Cloud GPU Price Index · week of October 19, 2026 · provisional
Cloud GPU Price Index
$3.53per H100 GPU-hour · the median of 36 listed on-demand rates from 34 providers
Listed rates, not availability. Read . About this figure ↓ · Receipts and how to reproduce it · Renting H100s for a year or more? Reserved quotes ↓
- Base-100 index105.0week of September 7, 2026 = 100
- Middle half$2.99–$4.3925th to 75th percentile of the same rates
- Specialist GPU clouds$3.2332 quotes · 30 providers
- Hyperscalers$9.904 quotes · +207% vs specialists
To quote: in the week of October 12, 2026 (final), the Cloud GPU Price Index stood at $3.53 per H100 GPU-hour, the median of 36 listed on-demand rates from 34 providers — GPU Economy, gpueconomy.com/price-index. A final week never changes; the current week is provisional.
On this page: hour, week and month · about this figure · weekly series · method · spot and commitments · monthly average · every hour · cite · the providers behind it, compared: /compare
One index, several windows
| Window | USD / GPU-hour | Vintage | Coverage, as of |
|---|---|---|---|
| This hour | $3.53 | written once, never rewritten | 36 quotes · 34 providers · 2026-10-23 21:40 UTC · receipt |
| This week, October 19 – October 25 | $3.53 | provisional | 36 quotes · 34 providers · receipt |
| Last final week, October 12 – October 18 | $3.53 | final, never recomputed | 36 quotes · 34 providers · receipt |
| September 2026, monthly average | $3.883267 | final, frozen 2026-10-01 00:42 UTC | 589 points over 720 hours · open panel, before the panel rule |
| October 2026, monthly average so far | $4.12257 | provisional | 555 points over 550 hours · fixed panel, 17 pairs from 15 providers |
October 2026’s average is taken over a panel fixed before the month began: 17 pairs from 15 providers. 19 of this week’s 36 quotes are not in it — they were read in fewer than 90% of the hourly points of the two weeks the panel was decided on, and can enter next month’s panel. 18 of those 19 list below the month’s $4.12257, which is why the month runs above this week’s $3.53. Rules: specification.
About this figure
The median publicly listed on-demand price of one NVIDIA H100 GPU-hour, across the providers tracked here. One number for “what does an H100-hour cost right now”.
For the week of , the Cloud GPU Price Index stands at $3.53 per H100 GPU-hour, the median of 36 publicly listed on-demand rates from 34 providers, read . That week is still running, so this value is provisional: it is recomputed on every hourly scrape and settles when the week closes. Every rate is one GPU for one hour - a node price divided by its card count (why, with an example).
How this series was made. The series starts with the week of September 7, 2026, the first week published. Every weekly value is computed from reads made in that same week - the newest read of each provider and region, counted only if it was confirmed within 48 hours of the week's last scrape - and published with its receipt, the list of those quotes. No value was entered by hand, reconstructed after the fact or carried over from an earlier week. Closed months are published as they were first published and never recomputed. The reconstructed history lower on this page is not part of it: price-index-history.csv is a separate series under its own name, rebuilt after the fact from a public price catalog, whose repository is named at https://gpueconomy.com/methodology#second-hand-history, for the weeks before September 2026. It is second-hand, and it is never merged into or back-filled with this one. More on /methodology.
80.2 index, base = 100
Listed rates against January 2025. 20% below the week of January 6, 2025, measured only on providers present week to week. Reconstructed series, through the week of October 12, 2026.
By operator class
| Operator class | Median / GPU-hour | Quotes | Providers |
|---|---|---|---|
| Hyperscalers — AWS, Azure, Google Cloud, Oracle | $9.90 | 4 | 4 |
| Specialist GPU clouds | $3.23 | 32 | 30 |
| Marketplaces | not in the index | — | — |
Hyperscalers list the H100 at a median 207% above specialist GPU clouds this week: $9.90 across 4 hyperscalers against $3.23 across 30 specialist clouds. Marketplaces have no value by construction, not for want of coverage: a marketplace floor moves with supply and is not a rate card, so it never enters the index (on 2026-09-11 one Vast.ai window read $53.34 for an RTX 5090 GPU-hour). The class of every provider is in providers.json; each week's cut is by_operator_class in price-index.json and four columns in price-index.csv.
| Figure | Value | As of | Providers | Source |
|---|---|---|---|---|
| Current median | $3.53 | week of October 19, 2026 (provisional) | 34 | our hourly scrape |
| Last month-end median | $4.25 | week of October 12, 2026 | 6 | reconstructed (public price catalog) |
Do not read the gap between those two as a price move. They come from different panels (34 providers against 6) and different sources, so the difference is mostly who is counted, not what anyone charges. The change that is a change is the base-100 index, which only moves when a provider changes a listed price: from the end of September 2026 to the week of October 12, 2026 it went from 78.4 to 80.2 (+2.3%). Our own series starts at 100 in the week of September 7, 2026; it will carry its own month-over-month once it has two closed months.
Two numbers per week: the dollar median and a base-100 index
The dollar median answers “what does an H100-hour cost this week”. It moves when the panel changes as well as when prices do — a cheaper provider joining lowers it without any rate card moving. The second column, index (base = 100), is built to move only on price changes: Chain-linked index, base week = 100: each week's link is the geometric mean of week-over-week price ratios over the (provider, region) pairs quoted in both weeks; the index is the cumulative product. A provider or region joining or leaving contributes no ratio and cannot move it. Cite the median for a level and the index for a change.
Weekly series
One point per ISO week (Monday start). Download: JSON (current value, constituents with source links, series) ·CSV (series only). Every week has a receipt — the quotes that entered its value, each with its source and read time — and the command that turns it back into the number: reproduce the index. On your own page: embed this week’s value (/embed/price-index, an iframe, no script).
| Week of | Median USD / GPU-hour | Middle half (p25–p75) | Quotes | Providers | Index (base = 100) | Weighting check | Vintage | Receipt |
|---|---|---|---|---|---|---|---|---|
| $3.53 | $2.99–$4.39 | 36 | 34 | 105.0 | $3.42 | provisional | 2026-W43.csv | |
| $3.53 | $2.99–$4.39 | 36 | 34 | 105.0 | $3.42 | final | 2026-W42.csv | |
| $3.49 | $2.99–$4.39 | 33 | 31 | 104.2 | $3.34 | final | 2026-W41.csv | |
| $3.49 | $3.00–$4.40 | 31 | 29 | 102.6 | $3.25 | final | 2026-W40.csv | |
| $3.70 | $2.99–$4.41 | 26 | 24 | 100.4 | $3.70 | final | 2026-W39.csv | |
| $3.90 | $2.89–$5.95 | 21 | 19 | 100.1 | $3.90 | final | 2026-W38.csv | |
| $3.92 | $3.12–$6.00 | 20 | 18 | 100.0 | $3.92 | final | 2026-W37.csv |
Method
Median, not mean: a single provider’s outlier listing should not move the index. One quote per provider and region: the newest listed rate of each pair in the week, so a provider that lists several regions enters the median once per region (this week Paperspace gives 2 of the 36 quotes). As a check, the same quotes taken as one median per provider and then the median across providers give $3.42, $0.11 below the index. The index keeps its definition; the check shows what the weighting rule costs, and a gap opening between the two is the sign that one provider’s regions, not prices, moved the median. On-demand rates only — spot capacity is too volatile to scrape meaningfully and reserved pricing answers a different question. Node-priced offerings are divided by their GPU count before entering the median. Committed-use rates that are published — Google Cloud’s and Azure’s one- and three-year prices — are listed on each card’s page but stay out of the index; negotiated contract rates are not public and are not represented anywhere.
Known limitations
Where this figure is weaker than it looks, named here rather than left to be found. None of it is hidden in the data: every week's quotes are published as a receipt, and the ones kept out carry the rule that excluded them (reproduce the index).
- Listed rates, not paid rates. Everything here is a price a provider publishes. What a buyer actually pays after a negotiated contract, a committed discount or credits is not public and is not in this number; the committed rates that are published are measured separately (spot and commitments against on-demand).
- A provider with several regions votes several times. The median is taken over one quote per provider and region; this week the largest contributor gives 2 of the 36 quotes (concentration index 309). The same quotes taken as one median per provider give $3.42, $0.11 below the index — the check printed in the table above and in price-index.json.
- Two markets in one median. 4 hyperscalers list a median $9.90 and 30 specialist clouds $3.23 — a gap of 207%. A single median across both sits between two groups that sell different things; read it with the split, which every week carries as
by_operator_class. - The panel is who publishes, not who sells most. A provider enters the week its listing is first read, with no weight for its size: a large cloud and a small one count the same, and capacity, utilisation and how much is actually rented at a listed rate are not something we can see.
- One card, one product class. H100 SXM on-demand, in USD. Marketplace floors, non-USD listings and spot are tracked on the site and kept out of this figure (method), so the index says nothing about the cheapest hour available anywhere.
At a glance
| H100 SXM, this week | $3.53 / GPU-hour | Median of 36 listed on-demand rates from 34 providers (provisional) |
|---|---|---|
| Weighting check | $3.42 / GPU-hour | The same quotes as the median of each provider’s own median, $0.11 below the index; a check on the weighting rule, not a second index |
| Hyperscaler premium | +207% | $9.90 across 4 hyperscalers against $3.23 across 30 specialist clouds |
| Spread to cheapest | −49% | $1.79 on Cudo Compute (no-kristiansand-1) against the median |
| Spread, cheapest to dearest | 518% | $11.06 on Microsoft Azure is that much above the cheapest listed rate |
What the numbers say
- The median publicly listed on-demand rate for an NVIDIA H100 SXM is $3.53 per GPU-hour in the week of October 19, 2026, across 34 providers.
- Specialist GPU clouds list the H100 at a median $3.23 per GPU-hour, 67% below the $9.90 median of AWS, Azure, Google Cloud and Oracle.
- The spread within one card is wide: $1.79 to $11.06 per H100 GPU-hour, a factor of 6.2, for the same listed on-demand product class. Tiers and scopes are named in the table below.
- How much reading is behind this: 745,042 observations since , across 100 providers and 80 GPU models, in 1,147 hourly runs. One observation is one read of one listed rate - the same rate read every hour counts every hour - and each links to the page it was read from. It is not a deal, not a card free to book, and no figure here is computed from the count.
- What these figures do not measure: negotiated or committed-use contracts, spot capacity, marketplace floors, and quotes listed in a currency other than USD. Each of those is tracked on the site but kept out of the index. Nor availability: a quote count counts rate cards, one per provider and region, not GPUs free to book.
- What it does include: community-cloud and serverless list prices (today RunPod Community Cloud and Modal). Each is a fixed hourly rate the provider publishes for one H100, not a marketplace floor, so the index takes it; the card median elsewhere on the site leaves community capacity out, because it ranks rate cards you can hold. Leaving the one community quote out would have moved each published week by about $0.025. Stated on September 27, 2026; no rule changed.
Spot and commitments against on-demand
As of , against the same provider’s own on-demand rate, spot rates run a median 50% lower (58 rate pairs at 11 providers), 1-year commitments run a median 15% lower (38 rate pairs at 7 providers) and 3-year commitments run a median 30% lower (22 rate pairs at 5 providers).
| Market | Median vs on-demand | Range | Rate pairs | Providers |
|---|---|---|---|---|
| Spot | −50% | −82% to −10% | 58 |  Microsoft Azure,  CoreWeave,  Verda,  Daytona, Elice Cloud,  Hyperstack,  Lyceum,  MimicPC,  Paperspace,  Vultr,  Yandex Cloud |
| 1-year commitment | −15% | −45% to −1% | 38 |  Atlantic.Net, Microsoft Azure,  Bitdeer AI,  Civo,  Google Cloud, Paperspace,  Seeweb |
| 3-year commitment | −30% | −63% to −6% | 22 | Atlantic.Net, Microsoft Azure, Bitdeer AI, Civo, Google Cloud |
A discount is not a saving by itself: a commitment is paid whether the GPU runs or not, and spot capacity can be reclaimed mid-run. None of these rates enters the Price Index, which takes on-demand rates only. Every committed rate, card by card, is on reserved pricing.
Renting the H100 SXM for a year or more?
The cheapest committed H100 SXM rate each provider publishes, by term. A commitment is paid whether the GPU runs or not.
| Term | Per GPU-hour | Provider | vs its on-demand |
|---|---|---|---|
| 1-year | $1.80 €1.61 listed · ECB 1.1193 on October 23, 2026 read | Seeweb cheapest of 6 providers | −15% ($2.12) |
| 3-year | $1.79 read | Bitdeer AI cheapest of 4 providers | −24% ($2.36) |
| Other commitment commit | $2.72 read | Hyperstack cheapest of 2 providers | — |
Across every card, a 1-year rate runs a median −15% against the same provider’s on-demand rate (7 providers); a 3-year rate −30% (5 providers). Every committed rate, card by card.
Get quotes for the H100 SXM. A person at GPU Economy reads your request and passes it to up to three providers that list a rate for the card, the cheapest listed first; the quotes come from them, directly to you.
We may be paid a fee if a request becomes a contract; who receives it follows the listed rates, never a fee, and it changes no published rate or the Price Index (how, privacy).
The monthly average
The arithmetic mean over a calendar month (UTC) of the hourly index, each hour taken as the median of the month’s fixed panel at that scrape (hours with fewer than 8 panel providers left out); published when such hours cover at least 75% of the month, frozen at the first scrape of the next month and never recomputed. A month’s panel is every (provider, region) present in at least 90% of the published hourly points of the 14 days before the 24th of the previous month, 00:00 UTC, decided by the database at the first scrape after that instant and never changed; a provider that qualifies later enters the next month. A closed month is a statement about that month and is never revised; the running month is provisional and moves every hour. September 2026 predates the panel rule and is averaged over every hourly point, marked open. How it may be relied on: specification; on one page: factsheet. File: price-index-monthly.csv.
| Month | USD / GPU-hour | Vintage | Hourly points over hours | Panel |
|---|---|---|---|---|
| October 2026 | $4.12257 | provisional | 555 over 550 (100%) | fixed, 17 pairs |
| September 2026 | $3.883267 | final, frozen 2026-10-01 00:42 UTC | 589 over 720 (82%) | open (before the panel rule) |
The index every hour
At every hourly scrape: the median of the newest USD on-demand H100 SXM listed rate of each (provider, region), read within the 48 hours before that scrape, marketplace floors excluded; published when at least 8 providers entered, withheld (no value) otherwise. The same rule as the weekly index, applied at each scrape instead of at the end of the week. A week's value above is its last hourly point, rounded to the cent. Each point is computed once, by the database, when its scrape is stored, together with the quotes it was computed from, and never rewritten. The newest: $3.53 per GPU-hour at 2026-10-23 21:40 UTC, 36 quotes from 34 providers; 1146 points since 2026-09-06 18:45 UTC.
Files: price-index-hourly.csv (every point) and one receipt per week, every quote of every point with its read time and source — this week’s is 2026-W43.csv. How the series may be relied on: specification.
Noted on October 1, 2026: the ten points from 00:42 to 09:39 UTC that day read Nebius’s H100 at $3.85, from the column its page had just replaced - the page listed $4.50 “effective October 1, 2026” beside it, and our parser read the old column until 10:39 UTC. The points are not rewritten, as no point is; at the next point the open index moved from $3.5585 to $3.627, from the new Nebius figure and one new quote. October’s monthly average was not moved: on its fixed panel the median held at $3.99 either way.
| From the scrape of | USD / GPU-hour | Before | Quotes | Providers | What moved it (+ entered, − left, rate changed) |
|---|---|---|---|---|---|
| 2026-10-15 01:44 UTC | $3.53 | $3.525 | 36 | 34 | GPU.ai $3.56 → $3.86 |
| 2026-10-15 00:41 UTC | $3.525 | $3.53 | 36 | 34 | GPU.ai $3.72 → $3.56 |
| 2026-10-14 22:41 UTC | $3.53 | $3.49 | 36 | 34 | − TensorPool $1.99; GPU.ai $3.99 → $3.76 |
| 2026-10-14 18:40 UTC | $3.49 | $3.25 | 37 | 35 | GPU.ai $1.93 → $3.99 |
| 2026-10-14 17:40 UTC | $3.25 | $3.49 | 37 | 35 | GPU.ai $3.99 → $1.93 |
| 2026-10-14 08:40 UTC | $3.49 | $3.25 | 37 | 35 | Verda $4.087 → $4.1275; GPU.ai $3.12 → $3.99 |
| 2026-10-14 07:39 UTC | $3.25 | $3.49 | 37 | 35 | GPU.ai $3.68 → $3.12 |
| 2026-10-13 18:40 UTC | $3.49 | $3.25 | 37 | 35 | GPU.ai $1.75 → $6.01 |
| 2026-10-13 17:40 UTC | $3.25 | $3.49 | 37 | 35 | GPU.ai $3.76 → $1.75 |
| 2026-10-13 16:40 UTC | $3.49 | $3.37 | 37 | 35 | + Barrack AI $3.59; GPU.ai $3.99 → $3.76 |
| 2026-10-13 10:41 UTC | $3.37 | $3.49 | 36 | 34 | + HPC-AI.COM $2.20; GPU.ai $3.99 → $3.84 |
| 2026-10-13 08:40 UTC | $3.49 | $3.25 | 35 | 33 | GPU.ai $2.40 → $3.99 |
Another H100 index printing a different number for the same week is usually measuring something else: a transacted price instead of a posted rate card, a panel without the hyperscalers, a mean instead of a median -why two H100 indices differ. Two of them, Silicon Data’s and Ornn’s, are becoming settlement prices for exchange-traded futures; this index is a median of public rate cards, with a specification for anyone who wants to rely on it, and settles nothing itself: indices that settle futures.
USD-listed rates only: quotes listed in another currency (EUR, RUB, JPY, ...) are shown on the site with the dated official rate but stay out of the index, because a converted figure would move with the exchange rate, not the GPU market. Marketplace floors — the median of the five cheapest verified offers on Vast.ai at scrape time — are shown on the site with a marketplace floor tag but do not enter the index, because a floor moves with supply and is not a rate card. Coverage guard: a week is published only when at least 8 providers contributed, so the series never reads a change in the panel as a change in the market. The series therefore starts in the week of September 7, 2026, the first week that met the guard (20 quotes from 18 providers); earlier weeks fell below it and are not shown. The quotes and providers columns below are the honesty signal — read a point against its coverage.
Prices in this week’s median
Every quote in the panel, each linked to the page it was read from. The Deploy button beside a rate opens that provider’s own site; it changes neither the order of this table nor which quotes are in it (link policy). Added 2026-10-15: this page is the first one readers arrive on from an assistant’s answer, and it was the one page of rates with no way on to the provider.
Historical series, reconstructed from a public price catalog
Our own series starts in the week of September 7, 2026. For the time before it we publish a second-hand series rebuilt from a public, Apache-licensed price catalog (its source is named in the methodology) that bots rewrite every few hours: for each ISO week we take the cheapest per-GPU on-demand H100 SXM instance per provider as the catalog listed it at the end of that week, and publish the median across providers when at least 4 contributed. We vouch that the catalog published each number on that date — every row links to the exact commit — not that the provider’s own page did. It is a different panel from the series above (the catalog covers at most 7 of our providers) and is never merged with it.
From the week of to the week of , the reconstructed median moved from $7.24 to $4.25 per H100 GPU-hour, with the panel growing from 4 to 6 providers (between 4 and 7 over the period). Read each point against its provider count: a cheaper provider joining the catalog lowers the median without any listed price moving. The composition-proof view of the same period is the base-100 index: with the week of January 6, 2025 = 100, it stands at 80.2 in the week of October 12, 2026 — a 20% fall in listed H100 rates among providers present week to week, with the panel’s growth taken out.The catalog’s cheapest region is taken for each provider, so a point can sit below our own figure for the same provider (we read US regions for the hyperscalers); Azure is left out because its catalog file stopped refreshing in 2024. Download: CSV · also in JSON as historical_series. The same catalog, one card at a time, with a like-for-like table per provider: NVIDIA A100 40GB · NVIDIA A100 80GB · NVIDIA B200 · NVIDIA B300 · NVIDIA GH200 · NVIDIA H100 NVL · NVIDIA H100 PCIe · NVIDIA H100 SXM · NVIDIA H200 · NVIDIA L4 · NVIDIA L40 · NVIDIA L40S · NVIDIA RTX 4090 · NVIDIA RTX 5090 · NVIDIA RTX 6000 Ada · NVIDIA RTX A6000 · NVIDIA RTX PRO 6000 Blackwell · NVIDIA V100.
Every week of it, with its base-100 value: H100 SXM price history · CSV.
An AI agent can read this index as a tool: price_index on our free MCP server, in the Claude Connectors Directory and one click away in Cursor.
Cite this page
Published under CC BY 4.0: use and redistribute with attribution and a link. Cite the linked source for a price itself, and cite us for the normalization, the index and the comparison. Cite a week and its vintage, not a day: the week of October 19, 2026 is provisional. A provisional week is recomputed every hour and settles when the week closes; a final week never changes. Dated snapshots are archived as a dataset with a DOI, 10.5281/zenodo.22706541, which always resolves to the latest one: cite this page for the current figure, and the DOI with its version for a figure frozen as you used it.
| Plain | GPU Economy, "Cloud GPU Price Index", October 23, 2026. https://gpueconomy.com/price-index |
|---|---|
| Markdown | [GPU Economy, "Cloud GPU Price Index" (October 23, 2026)](https://gpueconomy.com/price-index) |
| APA | GPU Economy. (2026-10-23). Cloud GPU Price Index. Retrieved from https://gpueconomy.com/price-index |
| Dataset (DOI) | GPU Economy. Cloud GPU Price Index [Data set]. Zenodo. https://doi.org/10.5281/zenodo.22706541 |
Data updated: