Project Rainier
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,554 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 2, 2026
Sources
9 citations
Review status
Source-backed
Revision
v1 · 1,554 words
Add missing citations, update stale details, or suggest a clearer explanation.
Project Rainier is a distributed artificial intelligence supercomputer built by Amazon Web Services around its in-house AWS Trainium accelerators, created principally to train and serve the Claude models of Anthropic. Announced at AWS re:Invent in December 2024 and declared fully operational on 29 October 2025, the system links hundreds of thousands of second-generation Trainium2 chips, nearly 500,000 at launch, across multiple data centers in the United States, with the largest campus near New Carlisle, Indiana. AWS describes it as one of the world's largest AI compute clusters and the biggest deployment of its own AI silicon to date.[1][2][3]
Project Rainier is structured as an Amazon EC2 UltraCluster, a single logical machine assembled from many Trainium2 servers spread across more than one physical building. At the December 2024 unveiling, AWS and Anthropic described it as "an EC2 UltraCluster of Trn2 UltraServers" containing "hundreds of thousands of Trainium2 chips" and offering "more than 5x the number of exaflops used to train their current generation of leading AI models."[4][1] When AWS announced completion in October 2025, it reported nearly 500,000 Trainium2 chips already in service and said the buildout represented a roughly 70 percent increase over any AI computing platform the company had previously operated.[1][2]
The project is named after Mount Rainier, the 4,392-metre stratovolcano visible from Seattle, Amazon's home city, on a clear day. Unlike many large training systems that concentrate accelerators in a single hall, Rainier is deliberately spread across several sites and stitched together over the network, an approach that lets Anthropic treat geographically separated facilities as one training fabric.[3][5]
Rainier is the physical expression of a deepening commercial relationship between Amazon and Anthropic. The two companies began working together in September 2023, and on 22 November 2024 Amazon announced an additional $4 billion investment that brought its total stake to $8 billion and named AWS as Anthropic's "primary training partner" in addition to its primary cloud provider. As part of that agreement Anthropic committed to using AWS Trainium and Inferentia chips to train and deploy its future models, and engineers from Anthropic worked with Amazon's Annapurna Labs to optimize the Trainium software stack, including writing low-level kernels that interface directly with the silicon.[6][7]
Anthropic uses Rainier to build and run Claude, and AWS says the cluster gives the company more than five times the compute it used to train its earlier model generation. The relationship continued to expand after launch: in April 2026 Amazon agreed to invest up to a further $25 billion in Anthropic as part of a broader AI infrastructure arrangement.[1][8]
The cluster is built up in a hierarchy. The base unit is the Trainium2 server, which holds 16 Trainium2 chips. Four of those servers are bound together with the high-speed NeuronLink interconnect to form a Trn2 UltraServer of 64 chips. Tens of thousands of UltraServers are then connected through the Elastic Fabric Adapter (EFA) network to form the UltraCluster that spans multiple data centers.[1][4]
| Building block | Composition | Notes |
|---|---|---|
| Trainium2 chip | Single accelerator | Performs trillions of calculations per second; AWS claims 30 to 40 percent better price-performance than comparable GPU-based systems [1][5] |
| Trainium2 server | 16 Trainium2 chips | Physical server node [1] |
| Trn2 UltraServer | 4 servers, 64 Trainium2 chips | Chips linked by NeuronLink; 83.2 peak petaflops per UltraServer [4] |
| EC2 UltraCluster (Project Rainier) | Tens of thousands of UltraServers | Connected across data centers via Elastic Fabric Adapter (EFA) networking [1][4] |
| Total at launch (Oct 2025) | ~500,000 Trainium2 chips | More than 5x Anthropic's previous training compute; ~70% larger than any prior AWS AI platform [1][2] |
| Planned by end of 2025 | Over 1,000,000 Trainium2 chips | For both training and inference workloads [1][2] |
Networking is handled at two tiers. Within an UltraServer, NeuronLink (shown as blue cabling in AWS imagery) ties the 64 chips into a single high-bandwidth domain, while EFA (yellow cabling) connects UltraServers to one another and across buildings. Alongside Trainium2, Anthropic's workloads on the system also draw on Amazon's Graviton CPUs and Inferentia inference accelerators.[1][5]
The flagship campus sits on roughly 1,200 acres of former farmland near New Carlisle, in St. Joseph County, Indiana. AWS has put about $11 billion into the Indiana development, described as the largest capital commitment in the state's history, and the site is planned to grow to 30 data center buildings, each exceeding 200,000 square feet. Seven of those buildings were operational as of October 2025. The project has created more than 1,000 jobs with wages above the county average.[2][9][3]
AWS has not published a full list of the other locations. The company has said Project Rainier spans multiple data centers across the United States and that it operates hundreds of data centers across four US regions; the Indiana campus is described as one of the Rainier sites rather than the whole of it.[1][3]
The Indiana campus is engineered to draw up to 2.2 gigawatts of power at full buildout, a scale AWS has compared to the electricity consumption of roughly 1.6 million homes. AWS has emphasized efficiency measures across the design: new mechanical components projected to cut energy consumption by up to 46 percent, concrete formulated to reduce embodied carbon by 35 percent, vertical power delivery, and a cooling approach that combines outside-air cooling with closed-loop, direct-to-chip liquid cooling. The company also cites a water-use effectiveness of 0.15 litres per kilowatt-hour, against an industry figure it places near 0.375.[1][3][9]
| Milestone | Date | Detail |
|---|---|---|
| Initial Amazon-Anthropic partnership | September 2023 | First investment and cloud agreement [6][7] |
| Investment raised to $8 billion; AWS named primary training partner | 22 November 2024 | Additional $4 billion committed [6][7] |
| Project Rainier announced | December 2024 (re:Invent) | Unveiled as an EC2 UltraCluster of Trn2 UltraServers [4][1] |
| Fully operational | 29 October 2025 | ~500,000 Trainium2 chips across multiple US data centers [1][2] |
| Target scale | End of 2025 | Plan to exceed 1,000,000 Trainium2 chips [1][2] |
Project Rainier is one of the clearest examples to date of vertical integration in AI infrastructure. By designing the accelerator (Trainium2), the server and interconnect (UltraServer and NeuronLink), the network (EFA), and the data centers themselves, AWS controls the full stack underneath a frontier training run. Ron Diamant, the AWS engineer who leads Trainium, framed the advantage in those terms: "When we build our own devices, we get to optimize across the entire stack to really compress engineering time."[3]
The cluster is also notable as a large-scale demonstration that frontier models can be trained on hardware other than Nvidia GPUs, which have dominated AI training. AWS has characterized the deployment as the world's largest cluster of its own AI chips, and the project is widely read as part of a strategy by hyperscalers to reduce dependence on Nvidia by scaling custom silicon. Amazon's $8 billion commitment to Anthropic, ahead of Google's reported $3 billion stake at the time, ties that silicon strategy directly to a leading model developer.[1][2][9]
Several details of Project Rainier remain undisclosed or are stated only at the level of AWS marketing claims. The company has not released a verified aggregate performance figure for the full cluster, an authoritative total power draw across all sites, or a complete list of locations beyond the Indiana campus; the "5x compute," "70 percent larger," and "30 to 40 percent better price-performance" figures come from AWS itself rather than independent benchmarking.[1][5] The plan to surpass one million Trainium2 chips by the end of 2025 was an AWS projection made at activation rather than a confirmed installed count.[1][2]
Large data center campuses of this kind have also drawn scrutiny over local water and electricity use, and the Indiana site's projected 2.2-gigawatt demand places it among the most power-intensive computing facilities in the country, which is part of why AWS publicizes its cooling and embodied-carbon measures.[3][9] Separately, the successor accelerator, Trainium3, was developed in collaboration with Anthropic and previewed with claims of roughly four times the performance and about 40 percent better energy efficiency than Trainium2, indicating that Rainier-class systems are expected to be refreshed with newer silicon over time.[2][9]