Finance teams negotiating cloud contracts have begun seeing a new line in their terms: Amazon will pay for Anthropic’s models by the token, rather than by the hour of compute the models consume. The change, which takes effect next year, is small in appearance and significant in consequence, and people familiar with the matter say it could raise Amazon’s cost of reselling Anthropic’s technology to its cloud customers.
The arrangement governs how Amazon pays for the Claude family of models that it offers through its Bedrock platform and uses across its own products. Under the current structure, Amazon’s payments to Anthropic are tied to the compute capacity it makes available, a cost that rises and falls with infrastructure, not with how much customers actually use. The new metering ties payment to consumption, meaning every conversation, every document summary and every line of generated code that runs through Claude on Amazon’s infrastructure carries a direct charge.
The shift reflects a broader movement in the AI industry toward usage-based accounting. Model makers have spent the past year experimenting with pricing that matches the economics of inference, the stage where a trained model answers questions and does work, and token-based billing has become the industry standard for selling AI to customers. Amazon applying that same yardstick to what it pays its own supplier is a quieter version of the same trend, but the implications are larger because of the sums involved.
Amazon has invested roughly $8 billion in Anthropic since 2023, making it the startup’s largest financial backer, and the two companies are deeply intertwined. Anthropic trains and runs much of its operations on Amazon’s infrastructure, and Claude is one of the most popular model families available through Bedrock. The relationship has already drawn scrutiny from regulators: Britain’s competition authority reviewed the investment and closed its inquiry without action in 2024, while European officials have examined whether the companies’ arrangement gives Amazon privileged access to a rival’s roadmap.
The pricing change could cut both ways. If Amazon pays per token and its customers’ usage grows faster than expected, its cost base rises with demand, a margin squeeze that would be new for a cloud provider accustomed to the economies of scale in compute. Alternatively, Amazon could pass the costs through, adjusting what it charges Bedrock customers, a move that would ripple through the already competitive market for AI application programming interfaces. A person close to Amazon said the company expects to absorb part of the increase and share the rest with customers over time.
The switch also signals something about the direction of AI economics: inference, not training, is becoming the dominant cost. Training runs are large but discrete; inference is continuous, growing with every user and every agent deployed. As reasoning models, which think step by step before answering, become standard, the tokens consumed per task have climbed, and the cost of serving them has become the industry’s central accounting problem. Token metering forces every buyer and seller in the chain to confront that cost directly.
For Anthropic, token-based revenue from Amazon is more predictable than compute-hour billing, because it tracks the actual value the models deliver. The company has pushed its own pricing toward usage-based models, and aligning its largest customer’s payment structure with its own billing is consistent with that approach. For the cloud market, the change shows that the largest buyers of AI are starting to price the technology the way they price electricity, by the unit consumed.
The new metering could also affect how aggressively Amazon markets Claude against its own models and against rivals such as Google’s Gemini. If every token Amazon sells carries a cost from Anthropic, the incentive to steer customers toward models with lower per-token economics grows. Cloud customers, in turn, are already watching: several large enterprises told their procurement teams to track token consumption in the new contracts, according to a person familiar with the matter.
Anthropic’s relationship with Amazon has never been exclusive. The startup struck a separate agreement with Google in 2025 to train models on Google’s tensor processing units, and it sells its models directly to enterprises as well as through cloud marketplaces. The token-based structure does not change those other relationships, but it does change the economics of the largest one, and competitors will be watching whether other cloud providers follow with similar metering of what they pay model makers.
Executives at Anthropic declined to comment on the specifics. Analysts said the arrangement is best read as the AI industry maturing from a capacity race to a usage economy, where the winners are measured not by how much compute they own but by how efficiently they convert tokens into revenue. For Amazon, the next year will show whether that efficiency comes at a price it is willing to pay.
