On Monday, Amazon announced it would invest up to an additional $25 billion in Anthropic, bringing Amazon’s total committed investment in the company to roughly $33 billion. In exchange, Anthropic agreed to spend more than $100 billion on AWS over the next decade, running Claude on five gigawatts of Amazon’s own Trainium chips. As a structure, the Anthropic-Amazon relationship is now larger than the Microsoft-OpenAI relationship. Microsoft’s reported cumulative investment in OpenAI sits around $13 billion, with a compute commitment that has never carried a public price tag the size of Anthropic’s. Amazon has printed a bigger number on the same template.

The elegant thing about the template is that the numbers net out in Amazon’s favor. Amazon gives Anthropic $25 billion. Anthropic gives Amazon $100 billion back, just reshaped as AWS revenue. The AWS revenue shows up on AWS’s income statement. Anthropic, meanwhile, used the word “strain” in its announcement, explaining that enterprise and consumer demand for Claude has produced “inevitable strain” on its infrastructure, which is what a lab says when its servers are on fire and it needs dilutive capital to put out the fire. It is not what a lab says when it is executing a strategic masterstroke.

The five gigawatts of Claude running on Trainium instead of Nvidia is its own story. For AWS, it is the biggest public validation yet of the multi-year bet that custom silicon can serve a frontier model at frontier scale. For Nvidia, it is a datapoint that hyperscalers will redirect frontier demand to in-house chips when the dollar amounts get large enough. For TSMC, whose foundry still makes those Trainium chips, it is more or less neutral. Both of those things are interesting. But there is a third and even more interesting thing.

Jeff Bezos personally, separately, is finalizing a $10 billion round at a $38 billion valuation for a company called Project Prometheus, backed by BlackRock and JPMorgan. Prometheus has about 120 staff, most of them poached from OpenAI, xAI, Meta, and DeepMind, and it is run by Vikram Bajaj, formerly of Google X. Prometheus is focused on “physical AI,” which is the industry’s term for systems that learn by interacting with the real world rather than by reading the internet. Target applications include manufacturing, aerospace, robotics, drug discovery, and, specifically, logistics.

Logistics. Manufacturing. Robotics. Those are the problems that a certain large retailer, which Bezos used to run, spends approximately all of its time and operating budget trying to solve. Amazon operates somewhere around a million warehouse robots and employs roughly 1.5 million people, most of whom it would prefer to automate. A physical-AI company that actually worked would have precisely one enormous, obvious, captive first customer. It is Amazon.

So here is the picture. Amazon the public company has placed the largest single bet in the world on the proposition that the future of AI is a language model from San Francisco running on AWS chips. Bezos the individual has placed a $10 billion bet on the proposition that the important AI of the 2030s is the one that moves boxes. He has done this at the same time, using the same pool of talent, out of the same capital environment, and in his first operational technology role since he stepped back from Amazon in 2021.

The two bets are complementary. Frontier language models handle the cognition layer. Physical-AI systems handle the embodiment layer. Amazon buys the first layer wholesale from Anthropic and, in time, the second layer from Prometheus, and in both cases captures the infrastructure economics because the compute runs through AWS. Bezos owns a slice of the second bet personally. Amazon owns a slice of the first. One way or another, the compute rents accrue to Amazon.

Sources: Amazon newsroom, CNBC, TechCrunch, Axios, Capacity Global (summarizing FT), The Next Web

–
By the Control Plane Editorial Team