Home โ€บ Blog โ€บ Project Rainier
AI Factory Case Studies ยท Part 4 ยท Project Rainier

Project Rainier: The AI Factory Built Without NVIDIA

Every site so far in this series runs on NVIDIA GPUs. Project Rainier does not. Amazon filled farmland near New Carlisle, Indiana with nearly half a million of its own Trainium2 chips and handed the cluster to Anthropic to train Claude. This case study looks at what it means to own every layer from chip to cloud, how a utility created a company just to power data centers, and why residents near the site started checking their wells.

FL
FrontierAGI Team
โš ๏ธ How this was researched (as of 26 Oct 2026) Chip and activation figures come from Amazon's own announcement of Project Rainier. Power comes from NIPSCO, NiSource and coverage of the Indiana deal; incentives from Inside Indiana Business, DCD and WVPE; water from WNDU and a Notre Dame student investigation. Size estimates come from Epoch AI. Amazon has not published megawatts. One disclosure is worth noting up front: this site is written partly with Anthropic's Claude, the model trained on Project Rainier. We have tried to treat Amazon and Anthropic exactly as we treated xAI, OpenAI and Microsoft. See Reading the numbers.

1. Case file

โ›ฐ๏ธ
Project RainierNew Carlisle, St. Joseph County, Indiana ยท second campus near Canton, Mississippi
OwnerAmazon Web Services
Main userAnthropic (Claude)
Chips~500k Trainium2 at activation (AWS)
Building blockTrn2 UltraServer ยท 64 chips
Investment$11B campus + $15B NW Indiana expansion
Site~1,200 acres ยท ~30 buildings planned
IT power now (est.)~910 MW ยท Epoch AI
Full build (est.)~1.9 GW IT by 2028 ยท Epoch AI

AWS said in October 2025 that Anthropic would be running Claude on more than 1 million Trainium2 chips across Project Rainier and Amazon Bedrock by the end of 2025. Some trackers still list 1 million as a target โš ๏ธ. Press reports of the full build put grid draw at about 2.2 GW.

2. Build timeline

Project Rainier, 2024 โ†’ 2028

Solid dots are confirmed by Amazon, a utility or a public body, hollow dots come from trackers or the press, โš ๏ธ marks claims we could not confirm and dashed dots are planned. Click a dot for its source.

Click a milestone to see where it comes from.

Amazon announced an $11 billion data center campus in St. Joseph County in April 2024, after acquiring farmland around New Carlisle through a company that did not name it. Construction began in September 2024, and according to press reports seven buildings were running within about seven months. AWS declared Project Rainier active in October 2025, less than a year after its public unveiling. In November 2025 Amazon announced a further $15 billion for new campuses in Northwest Indiana, adding 2.4 GW and 1,100 jobs.

3. The bet: own every layer

NVIDIA sells the most sought-after chips in computing, and it charges accordingly. A frontier lab that buys NVIDIA GPUs pays NVIDIA's margin on every chip, and waits in NVIDIA's queue. Amazon's bet is that its own chips, designed by its Annapurna Labs unit, can deliver enough training performance at lower cost to make up for NVIDIA's software lead, and that owning every layer gives it control over supply and price.

๐Ÿ’ก Why Anthropic matters to the betA custom chip is only as good as the software and models that run on it. Anthropic is Rainier's anchor customer and works closely with Annapurna Labs on Trainium. That gives Amazon a frontier workload to prove the chip on, and gives Anthropic a large, dedicated cluster. It also ties Anthropic more closely to one chip family, though it also trains on Google TPUs and NVIDIA GPUs.

4. Who owns each layer

The stack, site by site

Pick a site. Each layer is coloured by whether the site's owner built it in-house or bought it from someone else. Simplified.

5. The chips

Trainium2 is built for training and serving large models. AWS groups 16 chips into a server, then joins four servers into a Trn2 UltraServer of 64 chips with its NeuronLink interconnect, so they act as one large accelerator. UltraServers are linked across buildings by Amazon's Elastic Fabric Adapter (EFA) network into an UltraCluster. For comparison, NVIDIA's GB200 NVL72 joins 72 GPUs in one rack. Head-to-head performance depends heavily on the workload and the software, and neither company publishes a neutral comparison; see custom silicon vs NVIDIA.

Chip count

How many Trainium2 chips Claude runs on, by source. Toggle between what AWS says and how trackers describe it.

6. Spread, not stacked

Fairwater packs GPUs into dense two-storey halls at about 140 kW per rack. Rainier went the other way: a cluster spread across many buildings on a large rural site, and across states. Reports describe the Indiana campus as the primary site, with a smaller campus near Canton, Mississippi, linked by high-speed networks. Spreading out is easier to cool and to build in phases, but pushes more traffic onto the network between buildings, so the cluster design leans on the EFA network rather than on very large single racks.

Two ways to build a big cluster

Dense (Fairwater)~140 kW racks ยท two storeys ยท short cables ยท liquid cooling throughout ยท one site sized for the grid connection
Spread (Rainier)~1,200-acre campus ยท ~30 buildings planned ยท 64-chip UltraServers joined by EFA ยท phases added building by building ยท second campus in another state

7. Power: a utility built for data centers

Rainier gets its power from NIPSCO, the northern Indiana utility owned by NiSource. To serve Amazon's growth without loading the cost onto households, NIPSCO set up a separate subsidiary, GenCo, to own new generation for data centers. Circle of Blue called it the first utility dedicated to data centers in the United States. The November 2025 package includes two new gas turbines with a combined 2.6 GW and a 400 MW / 1,600 MWh battery. Amazon pays for any new transmission, and NIPSCO estimates existing customers will save about $1 billion over 15 years. In April 2026 NiSource said it was speeding up energization of Amazon's sites.

How the GenCo arrangement works

Tap a box. Contract terms are summarised from public announcements.

Like Colossus and Abilene, the new supply is mostly gas; unlike them, the turbines belong to a regulated utility rather than sitting inside the data center fence. Related reading: the power bottleneck.

8. Water and taxes

Water. In May 2025 WNDU reported that residents near New Carlisle said their wells were running dry while the Amazon campus and a nearby GM/Samsung battery plant were under construction, and that the Indiana Department of Natural Resources was investigating. A Notre Dame student investigation estimated peak demand on the local aquifer could reach up to 35.5 million gallons a day, about 80% of its safe yield โš ๏ธ. County officials have said water use will be capped and kept within local standards; opponents dispute what cooling system Amazon uses. We found no final state finding on the wells.

Taxes. St. Joseph County approved a 10-year, 50% real property tax abatement on 870 acres and an 85% exemption on personal property such as servers. Officials said Amazon would still pay more than $722 million in taxes over 35 years. In March 2026 some council members asked Amazon to consider giving back part of roughly $4 billion of abatements over 35 years; Amazon had not responded publicly.

2024 County approves 50% / 85% abatements for the $11B campus.
May 2025 Residents report dry wells; state DNR investigates.
Nov 2025 GenCo power package announced with the $15B expansion.
Mar 2026 Council members ask Amazon to renegotiate part of its tax break.

9. Four bets compared

ColossusAbileneFairwaterRainier
The betSpeedOutside capitalNetwork of sitesOwn silicon
ChipsNVIDIA H100 โ†’ GB300NVIDIA GB200NVIDIA GB200/GB300Amazon Trainium2
Size (Epoch est.)~946 MW~421 MWAtlanta ~636 MW~910 MW
PowerGrid + own turbinesGrid + gas microgridUtility gridUtility via data-center-only GenCo
Main controversyTurbine permitsTax breaks, delaysUtility gas plantsWells, tax breaks
No NVIDIAHalf a million custom chips at activation
Own the stackChip, server, network, building, cloud
A utility for data centersGenCo keeps costs off household bills, in principle
Water mattersWells, aquifers and cooling choices
VerdictRainier is the strongest test yet of whether frontier AI needs NVIDIA. Amazon built a cluster at the same scale as the others on its own chips, in about a year, for a frontier lab. What it does not yet show is how its cost and performance compare in practice, because neither Amazon nor Anthropic publishes that. Next, Meta's Prometheus and Hyperion take a different route: tents for speed, then a campus the size of a small city.

10. Reading the numbers

From Amazon
Nearly 500,000 Trainium2 chips at activation (Oct 2025); more than 1 million Trainium2 chips for Claude by end of 2025 (stated as a plan; some trackers still treat it as a target โš ๏ธ); $11B + $15B; 2.4 GW and 1,100 jobs in the expansion; 64-chip UltraServers.
From the utility
GenCo; two gas turbines totalling 2.6 GW; 400 MW / 1,600 MWh battery; ~$1B customer savings over 15 years (NIPSCO estimate).
Estimates
~910 MW IT power now and ~1.9 GW by Q1 2028 are Epoch AI estimates; ~2.2 GW full grid draw and 30 buildings are press figures.
Contested
Dry wells (DNR investigation, no final finding found); aquifer stress estimate from a student investigation โš ๏ธ; cooling type disputed by opponents.
Illustrations
The stack-ownership chart simplifies who builds each layer.