Compute Futures: Silicon Data’s GPU Prices

Silicon Data publishes GPU rental price benchmarks that express the cost of covered GPU capacity as standardized hourly rates.

GPUs are processors used for tasks such as running AI models. One GPU-hour means using one GPU for one hour, so a rental price per GPU-hour describes the cost of that capacity over time.

Synoptic’s GPU pricing data includes provider-price updates that customers can use in their own forecasts between Silicon Data’s benchmark readings.

How do Silicon Data’s GPU price indexes work?

Silicon Data’s GPU benchmark family covers NVIDIA H100, H200, A100, B200 and B300 GPUs, plus AMD MI300X. It also publishes separate H100 and A100 hyperscaler readings.

Its governance document, section 2, specifies publication once each business day by 21:00 UTC. The settlement benchmarks use verified rental transactions and executable offers; static marketing rate cards and unverified feeds are excluded.

The published methodology follows four steps:

1. Convert rental prices into USD per GPU-hour, accounting for the number of GPUs and rental duration.

2. Filter incomplete or invalid records and check unusual price movements and outliers.

3. Adjust for differences in rental terms, location and hardware so prices are comparable.

4. Average the adjusted observations for each provider, then combine those provider values using weights.

CME plans to use Silicon Data’s H100 and B200 neocloud benchmarks as reference prices for its proposed compute futures, contracts tied to GPU rental costs. These two benchmarks cover the following provider segments:

Neoclouds are cloud providers focused on GPU computing. Hyperscalers include AWS, Azure and Google Cloud. Match a provider update to the benchmark segment it belongs to before using it in a forecast.

Where do provider-price updates fit?

Silicon Data publishes a benchmark reading on its schedule, while providers can change their offers at any time between updates. Synoptic’s real time data provides customers with information they can use to estimate the next benchmark value before it is published.

For example, suppose a covered neocloud provider changes its H100 rental price after one reading and before the next. A customer could assess that update alongside other observations in its forecasting model. Whether it affects the official index depends on Silicon Data’s provider coverage, input rules and adjustments.

What does Synoptic provide?

Synoptic provides real-time GPU pricing data from 60+ providers, including hyperscalers, neoclouds and marketplaces. Customers use this data to estimate the next GPU price benchmark before it is published.

Request a sample of Synoptic’s GPU pricing data and test it with your own models.

How does this relate to compute futures?

One application of Silicon Data’s benchmarks is CME’s proposed compute futures. These contracts reference the H100 and B200 neocloud rental indexes and settle in money rather than delivering GPU computing time.

CME’s compute futures remain pending regulatory approval, with review extended to January 2027. The benchmark, a provider’s rental quote and a futures contract’s price remain distinct pieces of information.