On the evening of Oct 10, an agent running inside goose, the open-source desktop agent, bought ten GPU machines from Kilawatt Cloud, one request at a time. Each was an NVIDIA RTX A4000 booked for 20 minutes at four cents. There was no account, no invoice and no API key handed over in advance. The agent asked for a price, signed a USDC payment authorization, and a settlement service moved the money on Base.
What happened
- All ten machines came up and shut down on schedule, running 1,202 to 1,240 seconds against 1,200 booked.
- Not one refund.
- The health agent inside the machines logged 195 readings, and every one said healthy.
- Peak GPU temperature across all ten runs was 73 C.
- Total paid: $0.40, in ten settled USDC payments.
How it works
Kilawatt publishes an MCP server with x402 tools. The agent quotes a price, checks it against local limits, signs the payment, pays, and gets back a job ID and receipt. You can see the discovery manifest yourself:
curl -s https://www.kilawattcloud.dev/.well-known/x402
The temperatures
Nine of the ten runs peaked on the first reading, then settled at 44 to 48 C within about ten minutes. These are not load temperatures. The job request has no training command, so the GPUs were close to idle.
What this does not prove
This is a small test, on our own money. The wallet is ours, so it is not customer revenue and not a customer result. No training workload ran. Earlier in our own testing, 41 percent of launches (178 of 439, since Sep 21) never became ready, and every one of those was refunded. These ten are a clean run on one GPU model from one supplier, Vast.ai, and should not be read as the platform average. The report says all of this in plain words.
Check it yourself
Every payment hash and job ID is in the full report, so you can look the payments up on Basescan. The report has six charts, the proof tables and the limits:
https://kilawattcloud.dev/super-intelligence
Look for "x402 Goose Stability Run: Ten GPUs, No Refunds" under MCP / Agent Layer.
If you want to run a test against the endpoint, or you find something we got wrong, tell us: hello@kilawattcloud.dev
Independent test by Kilawatt Cloud. Not affiliated with or endorsed by Block or the goose project.
Top comments (1)
Good to see the receipts published with the claims. One thing I checked against our own data: of the wallets that paid an x402 Bazaar seller on Base in the last 7 days, 1.2% are bound to an ERC-8004 agent identity (as of 11 Oct). So an agent buying compute like this is, as far as the chain can tell, almost always an anonymous wallet. Was the goose wallet registered as an agent anywhere, or is your report the only place that links those ten payments to goose?
I'm also curious where the "local limits" the agent checked the quote against live: in goose's config, in the MCP server, or in the wallet itself? That decides whether a prompt can talk the agent past them.