On the evening of October 10, an autonomous agent running inside goose, the open-source desktop agent, successfully purchased ten NVIDIA RTX A4000 GPU machines from Kilawatt Cloud. The transaction was executed entirely by the agent, with no human intervention, no pre-existing account, and no API key handed over in advance. Each GPU was booked for 20 minutes at a cost of four cents, settled in USDC on the Base network.

The Transaction Mechanics

The agent utilized an MCP server published by Kilawatt Cloud equipped with x402 tools to facilitate the purchase. The workflow involved the agent quoting a price, checking it against local limits, signing a USDC payment authorization, and receiving a job ID and receipt upon settlement. A settlement service moved the funds on Base, completing the loop from request to resource allocation without traditional billing infrastructure.

Performance and Stability Data

All ten machines launched and shut down on schedule, running for durations between 1,202 and 1,240 seconds against the 1,200 seconds booked. There were zero refunds, indicating successful delivery across all instances. Internal health agents logged 195 readings, all of which reported healthy status, with peak GPU temperatures reaching 73ยฐC. Notably, nine of the ten runs peaked on the first reading before settling at 44 to 48ยฐC within ten minutes, suggesting the GPUs were largely idle as no training workloads were assigned.

Limitations and Context

Kilawatt Cloud explicitly notes that this test used their own wallet, meaning it does not represent customer revenue or results. The sample size is small and limited to one GPU model from a single supplier, Vast.ai. In earlier internal testing since September 21, 41 percent of launches (178 out of 439) failed to become ready, though all were refunded. This specific run of ten should not be interpreted as the platform's average reliability.

Key Takeaways

  • Autonomous agents can now independently negotiate, pay, and provision cloud GPU resources using stablecoins.
  • The integration of MCP servers with x402 tools enables trustless payment authorization without human oversight.
  • Settlement occurred on Base with USDC, demonstrating the viability of Layer 2 networks for micro-transactions in AI infrastructure.
  • Kilawatt Cloud provides full transparency via Basescan payment hashes and a detailed report at kilawattcloud.dev/super-intelligence.

The Bottom Line

While the technical execution of this autonomous purchase is a significant milestone for agent-native infrastructure, the results must be viewed with caution due to the lack of training workloads and Kilawatt's own admission of a 41 percent failure rate in broader testing. This is a proof of concept for payment rails, not a benchmark for production reliability.