An AI agent can now buy inference one request at a time. The first production economics are tiny—and important.

AWS says Incarna put pay-per-inference into production with BlockRun and AgentCore Payments. During beta, agents processed more than 1,000 individually settled payments ranging from $0.001 to $0.05 per inference call, with spend limits enforced outside the model.

The signal

AWS says Incarna has put a pay-per-inference flow into production using Amazon Bedrock AgentCore Payments and BlockRun. During beta, agents processed more than 1,000 payments ranging from $0.001 to $0.05 per model call, each settled individually on-chain.

A machine is buying another machine’s intelligence

The agent encounters a priced inference request, checks it against an infrastructure-enforced spending limit, signs the x402 payment and receives the model inference. The customer owns the wallet and grants delegated signing authority; the model cannot raise its own limit.

The economics are tiny by design

Individual payments of a fraction of a cent to five cents will never look impressive beside enterprise procurement values. That is precisely why autonomous settlement matters: human approval would cost more than the service being purchased.

What is proven

AWS reports that Incarna completed the AgentCore Payments integration in three days, versus an original two-to-three-month scope, and ran more than 1,000 real payment events during beta. The production flow enforces session budgets outside the model.

What this does not prove

The sample is still small. We do not have aggregate spend, task-level ROI, repeat-agent counts, provider revenue, or evidence that purchasing external inference improves outcomes enough to justify the added cost across large workloads.

What we watch next

Payments per agent session, inference spend per successful task, model-selection economics, provider revenue, repeat machine buyers, aggregate volume and whether dynamic purchasing measurably improves task completion or cost efficiency.