The same GPU costs up to 2× more depending on where you rent it

On the day we wrote this, an H100 PCIe rented for $2.00/hr on one tracked provider and $3.29/hr on another — a 1.6× spread on identical silicon. The A100 80GB SXM spread 1.8× ($1.59 to $2.79). These aren’t stale listings; they’re simultaneous, bookable list prices, and some version of this spread appears in our provider matrix every single day.

Why doesn’t it arbitrage away? Because a GPU-hour isn’t actually a commodity — yet. The spread decomposes into real differences: reliability tiers (a marketplace host and an enterprise SLA are different products), interconnect and cluster context (the same card in an InfiniBand pod is worth more than in a lone box), switching costs (data gravity, image rebuilds, quota relationships), and plain information friction — most renters simply don’t see today’s table.

The interesting question is which part of the spread is real quality difference and which part is friction that transparency erodes. Our read of the record so far: the floor moves first on marketplaces, then list-price providers follow in discrete cuts (the events feed catches each one). The spread compresses slowly — but it compresses.

Practically: if your workload tolerates interruption, the spread is free money (spot runs ~40–50% under on-demand on top of choosing the cheap provider — the calculator multiplies it out for your run). If it doesn’t, you’re paying for the tier, not overpaying for the GPU — and now you know exactly how much.