Introduction
What GPUQuant publishes, where each number comes from, and how the benchmarks, the settlement indices and the tradable instruments relate to each other.
GPUQuant publishes the price of NVIDIA datacenter compute: what it costs to rent one H100, B200, H200 and A100 for one hour, month by month, from the price lists the cloud providers publish themselves. Beside those benchmarks it shows the indices the CME compute futures settle against, the forward curve on them, and the financial instruments that carry exposure to compute. These pages state how each number is made and what it can and cannot be used for.
The numbers on the site#
| Number | What it is | Computed by |
|---|---|---|
| GPUQuant Reference Price | The list price of one GPU-hour on an eight-GPU node, averaged across Amazon Web Services, Microsoft Azure, Google Cloud and Oracle Cloud Infrastructure. Monthly. | GPUQuant |
| Neocloud index | The same GPU-hour across 16 specialist GPU clouds, from their published rates. Monthly. | GPUQuant |
| Silicon Data indices | Daily realized rental prices, modelled by Silicon Data. SDH100RT and SDB200RT are what the compute futures settle against. | Silicon Data |
| Forward curve | What the market expects the settlement index to be, out to three years. | Silicon Data |
| Residual value | What one physical GPU is worth, derived from the forward curve. | GPUQuant |
| Instruments | Perpetuals and futures with exposure to the compute economy, priced live by their venues. | The venues |
Two tiers, never averaged#
The hyperscalers and the specialist GPU clouds sell the same hardware at prices that differ by roughly a factor of three. GPUQuant keeps them as two figures, the reference price and the neocloud index, computed the same way and shown side by side. A single blended number would hide the one fact a reader most needs.
Both are published list-price indicators. Neither is a transaction-weighted market price, neither is a settlement index, and neither reflects negotiated or committed-use discounts. The settlement indices are Silicon Data's, and the site says whose number is whose everywhere one appears.
What GPUQuant is not#
- Not the CME settlement index. The reference price and the neocloud index are calculated independently from published list prices. The contracts settle against Silicon Data's indices, which GPUQuant displays under licence and does not compute.
- Not affiliated with CME Group, Silicon Data, any cloud provider or any trading venue.
- Not a broker or a custodian. Where an instrument can be traded from the site, the order is signed in your own wallet and executed by the named venue, which holds the collateral. GPUQuant holds no funds and no keys.
- Not a record of transactions. A list price is what a provider publishes, not what every customer pays.
- Not investment advice.
Where to start#
- If you trade, or plan to
- Compute futures explains the contracts and what one represents. Silicon Data indices explains what they settle against and how those figures relate to the benchmarks here.
- If you cite a number
- The reference price and the neocloud index state the method. Data sources says where each price is read from, and Coverage shows what is stored right now, including everything excluded and why.
- If a chart looks wrong
- Time and coverage covers flat lines, held prices, missing providers and a change that reads N/A. The common questions are shorter still.
Reading the docs by machine#
Every page here is also served as markdown at the same address with .md appended, for instance /docs/reference-price.md. The index is at /llms.txt and the whole documentation in one file at /llms-full.txt. Both forms are rendered from the same source, so they cannot disagree.