All news

Amazon's Anthropic stake turns cloud capacity into lab equity

A stack of dark graphite server modules with a cobalt porcelain seedling growing from the top one

Amazon commits up to $4 billion to Anthropic, which moves its main workloads onto AWS and Amazon chips. Clouds now buy their way into the model layer.

Amazon has agreed to invest up to $4 billion in Anthropic, the maker of the Claude models, for a minority stake, starting with $1.25 billion and holding an option on the rest. The deal, struck on 25 September, makes AWS Anthropic's primary cloud provider for mission-critical workloads, including safety research and the development of future foundation models, which Anthropic will build, train and deploy on Amazon's own Trainium and Inferentia chips.

Anthropic commits to long-term availability of its future models to AWS customers through Amazon Bedrock and says its corporate governance does not change. The agreement mirrors Microsoft's partnership with OpenAI and gives Amazon a frontier lab of its own. It also settles who now pays for the frontier. Training frontier models costs more than venture rounds alone can fund, and the companies that sell compute have become the natural investors.

A cloud's investment largely comes home. Much of the money returns as cloud revenue, so the supplier gets equity, a long-term customer and a showcase for its own chips in one contract. Lab valuations should be read accordingly, because part of the price is paid by strategic buyers who value the compute commitment, not only the shares. The same mechanism concentrates the frontier around a few clouds. Financial investors in model labs now compete with corporations that can pay in capacity, which pushes them toward the layers above the model, where returns depend on product rather than on access to chips.

Distribution follows the money. The clouds are becoming the main channel for models: a startup on AWS can reach Claude through its existing contract, security review and bill, and the same holds for OpenAI models on Azure. Trying several models and switching between them gets cheaper, which rewards products that are model-agnostic. Anthropic's move to Trainium and Inferentia is also a serious test of an alternative to Nvidia for frontier training. If it works, compute gets more competitive for everyone else, and founders who keep their stack portable collect the gain.

Little of this is locked in yet. Only the first $1.25 billion is committed, and whether Amazon takes up the option to the full $4 billion will show how the first phase performs. The chip bet faces its own verdict when the first Claude models trained on Amazon chips are compared with those trained on Nvidia hardware. The sharpest risk to the arrangement sits with regulators, and the open question is whether competition authorities begin to treat cloud investments in model labs as a form of control.

The frontier is now financed by its suppliers. That is why Fanaura's thesis for the layer above the model favours AI-native companies that work with any lab and any cloud and gain from falling compute prices rather than paying for the race. Clouds are paying for the labs in capacity, and the cheapest seat in that arrangement belongs to the companies built on top.

For founders

Building a model-agnostic AI-native product that gains as compute gets cheaper? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.