Cloud GPU Price Index — factsheet
As of · method 1.33.64 · specification 1.3
The median publicly listed on-demand rental price of one NVIDIA H100 SXM GPU-hour, in US dollars, across the cloud providers GPU Economy reads every hour — with every quote behind every value published, so anyone can recompute it.
| Week of | USD / GPU-hour | Vintage, coverage |
|---|---|---|
| November 2, 2026 | $3.25 | provisional · 39 quotes, 37 providers · receipt |
| October 26, 2026 | $3.39 | final · 40 quotes, 38 providers · receipt |
| October 19, 2026 | $3.39 | final · 38 quotes, 36 providers · receipt |
| October 12, 2026 | $3.53 | final · 36 quotes, 34 providers · receipt |
November 2026 panel (31 pairs, CSV): Atlas Cloud, Amazon Web Services, Microsoft Azure, Bitdeer AI, Canopy Wave, Civo, CoreWeave, Crusoe, Cudo Compute, Verda, Denvr Dataworks, Google Cloud, Geodd, GPU.ai, HexGrid, Hyperbolic, Hyperstack, Jarvislabs, Lambda, Massed Compute, Modal, Nebius, Neysa, Oracle Cloud (OCI), Paperspace, RunPod, Taiga Cloud (Northern Data Group), Thunder Compute, Together AI.
Concentration, week of November 2, 2026: the largest provider gives 2 of 39 quotes (Paperspace); HHI 283.
How it is built
- Sources: each provider's own public pricing page or public API, read every hour by our parser. No provider submits a figure; nothing is entered by hand.
- Inclusion: on-demand H100 SXM rates listed in USD, one per provider and region, confirmed within 48 hours; marketplace floors excluded; published only with at least 8 providers.
- Three windows, one rule: the index at every hourly scrape; the week, its last hourly point; the month, the average of every hour over a panel fixed before the month begins (November 2026: 31 pairs from 29 providers).
- Frozen: an hourly point is written once, when its scrape is stored; a closed week and a closed month are never recomputed, not for a late source and not for a later rule.
Why it can be audited
- Every value has a receipt: the quotes behind it, each with its source URL and the time it was read — weekly receipts, hourly receipts, monthly panels. Commands that recompute a week, an hour and a month: /price-index/reproduce.
- No discretion: the panel and the monthly value are computed by the database from stored readings, by published rules; the rules are versioned and every change is dated in the method changelog.
- Archived: DOI 10.5281/zenodo.22706541. Own history since September 7, 2026 (59 days); a separately named series reconstructed from a public price catalog goes back to January 2025.
- Listed rates, not transactions: it measures what providers publicly ask, which anyone can check on the linked page — not negotiated or transacted prices, and not availability.
Use
Free under CC BY 4.0 for reading, citing, research and products: price-index.json, hourly CSV, monthly CSV, panel. Settling a contract, option, token or other instrument on it is free too; a settlement licence adds binding terms, the right to name the index as the reference, signed values and custom indices: /data#settlement. Full rules: specification. Contact: data@gpueconomy.com.
Disclosures
GPU Economy sells no GPU capacity, issues and settles no contract on the index, and holds no position in any market it measures (publisher statement). It holds referral links with Massed Compute, RunPod, Thunder Compute, providers in the current panel: a referral changes no price, inclusion or order, and panel membership is decided by the rule above alone (link policy). Since October 1, 2026 it passes reserved-quote requests to providers at a reader’s request and may be paid a fee by a provider when one becomes a contract; requests are routed by listed rates, and a fee changes no rate, panel or value (disclosure).
Data updated: