What this is
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.
Dataset status
The market snapshot and the core surveys refresh independently. Their stamps stay separate everywhere on the site.
Market snapshot
Forward curves, the terminal, and the market rail. Rebuilt by the forward-market pipeline from Kalshi ladder snapshots; this pill flags itself stale past 26 hours.
Live quotes
Machine-fetched quote rows and the observation ledger they feed. Rewritten by the weekly quote refresh, which also bumps the core stamp; flags itself stale past 10 days.
Core data
Manual quote survey, hardware specs, token prices, financials, and buildout claims. Refreshed by hand-survey sweeps; appends “re-survey due” past 60 days.
| Dataset | As of | Coverage | Status |
|---|---|---|---|
| Market-implied curves | Instrument anchor + monthly tenors to Aug ’27 | 64 quality-passed / 6 thin or interpolated | |
| GPU rental quotes | Provider list and marketplace rates | Manual survey 2026-07-06 + 11 automated rows 2026-07-23 | |
| Inference API pricing | Input, output, cache, and batch list rates | Current survey | |
| Hardware, capex & buildout | Specs, filings, guidance, and cluster claims | Core dataset refresh |
Coverage & gaps
What each tracked accelerator actually has behind it — and where the record is thin. Every count derives from the datasets, so this matrix cannot flatter itself.
16 chips quoted · 19 providers · 5 with forward curves · 16 with exact history
| Chip | Rental quotes | Classes quoted | Forward curve | Exact history | Badge | Spec sheet |
|---|---|---|---|---|---|---|
| NVIDIA | ||||||
| V100 SXM2legacy | 1 quote | marketplace | not tracked | shallow — 1 date since 2026-07-06 | no badge | spec sheet available |
| A100 SXM 80GBlegacy | 5 quotes | specialist | to Aug ’27 | shallow — 1 date since 2026-07-06 | badge available | spec sheet available |
| H100 SXM | 14 quotes | hyperscaler + specialist + marketplace | to Aug ’27 | shallow — 4 dates since 2026-07-06 | badge available | spec sheet available |
| H200 SXM | 9 quotes | hyperscaler + specialist + marketplace | to Aug ’27 | shallow — 4 dates since 2026-07-06 | badge available | spec sheet available |
| B200 (HGX) | 10 quotes | hyperscaler + specialist + marketplace | to Aug ’27 | shallow — 4 dates since 2026-07-06 | badge available | spec sheet available |
| B300 (Blackwell Ultra) | 4 quotes | hyperscaler + specialist + marketplace | not tracked | shallow — 2 dates since 2026-07-06 | badge available | spec sheet available |
| Rubin (VR200)announced | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| GeForce RTX 4090 | 3 quotes | specialist + marketplace | not tracked | shallow — 4 dates since 2026-07-06 | badge available | spec sheet available |
| GeForce RTX 5090 | 3 quotes | specialist + marketplace | to Aug ’27 | shallow — 4 dates since 2026-07-06 | badge available | spec sheet available |
| AMD | ||||||
| Instinct MI300X | 4 quotes | hyperscaler + specialist | not tracked | shallow — 1 date since 2026-07-06 | badge available | spec sheet available |
| Instinct MI325X | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| Instinct MI355X | 3 quotes | hyperscaler + specialist | not tracked | shallow — 1 date since 2026-07-06 | no badge | spec sheet available |
| Instinct MI400announced | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| TPU v5e | 1 quote | hyperscaler | not tracked | shallow — 1 date since 2026-07-06 | badge available | spec sheet available |
| TPU v5p | 1 quote | hyperscaler | not tracked | shallow — 1 date since 2026-07-06 | no badge | spec sheet available |
| TPU v6e (Trillium) | 1 quote | hyperscaler | not tracked | shallow — 1 date since 2026-07-06 | badge available | spec sheet available |
| TPU v7 (Ironwood)ramping | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| AWS | ||||||
| Trainium2 | 1 quote | hyperscaler | not tracked | shallow — 1 date since 2026-07-06 | no badge | spec sheet available |
| Trainium3 | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| Microsoft | ||||||
| Maia 200 | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
| Intel | ||||||
| Gaudi 3 | not tracked | — | not tracked | not tracked | no badge | spec sheet available |
Also quoted in the rental survey but outside the spec sheet: GB200 NVL72, L40S — rack-scale configurations and chips not in the accelerator database.
GPU rental prices
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: hyperscalers (AWS, Azure, Google Cloud, Oracle), specialist clouds (Lambda, CoreWeave, Nebius, Crusoe, Together, RunPod, Hyperstack, Voltage Park, Modal, …), and marketplaces (Vast.ai, SF Compute, Salad) whose open bids set the market floor, often on community hardware with fewer guarantees.
Hand-surveyed quotes were fetched from provider pricing pages or Vantage instance listings on 2026-07-06; 11rows with machine-readable primary sources (Vast.ai’s marketplace search API, the feeds behind AWS’s pricing pages, Azure’s Retail Prices API, Vultr’s public plans API) are re-fetched for the reviewed current table by a weekly full-snapshot pipeline, last on 2026-07-23. 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 exact received JSON text, 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.
Forward curves & market-implied prices
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 for B200, H200, H100, and A100 — not a current spot quote. RTX 5090 has no weekly series, so its first point is explicitly the front-month monthly-average event. The rest of the strip uses monthly-average events through Aug ’27. The implied level for a tenor is the strike where the ladder’s yes-probability from bid/ask midpoints crosses 50%: the market’s median expectation, not an executable quote.
Monthly raw medians are transformed with two weighted [1,2,1] smoothing passes from the third monthly tenor onward; the first two monthly values are unchanged. The complete methodology key is kalshi-ladder-median-50-v1+smooth-121x2-after-front-v1.
This snapshot contains 64 quality-passed implied medians, 6 thin-ladder points, and 0 interpolated points. Thin or interpolated points render hollow. “Quality-passed” describes the underlying ladder, not an observed rental price: every plotted value remains reconstructed and non-executable.
Superseded snapshots stay in a dated archive — currently 28 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; they are displayed as legacy and never compared with the current weekly anchors. “Prior snapshot” movement requires the same event ticker and methodology, and reports the actual elapsed time rather than calling the gap 1d.
1 settled month captured — the forward-accuracy scorecard publishes at 3.
Why these markets exist
“Compute right now is where oil was before NYMEX — traded only via OTC deals … the industry will need a similar derivative market.”
01 · Through 2024
OTC
Multi-year cloud commitments and private neocloud deals. Terms are bespoke, prices opaque, and there is no common reference rate.
02 · 2024–25
Indices
Daily GPU-rental benchmarks from Silicon Data and Ornn give compute a printed spot reference.
03 · July 2026
Forward curvesHere
Kalshi event-contract ladders produce a market-implied monthly term structure — the curves tracked here.
04 · Oct 2026
Futures
CME lists Silicon Data H100 and B200 rental-index futures on NYMEX for an Oct 5 first trade, pending CFTC review; ICE with Ornn remains announced.
05 · Next
Options
Options, perpetuals, and structured capacity finance would complete the risk-transfer stack.
Venue monitor
Where compute risk can trade
Status record ·
Kalshi
Weekly and monthly-average GPU rental ladders across B200, H200, H100, A100, and RTX 5090.
- Instrument
- Event contracts → implied curves
- Benchmark
- Ornn hourly index
- Curves launched
CME Group
Silicon Data H100 and B200 Rental Index futures (GPU1/GPU2) on NYMEX: 730 GPU-hours per contract, quoted in USD per GPU-hour, 36 monthly tenors, financially settled.
- Instrument
- Cash-settled compute futures
- Benchmark
- Silicon Data H100 / B200 rental indices
- First trade date
ICE
Planned contracts on transaction-based OCPI benchmarks for H100, H200, B200, and RTX 5090 capacity.
- Instrument
- USD cash-settled GPU futures
- Benchmark
- Ornn Compute Price Index
- Announced
“Live” means listed contracts were available in the archived source record. “Scheduled”, “announced”, and “pending review” are not launch confirmations; status is intentionally frozen to the date above until the source record is refreshed.
Price signals & reference rates
Four kinds of numbers now get called “the price of compute”. They measure different things, revise differently, and only some can be independently checked.
| Signal | Exemplar & record | What it measures | What it hides | Revision policy | Auditability | Our nearest series |
|---|---|---|---|---|---|---|
| Posted list / ask surveys | This site's quote table and its citable class aggregates (e.g. ACT.H100.SPEC-MED@primary-on-demand-v1)Rental-price methodology (surveyed ) | The prices providers publicly ask: on-demand list rates and open marketplace asks, normalized to USD per GPU-hour with each row's billing model retained. | What anyone actually pays — negotiated commitments, private discounts, and whether posted capacity is really available at the posted price. | A provider can change or delete a list price at any time; our dated observations are kept immutable in the price ledger, but the posted price itself carries no finality. | Open — every quote links its source page and observation date, and verified ledger rows pin the raw payload digest to durable evidence. | This is our signal: the ACT.* class aggregates (posted list/ask prices, not transactions). View our nearest series for Posted list / ask surveys → |
| Transaction-VWAP indices | Ornn Compute Price Index (OCPI); Silicon Data H100 / B200 rental indicesICE announcement (announced ) · SER-9785 listing notice (first trade date ) | Volume-weighted prices computed by an index provider from actual rental transactions — a spot reference for what trades, not what is listed. | The trade tape behind the print: constituent transactions, panel composition, and weighting are the provider's proprietary inputs, so a print cannot be independently recomputed. | Set by each provider's own index methodology; we do not track or republish their correction terms here. | Proprietary — index levels are licensed data. This site links the public announcements only and republishes no index values. | None — this site publishes posted prices and market-implied levels, never transaction or settlement data. |
| Exchange settlement prices | CME Group Silicon Data H100/B200 rental-index futures (scheduled — first trade date Oct 5, 2026, pending review); ICE × Ornn GPU futures (announced May 19, 2026, pending review)SER-9785 listing notice (first trade date ) · ICE announcement (announced ) | The price at which cleared futures settle: a regulated, tradable consensus at a defined timestamp, cash-settled against the underlying index. | Everything off-exchange — physical availability and bespoke terms — and, until a first trade prints, everything else too. | A settlement published under an exchange rulebook is final for margining — the opposite end of the spectrum from a list price a provider can edit at will. | Public print, proprietary underlying — settlement values are exchange-published, but they cash-settle to licensed index levels (see the sources). | None — this site publishes posted prices and market-implied levels, never transaction or settlement data. |
| Event-market implied levels | Kalshi compute event ladders — the exact source of this site's forward curvesKalshi compute markets (curves launched ) | Where bid/ask midpoints on yes/no price brackets put the market's median expectation for a future rental price — a probability-implied level, moving while the books trade. | An executable rental price: implied medians are reconstructions from thin, bounded ladders — nobody rents a GPU at the implied level. | Implied levels move continuously until each event resolves; our snapshots are dated, archived immutably, and compared only under an identical methodology key. | Open — the ladders are public order books, and our snapshots retain strike, spread, volume, open-interest, and event-ticker evidence per point. | These are our forward curves: the Kalshi-implied strip on the forward board, labeled with per-point ladder quality. View our nearest series for Event-market implied levels → |
What settles where
| Venue | Reference rate | Signal type | Our nearest comparable |
|---|---|---|---|
| Kalshi | Ornn hourly index | Transaction-VWAP indices | Our forward curves are reconstructed from these exact ladders — the implied strip, not the settlement index. View our nearest comparable for Kalshi → |
| CME Group | Silicon Data H100 / B200 rental indices | Transaction-VWAP indices | Nearest comparable, not the same signal: ACT.H100.SPEC-MED@primary-on-demand-v1 and ACT.B200.SPEC-MED@primary-on-demand-v1 track posted prices for the same chips. View our nearest comparable for CME Group → |
| ICE | Ornn Compute Price Index | Transaction-VWAP indices | Nearest comparable, not the same signal: our ACT.* posted-price aggregates cover the announced contract chips (e.g. ACT.H200.SPEC-MED@primary-on-demand-v1). View our nearest comparable for ICE → |
Every venue above settles to a proprietary transaction index. Our open series sit next to those rates — same chips, same unit — but they are posted-price surveys, never the settlement value.
Read this first
What our numbers are not
- Not transaction prices. We publish posted list/ask prices and market-implied levels; no executed-trade data enters any series.
- Not a settlement benchmark. ACT.* aggregates are versioned survey statistics (posted list/ask prices, not transactions) — nothing here is licensed for, or suitable as, a contract's reference rate.
- Not executable quotes. Forward-curve values are probability-implied medians reconstructed from event ladders; no one will rent to you at those levels.
- Not a republication of third-party indices. Ornn and Silicon Data index values never appear on this site — we record names, dates, and public announcements only.
Comparison conventions
The rules that keep unlike hardware, APIs, and estimates from becoming false equivalences.
Hardware
Dense FLOPS
Every throughput figure is dense tensor throughput. NVIDIA and AMD datasheets headline 2×-higher “with sparsity” figures that assume 2:4 structured-sparse weights. BF16/FP16, FP8, and FP4 stay separate; cost-efficiency charts use dense BF16 unless labeled otherwise.
Inference
Blended token price
The single-number price is (3 × input + output) ÷ 4, a typical 3:1 read/write mix. Batch tiers (~50% off), cached input (~90% off), long-context surcharges, and regions shift real costs. “GPT-4-class” means matching original GPT-4 on MMLU, ≈86.4.
Provenance
Estimates policy
Anything not traceable to a primary source carries an est badge: training compute, cluster power or chip counts, some street prices, and aggregator quotes. Company financials come from filings and earnings releases, with fiscal mapping disclosed alongside the charts.
Evidence & sources
Primary sources anchor the datasets; secondary research is used only where first-party disclosure is incomplete.
Provider pricing
GPU rental quotes
Lambda, RunPod, CoreWeave, Nebius, Together, Hyperstack, Crusoe, Modal, Voltage Park, SF Compute, Vast.ai, AWS, Vantage, and Google Cloud TPU.
Official APIs
Model pricing
Filings & research
Financials and buildout
Earnings releases and SEC filings from NVIDIA, Microsoft, Alphabet, Meta, Amazon, Oracle, and CoreWeave; Epoch AI for training-compute estimates; company announcements and sector reporting for clusters; getdeploying and AIMultiple for market medians.
Evidence & release log
A concise record of the stamps driving the current build.
- Market-structure record refreshed
Venue monitor updated from CME notice SER-9785: Silicon Data H100/B200 rental-index futures on NYMEX are scheduled for a first trade date of , pending CFTC review.
- Market snapshot refreshed
Forward ladders and archived-session comparisons use this independent market timestamp.
- Core survey refreshed
Rental, hardware, inference, financial, and buildout surfaces use the core dataset stamp unless a panel says otherwise.
What counts as evidence?
A primary price page, filing, earnings release, technical datasheet, or market observation is treated as reported evidence. Reconstructions, third-party medians, and incomplete public claims are labeled inferred or est at the point of use.
Update cadence & contact
Core data was last refreshed 2026-07-23. Rental quotes and token prices are refreshed on a regular sweep; financials update each earnings season; hardware and cluster entries as news lands. Spotted an error or a missing provider? data@aicomputetracker.com.
Datasets and open data feeds are licensed CC BY 4.0 — reuse freely with the attribution “AI Compute Tracker (aicomputetracker.com), CC BY 4.0”.
Nothing on this site is investment advice. Trademarks belong to their owners; prices belong to the providers; mistakes belong to us.