Hourly USD rental rates for the GPUs the AI economy runs on — measured across independent venues, aggregated and manipulation-resistant, then printed on-chain every ten minutes. Free to read, free to verify. No key.
| Index | Last print | 24h | Sources | Grade | Coverage 24h | Status |
|---|---|---|---|---|---|---|
| loading live data… | ||||||
Coverage counts ten-minute prints published against prints expected over the trailing day. A missing print is never interpolated: when the pipeline cannot compute an honest value, the slot stays empty. Settlement grade requires five independent sources — below that, an index publishes for display only.
Each index observes public rental markets and reduces them to one number per tier, every ten minutes, under the following clauses. Parameters marked [param] are governed behind a timelock — 1 hour on this testnet deployment; 48 hours is the guarded-launch target, not today's value.
Offers and executed-lease prices are drawn from each venue's public API. An executed price carries a 2× weight over a listed offer — a filled trade is better evidence than an ask.
Observations are reduced to a versioned hardware specification — for the flagship, H100 SXM 80 GB, on-demand, cluster-capable. Other form factors are scaled to that specification, and quotes that are implausible or stale are discarded.
Observations far from the cross-source consensus are dropped. Wash prints, fat fingers and decorative asks die here.
Observations weight by cluster size and kind, and no venue may carry more than a capped share of the total weight.
A signer quorum posts each value on-chain, bounded by a per-mark deviation breaker and a dispute window. Settlement never reads a single print: it takes a 72-hour TWAP of the published series, subject to minimum coverage.
Below 4 independent sources an index goes STALE — visibly, on-chain. It is never estimated or interpolated. A missing print means we did not have an honest number, and the record keeps the hole. Every other clause depends on this one.
In full: a source that fails a single fetch keeps its last good observation for up to 3 hours, and that carried reading counts toward the 4-source floor — one flaky API must not flap a whole tier STALE. Past 3 hours the source drops out for real and the floor applies with no carried readings at all. Carrying is bounded, per-source and recorded; what is never carried is the published number itself.
Independent venues with machine-readable public pricing. Live counts update on the sheet above; this ledger records which venues carry which silicon.
| Venue | Kind | Carries |
|---|---|---|
| Vast.ai | listed offers | H100 · 4090 · 5090 · A100 |
| RunPod | listed prices | H100 · H200 · A100 · L40S · B200 |
| Akash | listed price stats | H100 · A100 |
| io.net | listed tier prices | H100 · H200 · A100 |
| Cudo Compute | listed machine prices | H100 · A100 · L40S |
| Vultr | cloud plan sheet | L40S |
| DataCrunch | instance price sheet | B200 · H200 · H100 · A100 · L40S |
| Clore.ai | P2P marketplace asks | 5090 · 4090 |
| Shadeform | multi-cloud price sheet | B200 · 5090 |
| OCI | public price list | B200 · H200 |
A venue is admitted when it publishes machine-readable pricing, survives the consensus band against the existing panel, and is independent of existing sources. One broker (Shadeform) is admitted, for supply no direct adapter reaches; every cloud we already poll ourselves is stripped from its rows before they enter the panel, so no operator is counted twice. Non-USD venues wait for FX normalization.
No key, no auth, CORS enabled. Please cache for sixty seconds; if you build something real on this data, tell us — the index is the product.
TESSELEXGPU derivatives venuemarkUsd), staleness
flag, source count and 24-hour statistics. The composite and the H200−H100 basis
spread ride along.?since= / ?until= (unix seconds),
?index=, ?limit=, and ?interval=1d for daily
closes (last print per index per UTC day). The response carries a meta
block stating the interval, the retention and the default window, so a consumer never
has to infer what it received — {ts, index, usd,
sources} — timestamps on exact ten-minute boundaries, one print per boundary,
gaps preserved honestly.