Amazon starts paying Anthropic by the token, and disputes that this costs more

Amazon and Anthropic have restructured how AWS pays for Claude. From next year, the billing shifts from compute-hours (reserved GPU capacity) to tokens (units of model output). The Information reported the change on 30 June; Amazon disputes the accompanying claim that it will raise costs. Both things can be true, and the structural story sits in the gap between them.

What actually changed. Under the old arrangement, Amazon paid for reserved compute — GPU and TPU capacity, billed by the hour, whether or not the silicon was doing work. Under the new arrangement, Amazon pays per token generated or processed. Token pricing scales with inference volume; compute-hour pricing scales with reservation. That is a different economic object, and the switch was negotiated as part of the broader partnership refresh tied to Amazon's additional $25bn commitment to Anthropic in 2026, on top of the earlier $4bn stake.¹

Neither party has disclosed the per-token rates or the volume-tier schedule. That absence matters, because it is the entire ballgame.

What the frame predicts. Inference economics (the cost of running models, rather than training them) is the binding constraint at the frontier now, and the unit that matters is the token. When a hyperscaler moves its internal accounting from capacity to output, it is admitting the same thing the market has been saying for eighteen months: the scarce, priced good is inference work, not the machines that do it. Buying tokens is buying the finished thing. Buying compute-hours is buying an option on the finished thing.

For Amazon, the immediate consequence is that cost now tracks product usage. Alexa+, Kiro (Amazon's AI developer tool), and the Claude surface inside AWS Bedrock all consume tokens directly. If those products scale, the bill scales with them. Under the old regime, growth in Alexa+ usage against reserved capacity was, at the margin, free — Amazon had already paid. Under the new regime, it is not.

Who this favours. This is a shift in bargaining power toward Anthropic, and the shape of it is worth being precise about. Compute-hour pricing puts idle-capacity risk on the vendor: if Amazon over-reserved, Anthropic still got paid, but the marginal token was cheap for Amazon to produce once the reservation was sunk. Token pricing puts volume risk on the buyer: Amazon now pays for every token its products actually generate, and Anthropic captures revenue proportional to how much Claude gets used across the AWS estate.

The $25bn investment reads differently in that light. One tempting reading is that Amazon paid $25bn and got, among other things, a worse pricing model. A more careful reading is that the $25bn is the reason Anthropic agreed to a token schedule Amazon can live with — that the investment bought preferential volume tiers rather than the previous capacity arrangement. Both readings are consistent with what has been reported. Neither can be settled without the rate card.

Amazon's denial, read carefully. Amazon disputed the "could raise costs" framing in The Information's report. It did not deny the pricing model change.² That is a specific kind of denial, and it is structurally coherent.

At sufficient inference volume, per-token rates negotiated with your largest investor can plausibly come in below the blended cost of reserved compute-hour capacity, particularly if Anthropic's inference efficiency (compute consumed per token) has improved since the original deal. Token pricing also eliminates the deadweight cost of idle reservations, which is not a trivial line at hyperscaler scale. If Amazon's utilisation on the old contract was, say, 60%, then a token rate priced against 100% utilisation is competing with a compute-hour rate effectively inflated by the idle 40%. It is entirely possible for Amazon's total bill to fall while the marginal price of a token, on paper, rises.

So the honest position is: Amazon is probably not lying, and The Information's sources are probably not lying. They are comparing different things. The list-rate comparison points one way; the blended, utilisation-adjusted comparison points the other. Neither has been published, so neither can be checked.

What this is a case of. The metering logic here is the same one that restructured enterprise software over the last three years. Vendors moved customers from per-seat pricing (a fixed fee per human user) to consumption pricing (a fee per unit of work done), and the effect was to compress the margin of any reseller sitting between vendor and end customer. When the reseller was priced on capacity and re-sold on capacity, the arbitrage was stable. When the vendor re-priced the reseller on consumption, the arbitrage compressed unless the reseller also re-priced downstream.

Amazon is the reseller in this arrangement. Anthropic is the vendor. AWS Bedrock is the downstream product. The question, and it is the question that determines whether Amazon's denial ages well, is whether Bedrock's token pricing to enterprise customers gets tightened to match. If it does, Amazon has passed the structural risk through. If it doesn't, Amazon is absorbing it.

What to watch.

• Bedrock's public token pricing for Claude models over the next two quarters. A quiet tightening would suggest Amazon is passing the new cost basis downstream.
• Alexa+ and Kiro unit economics, if Amazon discloses them. These are the internal consumers most exposed to the switch.
• Anthropic's disclosed revenue mix. If AWS-routed revenue grows faster than direct API revenue in 2027, the token schedule is doing work for Anthropic.
• Whether any of the other frontier labs sitting on hyperscaler compute contracts (OpenAI-Microsoft, Google's internal Gemini accounting) renegotiate onto token-equivalent terms. The Amazon-Anthropic deal is the first public instance of this reshaping; it is unlikely to be the last.

The interesting thing about this restructuring is not the headline cost direction, which cannot be determined from public data and which both parties have incentives to spin. The interesting thing is that the largest hyperscaler-to-lab commercial relationship on record has quietly changed its unit of account, from the machine to the output of the machine. That is inference economics arriving in the plumbing. Everything downstream of it — Bedrock pricing, hyperscaler margin at the AI layer, who captures the surplus as Claude usage scales — follows from that change of unit.

Footnotes