Agent surveys four GPU suppliers and buys B300 capacity in USDC via x402

On October 2, 2026, an autonomous agent with no account, no API key and no human in the loop located, paid for and received a two-GPU NVIDIA B300 machine, according to a write-up published by Kilawatt Cloud on dev.to. The payment, $2.97 in USDC on Base, was a real transaction against a production endpoint rather than a simulation.

Surveying four suppliers

Before spending anything, the agent read live inventory from four GPU suppliers. Only RunPod listed any B300 capacity at the moment of the test, offering machines of up to two cards in limited quantities. Vast.ai, Lambda and Hyperstack all showed none in stock. Because no supplier offered a four- or eight-card B300 machine, the agent did not attempt requests at that size.

The test was built around x402, a payment protocol intended for machine-to-machine transactions. The whole flow — checking inventory, requesting the machine, paying and receiving it — was handled by the agent on its own.

The booking

The agent requested RunPod's two-GPU B300 machine and paid in USDC on Base. Both GPUs were booked, and the machine became usable 17 seconds after payment. The supplier's own machine report confirmed that two B300 GPUs were actually present. The instance then ran out its ten-minute booking and shut down on schedule: Kilawatt Cloud's platform logged the shutdown at the ten-minute mark, and RunPod stopped listing the machine afterwards.

Supply, not the platform, was the bottleneck

Kilawatt Cloud says its platform can request B300 jobs of up to eight cards per request, so the ceiling on what could run was set by the market rather than the tooling. The company frames this as normal behaviour for a newly released GPU generation: most early capacity gets tied up in long-term reservations, leaving little behind for on-demand buyers. Its takeaway is that anyone needing guaranteed B300 volume should treat on-demand marketplaces as the wrong channel and pursue dedicated capacity agreements instead.

What the test does not show

Kilawatt Cloud flags several limits itself. Inventory was read immediately before the request, and reading stock is not the same as holding it, so availability could look different an hour later. A single two-GPU job is one live data point rather than a benchmark, and it says nothing about larger B300 configurations. The figures also come from Kilawatt Cloud's own platform records plus the supplier's machine report, and the test was run by the company whose platform is being demonstrated — worth keeping in mind when weighing the numbers.

Why it matters

The notable part is not the GPU count but the shape of the transaction. An agent discovered what capacity existed, paid for it in a stablecoin and had hardware delivered — without an account, an API key or a human approving anything, and within seconds. If compute can be bought this way, procurement becomes something agents can do opportunistically, shopping spot capacity across providers the way price-comparison bots shop markets, instead of something that requires contracts, logins and pre-arranged keys.

The binding constraint right now is supply. On the day of the test, only one of four suppliers had any B300s available on demand, capped at two cards per machine. Kilawatt Cloud says it will re-check all four suppliers on a regular cadence, run more two-GPU B300 jobs as stock allows, and test larger single-machine configurations — publishing results even when they fail — as soon as any supplier lists them.