Why we built Compute Signal

Compute Desk publishes the daily, blended GPU price indexes that compute futures on Nodal Exchange and Architect’s AIX will financially settle against. To build those benchmarks, our team collects rental listings, settled transactions, hardware prices and token prices every business day, and cleans them against one set of definitions.

Compute Signal opens that work to the teams who buy, sell and finance compute. It runs on the same data, the same GPU catalog and the same quality checks as the indexes.

One supply chain, priced in pieces

The AI supply chain prices compute three times. An operator buys the hardware. A neocloud or hyperscaler rents it out by the GPU-hour. An inference company sells those hours on as tokens.

Each price pulls on the others. Token prices cap what an inference company can pay for a GPU-hour. Rental prices set how fast a hardware purchase pays back. Hardware prices put a floor under a rental offer. When one layer moves, a buyer or lender wants to know who moved first and whether the other layers will follow.

Most teams answer that question from separate sources, each with its own GPU names, regions and dates. Even within one layer, the numbers mislead. Many price trackers average a hyperscaler list price, a neocloud ask and a settled deal into a single “H100 price” that no deal cleared at. A median of three quotes looks the same on screen as a median of three hundred, and nothing tells the reader which one they hold.

The whole market on one set of definitions

Compute Signal puts the four layers of the supply chain side by side, on one GPU catalog, one set of region and cohort groupings and one daily cadence. An H100 in the token view is the same H100 in the hardware view.

Rental prices. $/GPU-hr by GPU, region, contract type and term, updated daily, with availability alongside. Neocloud and hyperscaler cohorts stay separate.

Transacted prices. Where deals cleared, as anonymized aggregates with medians and term curves. Signal never shows a single traceable deal.

Hardware prices. What the metal costs to own, by GPU family, condition and region. Signal works out buy versus rent across commitment length and hours of use per day.

Token economics. LLM API pricing, throughput against price, and daily token volumes by model as a proxy for where demand is going.

Published indexes. Hopper US and Blackwell US benchmarks, each as a Blended Transaction Index and an Informational Index of listed neocloud prices. Each index has a factsheet in the app and publishes every business day at the 16:30 New York close.

We hold every layer to the rules we apply to the indexes. Signal draws a hyperscaler list price, a neocloud ask, a private offer and a settled price as separate populations on one axis, so a reader sees where the market clusters and where deals cleared. The one published series that blends them says so on its label. Signal prints a median only when enough observations sit behind it, and every series shows its as-of date and how many vendors contribute to it.

Ask for the chart you need

Compute Agent turns a question into a chart, drawn from the same data as the rest of Signal. Ask for H100 availability against listed prices, or the published index against neocloud asks. The model picks the series, our server fetches and fills every number, and the chart engine that draws the rest of the product draws the answer. Nobody has to wait for us to publish the comparison they want.

Who we built it for

On the physical side, neocloud and hyperscaler pricing teams use Signal to reprice an offer against their cohort. AI labs and inference companies grade a quote before they sign it. Operators decide whether to rent the hours or buy the hardware.

On the financial side, lenders mark GPU-backed collateral every reporting period. Funds and analysts entering compute markets need a dated mark they can cite. Risk teams follow the benchmarks that compute futures will settle against.

Teams that serve tokens sit on both sides. They earn in tokens and pay in GPU-hours, and Signal puts both series on one screen.

Compute Signal is live at app.compute-desk.com. For the futures that will settle on our indexes, read the announcements from Nodal Exchange and Architect.