Methodology & sources

AI Compute Tracker (aicomputetracker.com) tracks the economics of AI compute: what GPUs rent for, what the chips can do, what a million tokens costs, and how much capital is pouring into the buildout. It exists because these numbers are scattered across provider pages, earnings decks, and research posts — and because most public comparisons quietly mix incompatible units.

16 chips quoted · 16 providers · 5 with forward curves · 16 with exact history

Rental prices are normalized to US$ per GPU-hour (per chip-hour for TPUs/Trainium). Every row retains its billing model: on-demand provider list, capacity block, serverless, variable marketplace ask, or secondary snapshot. Multi-GPU instance prices are divided per accelerator; unlike billing models are never joined into one history series.

Providers are grouped into three classes, defined in the glossary: hyperscalers (Oracle OCI, AWS, Google Cloud and Azure), specialist clouds (RunPod, CoreWeave, Crusoe, Hyperstack, Nebius, Lambda, Together AI and Voltage Park), and marketplaces (Vast.ai and Salad), whose open asks set the marketplace floor, often on community hardware with fewer guarantees.

Hand-surveyed quotes were fetched from provider pricing pages or Vantage instance listings on 2026-10-14; 11 rows with machine-readable primary sources (Vast.ai’s marketplace search API, the feeds behind AWS’s pricing pages and Azure’s Retail Prices API) are re-fetched for the reviewed current table by a weekly full-snapshot pipeline, last on 2026-10-14. Healthy sources refresh independently. When a Vast.ai book lists too few offers to use, the chip takes its most recent marketplace-tape capture from the last 7 days, dated with that capture's own time, if it is newer than the row it replaces; otherwise, and whenever a source is unavailable, the pair retains its original price and observation date. A pair its source explicitly retires (a Vultr plan that is not deployable on demand or is no longer listed) is removed instead of retained. This does not renew the manual survey date. Values marked est come from same-week aggregator snapshots. Historical series are indicative quarterly midpoints reconstructed from list prices, launch announcements, and archived snapshots — treat them as ±15%.

The immutable price ledger samples marketplace books every four hours and machine-readable provider lists daily. Each successful observation retains its received JSON text (a Vast.ai book is now kept only as an allowlist projection without host identifiers, plus the SHA-256 of the full response), payload digest, source identity, methodology, and observation time. Partial failures become visible gaps; hand-surveyed rows are added only on real survey dates and are never copied forward. The evidence ledger renders this chain end-to-end — snapshot IDs, checksums, verified-row counts, and the pinned raw-evidence references.

The forward-curve board is built from Kalshi event-contract ladders on GPU rental prices referenced to Ornn’s hourly index. The first point is a future exact-time weekly event when a usable weekly ladder exists — not a current spot quote. If the weekly book is unavailable, the first point is explicitly the front-month monthly-average event. Its instrument and event identity are recorded per chip; incompatible anchors are never used to calculate movement. The rest of the strip uses monthly-average events through Aug ’27. Each implied level is the strike where a ladder’s yes-probability crosses 50%, and the resting bids and asks bracket it: a bid at one strike is a floor on every lower strike’s probability and an ask a ceiling on every higher one, so the bid-only and ask-only envelopes each cross 50% somewhere. The bracket is those two crossings, each interpolated linearly between the listed strikes around it; strictly, the quotes bound the median only to the outer of those strikes. A side no quote bounds is open. That is the market’s median expectation, not an executable quote — labeled Kalshi-implied wherever it headlines. Each Kalshi compute contract settles on Ornn’s executed-trade index — not comparable with posted list prices.

How each strike is read. The level reads only usable quotes: a strike with a bid above 0¢ and an ask below 100¢ counts as two-sided, but is read at its midpoint only when no wider than 20 points with size on both sides, and a strike without such a quote only at a last trade with 24-hour volume. A weighted isotonic fit (pool-adjacent violators, each quote weighted by its tightness) keeps them monotone — it can raise a level as well as lower it, unlike a running minimum — and nothing is smoothed: each monthly value is its own ladder’s estimate. With fewer than two usable quotes, or a fit that never crosses 50% within its own strikes or crosses it outside the bracket, no quote sets the level: a closed bracket gives its midpoint, and an open one only a placeholder (its closed side, or the middle of the listed strikes when neither side is closed) that the site never shows as a price — it shows the bracket. In this snapshot 7 of 55 ladders are read from their fit of 33 usable quotes, and 34 brackets are open on at least one side. A point passes — it is identified — only with both sides of its bid/ask bracket inside the listed strikes, a bracket no wider than 20% of its median, and at least 2 two-sided strikes, 1¢/99¢ placeholders excluded, quoted between 20% and 80%; 7 of 60 points pass this snapshot, and 84 of the snapshot’s 213 two-sided strikes are 1¢/99¢ placeholders. Only identified points draw a line and a band, and only they feed moves, carry, spreads and curve shapes.

This snapshot contains 7 quality-passed, 53 thin-ladder and 0 interpolated points. The lines and bid/ask bands join only quality-passed points; gaps are hollow at their level, unplotted without one, and no headline reads a move, carry, spread or curve shape from a gap. “Quality-passed” describes the underlying ladder, not an observed rental price: every plotted value remains reconstructed and non-executable. The complete methodology key is kalshi-ladder-median-50-v2+unsmoothed-v1.

Superseded snapshots stay in a dated archive — currently 34 sessions — that powers the curve time machine. Each snapshot persists its method and exact first-point instrument identity. The first 12 legacy sessions use a different monthly methodology. 13 sessions first published under an earlier estimator were re-derived under the current one on Oct 15, 2026 from the raw ladders retained when they were captured; each records the re-derivation. 7 sessions without retained ladders keep the estimator they were published under, labeled as such. Every session is displayed with its own methodology and never compared across one. “Prior snapshot” movement requires the same event ticker and methodology, and reports the actual elapsed time rather than calling the gap 1d.

Every tracked venue, its status and each dated market-structure event have a page of their own: GPU compute futures & markets.

1. 2026-10-14
   Market-structure record refreshed

   Latest event: Kalshi — No weekly ladder open; monthly ladders open (2026-10-11). This site’s archived Oct 11 session (anchor per chip).

2. OCT 16 2026 · 11:51 UTC
   Market snapshot refreshed

   Forward ladders and archived-session comparisons use this independent market timestamp.

3. 2026-07-23
   Last complete core survey

   Later domain reviews are tracked separately. The latest is 2026-10-15; retained inputs keep their own dates.

Each domain’s review target and next due date are in the dataset-status table; a partial review never moves a due date. The complete survey is due every 60 days (last: 2026-07-23, overdue since 2026-09-21; latest domain review: 2026-10-15). Spotted an error? Email data@aicomputetracker.com or use “Report a wrong price” on a quote row; material corrections are logged on the corrections page.

The site’s own text and data compilations are licensed CC BY 4.0; reuse them with the attribution “AI Compute Tracker (aicomputetracker.com), CC BY 4.0; includes data from Epoch AI (epoch.ai), CC BY 4.0”. Third-party source data — such as Kalshi market data, Vast.ai marketplace data, and providers’ list prices — remains subject to its sources’ terms.

Nothing on this site is investment advice. Trademarks belong to their owners; prices belong to the providers; mistakes belong to us. The terms of use say what you can rely on.