An optional per-call settlement lane (x402) for Serverless endpoints

[dhyabi2]

Motivation

RunPod Serverless is billed per second of compute — from the moment a worker starts until it
fully stops, rounded up, on the GPU tier in play (e.g. ~$0.00016/s on A4000 through ~$0.0024/s
on B200). With Flex workers scaling to zero when idle, the unit you actually charge for is the
individual job a caller submits to an endpoint: one /run or /runsync request, one worker's
execution time, a cost that lands in fractions of a cent up to a few cents for a single
inference.

That is exactly the tier where the settlement rail outweighs everything else. When the caller
is an AI agent invoking one endpoint once, a per-call cost of fractions of a cent cannot
sensibly carry an API key, a prepaid balance, a minimum top-up or a card on file. The smaller
the payment, the more the rail itself matters — and today there is no per-call, no-account,
feeless way for an autonomous agent to pay for a single Serverless job on RunPod.

Proposed solution

Offer the x402 protocol (HTTP 402 Payment Required) as an optional settlement path on
Serverless endpoints, paired with a feeless settlement network (Nano/XNO — zero
transaction fees, sub-second finality). The flow:

1. An agent calls an endpoint tokenlessly; the endpoint answers 402 with a pay_to address
   and an amount.
2. The agent settles in XNO and retries with the receipt in the request.
3. RunPod verifies (a short on-chain check), runs the job, and returns the result — one HTTP
   round-trip, no account, no card, no minimum.

The billing you already do gives you the exact number to quote: the per-second GPU rate times
the job's expected execution time is a natural maxAmountRequired for that endpoint and GPU
tier. Because the RunPod caller side is an SDK (runpod-python), a small client helper that
mounts an x402 payment handler would give existing users a copy-paste-able shape to run today.

Alternatives

• Prepaid credit / top-up balances: workable, but they put a top-up and minimum-balance step
  between an agent and a single cheap job — disproportionate at sub-cent costs.
• Card on file per human account: not a path an autonomous agent can walk.
• API-key gating: fine for a single operator, but it does not scale to dozens of agent callers
  each paying for their own invocation.
  The x402 + feeless lane is additive: keep every existing model, add the no-account per-call
  path for agents.

Additional context

x402 is an open, neutral standard for internet-native per-request payments (https://x402.org).
Feeless settlement means a job priced in fractions of a cent settles for the value of the call
itself, same round-trip, no fee floor. Happy to put together a working reference handler
against a Serverless endpoint (or a client snippet in runpod-python) if that would help the
discussion.