Price indices, compared
How the GPUQuant reference price, Silicon Data’s indices and Ornn’s Compute Price Index each arrive at a dollar figure for an hour of the same GPU, why they differ, and what the site shows of each.
The market-data page groups price indices by what they measure, not by who publishes them. Under each GPU sit every figure GPUQuant can show for an hour of it, from three different kinds of source, and each row names whose number it is. They are not three estimates of the same thing that ought to agree. They are three different questions, and the gap between the answers is the information.
Three numbers for one GPU-hour#
| Row | Source | What it is | Reading |
|---|---|---|---|
| Hyperscaler list, Neocloud list | GPUQuant | The price a provider publishes on its own price list, read from that list. | Monthly, held forward |
| Neo-cloud index, Hyperscaler index, Blended index | Silicon Data | A modelled estimate of what capacity rents for, across deals GPUQuant cannot see. | Daily |
| Transacted | Ornn | A volume-weighted average of the prices rentals actually cleared at. | Daily, settled 20:00 UTC |
List prices: the GPUQuant reference#
The GPUQuant rows are the site's own figures: the on-demand price each provider publishes for the eight-GPU node, collected daily, normalized to one GPU-hour and averaged in two steps, per region then per provider. Every figure has a URL behind it and can be checked against the provider's page. It is the ceiling of the market, the price before any negotiation, and it moves a few times a year on month boundaries. The method is on the reference price and neocloud index pages.
Modelled: Silicon Data#
Silicon Data's indices are standardized estimates of what NVIDIA capacity rents for across neo-clouds, hyperscalers, colocation and private platforms, modelled from transaction data by a method GPUQuant cannot reproduce. They cover negotiated and committed-use deals, which is why they sit well below a list price. Two of them are what the CME compute futures settle against, so they are the market's reference whether or not one agrees with the method. The Silicon Data page describes each index, its segment, and the forward curve.
Transacted: Ornn#
The Ornn Compute Price Index is built from the prices GPU rentals actually cleared at: Ornn's own words are a “volume-weighted, winsorized average of transacted GPU rental prices”, normalized across configuration, provider and region, and settled once a day by 20:00 UTC. It is the transacted view of the same market the other two describe from a price list and from a model, and it is carried on the Bloomberg Terminal. Ornn publishes the method at data.ornn.com/methodology.
Reading them side by side#
Expect the list price on top, the transacted index below it, and the modelled index lowest, and expect the order to hold. When it does not, something has changed: a provider cutting its list price closes the gap from above; a transacted index rising toward list means capacity is tight and discounts have gone; a modelled index moving while the other two sit still is a change in the estimate rather than in the market. The compare control overlays any rows in a group, and the change view puts them on one percentage axis so a $12 list price and a $3 index can be read together.
The forward curve and the residual value#
Under H100, B200 and A100 two more subsections sit beside the price indices, because they are about the same GPU and belong under its name. The forward curve is Silicon Data's: what the market expects its settlement index to be from spot out to three years, drawn with tenor along the axis rather than time, so it is compared day against day rather than overlaid on a price. The residual value is GPUQuant's own derivation from that curve, the discounted value of what one GPU can still earn over its remaining life, in dollars per GPU rather than per hour. Neither is a price index, and neither can be overlaid on one; the compare control offers only rows in the same unit on the same kind of chart. Both are described on the Silicon Data page.
The token price indices#
The same idea applies to inference. Silicon Data's token index is a blend across labs of the price of a million tokens, weighted by spend; Ornn's token price index is the same quantity per lab, volume-weighted, settled daily. Beside each other they show whether a move in the blend is a price change or a shift in which labs' tokens are being bought.
What the site shows of each#
GPUQuant's own rows are complete histories. Silicon Data's rows are the newest 7 days, under a written display licence. Ornn's rows are the trailing 92 days for GPUs and the trailing month for tokens, read from Ornn's public API with Ornn's permission and shown with Ornn named as the source. None of it is redistributed, and none of the third parties' data is stored beyond what their own public endpoints serve.