Amazon and Anthropic are rewriting the financial plumbing of one of the largest AI partnerships in history. Starting in 2027, Amazon will pay for Anthropic’s AI models, including Claude, based on tokens consumed rather than compute hours used on AWS.

The $100 billion relationship gets a new price tag

The pricing shift is part of a renegotiated agreement that follows a massive expansion of the Amazon-Anthropic collaboration announced in April 2026. Under that deal, Anthropic committed over $100 billion to AWS over 10 years for up to 5 gigawatts of compute capacity.

The commitment includes nearly 1 gigawatt of Trainium chips, Amazon’s custom silicon, expected to be deployed by the end of 2026. Amazon, for its part, has invested an additional $5 billion in Anthropic, with the potential for another $20 billion on top of that. Prior investment by Amazon in Anthropic had already totaled $8 billion before these recent rounds.

Why tokens change the math

In the AI world, a “token” is roughly a fragment of text, about three-quarters of a word. Every time you send a prompt to Claude and get a response, both the input and output are measured in tokens. Under token-based pricing, you pay for what the model actually processes, not how long the hardware runs.

Amazon’s spokesperson has stated that the partnership remains multifaceted, pushing back against the narrative that costs are set to rise. But the concern is real enough that reports suggest Amazon is considering hedging its bets by exploring alternatives, including OpenAI’s models and Amazon’s own Nova lineup of AI tools.

A broader industry pivot

OpenAI already prices its API access on a per-token basis. Google charges for Gemini API usage the same way. The logic is straightforward: token pricing aligns cost with value delivered. If a model generates a useful 2,000-word report, the provider gets paid proportionally. If it answers a simple yes-or-no question, the charge is minimal.

What this means for investors

For Amazon shareholders, the key question is whether token-based pricing increases the company’s cost of deploying Claude across its ecosystem. If it does, the $5 billion to $25 billion investment in Anthropic starts looking less like a strategic moat and more like an expensive dependency. Amazon’s exploration of OpenAI models and its own Nova alternatives suggests internal teams are already running those calculations.

With the token-based transition not scheduled until 2027, there’s time for both companies to adjust. But the negotiation itself reveals something important about where AI economics are headed: toward a world where every word generated by a model has a price, and the companies consuming those words are starting to count very carefully.