On October 2, 2026, we ran a live verification of the x402 pay-per-job service on Kilawatt Cloud. No simulation. An autonomous agent asked for GPU capacity, paid in USDC on Base, and got machines back with no account, no API key, and no human intervention.
We are sharing the full results, including the part that did not go perfectly, because the failure is the most interesting part.
The test
Three requests for 8x NVIDIA A100 each, 24 GPUs in total. Each request was paid for on Base, ran for ten minutes, and went to the production endpoint. Every payment was a real USDC transaction ($1.55 per job, $4.65 in total).
Under standard routing, all three requests were served by a single supplier, Vast.ai.
Results
| # | Supplier | Config | GPUs booked | GPUs ready | Time to ready | Paid | Outcome |
|---|---|---|---|---|---|---|---|
| 1 | Vast.ai | 8x A100 | 8 | 8 | 4m 17s | $1.55 | Delivered |
| 2 | Vast.ai | 8x A100 | 8 | 0 | Never became healthy | $1.55 | Not delivered, refunded |
| 3 | Vast.ai | 8x A100 | 8 | 8 | 4m 14s | $1.55 | Delivered |
- Delivery rate: 2 of 3 requests, which is 16 of 24 GPUs.
- Time to ready: both successful jobs reached a healthy state in about 4 minutes 15 seconds and ran their full ten-minute booking.
- The failure: the third machine never became healthy. The platform detected it and refunded the $1.55 payment automatically, with no manual action. The refund was issued as account credit, not as USDC sent back to the wallet.
Why we are publishing the failure
Anyone can show a demo that works. What matters for agent-driven payments is what happens when a machine does not come up, because an agent has no one to complain to. The money is gone the moment the payment settles on-chain, so the refund path has to be automatic.
Limits of this test
- Small sample: three requests is not a benchmark. Treat these numbers as one live data point.
- Ready time: measured from placement to the first healthy status check, and checks run roughly once per minute, so figures have about a minute of resolution.
- Platform records: the figures come from Kilawatt Cloud's own records. We did not get independent confirmation from the supplier that each machine was terminated.
- Single supplier: standard routing sent all three jobs to the same place, so this test tells you about one supplier only.
Read the full reports
Every live verification run we do is published in full, with the provider, GPU count, result, time and cost for each job, including the ones that failed. You can read all of them on our Super Intelligence page:
https://kilawattcloud.dev/super-intelligence
If you are building agents that need compute on demand, we would like to hear what you would want to see in a test like this.
Kilawatt Cloud: https://kilawattcloud.dev
Top comments (1)
Việc kết hợp thanh toán crypto trực tiếp cho tài nguyên compute là một hướng đi rất thực tế để giải quyết bài toán scaling tự động. Tôi từng gặp rắc rối khi các agent chạy quá giới hạn budget dự kiến do vòng lặp vô tận, dẫn đến việc cháy tài khoản API hoặc hóa đơn cloud tăng vọt trong đêm. Việc dùng USDC giúp kiểm soát dòng tiền chặt chẽ hơn vì tính chất programmable của nó, nhưng rủi ro lớn nhất vẫn nằm ở việc xử lý các giao dịch bị treo hoặc lỗi mạng ngay tại thời điểm agent cần tài nguyên nhất. Hy vọng phần tiếp theo sẽ nói kỹ về cách các bạn xử lý logic rollback khi giao dịch on-chain gặp vấn đề.