Introduction
GPUQuant publishes one number per GPU per month: what it costs to rent one NVIDIA GPU for one hour at published cloud list prices. Everything else on the site exists to make that number checkable.
Currently 4,381 normalized observations across 106 regions and 52 months, back to May 2022.
Understanding GPUQuant
The central quantity is the GPUQuant Reference Price: the median of the per-provider regional medians for one GPU in one month. It is a provider-balanced, cross-region benchmark, and it always carries the provider count and region count that produced it.
Four things follow from what the underlying data actually is.
- It is a price list, not a market. There is no volume, no bid or ask, no order book and no intraday movement. Candlesticks, tickers and live indicators would all misrepresent it, so GPUQuant uses step lines.
- Region is part of the price. The same H100 costs materially different amounts in different regions, so no figure is shown without its regional scope, and the regional range is published alongside the median.
- Coverage is uneven and that is real. Providers do not all sell every GPU, and series start when the hardware launched, not when the chart begins. Short series and absent providers are shown as they are.
- Variants are different products. An H100 NVL 94GB is not a cheaper H100 80GB, and an A100 40GB is not an A100 80GB. GPUQuant benchmarks one canonical variant per model and lists the exclusions.
The unit
One GPU-hour is one NVIDIA GPU, rented for one hour, on Linux, on demand, with no commitment. One GPU-month is 730 of those hours: a month of continuous access to a single GPU.
The 730-hour month is the convention compute contracts use, including the planned rental-index futures, which is why GPUQuant uses it rather than an actual calendar month length. Switching a chart between $/GPU-hour and $/GPU-month changes nothing but the multiplier.
Where to go next
- Data sourcesHow each provider publishes prices and what each feed can and cannot give.
- CalculationsThe formulas, with worked examples for all three providers.
- Full methodologyThe complete definition, including known limitations.
- Data coverageCoverage matrix, ingestion health and every excluded SKU.
- Compute futuresWhat the planned H100 and B200 contracts are.
There is no public developer API in v1.
Frequently asked questions
- Why is the H100 price here higher than the price I see advertised elsewhere?
- GPUQuant measures the three largest cloud providers at published list price. Specialist GPU clouds and marketplaces frequently sell the same hardware for a fraction of that, and large hyperscaler customers negotiate below list. This benchmark is deliberately a hyperscaler list-price indicator, which is the most verifiable series available and the closest public analogue to a cash-market reference.
- Is this real-time?
- No, and it should not be. The underlying prices are published price lists that change in discrete steps, sometimes staying still for a year. The finest resolution the sources support is monthly, so monthly is what GPUQuant publishes. Nothing here ticks.
- Why is a chart flat for months at a time?
- Because the price did not change. A list price that holds is the normal case, not a loading failure. Charts hold the last published value and mark held months in the tooltip.
- Why does a provider line simply stop, or never appear?
- Because that provider publishes no eligible offering for that GPU. A missing provider is absent, never zero and never interpolated. The reason is shown in the legend and on the provider comparison table.
- Why is the twelve-month change sometimes N/A when the chart clearly shows a price?
- Because the two endpoints were not built from the same providers. If a provider entered the series during the year, the difference between the endpoints would partly measure that change in coverage. GPUQuant rebuilds both endpoints from the providers present in both months, and when none span the period it declines to quote a number.
- Why is the newest H200 price sometimes below the H100 price?
- That is what the published prices say. H200 is newer silicon with better throughput per dollar, and its regional footprint is smaller and concentrated in cheaper regions, while H100 is sold in many more regions including expensive ones. Both effects push in the same direction. It is a real feature of list pricing, not an error.
- Does GPUQuant let me trade compute futures?
- No. GPUQuant is an analytics product. It does not provide, facilitate or route access to trading of any kind, and it publishes no futures prices.
- Do I need an account?
- No. There are no accounts, no login and no per-user data. Everything on the site is public.