How every number here is derived.

ComputeX Labs is a price index and order book for decentralized GPU compute, quoted in USDC and settled on Solana. It does not own hardware. It reads the networks that do, normalises their catalogues onto one set of GPU classes, and republishes the result as a single comparable book.

Two networks are indexed today: Nosana, which runs its market registry on Solana, and Akash, a provider-bid marketplace on Cosmos. Both publish open, unauthenticated APIs. Everything you see is rebuilt from those feeds every 150 seconds.

Nothing is scraped and nothing is estimated from a third-hand aggregator. These four endpoints are the whole data supply:

Nosana publishes two different numbers per market, and picking the wrong one understates the buyer's cost. usd_reward_per_hour is what the host earns after the network fee. The price a client actually pays is denominated in NOS, so it has to be converted at spot:

Akash returns min, max, average, weighted average and median across every live provider bid for a model. ComputeX uses the median. The minimum is a single provider's floor and is not generally available; the median is the rate a buyer can actually expect to clear at.

Availability is reported directly by Akash as leased-versus-total GPUs per model, and is passed through unmodified.

The two networks name the same silicon differently — "NVIDIA A100 80GB Community" against model a100 with ram 80Gi. Both are mapped onto one catalog of GPU classes so a comparison is like-for-like, with memory size used to separate variants that share a model number.

Nosana sells some capacity as whole multi-GPU boxes: an 8×H100 node is one market with one price. Those are divided down to a per-GPU rate so they sit on the same axis as a single card, and the table tags them so you can see which rows are nodes.

Hardware that is not in the catalog is not dropped. It passes through under a derived label, so a GPU that appears on either network tomorrow shows up in the index without a code change.

Low, median and high are computed across every market quoting that class at that moment, on both networks. The Nosana and Akash columns are each network's own cheapest live offer — the one a buyer would actually take.

The gap is the difference between those two cheapest offers as a percentage of the more expensive one. It only exists for classes both networks quote; where one is silent, the column is empty rather than filled with a substitute.

Price per compute-unit divides the hourly rate by a relative throughput index. That index is the one number on this site that is a judgement rather than a measurement, and it is labelled as an index everywhere it appears.

No DePIN network publishes a historical price archive, so there is nothing to backfill from. The charts are built from ComputeX's own observations: every refresh appends one point to a rolling series held in Netlify Blobs.

A freshly deployed instance therefore has a short chart, and it says so instead of drawing an invented trend. The series fills in from the first deploy onward.

Jobs are not provisioned. Posting to the book records a spec at the price it was quoted and makes it publicly visible; no hardware is reserved and nothing is charged.

There is no regional pricing and no term discount. Neither feed publishes either, so applying one would make the quote a guess. A region selection is recorded with the job as a routing preference and explicitly does not move the price; a longer term multiplies the same hourly rate.

Deposits are real mainnet USDC transfers, but they are only enabled once a settlement address is configured. Until then the deposit panel is inert and says so.

Quotes are denominated in USDC. Funding a balance is an ordinary SPL transfer signed in your browser and broadcast by our own endpoint, so no RPC credential is ever exposed to the client and no public RPC has to accept a browser origin.

Solana is the rail for the obvious reason: per-job settlement has to cost a fraction of the job. It is also, at present, where the cheaper GPU network happens to live.

Upstream APIs are public and unauthenticated, so ComputeX calls each one once per cache window for the whole site rather than once per visitor. Market data has a 150-second TTL; Solana telemetry has 20 seconds.

If a source fails, the previous snapshot is served and flagged stale rather than showing an error. If both price sources fail at once, the book refuses to serve rather than showing a half-empty market.