Amazon, Google, Microsoft, Meta Q2 earnings: The AI CapEx ROIC is bad thesis is DEAD

Hey everyone,

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:

»A lot of the expense, especially you see it in CapEx, you’ve seen our CapEx really pivot toward what I would call and do call short-lived assets, which really, right, that CPUs and GPUs that have relatively shorter lead times. And so if the demand environment changes, you just slow down what is, in fact, the largest component, right, and the driver of COGS. The investment into land and data center builds is actually quite flexible, right?«

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

»There are 2 major parts of the investment, the data centers and the servers and networking equipment that go into them. These have different capital cycles. Data center capital is spent starting 2 years before we can put servers into them to start monetizing. Once a data center opens with servers plugged in, we start generating significant revenue right away and then get to monetize these data centers for 30-plus years without having to spend that start-up capital again.«

»Servers and networking equipment operate on a shorter cycle. We typically purchase these a few months before putting them into service, so we have strong visibility into customer demand before we trigger the spend. If the demand isn’t there, we won’t spend the capital. For servers and networking equipment, on average, it takes a little less than 3 years to break even on that investment. The servers currently have a useful life of at least 5 to 6 years, and most of our AI capacity these days is being contracted for at least 5-year terms.«

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:

»Therefore, our longer-term capacity strategy aims to give us the flexibility to continue growing compute in 2028 and beyond by laying down data center and network foundations to accommodate future server decisions. The long-lived nature of these assets inherently provides the flexibility that will make it possible to adjust our investment to the pace of AI adoption.«

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

»Look, we are working off a disciplined ROIC framework here. Obviously, we, to the extent that our input costs going up to us, we reflect that in our ability to price our solutions and see returns there.«

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:

»But we also want to make sure we have a resilient, not just growth outlook but also a resilient balance sheet and a strong balance sheet, a healthy balance sheet, which is the rationale behind expanding into the equity markets. At this point, we’re not planning to go back to the equity markets with the exception of, as you recall, part of our equity offering was the ATM.«

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:

»We expect to recognize just over 50% of the total backlog as revenue over the next 24 months.«

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:

»The increase in the range is primarily due to an acceleration in the delivery of capacity to meet growing demand.«

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:

»For the full year, our cloud revenue surpassed $214 billion with nearly 90% from customers outside of frontier model companies.«

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:

»Commercial remaining performance obligation grew 84% to $678 billion. All sequential commercial RPO growth was driven by commitments from customers outside of frontier model companies. And RPO increased 25% when excluding OpenAI. RPO, including OpenAI, has a weighted average duration of 2.3 years and roughly 30% will be recognized in revenue in the next 12 months, up 37% year-over-year.«

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:

»When it comes to knowledge work, we now have over 30 million paid Microsoft 365 Copilot seats, with net seat adds more than doubling quarter-over-quarter. Over the last 3 quarters, user satisfaction scores have doubled and are now at an all-time high.«

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:

»The number of conversations per user nearly doubled year-over-year. Average weekly engagement is on par with Outlook and Teams.«

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:

»We are building a new model system where the harness, context, memory, and action space are separate from any one model family, thereby moving the frontier on the cost-to-outcome curve. And it’s not just about cost. It also has the added benefit of business continuity and resilience because every model is substitutable.«

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.

---

---

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.

»Our chips business now has an annual revenue run rate of over $25 billion, growing triple-digit percentages year-over-year. Our AI revenue run rate climbed significantly quarter-over-quarter, and is now also over $25 billion, growing triple-digit percentages year-over-year.«

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:

»We’re seeing strong growth across both AI and non-AI, what we call core, and growth in one is driving growth in the other. Growth in AI drives core because post-training reinforcement learning and agent tool use is mostly done on CPUs versus AI accelerators.«

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:

»As customers invest in AI, we see a corresponding increase in core consumption. We expect this relationship to strengthen over time as more AI workloads move into full-scale production.«

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.

»The profitability you’re seeing from AWS isn’t random, it’s a result of disciplined efficiency gains, capacity optimization, which we benefited quite a bit from in Q2, and always closely managing our fixed costs.«

And on the AI workloads specifically:

»We see the margins and returns in AI tracking what we saw with core at the same point of evolution, actually, a little ahead.«

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:

»In fact, the demand we already have for 2028 is striking. And remember, enterprises are still very early in using inference at scale in their current production applications. We long believed AWS could become a few hundred billion-dollar revenue business and now believe it’ll be at least double that, and very possibly be $1 trillion annual revenue business for us in time, with very appealing accompanying free cash flow and return on invested capital.«

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:

»On Instagram, global time spent this quarter grew double digits year-over-year, largely driven by improvements to our feed and reels recommendations. On Facebook, video time spent increased 9% globally year-over-year and over 10% within the U.S. and Canada, where it was driven by ranking improvements. We are finding that LLMs are increasingly capable of delivering ranking and recommendations gains.«

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:

»And soon, it will go further, including suggesting ways to grow your business, giving you competitive intelligence and real-time insights into what’s working and what’s not. And over time, we’d like to build this into a business-in-a-box service that can help you start and run a whole business using Meta’s platforms. In terms of how we will monetize these, we have a mix of subscriptions, volume-based pricing, and I expect that we’re going to continue to evolve more of these products to be like our ad systems where businesses only pay us when we achieve results for them.«

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

2. Compute revenue - direct.

»We’re getting a lot of offers for compute at a significant premium over what we paid for it.«

And the mechanism Zuck laid out is genuinely novel:

»Over time, that will let us run an efficient auction over our compute, similar to how we do that for advertisers today.«

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.

»Now in terms of running the business, obviously, a common trade-off that we need to make is around how much do you monetize something today versus develop future assets for the future. And I think that it’s always a portfolio, right? It’s not like you don’t want to only do long-term things and do no -- like -- and not kind of prove the markets have that exist in the near term, but I also think it would be foolish to basically just sell all of the compute and take a short-term profit«

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:

»I expect it to become a lot easier to ship new apps. So we are planning to build out more ideas and use our recommendation systems to scale them to the people who will find them interesting, as we’ve done with Threads.«

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:

»A lot of people view the surface layer of we build some social media apps and we have an ad business. We are really a full-stack technology company. We built our own data centers, our own infrastructure, our own chips, our own low-level software. ... It just seems to me pretty clear that having kind of sovereignty over building your own models is going to be an important part of that stack going forward.«

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:

»I get that this is sort of a big bet across the industry. My personal bet is that the people who invest in this are going to be rewarded and feel very good over time.«

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