Amazon is placing a huge bet on Anthropic, agreeing to invest up to $25 billion — including a $5 billion upfront payment and as much as $20 billion tied to future milestones — while Anthropic promises more than $100 billion of AWS spending and up to 5 gigawatts of capacity for its Claude models.

Massive order, deep-pocketed partner Amazon has agreed to boost its backing of Anthropic with up to $25 billion of new investment, on top of about $8 billion it has already provided, the companies said in a joint announcement. The initial tranche will be $5 billion paid now, with as much as $20 billion available later if Anthropic hits certain commercial milestones tied to the deal. Anthropic also pledged to spend more than $100 billion on Amazon Web Services technologies over the next ten years, locking in capacity and hardware commitments that stretch well beyond the next business cycle. Key deal terms - Up to $25bn of new Amazon investment; $5bn initial payment, up to $20bn conditional on milestones. - Anthropic has secured up to 5 gigawatts of capacity to train and serve its Claude models. - Commitment to bring nearly 1 gigawatt of Trainium2 and Trainium3 capacity online before year-end. - Anthropic pledges more than $100bn of AWS spending over ten years. The agreement includes a commitment to bring nearly 1 gigawatt of Trainium2 and Trainium3 capacity online before the year is out, signalling both a near-term build-out and a long-term lock on Amazon's custom AI silicon. Andy Jassy, Amazon's chief executive, said the decision to run Anthropic's large language models on AWS Trainium reflects progress the two firms have made on custom silicon and infrastructure for generative AI. The headline figures underline how costly the race for AI compute has become. Amazon told investors earlier this year it expects significant capital spending in 2026, mostly on AI infrastructure, and hyperscalers are competing fiercely to secure customers and chip allocations. Tokens, pricing and a cautious calculus At the centre of the deal is a broader shift in how AI firms sell compute. Tokens — the words and characters that make up both prompts and model outputs — are the unit of usage that determines how much work a model does. Anthropic's latest published rates charge $5 per million input tokens and $25 per million output tokens, a price schedule the company says better ties revenue to actual consumption. Dario Amodei, chief executive of Anthropic, has warned of a "cone of uncertainty" when it comes to capacity planning: data centres take one to two years to build, so firms are committing billions now for demand they can't yet verify. "If you're off by a couple years, that can be ruinous," Amodei told a podcast in February, adding he gets the impression some competitors "haven't written down the spreadsheet" and are acting more on hype than on hard forecasts. That insistence on measurable usage helps explain why Anthropic has moved away from unlimited, flat-rate enterprise plans and toward per-token billing. The company has also throttled or cut off some third-party tools that were heavy token consumers, a step executives say protects margins and keeps infrastructure costs aligned with real customer value. Other industry figures have warned the token tally itself can be misleading. Jensen Huang, chief executive of Nvidia, said he'd be "deeply alarmed" if an engineer on a $500,000 salary wasn't using at least $250,000 worth of compute — a remark that illustrates how internal metrics can warp behaviour. Ali Ghodsi, chief executive of Databricks, put it more bluntly: "If your goal is to just burn a lot of money, there are easy ways to do that," he said, recommending firms focus on outcomes rather than raw token totals. Enterprise buyers are uneasy Enterprise buyers are raising similar concerns, saying CTOs and CIOs struggle to build strong return-on-investment cases for heavy token consumption.

Related Articles

"Anthropic's commitment to run its large language models on AWS Trainium for the next decade reflects the progress we've made together on custom silicon," said Andy Jassy, Amazon chief executive. The pact locks in capacity and cash that will shape the next phase of the AI compute race.

This article was created with AI assistance.