressively. But by Q4, it had recovered to 35.0%, and the full-year margin was 35.4%.

Here is my core argument: we are severely compute-constrained for the foreseeable future. Amazon has invested $131.8B in capex for 2025 and has guided to approximately $200B for 2026, predominantly for AWS infrastructure. The company added more than 1 gigawatt of data center capacity in Q4 alone and 3.9 gigawatts in the trailing 12 months, which is double what AWS had in total in 2022. And Andy Jassy expects to double power capacity again by the end of 2027.

Despite this massive buildout, demand continues to outstrip supply. Jassy noted on the Q1 call that GPU and motherboard shortages were limiting the pace of AI workload onboarding. Bedrock (Amazon’s managed AI service) reached a multi-billion-dollar annualized run rate with customer spend growing 60% quarter-over-quarter to a base of over 100,000 customers. Trainium2 is fully subscribed with 1.4 million chips deployed.

In this environment, there is no incentive for hyperscalers to engage in a pricing war. When every chip you install is immediately monetized, you don’t cut prices — you add capacity. Until compute supply catches up with demand (which I don’t expect before 2029 at the earliest), AWS can maintain mid-30%+ operating margins without sacrificing growth. The margin should hold around pre-AI era levels (AWS operated in the 28–35% range historically, with 2024 averaging 37%) because the scarcity dynamic supports pricing power.

Trainium and Custom Silicon are key things for long-term margins

This is a point I don’t think gets enough attention. NVIDIA’s gross margin sits at roughly 73–75%. Every cloud provider that is 100% dependent on NVIDIA for AI compute is paying that tax on every GPU. That cost flows through to the cloud provider’s cost of revenue and structurally limits the margin they can earn on AI workloads.

Amazon, through its Annapurna Labs subsidiary, has developed Trainium and Inferentia custom ASICs, as well as Graviton CPUs for general compute. Combined, these custom chips have surpassed a $10B annualized revenue run rate, growing at triple-digit percentages YoY. According to Amazon, Graviton provides 40% better price-performance than x86 processors and is adopted by 90% of AWS’s top 1,000 customers.

Trainium2 powers Project Rainier, the world’s largest operational AI compute cluster with 500,000+ Trainium2 chips, which Anthropic uses to train its Claude models. Trainium3 is in preview with broader volumes expected in early 2026, and Trainium4 is targeted for 2027.

I am sharing here the chart that we made some months ago in our detailed Amazon Trainium piece, where we calculated the manufacturing costs of Amazon Trainium, Google TPUs, and Nvidia’s Blackwell B200:

You can see the most significant difference: if it costs Amazon $ 3,000-$ 3,500 to produce a Trainium3 chip, it costs them $35k-$40k to buy an Nvidia B200 chip. Even though B200 is much more performant from a cost-of-ownership perspective, Trainium3 gives B200 a run for its money.

The margin math is straightforward. When you design and manufacture your own silicon (using TSMC for fabrication and a design partner like Broadcom, Marvell, MediaTek), your cost per unit of compute is significantly lower than buying merchant silicon from NVIDIA at a 73–75% gross margin. This gives AWS a structural margin advantage for AI workloads vs. a competitor that sources 100% from NVIDIA. It doesn’t mean AWS abandons NVIDIA (it still offers NVIDIA instances), but having an alternative lets AWS capture more of the AI value chain and maintain margin in ways that someone who is entirely dependent on NVIDIA simply cannot.

This key difference will prove even more important in the coming years, especially once demand/supply for compute is more in balance and the hyperscalers’ focus shifts from capturing revenue growth to profitability and customer optimization.

Traditional cloud demand is actually accelerating because of AI agents

There’s a narrative that AI is all that matters for AWS growth. That misses something important: AI agents themselves create enormous demand for traditional cloud services, as we already discussed in part in our The Forgotten Chip: CPU the New Bottleneck of the Agentic AI era article. Every AI agent needs storage (S3), compute (EC2, powered increasingly by Graviton), databases, networking, and monitoring. The more AI agents there are in production, the more traditional cloud infrastructure gets consumed.

The number of AI agents and their deployment is rapidly surging right now. Here is an alt provider that tracks the Model Context Protocol (MCP), an open-source standard for connecting AI applications to external systems.

The number of MCP servers being set up every month is growing exponentially, and the MoM pace is accelerating. The market is still in very early stages, as the current number of MCP servers is probably less than 1% of the API market. But the interesting thing was comparing where these MCP servers were being deployed with where API deployments are. Both Azure and GCP % of these MCP server deployments were lower compared to their API deployments, while AWS MCP deployments actually rose compared to API deployments:

While this data is not large enough yet, it could indicate that more smaller companies are on AWS than on the other two hyperscalers and that they are early adopters. The data in some form also shows the importance of AWS’s »legacy« cloud infrastructure, which is very much needed in the Agentic AI phase.

The demand for traditional AI infrastructure is skyrocketing, and you can see it from comments from CEOs of AMD and Intel, where they are basically sold out of CPUs, and you can see it when talking to other industry experts.

This is a former Amazon employee talking about the usage of the AWS S3 service (storage):

The key takeaway from this comment is the need to store this data for audit and safety purposes. Companies running these AI agents need clear, auditable trails of what an AI agent has done, so they can track and monitor as problems arise and fix them. Nobody wants to give an AI agent full permission to run freely across the company's stack and make changes and tasks that nobody has visibility into. This human visibility means services like Storage grow even more in usage.

A big tell was also the November OpenAI- AWS deal. The press release stated that OpenAI would access “hundreds of thousands of state-of-the-art NVIDIA GPUs, with the ability to expand to tens of millions of CPUs to rapidly scale agentic workloads.”

The GPU part is known, but the CPU part is the most interesting one. We need »legacy« cloud workloads and CPUs to enable the AI agent economy; that is just the way it is, and this is a big uplift for AWS, which has the largest fleet of optimized cloud services out there.

Amazon noted that more of the top 500 U.S. startups use AWS as their primary cloud provider than the next two providers combined. That startup and scale-up cohort is building AI-native applications that are heavily cloud-intensive.

The on-prem fallacy and SMB lock-in

Some investors argue that AI inference will eventually move to the edge or on-prem, killing the cloud growth story. Let me push back on this.

First, even if 90% of personal AI assistant use cases eventually run on edge devices (phones, laptops, local hardware), the remaining 10% that stays on cloud or on-prem infrastructure is still an enormous market. These are the “god-like AI” use cases: complex enterprise reasoning, multi-step agentic workflows, financial modeling, drug discovery, code generation at scale. These require the kind of compute density and model size that doesn’t fit on a phone. And these use-cases are the most profitable, as their outputs are the most valuable.

Second, on-prem AI infrastructure is radically more complex than anything businesses have managed before. Running an AI inference cluster on-prem means managing GPU and CPU servers, networking fabric, cooling systems, model deployment pipelines, and monitoring at a level of sophistication that most IT departments have never dealt with. For any small or medium-sized business, the cost and complexity of running your own AI infrastructure to have your “AI accountant” or “AI customer service agent” simply doesn’t make sense when you can rent it from AWS for a fraction of the upfront cost with zero operational hassle.

The cloud is the natural home for AI workloads for the vast majority of companies, and that reality isn’t changing anytime soon. If anything, as AI becomes more central to knowledge work, more companies will move to the cloud specifically to access AI capabilities they can’t build or run themselves.

The revenue trajectory of AWS

With all that in mind, I believe AWS, with its power capacity availability, which I already discussed in my previous articles, is well-positioned for multiple quarters of accelerating growth. I believe, despite AWS’s size, we will soon see the segment grow by +30% YoY. AWS also exited Q4 2025 with a backlog of $244B with a weighted average remaining life of 4.1 years. Capacity is being installed and monetized as fast as it comes online.

If AI agents truly absorb a meaningful portion of knowledge work over the next 5–10 years — and companies like Anthropic ($19B ARR rate up from $9B just two months ago) and OpenAI are building the models to do exactly that — then the total demand for cloud inference is going to be multiples of what it is today. Every AI-powered accountant, lawyer, engineer, customer service agent, and analyst running in the cloud creates recurring compute demand.

The Anthropic and OpenAI stakes are hedges

Besides the already mentioned segments, Amazon also has other important aspects such as Project Kupier, Subscription business, and stakes in Anthropic and OpenAI, which are now becoming increasingly important.

Amazon has invested approximately $8B in Anthropic (capped below 33% ownership) and recently announced a strategic partnership with OpenAI that includes an investment of up to $50B (starting with an initial $15B commitment, with the remainder tied to milestones and a potential OpenAI IPO).

Anthropic just closed a $30B funding round at a $380B post-money valuation in February 2026. If Amazon holds roughly 20% of Anthropic (estimates vary given the cap structure), that stake is worth $76B on paper. But in the last few weeks, Anthropic has accelerated its adoption and revenue growth so much that a $500B valuation for a company that will probably exit 2026 at a $50B ARR growing 5x YoY and disrupting the whole knowledge work economy is nothing extraordinary, which would add $100B of value or almost 5% of Amazon’s current market cap.

For OpenAI, the proposed $100B funding round would value the company at approximately $830B. Amazon’s $50B investment at those terms would represent roughly a 6% stake.

Combined, these stakes could be worth +$145B. And here’s the real value: in a world where Anthropic, OpenAI, and Gemini become the application layer, having significant stakes in two of those companies isn’t just financial investments. They are Amazon’s guarantee that the biggest AI consumers remain AWS customers. OpenAI has committed to spending $100B on AWS over the next eight years. Anthropic is using Project Rainier (500,000+ Trainium2 chips) for training. Both are locked in as massive cloud customers.

Valuation

Now let’s put the numbers together. I’m deliberately being conservative in places and factoring in serious disruption risk. Here are my numbers:

Read more