Amazon’s AI Resurgence: AWS & Anthropic’s Multi-Gigawatt Trainium Expansion

AWS has spent two years looking like the odd one out among the big cloud players. Azure rode OpenAI to the top of new cloud revenue, Google Cloud closed the gap with its TPU push, and Amazon just sat there, its custom EFA networking still trailing InfiniBand and RoCEv2 on raw performance and on user experience. Investors noticed. Amazon is the clear laggard among the major AI infrastructure stocks this year. But the story underneath the stock chart is more interesting than the stock chart itself.

The fix has a name: Anthropic. Dario Amodei's company has gone from roughly $1 billion to $5 billion in annualized revenue in 2025, the fastest ramp of any frontier lab this year. Amazon saw this coming as early as September 2023, when it put $1.25 billion into Anthropic, expandable to $4 billion. By November 2024 that had grown to another $4 billion, and Anthropic named AWS its primary training partner. Right now, three AWS campuses are in the final stretch of construction, adding more than 1.3 gigawatts of capacity built specifically to train Anthropic's models, at a pace of construction AWS has never hit before.

Here's the catch investors keep missing: none of that capacity is generating revenue yet. Trainium2, Amazon's third-generation custom chip, has had the usual yield hiccups you'd expect from a new silicon ramp, and these buildings are still coming online. On paper Trainium2 loses badly to Nvidia's GB200 — 667 teraflops per chip versus 2,500, and less than a third of the scale-up network bandwidth. Yet once you run the numbers through total cost of ownership rather than raw specs, Trainium2 turns out to be genuinely competitive on cost per token and cost per terabyte-per-second of memory bandwidth. And that specific metric, memory bandwidth per dollar, happens to map almost exactly onto what Anthropic needs for its aggressive reinforcement-learning-heavy training roadmap.

That's not a coincidence. Anthropic engineers have been embedded in Trainium's design process for a while now, effectively treating Amazon's chip division, Annapurna Labs, as their own custom silicon shop. That kind of tight hardware-software co-design is rare air, shared really only with Google DeepMind and its TPUs. Meanwhile Anthropic's massive inference workload, which is arguably even more important than training, still runs largely on Google Cloud, a relationship that predates the AWS deal and hasn't gone away. So AWS isn't winning all of Anthropic's business. It's winning the training half, and betting that a specialized, cost-efficient chip beats a faster but pricier one when you're running post-training at gigawatt scale.

If these three campuses ramp as expected by the end of 2025, SemiAnalysis reckons AWS growth could push past 20% year-over-year, enough to quiet a lot of the cloud-crisis talk. It's a bet on architecture fit over spec sheets, and on one customer's scaling obsession carrying an entire hyperscaler's AI narrative.