# The reference price > How the GPUQuant Reference Price for H100 and B200 is calculated from the list prices of AWS, Azure, Google Cloud and Oracle: what is measured, the five steps, and why each provider counts once. Updated 2026-09-06. Canonical: https://gpuquant.com/docs/reference-price The GPUQuant Reference Price is the list price of one GPU-hour on the hyperscalers: what Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure publish for renting an 8-GPU node, on demand, divided to one GPU, balanced so that each provider counts once. It is one number per GPU per month. ## What is measured GPUQuant tracks **first-party published prices only**. Aggregated third-party pricing services and GPU marketplaces are out of scope: a number that cannot be verified against a primary source cannot be defended. The measured price is, without exception: - USD per individual GPU-hour - Linux - On-demand, pay as you go - No spot or preemptible capacity - No reservations, committed use or savings plans - No Dev/Test or negotiated pricing - Shared tenancy, no dedicated hosts, no software licences The unit is the **8-GPU SXM or HGX node** on every provider, so the comparison holds the machine constant. How a published machine price becomes a per-GPU figure, and which variant of each GPU counts, is on [Normalization](https://gpuquant.com/docs/normalization). The reference price is one of two tiers. The other, the [neocloud index](https://gpuquant.com/docs/neocloud-index), is built the same way from 16 specialist GPU clouds, and the two are never averaged together. ## The five steps Region is a real pricing dimension: the same H100 costs materially different amounts in different regions, and GPUQuant never publishes a bare number without its regional scope. 1. Normalize every eligible offering to USD per individual GPU-hour. 2. Within each provider and region, take the **lowest** eligible price. Where a provider lists several equivalent nodes, for instance an Azure host with and without InfiniBand, the cheapest is the honest price of renting that GPU there. 3. Take the **average** of each provider's regional prices. 4. Take the **average of the available provider averages**. That is the GPUQuant Reference Price. 5. Record the minimum and maximum across every eligible regional price as the regional range, with the provider count, the region count and the exact provider set. ``` AWS regional prices → average = A Azure regional prices → average = A Google Cloudregional prices → average = G Oracle regional prices → average = O Reference = average(A, A, G, O) Regional range = min(all), max(all) ``` ## Each provider counts once Balancing at the provider level is the load-bearing step. Google Cloud publishes H100 in far more regions than AWS does, so a flat average across all quotes would be a Google Cloud index wearing a global label. Collapsing each provider to one figure first is what removes that weight, and it is why the final step does not also have to be a median. The median is still computed and stored beside each provider figure, and the two differ by around two percent wherever a GPU is sold in more than a handful of regions. ## With fewer providers With one provider the reference is that provider's own average, and the label says so. The provider count is stored with every published point and shown in the tooltip. A provider with no eligible offering for a GPU is absent from the average, never zero. When the contributing provider set changes between two months, the reference can move for a reason that has nothing to do with prices; every stored point records its provider set so that case can always be identified, and [Time and coverage](https://gpuquant.com/docs/history) states how the charts mark it. ## Calculation versions Every aggregate carries a calculation version, currently **v4**, so a change to the maths never makes old rows unexplainable. - **Version 1** published a median of the provider medians. With the three providers there were at the time, a median returns one provider's figure verbatim and discards the other two, so the largest provider could halve its price without the headline moving. - **Version 2** publishes the average of the provider figures, with the median kept beside it. - **Version 3** added Oracle as the fourth provider from August 2026. The maths did not change; the inputs did. ## Published and tracked Two GPUs are published: NVIDIA H100 80GB SXM and NVIDIA B200 180GB. They are the two CME Group and Silicon Data plan to list rental-index futures on, and each has its own page with the full series and the offerings behind it. GPUQuant prices and stores more NVIDIA datacenter GPUs than it publishes, and has done since the first snapshot. Those series keep growing whether or not they are shown, because Microsoft and Oracle publish no price history and a month not captured cannot be recovered. Publishing one of them is a flag; capturing one late is impossible.