Anthropic-Amazon Trainium expansion

The Anthropic-Amazon Trainium expansion is an enlarged compute and investment agreement between Anthropic and Amazon, announced on 20 April 2026, under which Amazon committed to invest up to $25 billion more in Anthropic and Anthropic secured up to 5 gigawatts of new capacity built on AWS Trainium chips to train and serve Claude, while pledging more than $100 billion to Amazon Web Services over the following decade.^[1][2] The deal builds directly on Project Rainier, the large Trainium2 cluster the two companies brought online in 2025, and it sits alongside a parallel agreement Anthropic struck with Google for Tensor Processing Units. Taken together, the two arrangements were among the largest bets, as of 2026, that frontier models can be trained and served at scale on custom silicon as well as on Nvidia GPUs.

Deal terms and scale

The April 2026 agreement has two halves: money flowing into Anthropic, and money flowing out of Anthropic to AWS.

On the investment side, Amazon put $5 billion into Anthropic immediately, with up to a further $20 billion tied to certain commercial milestones, for a headline figure of up to $25 billion of new capital.^[1][2] That is on top of the $8 billion Amazon had already invested across earlier rounds, so its total potential investment rises toward $33 billion. Reporting at the time noted that the $5 billion of fresh cash brought Amazon's invested total to $13 billion, with the remaining $20 billion contingent on future targets.^[3][2] The $5 billion was later counted in Anthropic's Series H round, announced on 28 May 2026 at a $965 billion post-money valuation; the round included $15 billion of previously committed investments from hyperscalers, among them $5 billion from Amazon.^[4]

On the spending side, Anthropic committed more than $100 billion to AWS technologies over ten years. In exchange it secures up to 5 GW of new compute capacity to train and run Claude.^[1][2] To put that number in perspective, 5 GW is a power budget on the order of several large nuclear reactors, and it is more than twice the roughly 2.2 GW that the flagship Project Rainier campus in Indiana is expected to draw when fully built.^[5] Reuters, which reviewed Anthropic's confidential IPO prospectus, reported in September 2026 that the filing showed about $110 billion of long-term infrastructure service obligations to Amazon, payable between May 2026 and April 2036 regardless of usage, under terms similar to Anthropic's Google agreement, which requires Anthropic to pay the difference if its actual spending falls short.^[6]

The companies said meaningful new capacity would arrive quickly, with significant Trainium2 capacity coming online in the second quarter of 2026 and scaled Trainium3 capacity later in the year. By the end of 2026 they expected nearly 1 gigawatt of Trainium2 and Trainium3 capacity to be running between them.^[1] The agreement explicitly spans current and future generations of Amazon's custom accelerators, naming Trainium2, Trainium3 and a forthcoming Trainium4, along with an option to buy later Amazon silicon and tens of millions of Graviton CPU cores for the surrounding infrastructure.^[1][2] It also included an expansion of inference capacity in Asia and Europe to serve Claude's international customers.^[1] The list of AI compute deals records the agreement as a contract value stated by Anthropic, beside Amazon's earlier equity investments and Anthropic's other compute deals.

Trainium and Project Rainier context

This expansion did not come from nowhere. Amazon and Anthropic began working together in September 2023, and on 22 November 2024 Amazon raised its investment to $8 billion and Anthropic named AWS its primary cloud and training partner. As part of that earlier deal, Anthropic said it was working with Amazon's Annapurna Labs on the development and optimization of future generations of Trainium, writing low-level kernels that interface directly with the Trainium silicon and contributing to the AWS Neuron software stack.^[7][8]

The physical result was Project Rainier, an Amazon EC2 UltraCluster of Trainium2 UltraServers that AWS declared fully operational on 29 October 2025 with nearly 500,000 Trainium2 chips spread across multiple United States data centers.^[9][10] One of its sites is a campus in St. Joseph County, Indiana, near New Carlisle, where AWS is investing $11 billion; by June 2025 seven buildings stood on the 1,200-acre site, out of around 30 that Amazon planned there over several years. The New York Times reported at the time that Project Rainier would also include facilities in Mississippi and possibly North Carolina and Pennsylvania.^[10][5] AWS described Rainier as one of the largest AI compute clusters in the world and said in October 2025 that Claude would be running on more than 1 million Trainium2 chips by the end of that year.^[9] By the time of the April 2026 announcement, Anthropic said it was using more than one million Trainium2 chips to train and serve Claude.^[1] The new 5 GW commitment is a multi-year pipeline of Trainium2, Trainium3 and Trainium4 capacity rather than a single named site.^[1]

AWS made Trainium3 UltraServers available on 2 December 2025, with up to 144 Trainium3 chips per UltraServer, and named Anthropic among the customers it said were cutting training and inference costs by up to 50% with Trainium.^[11]

Why Anthropic anchors on AWS custom silicon

The obvious question is why a leading model lab would tie itself so heavily to a chip family other than Nvidia's, which dominates AI training. The answer that both companies give is cost and the need for capacity.

Amazon designs Trainium in-house through Annapurna Labs, so its anchor tenant avoids, as CNBC put it, the premium margins of other chips.^[12] When Trainium2 instances became generally available in December 2024, AWS said they offered 30 to 40 percent better price performance than its GPU-based P5e and P5en instances.^[13] Amazon chief executive Andy Jassy framed the demand bluntly in April 2026, saying Amazon's custom AI silicon "offers high performance at significantly lower cost for customers, which is why it's in such hot demand."^[1][2] Anthropic said its run-rate revenue had passed $30 billion by April 2026, up from about $9 billion at the end of 2025, and that the strain of that growth had hurt reliability for some users at peak hours; by May it put run-rate revenue above $47 billion.^[1][4] For a company growing that quickly, securing cheaper compute at guaranteed volume matters as much as raw peak performance.

There is also a strategic logic to co-designing the hardware. Amazon said in April 2026 that Anthropic works closely with Annapurna Labs on developing and optimizing future Trainium chips, feeding back data from Claude training workloads, with the two engineering teams communicating on an almost daily basis about everything from low-level optimization to the architecture of next-generation chips.^[2] That tight loop means Anthropic helps shape the chips it will train on, which is a different relationship from buying GPUs off a shared roadmap. Dario Amodei, Anthropic's co-founder and chief executive, tied the buildout to demand rather than to any single vendor, saying "our users tell us Claude is increasingly essential to how they work, and we need to build the infrastructure to keep pace with rapidly growing demand."^[2]

Alongside the silicon, Amazon and Anthropic said the full Claude Platform would become available directly inside AWS, so that organizations already running Claude through Amazon Bedrock (more than 100,000 of them, by the companies' count) could reach the models without separate credentials or contracts.^[1][2] AWS made Claude Platform on AWS generally available on 11 May 2026.^[14] That distribution angle gives Amazon a commercial return beyond Anthropic's own compute spending. According to Reuters' account of the prospectus, sales through the Amazon and Google cloud marketplaces made up about 47% of Anthropic's 2025 revenue, about $2.16 billion, and a Reuters analysis put the distribution fees Anthropic paid the platforms at roughly $351 million.^[15]

Relationship to the Google TPU deal and multi-cloud strategy

The Amazon expansion is only one leg of Anthropic's compute strategy. The company has been explicit that it runs Claude across three chip families: Google's TPUs, Amazon's Trainium, and Nvidia's GPUs, matching workloads to the chips best suited for them.^[16][17]

The Google side is large in its own right, and is covered in Anthropic-Google TPU deal. On 23 October 2025 Anthropic agreed to expand its use of Google Cloud TPUs to as many as one million chips, bringing well over a gigawatt of capacity online during 2026 in a deal worth tens of billions of dollars.^[16][18] That arrangement was enlarged in April 2026, around the same window as the Amazon news. On 6 April Anthropic signed an agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity from 2027, which it later described as 5 GW, and on 24 April Google agreed to invest up to $40 billion in Anthropic, starting with $10 billion at a $350 billion valuation.^{[17][19][20][21]} Some venture investors were reportedly offering capital at a valuation of $800 billion or more in the same period.^[3] Reuters reported that the prospectus showed commitments of at least $111.1 billion to Google between April 2026 and July 2033, against about $110 billion to Amazon.^[6]

Running two hyperscaler deals at once is a deliberate hedge. By not depending on any single cloud or chip vendor, Anthropic protects itself against supply shortages, pricing power, and the risk that one platform stumbles on a hardware generation; Anthropic says the spread of platforms gives customers better performance and greater resilience.^[17] AWS remained the designated primary cloud provider and training partner as of May 2026, but the Google relationship gives Anthropic a second pool of non-Nvidia accelerators at gigawatt scale and a second large investor on the cap table.^[4] Anthropic's compute deals also extend beyond the two: by May 2026 it listed a strategic partnership with Microsoft and NVIDIA that includes $30 billion of Azure capacity, its $50 billion investment in American AI infrastructure with Fluidstack, and an agreement to use all of the compute at SpaceX's Colossus 1 data center.^[19]

Significance for non-Nvidia compute at scale

What makes the Trainium expansion notable is the size of the bet on silicon that Nvidia did not build. A commitment of up to 5 GW of Trainium-based capacity, layered on top of an existing million-plus chip deployment and a parallel million-TPU deal with Google, is a real-world test of whether a frontier lab can sustain training and inference largely outside the Nvidia ecosystem.

For Amazon, the deal validates a vertical-integration strategy that runs from chip design through servers, the NeuronLink and Elastic Fabric Adapter interconnects, and the data centers themselves.^[9] For the wider industry, it shows a sophisticated buyer committing hundreds of billions of dollars to platforms built on custom accelerators. Reuters reported that Anthropic's prospectus listed planned AI infrastructure spending of at least $518 billion over a decade with six partners, about 80% of it non-cancelable or payable regardless of usage and including the Amazon and Google commitments; the filing separately listed about $161.2 billion of Broadcom-related equipment lease obligations.^[6] The price-performance figures are company-stated rather than independently benchmarked, and the bulk of Amazon's $25 billion is contingent on milestones that may or may not be met.

Amazon's financial results reflect both sides of the relationship. Its second-quarter 2026 net income included $53.4 billion of non-operating pre-tax income, primarily from its investments in Anthropic. Amazon said at the same time that Anthropic and OpenAI had both made multi-year, multi-gigawatt commitments to Trainium, and Jassy said its custom-chip business had passed an annual revenue run rate of $25 billion and that Amazon was exploring selling Trainium chips to customers for use outside its cloud.^[22][23] The prospectus also describes the risks of the arrangement: according to Reuters, Anthropic told investors that Amazon, Google and Microsoft act at once as investors, customers, cloud providers, distributors and competitors, creating incentives that "may not be fully aligned" with its interests, and that its business could suffer if compute from third parties were "curtailed, repriced, or terminated."^[6]

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute, Anthropic, 20 April 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13Amazon and Anthropic expand strategic collaboration, About Amazon, 20 April 2026.
3. ^1 ^2 ^3Anthropic takes $5B from Amazon and pledges $100B in cloud spending in return, TechCrunch, 20 April 2026.
4. ^1 ^2 ^3Anthropic raises $65B in Series H funding at $965B post-money valuation, Anthropic, 28 May 2026.
5. ^1 ^2At Amazon's Biggest Data Center, Everything Is Supersized for A.I., The New York Times, 24 June 2025.
6. ^1 ^2 ^3 ^4 ^5Anthropic's $518bln AI buildout hinges largely on deals that cannot be canceled, filing shows, Reuters via Zawya, 29 September 2026.
7. ^1 ^2Anthropic raises another $4B from Amazon, makes AWS its 'primary' training partner, TechCrunch, 22 November 2024.
8. ^Powering the next generation of AI development with AWS, Anthropic, 22 November 2024.
9. ^1 ^2 ^3AWS activates Project Rainier: One of the world's largest AI compute clusters comes online, About Amazon, 29 October 2025.
10. ^1 ^2AWS activates Project Rainier cluster of nearly 500,000 Trainium2 chips, DataCenterDynamics, 30 October 2025.
11. ^Trainium3 UltraServers Now Available: Enabling Customers to Train and Deploy AI Models Faster at Lower Cost, Amazon, 2 December 2025.
12. ^Google and Anthropic announce cloud deal worth tens of billions of dollars, CNBC, 23 October 2025.
13. ^AWS Trainium2 Instances Now Generally Available, Amazon, 3 December 2024.
14. ^Claude Platform on AWS is now generally available, AWS, 11 May 2026.
15. ^Exclusive-Anthropic IPO prospectus lays bare deep dependence on Big Tech partners, Reuters via The Star, 29 September 2026.
16. ^1 ^2Expanding our use of Google Cloud TPUs and Services, Anthropic, 23 October 2025.
17. ^1 ^2 ^3Anthropic expands partnership with Google and Broadcom for multiple gigawatts of next-generation compute, Anthropic, 6 April 2026.
18. ^Google and Anthropic confirm massive 1GW+ cloud deal with up to 1 million Google TPUs, DataCenterDynamics, 24 October 2025.
19. ^1 ^2Higher usage limits for Claude and a compute deal with SpaceX, Anthropic, 6 May 2026.
20. ^Google to invest up to $40 billion in Anthropic as search giant spreads its AI bets, CNBC, 24 April 2026.
21. ^Google to invest up to $40B in Anthropic in cash and compute, TechCrunch, 24 April 2026.
22. ^Amazon.com announces second quarter results, About Amazon, 30 July 2026.
23. ^Amazon (AMZN) Q2 2026 Earnings Call Transcript, The Motley Fool, 7 August 2026.