H100 · $/hr on-demand
$–
What does it cost to rent an hour of an AI chip?
A live, free index of GPU rental cost — $/GPU-hour blended across providers for H100, H200, B200 and A100, on-demand and spot. The upstream input price to KOST's downstream token price — charted against Ornn's transacted index (the ICE futures reference, archived here daily), with the silicon spread over the power-cost floor.
Provider-median rental rates since 2024, on-demand vs spot. Spot (preemptible) is the dynamic leg — its discount to on-demand widens when GPUs are abundant and compresses when they're scarce.
Our list index against Ornn's transacted H100 index — the reference for ICE's compute futures, built from printed transactions. Ornn exposes only a rolling ~90-day window; we archive it daily so the permanent history stays free. The gap between the lines is the discount real deals get on posted prices.
For each GPU we take the median posted rate across every tracked cloud — not one provider's sticker. A representative price of the market, not a marketing number.
We publish both the guaranteed on-demand rate and the preemptible spot rate. The spot discount is a live scarcity gauge — it widens in a glut, compresses in a squeeze.
Provider-median list/spot rates weekly back to 2024, now charted against Ornn's transacted daily prints — we archive the rolling window daily so the permanent history stays free.
Free, provider-median, every tier — the price of an hour of AI in one number.