cloud
One of the things about being a cloud builder at the scale of the tech titans is that they control their own fates. They co-design and optimize their hardware and software stacks, wringing the best bang for the buck out of that iron so they can afford to make it at scale and then rent capacity on their infrastructure at a profit.
As we have pointed out a number of times, the big clouds – Amazon Web Services, Microsoft Azure, and Google Cloud – are more like original design manufacturers than not, creating fully integrated software stacks that also allow for multiple options within that stack if customers want to do that. This has vertical integration, but does not sacrifice horizontal optionality at any layer in the stack. While the big clouds allow you to have outposts – versions of their infrastructure that you can run in your datacenter – these outposts are still owned by them and managed by them. But the big AI model makers and probably more than a few large enterprises are going to want to get cheaper infrastructure by literally owning and running it themselves.
Google has already going down this road, allowing Anthropic to buy 3.5 gigawatts of TPU racks and install them in its own datacenters, starting in 2027. The TPUs systems are designed by Google but will be built by Broadcom, with both Google and Broadcom making some dough on the deal. For now, Anthropic is renting TPU capacity on Google Cloud, and that is definitely not the cheapest way to get AI compute engines. We have calculated that TPU capacity might cost on the order of $30 billion to $35 billion per gigawatts, which brackets the full system investment being made by Anthropic at somewhere between $105 billion to $122.5 billion, all in, including the datacenters, power, cooling, and iron.
Google had two choices when Anthropic said it wanted to buy TPU systems outright: Say yes, or potentially lose Anthropic, which actually has money and which is growing like crazy, as a customer. Anthropic has a tight relationship with AWS as well, and is of the 1.4 million Trainiums that AWS has deployed on its cloud as of the end of last year, Anthropic is training its Claude models and running inference against them on over 1 million Trainium2 chips. AWS and Anthropic have inked a deal to get an additional 1 gigawatt of Trainium2 and Trainium3 chips into the field by the end of 2026, which is somewhere between 500,000 and 600,000 XPUs. And just this month, the two companies inked a $100 billion deal to get 5 gigawatts of Trainium capacity between now and 2036, which is somewhere between 2.5 million and 3 million XPUs.
We do not think that Anthropic is going to rent all of that capacity, but rather wants to buy complete Trainium systems and park them in its own datacenters and stop paying the cloud premium. And we think that Anthropic will work with Marvell and Alchip, the chip shepherds that the Annapurna Labs division of AWS uses to get Trainium XPUs, Graviton CPUs, and Nitro DPUs into its own datacenters, to make racks of iron for it much as Google is allowing Broadcom to do the same with its TPU systems sold to Anthropic. What choice does AWS have? AWS can become an OEM for Anthropic or lose the business.
On a conference all going over the financial results for the first quarter of 2026, Andy Jassy, who ran AWS for a long time and who has been chief executive officer of Amazon for the past several years, danced around this idea, but didn’t deny it.
“On the question about Trainium and the notion of our selling racks over time, I do think that is very much a possibility,” Jassy explained. “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 are going to allocate to the existing demand and customers and how much we are going to save to sell as racks. And for our existing customers that we sell Trainium to, how many will be Trainium plus running on our cloud infrastructure versus just the chips themselves? But I expect over time, there is a good chance we are going to sell racks over the next couple of years.”
I think this is already baked into the Anthropic-AWS deal.
Moreover, some statements put out by AWS in the financial report surely make us think that AWS is pondering being a chip and system supplier as a complement to its selling virtualized compute engine capacity on its cloud. These statements were not in the press release or discussion with Wall Street, but sent separately by the PR people at Amazon. Here they are, and they are very interesting indeed:
- Amazon's chips business – inclusive of Graviton, Trainium, and Nitro – saw nearly 40 percent quarter-over-quarter growth in Q1, with an annual revenue run rate now over $20 billion, growing triple-digit percentages year-over-year. (I presume this is the revenue that Annapurna Labs books for internal sales to the AWS cloud and possibly to the Amazon parent company.)
- If the chips business were standalone and sold chips produced this year to AWS and third parties – as other leading chip companies do – the annual run rate would be ~$50 billion.
- Amazon now has over $225 billion in revenue commitments for Trainium, with an increasing number of companies betting on custom silicon.
- Trainium2 offers ~30 percent better price/performance than comparable GPUs; largely sold out. Trainium3 has been shipping since early 2026, with 30 percent to 40 percent better price/performance than Trainium2; nearly fully subscribed. Trainium4 is ~18 months from broad availability; much already reserved.
- Amazon Bedrock, used by over 125,000 customers, runs most of its inference on Trainium.
The upshot is that if AWS wants to cover its Trainium and Graviton and Nitro costs, it is going to have to sell systems to the likes of Anthropic and OpenAI, with which it has a deal for 2 gigawatts of Trainium gear, ramping in 2027 and representing maybe $60 billion to $70 billion in datacenter costs for OpenAI. And that makes it no different from any other OEM that is selling systems laden with Nvidia or AMD GPUs, where those companies get most of the profits and they get relatively little of the gravy.
The choice is less gravy or no gravy. And Microsoft Azure might be facing the same choice with its Cobalt CPUs and Maia XPUs.
With that as the backdrop, let’s go over the numbers for AWS for Q1 2026.
AWS revenue for the March quarter was $37.59 billion, up 28.4 percent. Operating income was $14.16 billion, up 22.6 percent and accounting for 37.7 percent of revenue. (This is precisely the margin that Anthropic and OpenAI cannot afford to pay.) AWS represented 20.7 percent of the overall $181.52 billion in Amazon sales, but 59.4 percent of the overall $23.85 billion in Amazon operating profit.
My best guess is that of the $45.17 billion in capital expenses that Amazon shelled out in the first quarter, just shy of $36 billion of that was for AI systems and another $3 billion was for more generic IT systems for AWS. The remaining $6.2 billion was for Amazon fulfillment center and transportation capex.
As always, I have some fun trying to figure out how much of AWS revenue is for core systems – servers, storage, and networking – and how much is for higher level platform and application software. Here is how I think this is playing out:
Here is another interesting chart that shows the growth of AWS systems (meaning servers, storage, and networking capacity) versus the advertising business at Amazon:
There is no reason these two revenue stream should correlate, but they do. We think the advertising business is reasonably profitable for Amazon, and that implies that the online store business probably us breaking even at best.
Amazon had $143.1 billion in cash and equivalents in the bank as it exited Q1 2026, and it can afford to invest heavily in AI systems for its own use and to rent – or sell – to customers. Amazon is on track to spend $200 billion on capex this year. Which sounds nuts, but there you have it.