No Nvidia needed: inside Amazon's massive AI data center built for Anthropic

01The Anthropic-Amazon compute deal in brief

The backstory here is one of the largest infrastructure commitments in the history of computing. In late 2024, Anthropic and Amazon announced a multi-year collaboration under which Amazon would invest and commit roughly $4 billion in AWS infrastructure for Anthropic, with AWS becoming the cloud provider Anthropic would run the "majority" of its workloads on. Amazon's project to deliver that compute, known internally as Project Rainier, went from announcement to enormous physical reality in a remarkably short time.

Anthropic itself is a heavyweight. According to Wikipedia's summary, the San Francisco-based public benefit corporation was founded in 2021 by former OpenAI members, including siblings Dario and Daniela Amodei, who serve as CEO and president, with the mission of promoting AI safety, and its flagship product is Claude, a family of large language models now used across consumer and developer products. AWS is betting that being Anthropic's primary infrastructure landlord makes it the default cloud for one of the few frontier model developers.

The Indiana campus is the concrete expression of that bet. Rather than renting Anthropic a patchwork of GPU instances scattered across existing regions, Amazon designed a purpose-built facility around its own accelerator, delivering what Anthropic has described as a very large fraction of the total compute it uses for training and inference. The deal also made Anthropic a partner rather than a customer in the ordinary sense: AWS gets early access to Claude models for its Bedrock service, while Anthropic gets a compute pipeline that does not depend on competing for constrained GPU supply.

02How big this data center actually is

The numbers attached to the site, reported by CNBC and other outlets with camera access, are difficult to picture. The campus spans well over a thousand acres of farmland. Rows of data halls stretch to the horizon in aerial footage, each one packed with Trainium servers staged by the thousands. Construction proceeded in phases, with buildings completed and brought online while crews kept pouring foundations for the next wave directly alongside them.

Capacity figures are the headline. Amazon has publicly framed the Rainier buildout as delivering on the order of a million Trainium chips of aggregate compute for Anthropic's use, a scale that, on raw accelerator count, dwarfs anything publicly disclosed for a single-purpose AI campus. Anthropic has said the resulting cluster gave it a step change in available training compute, enabling the frontier model releases it has shipped since the site came online.

To be precise about what is and is not verifiable: the "million chips" figure is an Amazon-stated target for the overall Anthropic compute program, not an audited count of chips physically humming in Indiana on any given day, and the exact number of buildings, megawatts, and server racks is not something either company publishes. What is beyond dispute is that the site is among the largest single-customer AI compute facilities on Earth, and that it was purpose-built for one customer's models. The chart below puts it alongside other headline AI campus projects by their publicly announced size.

03Trainium: Amazon's in-house AI accelerator

The chip at the heart of the campus is Trainium, Amazon's purpose-built machine learning accelerator. It is not a GPU, and that distinction is the entire point. Trainium is a dedicated training and inference engine, co-designed by AWS with its own software stack and its own server-level interconnect, so that enormous models can be distributed across tens of thousands of chips with high utilization.

Amazon's silicon capability did not appear from nowhere. Wikipedia's summary of Annapurna Labs describes it as Amazon's semiconductor division, established in Israel in 2011 and acquired by Amazon in January 2015, and now one of the world's largest fabless semiconductor companies and a top-tier customer of TSMC. Annapurna's product lines include the Nitro card that virtualizes and secures every EC2 server, the Graviton general-purpose Arm CPU, and Trainium, the AI accelerator line.

Each Trainium generation has pushed memory capacity and bandwidth upward, because large-model training is heavily memory-bound. Amazon says current-generation Trainium servers deliver on the order of hundreds of teraflops per chip for the low-precision math that dominates modern training, with an HBM memory capacity per chip that lets more of a model sit closer to compute. The engineering claim that matters most to Anthropic is cost: AWS has repeatedly stated that Trainium instances deliver meaningfully better price-performance for suitable workloads than GPU instances, and a dedicated campus built at hyperscale amplifies that margin.

What makes the Indiana site unusual is the software. Running a frontier training run on Trainium rather than Nvidia hardware requires a full distributed training stack: a compiler for the chip's architecture, cluster-level networking tuned for the model's parallelism strategy, checkpointing, failure recovery, and debugging tooling, all written and maintained in-house at Amazon and Anthropic rather than drawing on the vast CUDA ecosystem. That this stack now reliably trains frontier-scale models is arguably the most underappreciated achievement of the whole program, and one that no other cloud provider except Google can claim.

04Why hyperscalers build their own silicon

Amazon is following the logic that every large cloud operator has now accepted. When you buy merchant silicon, your cost structure includes the vendor's margin and a supply line you do not control, allocated among customers you are sometimes competing with. When you design your own chip for your own workload, you pay fabrication costs and your design team, and every efficiency you gain accrues to you alone.

The deeper motivation is workload specificity. Nvidia GPUs must remain general-purpose machines that serve thousands of customers with wildly different workloads. A hyperscaler's internal AI fleet runs a much narrower distribution of work: dense matrix arithmetic at low precision on known model families, orchestrated at a known cluster scale. Silicon specialized for that narrow distribution can be dramatically more efficient per dollar and per watt, which is exactly the trade a company operating millions of AI chips wants to make.

The chart below maps the major custom AI accelerator programs and the rough scale each has reached publicly. Google's Tensor Processing Unit, which Wikipedia's summary notes began internal deployment in 2015 and reached third parties via Google Cloud in 2018, remains the most mature; Amazon's Trainium fleet is now the largest by raw deployed count thanks to the Rainier campus; Microsoft's Maia line is earlier in its public scale-out.

05Power, land, and water: the physical footprint

A campus with a million AI accelerators is, first and foremost, an electricity project. Even with efficient custom silicon, aggregate demand runs into the gigawatt range when the site reaches full build-out, comparable to a small city or a large industrial plant. Securing that much power in one location, alongside the transmission to deliver it, is now the binding constraint on AI infrastructure everywhere, and it is why the site sits on wide rural land rather than near a metro fiber hub.

Cooling is the second physical story. At this density, air cooling alone does not suffice for the latest accelerator generations, so hyperscalers have moved to direct-to-chip liquid cooling: cold plates bolted to each chip, with liquid pumped through and heat exchanged against facility water. Liquid cooling dramatically raises rack power density, which is precisely what lets a campus reach a million chips without the buildings sprawling ten times wider.

Water and land are the quieter costs. Evaporative cooling towers consume millions of gallons of water per year at campuses this size, which has pushed Amazon and peers toward closed-loop and air-side economization where climate permits, and has made community relationships a genuine engineering input. The land itself, former farmland measured in square miles, carries substations, water treatment, fiber corridors, and logistics yards alongside the data halls. Building one of these is closer to constructing a utility than to racking servers in an existing warehouse.

06The Nvidia question: replace or complement

The framing in the video's title, "no Nvidia chips needed," is accurate for this campus but easy to overread. Amazon has not stopped buying Nvidia GPUs, and it is publicly committed to offering the newest Nvidia systems across AWS for the customers who want them, including its own AI teams. AWS operates the world's broadest cloud, and most of its AI customers run workloads that were built for CUDA-adjacent tooling and are not moving.

The relationship is better described as layering than replacement. Trainium handles the massive, predictable, Amazon-controlled workloads: Anthropic's training and inference, Amazon's own model development, and price-sensitive general training on SageMaker. Nvidia hardware remains the flexible frontier tier for the newest architectures and for customers who need the merchant ecosystem. The same pattern holds at Google and Microsoft: every hyperscaler that designs its own AI silicon still buys enormous volumes of GPUs.

For Nvidia, the concern is not losing Amazon as a customer but watching the highest-volume, most predictable workloads migrate to silicon it does not sell. A dedicated campus of a million accelerators that would once have been GPU orders is, by definition, volume that never arrives. Whether Trainium's price-performance advantage holds at the very frontier of model scale, against the newest merchant GPUs, is the empirical question that will decide how much further this migration goes.

07What it means for Anthropic's scaling plans

For Anthropic, the campus removes the constraint that has defined every AI lab's roadmap: compute. Training runs for frontier models are planned years ahead around clusters that may not exist yet, and labs that cannot guarantee capacity fall behind. With a dedicated million-chip pipeline, Anthropic has been able to plan model generations around a known, growing supply of accelerators rather than around allocations it must negotiate for.

Wikipedia's summary of Anthropic notes the company's scale and trajectory: founded in 2021 by former OpenAI members with a safety-focused mission, privately held with a reported valuation approaching a trillion dollars in 2026 and a widely reported IPO timeline. Compute security underwrites all of it. A lab planning a public offering must show investors a durable cost structure, and an owned-capacity pipeline on efficient custom silicon is a materially better story than renting constrained GPUs at market prices.

It also reshapes Anthropic's relationship with the other clouds. The company has taken investments and compute commitments from Google as well as Amazon, and sells its models through multiple clouds, but the Indiana campus means its center of gravity sits firmly on AWS Trainium infrastructure. That tight coupling is powerful while both companies' roadmaps align, and it is the kind of dependency both sides' lawyers have spent considerable time papering over.

08Limits, risks, and what is still unproven

Start with the performance claims. "Better price-performance than GPUs" is an AWS statement about specific benchmarks on specific workloads, not an independently audited result, and Trainium's advantages are largest on workloads that map cleanly onto its architecture. Anthropic's success in training frontier models on the campus is strong evidence the stack works, but public third-party comparisons against the newest Nvidia systems, on identical model families, remain thin.

Then there is concentration risk, which cuts both ways. Anthropic now depends on a single primary compute pipeline for the majority of its scaling. A facility failure, a chip supply interruption at TSMC, a power dispute, or simply a Trainium architecture decision that ages poorly would be felt immediately in Anthropic's roadmap. Diversification is the obvious hedge, and Anthropic has kept other options alive, but hedges are small next to a campus of this size.

Finally, the demand side. Everything above assumes that model training and inference demand keep growing fast enough to absorb campuses like this one. The industry is currently making that exact wager across a dozen multi-gigawatt projects simultaneously. If scaling economics shift, whether toward far more efficient models, or toward inference consolidating on fewer, cheaper chips, some share of today's announced capacity will end up as the most expensive empty buildings in the history of computing. Amazon and Anthropic are betting several years and billions of dollars that it will not come to that.