Amazon, Google, Microsoft, Meta Q1 earnings: AI profits are here, custom silicon is winning
Hey everyone,
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:
»The majority of the backlog is related to typical GCP contracts and we expect to recognize just over 50% of the backlog as revenue over the next 24 months«
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:
»Cloud operating income was $6.6 billion, tripling year-over-year and operating margin increased from 17.8% in the first quarter of last year to 32.9%«
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:
»Our enterprise AI solutions have become our primary growth driver for cloud for the first time. In Q1, revenue from products built on our GenAI models grew nearly 800% year-over-year.«
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:
»we’ll begin to deliver TPUs to a select group of customers in their own data centers in the hardware configuration to expand our addressable market opportunity.«
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:
»Turning to Search. AI continues to drive search usage, and queries are at an all-time high«
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:
»Our AI business surpassed $37 billion ARR, up 123%.«
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:
»Thanks, Brent. I think -- we’ve been talking about sort of where this AI business of ours has been in the cycle compared to even the cycle we saw with the cloud, which now seems very long ago. And how margins were actually better and they remained better in our AI business versus where we saw in the cloud transition, looking back.«
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:
»In knowledge work, it was another record quarter for Microsoft 365 Copilot seat ads, which increased 250% year-over-year, representing our fastest growth since launch. Quarter-over-quarter, we continue to see acceleration and now have over 20 million Microsoft 365 Copilot paid seats. The number of customers with over 50,000 seats quadrupled year-over-year and Accenture now has over 740,000 seats, our largest Copilot win to date.«
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:
»We have seen a surge in usage of our first-party agents with monthly active usage up 6x year-to-date. Copilot queries per user were up nearly 20% quarter-over-quarter. To put this momentum in perspective, weekly engagement is now at the same level as Outlook, as more and more users make Copilot a habit.«
Weekly engagement at the level of Outlook is a wild data point.
GitHub Copilot also continues to scale:
»We see this even with GitHub Copilot. Nearly 140,000 organizations now use GitHub Copilot and enterprise subscribers have nearly tripled year-over-year.«
And here is where it gets interesting - Microsoft is moving GitHub Copilot to a usage-based pricing model:
»And earlier this week, we announced our move to usage-based pricing model for GitHub Copilot as we align pricing to actual usage and cost«
»Microsoft Cloud gross margin percentage should be roughly 64%, down year-over-year, driven by continued investments in AI and increased GitHub Copilot usage. Just this week, we announced a business model transition in GitHub Copilot that will align pricing with usage and value that takes effect on June 1 of this year.«
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:
»Bookings growth was impacted by weaker renewals as customers balance spend between the traditional per-seat and the emerging seats-plus-consumption model.«
So enterprises are already adjusting their procurement around this. Hybrid pricing models are becoming the standard.
On capacity, Microsoft was, again, very direct:
»Even with these additional investments and continued efforts to bring GPU, CPU and storage capacity online faster, we expect to remain constrained at least through 2026. Despite these constraints, and the continued need to balance incoming supply, we expect Azure growth to show modest acceleration in the second half of the calendar year compared with the first half.«
And then this:
»I think in so many ways, this just reminds us of the last cycle. And when the TAM is so expansive and when shortages are generally, I think, growing seems to be the sentiment between supply and demand«
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:
»Over 10,000 customers have used more than one model on Foundry. 5,000 have used open source models, and the number who have used Anthropic and OpenAI models increased 2x quarter-over-quarter.«
»The majority of users leverage multiple models.«
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):
»We’re guiding for that to be better again in Q4. I think that’s where you’re starting to see, right? I think the thing that investors have been asking and Mark, you’re asking about is when we’ll start to see that show up in revenue growth. And I think that’s the first place you point to. We can also point to it, and I think you’ll start to see it in GitHub, right, where you see revenue growth rates and usage consumption models result in acceleration in the top line.«
In other words, the next two quarters are when Microsoft expects AI products to start showing up as accelerating revenue, not just bookings.
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:
»Starting with AWS, growth continued to accelerate, up 28% year-over-year, the fastest growth rate in 15 quarters, up $2 billion quarter-over-quarter, the largest Q4 to Q1 AWS revenue increase ever. AWS is now a $150 billion annualized revenue run rate business«
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:
»high-performance inference with the leading selection of frontier models in Bedrock, which saw 170% growth in customer spend quarter-over-quarter and processed more tokens in Q1 than all prior years combined.«
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:
»We saw nearly 40% quarter-over-quarter growth in Q1, and our annual revenue run rate is now over $20 billion and growing triple-digit percentages year-over-year, but this somewhat masks the size. If our chips business was a stand-alone business and sold chips produced this year to AWS and other third parties as other leading chip companies do, our annual revenue run rate would be $50 billion. As best as we can tell, our custom silicon business is now one of the top 3 data center chip businesses in the world, the speed at which we’ve gotten here is extraordinary. And we have momentum.«
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:
»And we now have over $225 billion in revenue commitments for Trainium.«
$225B in commitments just for Trainium is insane and was the most shocking number for me among all the earnings calls yesterday.
»Amazon Bedrock, which is used expansively by over 125,000 customers, runs most of its inference on Trainium and almost 80% of the Fortune 100 companies are using Bedrock.«
»While the largest number of AI chips we’re bringing in are Trainium, we continue to have a deep partnership with NVIDIA«
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:
»Different companies will offer different benefits for customers and the uniquely strong price performance that Trainium offers is compelling to our external and internal customers. For perspective, at scale, we expect Trainium will save us tens of billions of dollars of CapEx each year and provide several hundred basis points of operating margin advantage versus relying on others’ chips for inference.«
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:
»We continue to see customers increase cloud migrations and scale their use of AWS core services. Customers seeking the full benefit of AI are accelerating their transition to the cloud. We also see a strong correlation between AI spend and core growth. As customers spend more on AI, we see a corresponding demand increase in core.«
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:
»On the backlog, the backlog for Q1 is $364 billion. That does not include the recent deal that we announced with Anthropic for over $100 billion. There’s reasonable breadth in that as well. It’s not just 1 customer or 2 customers.«
$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:
»On the question about Trainium and the notion of our selling racks over time, I do think that’s very much a possibility. Always, we have to balance -- we have such demand right now for Trainium, and we have such demand from various companies who will consume as much as we make that we have to decide how much we’re going to allocate to the existing demand and customers and how much we’re going to save to sell as racks. ... But I expect over time, there’s a good chance we’re going to sell racks over the next couple of years.«
The most important framing on AI workloads from the entire AWS call:
»Most of the value companies derive from AI will be through agents«
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:
»On Instagram, the ranking improvements that we made in Q1 drove a 10% lift in reel time spent.«
»On Facebook, total video time increased more than 8% globally in Q1, the largest quarter-over-quarter gain in 4 years. Within the U.S. and Canada, ranking improvements we made drove a 9% increase in video watch time on Facebook in Q1.«
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:
»In the second half of last year, we began rolling out our new adaptive ranking model, which is an LLM scale adds recommender model that we use for inference. This model improves our inference ROI by routing requests to more compute-intensive inference models when it determines there is a higher probability of conversion.«
And then this longer segment, which I think is one of the most important pieces of technical commentary:
»Historically, we haven’t used larger model architectures like GEM for inference, because their size and complexity would make them too cost prohibitive. And the way we drive performance from those models is by using them to transfer knowledge to smaller, more lightweight models that are used at run time. The inference models are bound by strict latency requirements since they need to find the right ad within milliseconds, and that has, again, historically prevented us from meaningfully sizing up -- scaling up their size and complexity. But in the second half of last year, we introduced a new adaptive ranking model, which enables us to leverage LLM scale model complexity of 1 trillion parameters, and we made advances in the model architecture and codesigned the system with the underlying silicon, so it maintains the sub-second speed that is required to serve ads at scale. We also developed an approach that intelligently routes requests more compute-intensive inference models if it determines that there is a higher probability of conversion and that lets us drive both better performance and increase inference ROI.«
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:
»we are increasing our infrastructure CapEx forecast for this year. Most of that is due to higher component costs, particularly memory pricing, but every sign that we’re seeing in our own work and across the industry gives us confidence in this investment.«
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:
»That said, we are very focused on increasing the efficiency of our investments, and as part of that, we are rolling out more than 1 gigawatt of our own custom silicon that we’re developing with Broadcom, as well as a significant amount of AMD chips to complement the new NVIDIA systems that we’re rolling out as well.«
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:
»So you’re getting to a point where today, the models are still able to learn from people and then I think at some point, the models will have to improve themselves. And that’s how the growth is going to -- an improvement in the models is going to happen. And if you don’t -- if we don’t have an ability to do that, then we or anyone else, I think the companies that don’t do that are not going to be leading labs, then they’re not going to produce leading products. So I think that, that’s like -- that is a table stakes thing that we are focused on.«
»but then the model improvement, I think, is going to be something that’s going to go on for a very long time.«
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:
»That’s why we believe that we need to be a company that builds frontier models in addition to building the agents. And then in order to do that, you, of course, need to build your infrastructure in order to be able to do that well. So we’re undertaking this large investment to be able to do that top to bottom.«
»But I don’t hear any other labs out there talking about how they’re building an AI that’s really good at shopping. And I think that the reason for that is like not because shopping is the most important thing by itself, but because like empowering people to do the things that matter in their lives, whether that’s local or understanding social context, or shopping or personal health things or understanding what’s going on around them visually«
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.
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.