We let an AI agent buy its own GPU compute 18 times. It never once walked away empty-handed.

Description and use

An AI agent autonomously purchased GPU compute 18 times, automatically rerouting to a backup provider when RunPod failed, achieving zero dropped requests and minimal cost.

Problem

Automating GPU compute procurement without human intervention while ensuring reliability and cost efficiency.

Outcome

The AI agent completed 18 GPU compute purchases autonomously, with 89% of first attempts failing but automatically rerouted, achieving zero dropped requests and a total cost of $0.88.

Workflow

• 1. Agent formulates compute request.
• 2. Agent submits payment in USDC on Base.
• 3. RunPod processes request; if unavailable, agent reroutes to backup provider.
• 4. Compute instance is provisioned and returned to agent.
• 5. Agent logs success and repeats until 18 purchases are completed.

Evidence

18 real purchases, 18 completed jobs • 16 of 18 (89%) hit a failed first attempt — RunPod couldn't fulfill — and were automatically rerouted to a backup provider • Zero dropped requests • $0.88 in real settled payments

Source