# Silicon Data indices > The Silicon Data GPU rental indices the compute futures settle against, the forward curve, the RAM and LLM token indices, how each relates to the GPUQuant benchmarks, and what the site is licensed to show. Updated 2026-09-06. Canonical: https://gpuquant.com/docs/silicon-data Silicon Data publish daily indices of what NVIDIA GPU capacity actually rents for, modelled from transaction data across neo-cloud providers, hyperscalers, colocation markets and private rental platforms. Two of them are what the CME compute futures settle against. They are a different kind of number from anything GPUQuant computes: a GPUQuant figure is a reading of a published price with a URL behind it, and a Silicon Data figure is a standardized estimate from data GPUQuant cannot see, by a method it cannot reproduce. ## The GPU rental indices | Ticker | Index | Reading | Settles | | --- | --- | --- | --- | | SDH100RT | [Silicon Data H100 Rental Price Index](https://www.silicondata.com/products/silicon-index/h100) | Neo-cloud segment | Silicon Data H100 Rental Index Futures, NYMEX, listing 2026-10-05 | | SDB200RT | [Silicon Data B200 Rental Price Index](https://www.silicondata.com/products/silicon-index/b200) | Blend of every segment | Silicon Data B200 Rental Index Futures, NYMEX, listing 2026-10-05 | | SDA100RT | [Silicon Data A100 Rental Price Index](https://www.silicondata.com/products/silicon-index/a100) | Neo-cloud segment | No listed contract | Each index is the realized, standardized hourly rental price of the GPU in USD per GPU-hour. Silicon Data standardize for machine specification, rental term, platform performance and geography, remove statistical outliers and validate each day independently. The indices are computed **once a day**; there is no intraday value, so daily is the full resolution of the instrument rather than a compromise. Silicon Data also restate history from time to time, and a restatement is a new reading of an old date. ## A segment and a blend The two settlement indices are not the same kind of number, and that is the single most important thing on this page. - **The H100 index is a segment.** Silicon Data separate neo-cloud and hyperscaler pricing into distinct readings, and the ticker the contract references is the neo-cloud one. A hyperscaler H100 reading exists and the contract does not reference it. - **The B200 index is a blend.** It combines neo-cloud, hyperscaler, colocation and private-market observations into one reading, cannot be decomposed into a neo-cloud figure, and has no hyperscaler counterpart. So a like-for-like comparison against GPUQuant's own neocloud index holds for H100 and does not hold for B200, where the published figure already contains hyperscaler and colocation capacity. ## How they relate to the GPUQuant benchmarks | GPUQuant figure | Silicon Data figure | Comparable? | | --- | --- | --- | | [Neocloud index](https://gpuquant.com/docs/neocloud-index), H100 | SDH100RT, the neo-cloud segment | Same population. A list price against a realized price, so the GPUQuant figure sits above it. | | [Reference price](https://gpuquant.com/docs/reference-price), H100 | The hyperscaler segment of the H100 index | Same population, list against realized. | | Neocloud index or reference price, B200 | SDB200RT, the blend | Not like-for-like. The blend already contains hyperscaler, colocation and private capacity. | In every case the GPUQuant figure is a monthly list price and the Silicon Data figure is a daily modelled transaction price. They answer different questions and are shown as different products on the market-data page. ## The forward curve Silicon Data also publish a forward curve for each GPU: what the market expects the settlement index to be at each point from today out to three years, as a rate per GPU-hour. Tenor zero is that day's index, to two more decimal places, so the curve is the market's expected path for the settlement reference rather than a separate quantity. It anchors on the model's primary segment: neo-cloud where the index is segmented, the blend for B200. - **It is a surface, not a series.** Each day publishes a whole new curve. Fix a day and vary the tenor for the curve as it looked that day; fix a future month and vary the day for how the market revised its view of that month, which is the direct analogue of a futures contract's price history. - **Tenors run in quarter-month steps**, so a complete curve is 145 points from zero to 36 months. Two rates are published at each tenor: the term rate, the average out to that tenor, and the forward rate, the implied rate for that single period. They coincide at tenor zero and diverge after. - **Early curves were shorter.** A curve's length is a fact about the day it was published, not a validity check. ## Residual value Residual value is what one physical GPU is worth: the discounted value of what it can still earn over its remaining life. Silicon Data publish a residual value series behind their portal. **GPUQuant computes its own estimate** from the public forward curve, and the two are never presented as the same number. The model is the one Silicon Data describe on their product page: a continuous discounted cash flow over the forward curve across the GPU's economic life, with cash flow equal to the rental rate times utilisation times one minus an operating cost ratio. They publish no constants, so the life, the scale and the discount rate were fitted against their published output paired with the curve that fed each day, with the GPU's age taken from its release date rather than fitted. The fit is checked out of sample and the error against the publisher is stated on the chart. | GPU | Released | Fitted life | Mean error vs. publisher | | --- | --- | --- | --- | | H100 | 2023-03-21 | 79 months | 0.7% over 517 days | | A100 | 2020-11-16 | 86 months | 1.4% over 515 days | | B200 | 2024-10-15 | 77 months | 0.4% over 261 days | ## The RAM and token indices Two more Silicon Data indices settle nothing and are shown as context on either side of the rental price: memory is an input cost to the hardware, and token expenditure is the demand that fills the datacentre. **RAM Index (GDDR6)** A weighted basket of constituent GDDR6 products, reviewed every two months. An input cost to the hardware behind a rental price. Unit: USD, with no denominator stated by the publisher. Published daily on business days, 16:00 utc. **LLM Token Expenditure IndexSDLLMTK** What a million tokens of LLM inference costs to buy. The demand side of GPU rental, where the rental price is the supply side. Unit: USD per million tokens. > **Two things to know before quoting either** > > The RAM index states what it is a basket of but never what one unit is per, so it is comparable with itself over time and is not a quotable memory price. The site shows it as an index with no unit for that reason. > > The token index is a weighted blend of an open-weight series and a proprietary series that Silicon Data publish only to portal users. A move in the blend can be a shift in the mix between the two rather than a change in either price, so it says less on its own than it appears to. ## What the site shows Everything of Silicon Data's on the market-data page is read from their public charts and shown for the newest **7 days** only, under a written agreement with Silicon Data that covers display on this site and excludes redistribution, historical access and onward licensing. Nothing older than that window is served, the curve day picker is cut to the same window, and a reading obtained any other way is never displayed. Historical or broader access is a separate agreement with Silicon Data.