GPU rental prices: median $/GPU-hour by chip

GPU rental pricing is quoted publicly by dozens of providers and compared by almost nobody, because the quotes are not comparable as published: per-node against per-card, hourly against a committed year, reserved against interruptible. Gambit normalises every quote to dollars per GPU-hour and keeps a daily snapshot. This page is the free summary layer. The full grid (every provider, every term, the scatter and the history) lives on the DATA desk inside the terminal.

Median advertised price per GPU-hour on the latest price snapshot, 2026-07-31. The quote count is the number of provider-and-term quotes behind that median on that date; low and high are the extremes of the same sample.

The same chip prices differently by how long you commit. Gambit anchors the H100 curve on the on-demand reference of $5.50/GPU-hr and expresses every other tenor as a discount or premium to it. The curve is currently in contango with a slope of +76%: near-dated, interruptible capacity is cheap relative to guaranteed on-demand supply, so the curve rises out of spot.

Spot clears at $2.63, 52% below the on-demand reference, on 37 quotes, the deepest sample on the curve. The 12-month point at $4.65 (-15%) is the one most buyers actually transact at.

Every megawatt going into the ground is underwritten by an assumption about what an hour of compute will sell for. That number moves, it is quoted publicly by dozens of providers, and almost nobody writes it down over time. Gambit does, chip by chip, with the sample that produced it.

Medians travel with their quote count. A headline price with no sample size cannot be checked, and one seller's asking price is not a market. Every median here carries the number of provider and term quotes behind it, and the thin rows are flagged rather than quietly averaged in. That is the difference between a price you can act on and a price you can only repeat.

The curve is the forward view. Spot says what interruptible capacity clears at today. The committed tenors say what buyers will pay to be certain of supply for a year or for three, which is a more honest read on expected scarcity than any vendor roadmap.

It joins the other two layers. Rental prices are what compute earns, hyperscaler capex is what is being spent to supply it, and the datacenter buildout is what has actually been built and energised. Read together, they are the AI trade with its receipts attached. All three are free here, dated, and traceable to the filing or the quote they came from.

Coverage runs from Blackwell and Hopper datacenter parts down to consumer cards on rental marketplaces, and depth across it is very uneven: some chips carry dozens of quotes from hyperscalers, neoclouds and spot marketplaces, others carry one. That is the honest shape of this market, and it is why the count travels with every price on this page.

Normalisation. Every advertised price is converted to dollars per GPU-hour, so an 8-GPU bare-metal node and a single-card VM sit in the same column.

The median, not the average. One provider listing an eight-GPU node at a node price would drag a mean into nonsense. The median is robust to that, and the low and high columns show you the dispersion that the median hides.

Daily snapshots. The snapshot history behind this page covers 14 dates, 2026-07-28 to 2026-08-10. Medians here are struck on the latest price date, 2026-07-31.

What is not here. This is advertised rental pricing. It is not what a large buyer negotiates at, and it is not a settlement price: there is no clearing venue for GPU-hours to settle against.

This page publishes the nine chips carrying a live median on one snapshot date. The desk holds the rest: every chip it prices, the individual provider and term quotes behind each median rather than the summary, a term structure for any chip rather than the H100 alone, and the snapshot history the medians move through.

What are customers paying for intelligence?