Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,569 words
Add missing citations, update stale details, or suggest a clearer explanation.
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 roughly $25 billion more in Anthropic and to deploy up to about 5 gigawatts of new AWS Trainium capacity to train and serve Claude, while Anthropic pledged 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 represent one of the clearest bets so far that frontier models can be trained at scale on custom silicon rather than on Nvidia GPUs.
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 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 stake rises toward $33 billion. Reporting at the time noted that the $5 billion of fresh cash brought Amazon's invested total to about $13 billion, with the remaining $20 billion contingent on future targets.[3]
On the spending side, Anthropic committed more than $100 billion to AWS technologies over roughly ten years. In exchange it secures up to 5 gigawatts of new compute capacity dedicated to building and running 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 dwarfs the roughly 2.2 GW that the flagship Project Rainier campus in Indiana was engineered to draw at full buildout.
| Term | Detail |
|---|---|
| Announcement date | 20 April 2026 |
| New Amazon investment | $5 billion now, up to $25 billion total (rest milestone-based) [1][2] |
| Prior Amazon investment | $8 billion (across 2023 to 2024 rounds) [3][4] |
| Anthropic commitment to AWS | More than $100 billion over about 10 years [1][2] |
| New compute capacity | Up to 5 gigawatts of Trainium [1][2] |
| Online by end of 2026 | Nearly 1 GW of Trainium2 plus Trainium3 combined [1] |
| Silicon covered | Trainium2, Trainium3, Trainium4, future generations, plus Graviton CPUs [1][2] |
The companies said meaningful new capacity would arrive quickly, with significant Trainium2 systems coming online in the first half of 2026 and Trainium3 scaling 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 on later Amazon silicon and tens of millions of Graviton CPU cores for the surrounding infrastructure.[1][2]
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 named AWS as Anthropic's primary training partner. As part of that earlier deal, Anthropic agreed to train and deploy its models on Trainium and to help Amazon's Annapurna Labs tune the chip's software stack down to the level of low-level kernels.[4]
The physical result was Project Rainier, an Amazon EC2 UltraCluster that AWS declared fully operational on 29 October 2025 with nearly 500,000 Trainium2 chips spread across several United States data centers, anchored by an roughly $11 billion campus near New Carlisle, Indiana.[5][6] AWS described it as one of the largest AI compute clusters in the world. By the time of the April 2026 announcement, Anthropic said it was using more than one million Trainium2 chips to train and serve Claude, a figure that reflects Rainier plus additional capacity built out since.[1] The new 5 GW commitment is, in effect, the next several Rainiers: a multi-year pipeline of Trainium2, Trainium3, and Trainium4 buildouts rather than a single named site.
The obvious question is why a leading model lab would tie itself so heavily to a chip family other than Nvidia's, which still dominates AI training. The answer that both companies give is cost and supply.
Amazon designs Trainium in-house through Annapurna Labs, which lets it sell the chips to its own anchor tenant at prices it controls rather than at the margins Nvidia commands. AWS has claimed Trainium2 delivers roughly 30 to 40 percent better price-performance than comparable GPU systems, and Amazon chief executive Andy Jassy framed the demand bluntly, saying the custom silicon "offers high performance at significantly lower cost for customers, which is why it's in such hot demand."[2][5] For a company whose run-rate revenue had reportedly climbed past $30 billion in early 2026, up from around $9 billion at the end of 2025, securing cheaper compute at guaranteed volume matters as much as raw peak performance.[1]
There is also a strategic logic to co-designing the hardware. Anthropic engineers have worked alongside Annapurna on the Trainium stack since the 2024 deal, and Trainium3 was developed in collaboration with Anthropic. That tight loop means Anthropic helps shape the very 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 AWS's count) could reach the models without separate credentials or contracts.[1] That distribution angle gives Amazon a commercial return on its compute spend beyond Anthropic's own usage.
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, assigning workloads to whichever platform fits.[7]
The Google side is large in its own right. 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.[7][8] That arrangement was later enlarged in April 2026, around the same window as the Amazon news, with reporting describing a Google investment in Anthropic of roughly $40 billion and a post-money valuation near $350 billion, plus additional TPU capacity routed through Broadcom from 2027.[8] Some venture investors were reportedly willing to value Anthropic far higher still, in the range of $800 billion, in the same period.[3]
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. AWS remains the designated primary training partner, 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.
What makes the Trainium expansion notable is the size of the bet on silicon that Nvidia did not build. A 5 GW commitment to Trainium, 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. For the wider industry, it signals that the economics of custom accelerators have improved enough for a sophisticated buyer to commit hundreds of billions of dollars to them. The figures here are still 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. Even so, an agreement of this scale moves the question of whether non-Nvidia silicon can carry frontier AI from a research curiosity toward a settled commercial fact.