Methodology
Every number on this site is either collected from a provider source or computed by a formula we show. Nothing is typed in by hand, and nothing stale is presented as live.
Price check status (21 November 2026)
- Prime Intellect: check failed; last checked 31 October 2026. We could not verify the provider's current prices.
A failed check on one provider marks that provider's prices with their last check date, and those prices leave current comparisons after 48 hours. Our site health check fails only when a daily run has no matching completion record or most public-price providers are stale. The monitor also checks the published snapshot's age to catch a missed run or publication.
Where prices come from
Once per day we fetch each provider's own pricing source: a public API where one exists, the official pricing page otherwise. Current providers: Daytona, Lium, Impossible Cloud, Synteq, HyperAI, Runcrate, GPUwerk, RunPod, Lambda, Hyperstack, CoreWeave, Nebius, Verda (DataCrunch), Voltage Park, Crusoe, Jarvislabs, OVHcloud, Spheron, Microsoft Azure, AWS, Oracle Cloud, Hetzner, Together AI, fal, Modal, Vultr, Massed Compute, TensorDock, SaladCloud, Replicate, Fireworks AI, Baseten, DigitalOcean, GMI Cloud, Prime Intellect, Cerebrium, Koyeb, Thunder Compute, Latitude.sh, Hot Aisle, Scaleway, Google Cloud (expanding). AWS uses its own public on-demand and spot feeds; Azure uses its Retail Prices API. Earlier AWS observations from a third-party mirror are excluded at read time, with the original history retained. Each row stores the price, the source URL, and the exact fetch timestamp. Each offer shows the date its price was last checked.
Three kinds of numbers
- Sourced: a price collected from the provider, shown with its source link and verification date.
- Computed: derived values ($/hr per GB VRAM, break-even hours). The formula is printed next to the result. One provider-level case: Scaleway publishes prices only in EUR, so its rows are converted as
usd_hr = eur_hr × ECB daily EUR/USD reference rate(rate via the frankfurter.dev ECB mirror, refreshed with each scrape). Small day-to-day moves on Scaleway rows can therefore reflect the exchange rate, not a repricing. - Street (buy) prices: the buy-price tables on GPU pages come from public reseller product listings, re-checked daily; each listing carries a source and date. If a price cannot be rechecked, its earlier value stays with the date it was read. These are retailer observations, not manufacturer suggested prices.
- Never invented: if we can't verify a number, we say "no data" or flag the last updated value. A stale price presented as live is just a hallucination with a UI.
When a fetch fails
Providers change their pages. When a fetch or parse fails, that provider's rows are carried forward and visibly flagged with the date they were last verified: they never silently pretend to be today's price. Persistent failures get fixed or the provider is dropped.
The daily price record
Each day's prices are kept as recorded: 140 days since 5 July 2026. A recorded price is never edited; the corrections listed below are applied whenever the record is read. The cheapest price of every GPU on every day is in the downloadable history.
The GPU Rental Price Index (GRPI)
The GPU Rental Price Index turns the whole market into one number. It is chain-linked: base 100 on the first day of the record, then each day the level is multiplied by that day's link, the like-for-like change in the prices read that day. Each calendar month also has a reading, the average of its daily levels, on a page of its own.
The link compares configurations, not medians. Every on-demand configuration a provider lists (the same GPU variant, GPU count, instance and region) is compared with its own most recent price in the previous 7 days. A provider's ratios for one GPU are combined into one geometric mean, so each provider counts once for each GPU it rents, and the day's link is the geometric mean of those provider ratios. An offer that appears, disappears or changes its configuration has no earlier price of its own, so it never moves the level; a provider we could not re-read is left out that day, so a carried-forward price never counts as unchanged; and a price that goes back and forth returns the index to where it was when the same offers are compared on the day it rises and the day it falls. If offers join or leave in between, that price can count for more on one of those days than on the other, so the two moves need not cancel exactly.
About the base. The 2026-07-05 base snapshot covered 2 providers (Lambda, RunPod) and was cleaned once (removal of a RunPod $0.50 sentinel value), so the first link (2026-07-05 to 2026-07-06, +1.79%) reflects early-capture volatility rather than a settled market, and the provider universe reached full coverage (18 providers, all live) on 2026-07-08. The base anchors only the level, never a fixed provider set: the comparison widens as the record grows.
Tiers explained
Secure / on-demand and community / spot remain separate price tiers. The tier does not prove stock or a service guarantee. Reserved commitments and serverless rates also remain separate; neither counts as a spot observation. We show both and label every row, so the billing models can be compared explicitly. Prices are per GPU per hour and exclude storage, egress and networking.
Affiliate disclosure
Provider buttons use direct URLs. No commission is currently arranged. Prices sort the tables; tracking never changes the order. Lium listing links include campaign parameters so the provider can count referrals; no commission is arranged for those links. The cheapest listed rate wins whether or not the provider pays us anything, and no program we join asks for placement.
Revisions
13 October 2026: one market median, and daily prices by GPU. Every median the site prints for a GPU (the GPU pages' typical provider, the daily price by GPU, the per-family table, the hyperscaler and other-cloud medians and the monthly reports) now counts each provider's cheapest on-demand rate once and takes the median across providers. The per-family table and the hyperscaler medians used each provider's median rate before, so the same GPU could show two medians. The like-for-like index (GRPI) is unchanged.
This is a dated changelog for anything that changes how a published number is computed or which rows count toward it. Nothing here rewrites the ledger (rule #4: it is append-only). Every change below is either a read-time exclusion or a computation fix applied going forward.
- 2026-10-26: GRPI restated on matched configurations, with geometric means. Each day's link now compares every configuration with its own earlier price (within 7 days) and takes geometric means, each provider counting once per GPU. It used to be the arithmetic mean of each GPU family's median provider rate, day on day. Two things made the old level drift upward: an arithmetic mean of a rise and the fall back nets above 1 (an RTX 4090 rate that flipped about twenty times between two values added to the level each time), and a provider adding or dropping one configuration moved its median with no price changing. On 26 October 2026 the old method read 122.96, 23.0% above the base; the restated index read 109.72, 9.7% above. Every published level since 5 July 2026 is restated; the monthly readings began with this revision. The same configurations listed on both 24 July and 22 October rose 6.5% between those days, which the restated index (+6.8% over the same days) agrees with and the old one (+14.0%) did not.
- 2026-07-18: GRPI restated to per-family provider intersections. The index now takes its like-for-like provider set independently for each GPU family, instead of one global provider set for the whole index. This removed a composition leak where a provider joining or dropping a single family (while staying live overall) could move the level. Restating the series from the 2026-07-10 link forward moved published daily levels by roughly -0.02 to -0.38 points versus the previous calculation; the current level and the shape of the trend are unchanged in direction. See "How the index is built" above for the corrected method.
- Prices pause after 48 hours. Prices last read more than 48 hours ago leave current tables, rankings, headline figures, charts and downloads. During those two days, carried prices keep their last-checked badge and appear after fresh prices. A paused provider's page links to its own pricing page and gives the last successful check. Historical charts use successful observations on each recorded day; the original records remain in the price history.
- Errata: read-time exclusions. When a row is later found to be mislabeled or wrong (a tier mismatch, a stale price captured as fresh, and so on), we do not edit the ledger:
data/errata.jsonlists the excluded provider, GPU, kind and date range with a stated reason, and every live surface (movers, records, history, GRPI) applies that exclusion at read time. The raw snapshot file is untouched; only the derived views change.Provider GPU Dates Reason Vultr B200, H100 SXM (on-demand) 5 July 2026 to 17 July 2026 The pre-2026-07-18 Vultr pricing-page scrape could not distinguish spot from on-demand for these GPUs. The 2026-07-18 API migration revealed both deploy preemptible-only, so the earlier on-demand label was wrong. Reclassified to spot at read time so they no longer count as firm records or firm history. Vultr all GPUs 9 July 2026 to 17 July 2026 Vultr's fetch failed (HTTP 403) from 2026-07-09 through 2026-07-17, so every Vultr row in that window is last-good carry-forward, not a fresh daily observation. Excluded from firm-price surfaces, history, records, movers and the index at read time. The raw ledger rows are preserved. RunPod A40 (community cloud) 8 August 2026 Snapshot-scoped removal only. RunPod's official public GraphQL gpuTypes response at https://api.runpod.io/graphql, checked 2026-08-08T08:24:59Z, returned communityCloud=false while communityPrice remained 0.35 for A40. RunPod's official GraphQL guide at https://docs.runpod.io/sdks/graphql/manage-pods documents these pricing and availability fields. The adapter now requires the matching cloud flag to be true; this entry removes, and does not replace, the false tier from only the snapshot with this generated_at. RunPod RTX 3090 Ti (secure cloud) 8 August 2026 Snapshot-scoped removal only. RunPod's official public GraphQL gpuTypes response at https://api.runpod.io/graphql, checked 2026-08-08T08:24:59Z, returned secureCloud=false while securePrice remained 0.46 for RTX 3090 Ti. RunPod's official GraphQL guide at https://docs.runpod.io/sdks/graphql/manage-pods documents these pricing and availability fields. The adapter now requires the matching cloud flag to be true; this entry removes, and does not replace, the false tier from only the snapshot with this generated_at. RunPod RTX 5080 (secure cloud) 8 August 2026 Snapshot-scoped removal only. RunPod's official public GraphQL gpuTypes response at https://api.runpod.io/graphql, checked 2026-08-08T08:24:59Z, returned secureCloud=false while securePrice remained 0.59 for RTX 5080. RunPod's official GraphQL guide at https://docs.runpod.io/sdks/graphql/manage-pods documents these pricing and availability fields. The adapter now requires the matching cloud flag to be true; this entry removes, and does not replace, the false tier from only the snapshot with this generated_at. RunPod L4 (community cloud) 8 August 2026 Snapshot-scoped removal only. RunPod's official public GraphQL gpuTypes response at https://api.runpod.io/graphql, checked 2026-08-08T08:24:59Z, returned communityCloud=false while communityPrice remained 0.44 for L4. RunPod's official GraphQL guide at https://docs.runpod.io/sdks/graphql/manage-pods documents these pricing and availability fields. The adapter now requires the matching cloud flag to be true; this entry removes, and does not replace, the false tier from only the snapshot with this generated_at. AWS all GPUs 27 September 2026 The ec2.shop observation is a third-party compilation. AWS first-party on-demand and spot feeds replaced it on 27 September; the original ledger is retained, excluded from published statistics. AWS all GPUs 7 July 2026 to 26 September 2026 AWS observations came from ec2.shop, a third-party compilation. Excluded under the first-party source rule; original daily files remain unchanged. The replacement AWS feeds begin on 27 September. Together AI H100 (on-demand) 3 September 2026 to 5 October 2026 Together added a Preemptible column before On-Demand on its pricing page (https://www.together.ai/pricing) on or about 2026-09-03. The adapter took the first price after each GPU name as on-demand, so Together's preemptible price was recorded as on-demand from the 2026-09-03 snapshot on. The live page checked 2026-09-28 reads, per GPU per hour, Preemptible / On-Demand: H100 $1.99 / $3.99, H200 $2.99 / $5.99, B200 $4.09 / $8.19, B300 $4.99 / $9.99. Reclassified to spot at read time (exact price match, so a correctly recorded on-demand row is untouched); the adapter now reads the columns by header. The raw ledger rows are preserved. Together AI H200 (SXM) (on-demand) 3 September 2026 to 5 October 2026 Together added a Preemptible column before On-Demand on its pricing page (https://www.together.ai/pricing) on or about 2026-09-03. The adapter took the first price after each GPU name as on-demand, so Together's preemptible price was recorded as on-demand from the 2026-09-03 snapshot on. The live page checked 2026-09-28 reads, per GPU per hour, Preemptible / On-Demand: H100 $1.99 / $3.99, H200 $2.99 / $5.99, B200 $4.09 / $8.19, B300 $4.99 / $9.99. Reclassified to spot at read time (exact price match, so a correctly recorded on-demand row is untouched); the adapter now reads the columns by header. The raw ledger rows are preserved. Together AI B200 (on-demand) 3 September 2026 to 5 October 2026 Together added a Preemptible column before On-Demand on its pricing page (https://www.together.ai/pricing) on or about 2026-09-03. The adapter took the first price after each GPU name as on-demand, so Together's preemptible price was recorded as on-demand from the 2026-09-03 snapshot on. The live page checked 2026-09-28 reads, per GPU per hour, Preemptible / On-Demand: H100 $1.99 / $3.99, H200 $2.99 / $5.99, B200 $4.09 / $8.19, B300 $4.99 / $9.99. Reclassified to spot at read time (exact price match, so a correctly recorded on-demand row is untouched); the adapter now reads the columns by header. The raw ledger rows are preserved. Together AI B300 (on-demand) 25 September 2026 to 5 October 2026 Together added a Preemptible column before On-Demand on its pricing page (https://www.together.ai/pricing) on or about 2026-09-03. The adapter took the first price after each GPU name as on-demand, so Together's preemptible price was recorded as on-demand from the 2026-09-03 snapshot on. The live page checked 2026-09-28 reads, per GPU per hour, Preemptible / On-Demand: H100 $1.99 / $3.99, H200 $2.99 / $5.99, B200 $4.09 / $8.19, B300 $4.99 / $9.99. Reclassified to spot at read time (exact price match, so a correctly recorded on-demand row is untouched); the adapter now reads the columns by header. The raw ledger rows are preserved. Lium all GPUs (secure cloud) from 5 October 2026 Lium lists nodes run by independent hosts on its marketplace. The collector recorded its "secure" tier as secure cloud, which is the label for capacity a provider operates itself, so a single host's node became the headline on-demand price. These rows are marketplace hosts: they keep their prices and dates but no longer count as on-demand or secure cloud. The collector records them as marketplace hosts from 7 October 2026. - Same-day re-runs. A day's first complete check is that day's record, and it is never rewritten. If we check again the same UTC day (for example, to fix a mislabelled tier), the site's current prices and today's snapshot take the newer reading, and the rows it changed are kept with their time beside the record; the day's record itself stays as first written. The public repository copies today's snapshot once a day, so on a day with a re-run its file can hold the later reading. To cite a figure, give its date and the citation on the data page.
Configuration, availability and model estimates
GPU count, node totals, CPU, RAM, disk and region come from a published configuration. Missing fields say “not reported”; an absent node minimum is not treated as a confirmed single-GPU offer. A recent price does not establish available stock: stock is shown only where the provider's own data reports it. Vultr's plans API, for example, lists the locations each plan can be deployed in today, so a plan it lists in no location is shown as out of stock. RunPod's API gives a stock level for one GPU on Secure Cloud, on Community Cloud and on the two together, and the data centres holding Secure Cloud capacity: a Secure Cloud price with no level is shown as out of stock, a Community Cloud price only when neither tier has capacity (its own level is sometimes missing while the combined one shows stock at the community price), and a Secure Cloud price whose capacity sits in one location shows that place. Where RunPod grades the stock as low, medium or high, the row says so ("In stock: low"), as read at that day's check.
Monthly plan equivalents use 720 hours; listed annual Latitude plans use 8,760 hours. Model weights, supported KV-cache calculations, runtime headroom and training allowances are separated on each model page. They are estimates, not performance or compatibility claims.
Source observations and update dates are collected on the Sources page linked in the footer.