Amazon isn't picking winners in AI anymore. It's buying every ticket.
The company said Monday it's pumping up to $25 billion more into Anthropic. Five billion lands right away. The rest unlocks as Anthropic hits commercial targets. That brings Amazon's total Anthropic check to $33 billion, on top of a $50 billion OpenAI commitment from February.
Anthropic, now valued at $380 billion, agreed to spend at least $100 billion on Amazon Web Services over the next decade. That includes heavy use of Trainium, Amazon's own AI chips.
Amazon runs the world's biggest cloud. That business prints money, but only if someone keeps renting the compute. AI is the biggest compute demand of the last 15 years.
The cleanest way to guarantee that demand is to own the customer. So Amazon writes a huge investment check. Then the customer turns around and spends it back at AWS. Then Amazon earns the margin on the hosting, the chips, and the services layered on top.
Think of it like a landlord buying a stake in every restaurant on the block. Even if one closes, the rent keeps coming in.
Why is Amazon betting on both Anthropic and OpenAI when they're direct competitors? The simple read: Amazon doesn't know who wins. So it takes the rent from both.
That's a smart move when frontier AI is still a coin flip. It's less smart if one model lab pulls decisively ahead and sucks all the enterprise demand into its stack.
For now, both Anthropic and OpenAI are reportedly eyeing IPOs later this year. Amazon's two checks give it a front-row seat to both.
The deal leans heavily on Trainium, Amazon's in-house AI chip line. Anthropic agreed to use current and future Trainium chips on AWS, which pulls workloads away from Nvidia and keeps more margin inside Amazon's stack.
That's the real strategic piece. Nvidia sells to everybody. Amazon's job is to give AWS customers a reason to run on its silicon instead, and the cheapest way to do that is to hand a frontier lab $25 billion and ask them to train on Trainium at scale.
Anthropic's $100 billion AWS spend over 10 years works like a revenue guarantee for the cloud business. That's the bigger picture of hyperscaler-lab coupling: the labs get compute at cost, the hyperscalers lock in a decade of demand.
The Trainium piece is the one that squeezes Nvidia's margin over time. Every frontier-lab training run that lands on custom silicon is a training run that doesn't land on an H100 or a
B200. Amazon gets cheaper compute, Anthropic gets a longer runway, and Nvidia loses a small slice of the highest-margin workload on earth.
Anthropic's $100 billion AWS commitment isn't a pledge. It's a line item in the AWS growth model. That number flows straight into the earnings report the way a 10-year mortgage flows into a bank's balance sheet.
Amazon didn't bet on the AI race. It bet on the road they all drive on.
