Today we ran our biggest live test of Kilawatt Cloud so far. An autonomous agent paid for GPU jobs with x402 (USDC on Base), sent them through our gateway, and we counted what actually happened.
The numbers, straight from our database:
329 jobs attempted, each one paid for through x402.
200 jobs landed, meaning the machine reached a ready state. That's about 61 percent.
$17.73 paid in total, with $1.82 refunded automatically on jobs that never became ready.
Four GPU suppliers were in the routing mix.
We also ran 8 long stability jobs alongside the launches, and all 8 stayed healthy.
What this was, and wasn't: This was our own test, run from our own wallet. It is not customer traffic, and I'm not presenting it as demand. It is evidence that the payment, routing and refund path works under load, with real money and real hardware.
What we learned: The 129 jobs that didn't land are the most useful part of the data. Some suppliers had streaks of machines that never became ready, and one supplier we had to pause under a spend cap. Because the refund path worked, those failures cost us cents and not trust. That's the point of routing across multiple suppliers.
We started the day aiming for 100 landed jobs and ended at 200. Next run, we're aiming for 400.
If you build agents that need to buy compute on their own, I'd like to hear what breaks for you. The full run report, with the results by supplier, the hardware health data and the failures, is on our Super Intelligence page: https://www.kilawattcloud.dev/super-intelligence
Our MCP server is also on npm as kilawatt-mcp-server.
Top comments (1)
Số 61% success rate ở quy mô này khá thú vị — cho thấy failure chủ yếu do chỗ nào? Capacity limit, cold-start timeout, hay payment settlement race condition? Mình từng thấy pattern tương tự khi scale spot instances: phần lớn fail là do interruption notice đến trước khi job checkpoint kịp, không phải do code.
x402 payment flow có retry logic built-in không, hay agent tự handle idempotency key? Vì 129 job fail mà không retry tự động thì chi phí opportunity cost còn lớn hơn chi phí GPU.
Curious: agent có dynamic bidding strategy dựa trên real-time spot price history không, hay chỉ static max-bid? — found it via LabAgent, site: labagent .tech