Methodology changelog
A citation outlives the rule it was computed under. Every file we publish carriesmethodology_version — today 1.30.17 — top level inprice-index.json and prices.json, and as a column in every CSV and in each weekly receipt, because a CSV row is what survives copying. This page says what each version means. The rules themselves are onthe methodology page.
What the parts mean
- Patch — collection and coverage: a provider added or dropped, a parser fixed, a quote withheld. The definition of the published figure is untouched, so figures across a patch are comparable.
- Minor — new fields, a new cut of the same quotes, or a new rule about which reads count as current. Figures published before it stay valid.
- Major — the definition of the published figure itself. It should never appear here: a number that changes meaning gets a new series under its own name instead, so a figure someone cited never quietly becomes a different measurement. If this column ever says major, the series was renamed in the same change.
Changes
| Date | Version | Kind | What changed | Definition of the published figure |
|---|---|---|---|---|
| 1.30.17 | patch | Coverage: six names Clore.ai prints for cards this site already carries are now read rather than dropped - the A100 SXM4 40GB, the V100 SXM2 in both its sizes, the T4, the RTX A4000 and the RTX A6000. Five of the six are the mapping Vast.ai's module already makes, so the site's two marketplaces cannot read one name two ways; the sixth, the T4, the other marketplace has no name for, and NVIDIA shipped one T4 at one size. Each is a name plus a memory window, never a name alone. No figure is published by this today: none of the six books could carry a floor at the time of reading - nought to two free verified machines each, and every one of the 223 V100 SXM2 32 GB machines, the largest datacenter book on that marketplace, was rented or unverified - so each is recorded as withheld, and a card page shows nothing where a quote has never had a price. What the mapping buys is the next hour: a marketplace's book turns over, and a name nobody maps never becomes a figure. The definition of the published figure did not change. | unchanged | |
| 1.30.16 | patch | Coverage: Clore.ai's multi-GPU machines are now read as a second quote per card, the partition Vast.ai has always had - the machine's own price with its card count, region "Multi-GPU machines", never mixed into the single-GPU floor. Twelve node-priced quotes on the first run, and one card gains a source it needed: the RTX A2000, whose single-GPU book on this marketplace was one verified listing while ten whole machines of 6 to 12 cards stood beside it (Server #106239, $0.6965 an hour for 12 cards, $0.0580 a card). What had kept these machines unread was that a Clore host can let some of a machine's cards and keep the rest listed, which would make a node price something nobody can buy; the feed states it rather than leaving it to be inferred - false on a machine only ever rented whole, else a count of the cards still free - and a machine is priced only when every card in it is free. A machine holding two kinds of card is not any one card's node price and is read into neither book. The same two guards decide each book: a floor needs three machines, and a median over three times the cheapest ask is withheld as a thin market. The definition of the published figure did not change. | unchanged | |
| 1.30.15 | patch | Coverage: the NVIDIA GeForce RTX 3060 (12 GB) joins the catalogue, the cheapest card on the site and the second today whose two sources are both marketplaces - the Clore.ai floor $0.0678 over 6 verified single-GPU listings and the Vast.ai floor $0.0615 over 24, each the median of that marketplace's own five cheapest verified asks. No rate card on this site prices the card and RunPod's API carries no such name. Both books reported 12 GB throughout on the day, and both modules carry a memory window regardless, because the same name also covers an 8 GB card this site has no id for: a bare name would publish the smaller card as the larger one. The RTX 3060 Ti is a separate name and a separate book on both marketplaces. The definition of the published figure did not change. | unchanged | |
| 1.30.14 | patch | Coverage: the NVIDIA GeForce RTX 5070 Ti (16 GB) joins the catalogue with two sources, both of them marketplaces - the Clore.ai floor $0.2472 over 57 verified single-GPU listings and the Vast.ai floor $0.5363 over 5 verified offers, read within two minutes of each other. No other source on this site prices the card: RunPod's API carries no such name and it appears on no rate card we read. Each floor is the median of that marketplace's own five cheapest verified asks, so the 2.2x between them is two books of different hosts rather than a figure to average; the card page shows both and the reader can look either up. Laptop editions of the name are separate books on both marketplaces and never enter this one. The definition of the published figure did not change. | unchanged | |
| 1.30.13 | patch | Coverage: Taiga Cloud, the AI cloud of Northern Data Group, joins with two cards from its own pricing page - the H100 SXM at $2.60 an hour per GPU and the H200 SXM at $3.00. Both figures are printed as "from", and the page says why under each table ("Contact our team to unlock exclusive high-volume pricing"), so both are tagged as advertised floors rather than the price of a stated configuration, the tag Hyperstack's and GMI Cloud's "starting from" rates carry. The page prices each card twice, once as bare metal and once on demand, and these are two ways to take the same card by the hour rather than two markets, so the cheaper of the two is published and the row names which table it came from: the H100 is $2.60 either way, the H200 $3.00 bare metal against $3.20 on demand. The column heading is dollars per GPU-hour, so the figure is one card and not the eight-card machine the row describes. Taiga's vCPU table is not read: it prices processor cores and memory, not a GPU. Neither card gains a source it needed - the H100 SXM had 31 and the H200 40. The definition of the published figure did not change. | unchanged | |
| 1.30.12 | patch | Coverage: Vultr joins, read through its open plans endpoint api.vultr.com/v2/plans-metal, which answers a plain keyless request - Vultr's own website answers 403 to us on every path, the pricing page and the API reference alike, so the endpoint is where a reader checks these figures. The endpoint prices each bare-metal plan twice, on demand and preemptible, and says of each price whether the plan can be deployed that way; a rate is published only where it says it can. Today that is one on-demand plan, the four-card A100 80GB node at $9.589 an hour, $2.397 a card, and six preemptible ones, recorded as spot: the 8x H100 at $18.40 ($2.30 a card), 8x A100 SXM at $11.92 ($1.49), 8x L40S at $11.92 ($1.49), 8x MI300X at $14.80 ($1.85), 8x MI325X at $16.00 ($2.00) and 8x B200 at $25.60 ($3.20). Eight plans state an hourly cost while saying they cannot be deployed on demand - the 8x H100 at $23.92 among them - and those figures are not published as on-demand rates, because that is a label the source itself denies. A preemptible price equal to a plan's own on-demand price is the same figure twice rather than a second market and is left out, which is how the GH200, the four-card A100 and the 8x MI355X print theirs. Vultr's fractional cloud GPU plans, which sell slices such as 12GB of a 48GB A40, are not read: this site prices whole cards. No card gains or loses a source it needed - every card here already had three or more. The definition of the published figure did not change. | unchanged | |
| 1.30.11 | patch | Labelling: Together AI's reserved rates are now tagged "from" (advertised_from), because its GPU clusters page prints each one as "Reserved: Starting at" a figure per GPU-hour, above "Scale: 256 to 1,000+ GPUs" - a floor, not the price of a stated configuration. Until today they were tagged as instance prices; our price audit of 2026-09-29 found it, as it found the same with Hyperstack's "Starting from" reserved column on 2026-09-20. No figure moves: the reserved rates stay $3.19 for the H100 SXM, $6.79 for the B200 and $3.99 for the H200. Together's on-demand rates carry no such wording and keep their tag. Reserved rates do not enter the Price Index, so it is untouched. The definition of the published figure did not change. | unchanged | |
| 1.30.10 | patch | Coverage: the NVIDIA GeForce RTX 4070 Ti (12GB) joins the catalogue, a name RunPod's API has been reading and dropping since the card catalogue began. It has two sources today, each checkable where its row links: RunPod's API at $0.19 an hour in its Community Cloud, and the Clore.ai floor at $0.1991 over ten verified single-GPU listings, every one of them a 12GB machine, which is the memory RunPod reports for the card too. What changed is that Clore.ai became readable at all this morning; Vast.ai is not a third source, its book for the name holding one verified offer on 2026-09-27 and none on 2026-09-28. RunPod's Secure Cloud cell for the card is exactly $0.50, the placeholder its pricing page does not print, and it is skipped rather than published. A card with two sources shows a thin range, and says so. The definition of the published figure did not change. | unchanged | |
| 1.30.9 | patch | Coverage: Clore.ai joins as the site's second marketplace, and five cards that had only two sources now have three - the RTX 3070, RTX 3080 Ti, RTX 4080, RTX 4080 SUPER and RTX 5080. It is read on exactly the footing Vast.ai is: each figure is a trimmed floor, the median of the five cheapest verified, rentable, single-GPU listings, and it is withheld rather than published when the book holds fewer than three listings or the median runs more than three times the cheapest ask - which is why the RTX 3090 Ti and RTX A2000 read as a thin market on the first run, at one verified listing each. Verified is the marketplace's own badge, "rating above 4 with reviews", not a rule of ours. Clore prices a machine per day and its own table divides by 24 and adds the renter fee before showing a price, so the published figure is the listed price plus that fee over 24 hours; the fee is read from the marketplace on every run, 5% on 2026-09-29, and a run that cannot read it publishes nothing rather than guess. First figures, 04:40 UTC: RTX 4090 $0.3369, RTX 5090 $0.4856, RTX 3090 $0.1662, RTX 5080 $0.2927, RTX 4080 SUPER $0.3719, RTX 4080 $0.3063, RTX 3080 Ti $0.2188, RTX 3080 $0.1553, RTX 3070 $0.1194. Its bid-driven spot market is not read, and neither are multi-GPU machines, so no share of a larger machine is published as a machine. The definition of the published figure did not change. | unchanged | |
| 1.30.8 | patch | Coverage: the 32GB Tesla V100 is now a card of its own, v100-32, and three sources that had been reading its name and dropping it now publish it. The V100 this site has tracked since 2026-09-08 is the 16GB card, so a 32GB listing had nowhere to go; each of the three names the memory on its own page, which is what lets the two be told apart. CloudRift lists a V100 SXM3 at $0.28 an hour on demand and $0.25 for a month, beside the 16GB V100 SXM2 at $0.25 that stays the old card. Paperspace lists an eight-card V100-32G node at $9.20 an hour, $1.15 a card - the rate that was published as the 16GB card from 2026-09-08 until it was withdrawn on 2026-09-28. immers.cloud lists a four-card machine at 382.57 roubles an hour, 95.64 a card, $1.13 at the Bank of Russia rate of the day. No form factor is recorded for the new card: one source names SXM3 and the other two name none. Paperspace's other V100 row, printed as 'V100 / V100-32G', names either card and is published as neither. No existing figure changed hands: every quote that was on the site stays on the card it was on. The definition of the published figure did not change. | unchanged | |
| 1.30.7 | patch | Coverage: Cyfuture AI is no longer read, one day after it was added in 1.30.5, and its five rupee rates are withdrawn. The page that carries them (cyfuture.ai/pricing) answers 403 Forbidden to the machine that collects prices every hour, so not one of those figures was ever read on the site; it still answers a browser, and we do not work around a server that refuses us. Five cards lose this source: H100 SXM, A100 80GB, MI300X, L40S and RTX PRO 6000, each of which keeps every other source it had. No published figure changed, because none of Cyfuture's ever reached a page. The definition of the published figure did not change. | unchanged | |
| 1.30.6 | patch | Coverage: Selectel, a Russian cloud, is now read from its GPU page (selectel.ru/services/gpu/), which prices cloud servers by the hour in roubles on tariff cards. Only the cloud-server cards are read: the same page also prices managed Kubernetes node groups, at their own lower rate, and dedicated servers by the month, and neither is the product published here. Every figure on the page is printed as a 'from' rate and is labelled as such. One rate per card is kept, the machine with the lowest price per GPU, with its GPU count, converted at the Bank of Russia's official rate of the day. Three cards on 2026-09-28 at 84.41 roubles to the dollar: an H200 at 5,397.3 roubles an hour for eight cards ($7.99 a card), an RTX PRO 6000 at 579.94 for two ($3.44 a card) and an L4 at 69.88 ($0.83). As a rouble listing it does not enter the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.30.5 | patch | Coverage: Cyfuture AI, an Indian cloud, is now read from its pricing page (cyfuture.ai/pricing), which lists whole GPU instances by the hour in rupees, one table per card. Only the on-demand column is read - the 1, 6 and 12-month reserved columns are a commitment - and only from a cell printed in rupees, because the page's own currency switch reports a failure. One rate per card is kept, the instance with the lowest price per GPU, with its GPU count, converted at the European Central Bank's reference rate for the rupee. Five cards on 2026-09-28 at 95.98 rupees to the dollar: an H100 SXM at 2,552 rupees an hour for eight cards ($3.32 a card), an A100 80GB at 1,536 for eight ($2.00 a card), an MI300X at 2,125 for eight ($2.77 a card), an L40S at 485 for four ($1.26 a card) and an RTX PRO 6000 at 256.5 for one ($2.67). The page also prices a B300 with no on-demand rate, a 32GB V100, an MI325X whose stated memory is not that card's, and an Intel Gaudi 2; none of those is published. As a rupee listing it does not enter the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.30.4 | patch | Coverage: Cloud.ru Evolution, the largest Russian public cloud, is now read from the GPU tariff annex of its contract (cloud.ru/documents/tariffs/evolution/evolution-compute-gpu, version 260909), which prices 41 GPU virtual machines by the hour in roubles. The annex prints each rate with VAT, without VAT and the VAT itself; the figure published is the one with VAT, what a buyer pays. One rate per card is kept, the machine with the lowest price per GPU, with its GPU count, converted at the Bank of Russia's official rate of the day. Six cards on 2026-09-28 at 84.41 roubles to the dollar: an A100 40GB at 256.2 roubles an hour ($3.04), an A100 80GB at 732 for two cards ($4.34 a card), an H100 PCIe at 549 ($6.50), an H100 NVL at 646.6 ($7.66), an H100 (form factor not stated) at 1,708 for two ($10.12 a card) and a V100 at 244 ($2.89). The annex also prices an A100 PCIe without saying how much memory it has, on eleven cheaper rows; those are not published, because the card cannot be told from the page. As a rouble listing it does not enter the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.30.3 | patch | Coverage: immers.cloud, a Russian GPU cloud, is now read from its GPU page (immers.cloud/gpu/), which lists ready-made virtual machines per card by the hour in roubles. One rate per card is published, the machine with the lowest price per GPU, with its GPU count, converted at the Bank of Russia's official rate of the day; the company states it pays no VAT. Ten cards on 2026-09-28 at 84.41 roubles to the dollar, among them an H100 (form factor not stated) at 1,312.65 roubles an hour for four cards ($3.89 a card), an H200 at 839.97 for two ($4.98 a card) and an A100 80GB at 792.65 for four ($2.35 a card). As a rouble listing it does not enter the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.30.2 | patch | Coverage: Intelion Cloud, a Moscow GPU cloud, is now read from its server configurator (intelion.cloud/gpu_server_configurator/), the first source priced in roubles. The page prices the card alone, VAT included, with CPU, memory and disk added on top, so each rate is tagged as GPU only and shown as printed, converted at the Bank of Russia's official rate of the day. Four cards on 2026-09-28 at 84.41 roubles to the dollar: A10 27.38 roubles an hour ($0.32), RTX A5000 37.20 ($0.44), RTX 3090 50 ($0.59), A100 80GB 204.17 ($2.42). Its H100 SXM and H200 are not read: they cannot be ordered from the configurator and their own pages print different figures and say the card is not available. As a rouble listing it does not enter the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.30.1 | patch | Coverage: ionstream is no longer read. Its page prints one "from" rate per GPU-hour for the B200 ($2.69) and the B300 ($2.93) with no term, and we published both as on-demand - the cheapest of each card on the site. The company has since told us in writing that these rates assume a large part of the server paid up front and that the B200 is no longer available; the page states neither, so the rates are withdrawn rather than relabelled, and the figures in the letter are not published. The cheapest listed on-demand rate is now $3.69 for the B200 and $4.89 for the B300. The definition of the published figure did not change. | unchanged | |
| 1.30.0 | minor | A price listed in any currency can now be published (owner's decision): until today only USD and EUR. The listed figure is shown as listed, next to its USD figure at the day's official reference rate and that rate's date - the ECB rate (for a currency other than the euro, the cross of the two rates the ECB publishes against the euro), or the Bank of Russia's official rate for RUB and VND, which the ECB does not publish. Price changes are still detected on the listed figure, never on the rate, and only USD listings enter the Price Index. No quote changed today: sources priced only in roubles, yen, rupees or dong become readable from the next parser run. The definition of the published figure did not change. | unchanged | |
| 1.29.6 | patch | Coverage: the row Hyperstack prints as "NVIDIA H100 NVLink" is now read, the last of the six H100 and A100 rows on that page that was still being dropped. Its VRAM column says 80 GB, so it is the PCIe card on an NVLink bridge rather than the 94 GB H100 NVL, which is how the "NVIDIA A100 NVLink" row has been read since the page was first scraped. No published figure moves: Hyperstack prices the plain "NVIDIA H100" row lower in both sections ($2.50 an hour against $2.60 on-demand, $1.75 against $1.82 reserved), and the cheapest of the rows for a card is the one published, with the name it was read under shown beside it. What changes is that the rate is now read at all, so it becomes the quote if Hyperstack ever stops listing the plain row. The definition of the published figure did not change. | unchanged | |
| 1.29.5 | patch | Coverage: Together AI's HGX B300 card is now read, at $9.99 per GPU-hour on-demand. The card had carried "Contact us for pricing" and no figure; it now prints an hourly rate, and the page prints no "Starting at" reserved line beside it, so this is the first Together card published with an on-demand rate and no reserved one. The 270 GB HGX name is the eight-way SXM board, read as the same B300 the rest of the site tracks. The GB200 NVL72 card on the same page still says "Contact us for pricing" and is not read. The definition of the published figure did not change. | unchanged | |
| 1.29.4 | patch | Coverage: the row CloudRift lists as "H100" is now read, at $1.70 an hour on a one-month reservation. It had been dropped because the model column names no form factor; the row's own VRAM column says 80 GB, so it is published against the form-factor-neutral H100 card, which says exactly that and no more - the same footing Gcore's 1x H100 and Sesterce's "H100 - SXM5 / PCIe" are read on. Its On-Demand cell reads "Upon request" and publishes no figure, so the one-month rate is the only quote this row carries. CloudRift's "V100 SXM3" stays unread: 32 GB is another card, not another name for one we track. The definition of the published figure did not change. | unchanged | |
| 1.29.3 | patch | Coverage: ionstream joins as the 57th source read, with two rates from the line-up on its solutions page - B200 $2.69 and B300 $2.93 per GPU-hour. Both lines say "from", so both are tagged as advertised "from" prices rather than the price of a configuration, the same footing as GMI Cloud, Sesterce and Neysa. They are read as on-demand because the page names no term, no minimum and no commitment: its own words are "Pick the GPU that matches your workload and scale on demand". Each figure is already per GPU on the page, so nothing is divided; the eight-card node it describes carries no total price. The Vera Rubin card on the same page is "coming soon" with no figure and is not read. Both rates are the lowest this site reads for their card, which is what an advertised "from" is, and each is tagged so a reader can see it. The definition of the published figure did not change. | unchanged | |
| 1.29.2 | patch | Coverage: Hyperbolic joins as the 56th source read, with three rates from the card on its marketplace page - H100 SXM $3.19, H200 $3.99 and B200 $5.99 per GPU-hour. The column they sit under says "Starting From", so all three are tagged as advertised "from" prices rather than the price of a configuration, the same footing as GMI Cloud, Sesterce and Neysa: the page itself says the figures are refreshed weekly from the best available rates from suppliers on its platform. They are read as Hyperbolic's own rate card, not as a broker quoting another operator: no other operator's price appears anywhere on the page, and the comparison table that names AWS, Azure, CoreWeave, Fluidstack, Lambda Labs and RunPod prints multiples and no rates, so nothing in it can become a figure here. Each of the three sits inside the live range of its card. The definition of the published figure did not change. | unchanged | |
| 1.29.1 | patch | Coverage: Gcore joins as the 55th source read, with nine rates from its AI pricing page, which carries three separate GPU products. Its bare-metal clusters are priced per machine and published that way - EUR 18.16 an hour for 8x H100 SXM, EUR 10.16 for 8x A100 SXM, EUR 8.72 for 8x L40S PCIe, EUR 21.84 for 8x H200 SXM and EUR 48.00 for 8x B300 SXM - and a column is published only when the per-GPU figure derived from the node rate equals the one Gcore prints beside it. Everywhere Inference, an endpoint rather than a machine by the page's own words, adds EUR 1.08 for 1x L40S, EUR 1.26 for 1x A100 and EUR 1.53 for 1x H100 as serverless rates; Virtual GPU adds EUR 2.36 for 1x H100. The rates are listed in euro and published in euro beside the converted figure with the ECB rate and its date, so, like every EUR listing, they stay out of the Price Index. Which of the three products a rate comes from is shown on the row, because the same H100 costs a different amount in each. The 72x GB300 NVL72 column prints no rate at all and is not read. Only the one-GPU row of each per-configuration table is read; the larger rows are that same rate multiplied. The definition of the published figure did not change. | unchanged | |
| 1.29.0 | minor | A new market, shared: a GPU other tenants use too, first packet.ai's Dynamic plan ("shared infrastructure with the same peak performance and VRAM" at "roughly half the price"). Like community capacity it counts in no median rate card and no "cheapest" figure, and its risk tag is medium; it had been read as on-demand since 2026-09-26, when its RTX PRO 6000 at $0.66 became the cheapest card from 80 GB against $1.09 next. The Price Index is untouched - no shared quote is an H100 - and so is every other figure's definition; the change is which rates count as an on-demand rate card. | unchanged | |
| 1.28.29 | patch | Coverage: Deep Infra joins as the 54th source read, with five rates from the "Custom LLMs" table on its pricing page - A100 80GB $0.89, H100 $2.20, H200 $2.69, B200 $3.69 and B300 (270 GB) $4.89 an hour, each printed there as the provider's own "$X / GPU-hour". The rest of that page is per-token model pricing, which this site does not track, and the DGX clusters below the table are contact-sales; neither is read. The rates are recorded as on-demand instances: the page says "pay for uptime" with "minute granularity" and names no term, and nothing else on it is billed by the hour. "H100 80GB" states no form factor, so it is the neutral H100 and stays out of the H100 SXM index. All five sit at or just under the cheapest rate we read for their card, which is where an inference platform selling capacity by the minute belongs. The definition of the published figure did not change. | unchanged | |
| 1.28.28 | patch | Coverage: the NVIDIA RTX A2000 (6 GB) joins the catalogue as the smallest card on the site, with two sources - RunPod $0.12 an hour in its Community Cloud, read from RunPod's API as every RunPod row is and the cheapest rate it prints for anything, and the Vast.ai marketplace floor, $0.2037 an hour over 3 verified single-GPU offers on 2026-09-27, every one of them reporting 6 GB. RunPod's Secure Cloud cell for the card is exactly $0.50, the placeholder figure no RunPod page prints, and is skipped. The marketplace name covers two cards, so it is read below 8 GB: the cheapest ask under the bare name that morning was a 12 GB A2000 at $0.0521, a card this site has no id for. Three offers is the smallest book a floor is taken over here, and the same book was refused four hours earlier for a floor 3.6 times its cheapest ask, so this quote will read "thin market" on some hours - the guard working, not a gap. The definition of the published figure did not change. | unchanged | |
| 1.28.27 | patch | Coverage: AWS becomes a source for the NVIDIA RTX PRO 6000 Blackwell Server Edition (96 GB), at $3.3631 an hour for a g7e.2xlarge in us-east-1. G7e is the other half of the G7 generation mapped six hours earlier, and a different card: AWS's own G7e page says "G7e instances feature up to 8 NVIDIA RTX PRO 6000 Blackwell Server Edition GPUs with 768 GB of total GPU memory (96 GB of memory per GPU)" and its size table opens with "g7e.2xlarge | 1 | 96", so the GPU count is read and not derived. Three of its sizes carry one card and the smallest is the cheapest per GPU, as with every other AWS family read here ($3.9982 for g7e.4xlarge, $5.2682 for g7e.8xlarge). The figure sits among the dearest of the card's quotes, between Modal $3.0312 and Oracle $4.50; the Max-Q edition is a card of its own here and AWS prices no such size. The definition of the published figure did not change. | unchanged | |
| 1.28.26 | patch | Coverage: AWS becomes a fourth source for the NVIDIA RTX PRO 4500 Blackwell (32 GB), at $2.52 an hour for a g7.2xlarge in us-east-1. The G7 generation had sat unmapped in AWS's pricing JSON since it appeared; AWS's own G7 page names the card - "Up to 8 NVIDIA RTX PRO 4500 Blackwell Server Edition GPUs per instance with 32 GB memory per GPU" - and its size table gives one GPU for a g7.2xlarge, so the count is read and not derived. It is the dearest of the card's four quotes, beside EmpirioLabs $1.40, Massed Compute $0.92 and RunPod $0.72 in its Secure Cloud, which is where AWS sits for every card we read from it. The g7e generation is the RTX PRO 6000 Blackwell Server Edition and stays unmapped. The definition of the published figure did not change. | unchanged | |
| 1.28.25 | patch | Coverage: the NVIDIA A10G (24 GB) joins the catalogue with two sources - AWS at $1.006 an hour for a g5.xlarge in us-east-1, and Baseten at $1.2072 an hour, the $0.02012 a minute its pricing page prints for an "A10G 24 GiB VM". Both parsers had been reading the card and dropping it. It is AWS's own variant of the A10 and not the A10, which this site tracks separately and prices from $0.59 to $3.20 an hour; the two never share an id. The number of cards in a g5.xlarge is read off AWS's own G5 instance table, which prints one GPU of 24 GB for that size, and as with every AWS family only the smallest size is read, being one whole card and the cheapest per GPU. The g5g family stays out: it is the T4G, another card. The definition of the published figure did not change. | unchanged | |
| 1.28.24 | patch | Coverage: the NVIDIA GeForce RTX 3090 Ti (24 GB) joins the catalogue with two sources - RunPod $0.46 an hour in its Secure Cloud and $0.27 in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.2578 an hour over 5 verified single-GPU offers on 2026-09-27, every one reporting 24 GB, with a four-GPU machine at $1.7119 an hour for the node as a second quote. Neither RunPod cell is the $0.50 placeholder, so the card arrives with two rates rather than one. It is a different card from the GeForce RTX 3090, and because both hold 24 GB the names keep them apart rather than the memory: Vast.ai sells it as "RTX 3090 Ti" and RunPod as "NVIDIA GeForce RTX 3090 Ti", each with its own book and its own rate. Both parsers had been reading the name and dropping it. The definition of the published figure did not change. | unchanged | |
| 1.28.23 | patch | Coverage: the NVIDIA GeForce RTX 4080 SUPER (16 GB) joins the catalogue with two sources - RunPod $0.28 an hour in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.3511 an hour over 5 verified single-GPU offers on 2026-09-27, every one reporting 16 GB, with a two-GPU machine at $0.5356 an hour for the node as a second quote. RunPod's Secure Cloud cell for the card is exactly $0.50, the placeholder figure no RunPod page prints, and is skipped. It is a different card from the GeForce RTX 4080 published the day before, and because both hold 16 GB the names keep them apart rather than the memory: Vast.ai sells it as "RTX 4080S" and RunPod as "NVIDIA GeForce RTX 4080 SUPER", each with its own book and its own rate. Its Vast.ai book carried two verified offers on 2026-09-26, where a floor needs three, and five the following night. The definition of the published figure did not change. | unchanged | |
| 1.28.22 | patch | Coverage: the NVIDIA GeForce RTX 3080 Ti (12 GB) joins the catalogue with two sources - RunPod $0.18 an hour in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.1889 an hour over the five cheapest of 9 verified single-GPU offers on 2026-09-27, every one reporting 12 GB, with a two-GPU machine at $0.3756 an hour for the node as a second quote. RunPod does not offer the card in its Secure Cloud at all, so there is one rate and not two. It is a card of its own beside the GeForce RTX 3080 published the same hour - 12 GB against 10 - and the two are never folded together: each has its own name at both sources and its own book on the marketplace. The definition of the published figure did not change. | unchanged | |
| 1.28.21 | patch | Coverage: the NVIDIA GeForce RTX 3080 (10 GB) joins the catalogue with two sources - RunPod $0.17 an hour in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.1344 an hour over the five cheapest verified single-GPU offers on 2026-09-27, every one reporting 10 GB, with a three-GPU machine at $0.4230 an hour for the node as a second quote. RunPod does not offer the card in its Secure Cloud at all, so there is one rate and not two. Both parsers had been reading the name and dropping it. Vast.ai sells 10 GB and 20 GB machines under the one name - one of the 16 verified offers that hour reported 20 GB - so its book is read below 16 GB and the 20 GB card, which has no id here, is not priced under this one. The definition of the published figure did not change. | unchanged | |
| 1.28.20 | patch | Coverage: the NVIDIA GeForce RTX 3070 (8 GB) joins the catalogue with two sources - RunPod $0.13 an hour in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.0978 an hour over 10 verified single-GPU offers on 2026-09-26, every one reporting 8 GB, with a two-GPU machine at $0.3489 an hour for the node as a second quote. RunPod does not offer the card in its Secure Cloud at all, so there is one rate and not two. Both parsers had been reading the name and dropping it. The RTX 3070 Ti is a card of its own and is published by neither source. The definition of the published figure did not change. | unchanged | |
| 1.28.19 | patch | Coverage: the NVIDIA GeForce RTX 4080 (16 GB) joins the catalogue with two sources - RunPod $0.27 an hour in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.2810 an hour over 5 verified single-GPU offers on 2026-09-26, every one reporting 16 GB. Both parsers had been reading the name and dropping it since we started; what changed is the Vast.ai book, which carried no verified offer at 04:20 UTC and five at 16:12. RunPod's Secure Cloud cell for the card is exactly $0.50, the placeholder figure no RunPod page prints, and is skipped. The RTX 4080 SUPER is a different card, listed under its own name by both sources and published by neither: its Vast.ai book had two verified offers, and a floor needs three. The definition of the published figure did not change. | unchanged | |
| 1.28.18 | patch | Coverage: RunPod becomes the fourth source for the NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB), at $1.64 an hour in its Community Cloud, read from RunPod's API as every RunPod row is. NVIDIA names all three editions of the card in full, so the name RunPod carries - "RTX PRO 6000 Blackwell Max-Q Workstation Edition" - is this card and not the Server or Workstation edition we already read under another id. RunPod's Secure Cloud cell for it is exactly $0.50, the placeholder figure no RunPod page prints, and is skipped as every $0.50 cell is; the one rate left, $1.64, sits between the three the card's other sources published the same day - the Vast.ai community floor $1.34, CloudRift $1.40 for a month and EmpirioLabs $1.80 on demand. The definition of the published figure did not change. | unchanged | |
| 1.28.17 | patch | Coverage: Azure's RTX PRO 6000 Blackwell Server Edition is read - the two sizes Microsoft's size table (NC RTX PRO 6000 BSE v6 series) gives one whole GPU each, NC144ds_xl and NC144lds_xl, on demand, spot and at one- and three-year reservations, from the Retail Prices API (first read $5.50 an hour on demand, eastus2). The quarter- and half-card sizes stay out (a fraction of a GPU is not a GPU-hour, the A10 rule), as do the two-GPU NC288 sizes (the same price per GPU) and the sizes the table does not list; all are reported as unmapped. The card is not an H100, so the Price Index is untouched; the definition of the published figure did not change. | unchanged | |
| 1.28.16 | patch | Coverage: the NVIDIA RTX 5000 Ada Generation (32 GB) joins the catalogue with two sources - RunPod $0.83 an hour in its Secure Cloud and $0.49 in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.4548 an hour over 3 verified single-GPU offers on 2026-09-26, all reporting 32 GB. Both parsers had been reading the name and dropping it. It is a card of its own, neither the RTX A5000 (24 GB, Ampere) nor the RTX PRO 5000 Blackwell (48 GB), and the three are priced apart. The definition of the published figure did not change. | unchanged | |
| 1.28.15 | patch | Coverage: Leafcloud joins the rate cards - an Amsterdam provider that re-uses its machines' heat in buildings, listing in euro. Four figures on 2026-09-26, per GPU per hour and VAT excluded: RTX 6000 Blackwell EUR 2.76, H100 EUR 3.15, A100 80GB EUR 1.61 and A30 EUR 0.60, each shown beside the ECB conversion and its date as every euro listing is, and each outside the Price Index as every euro listing is. Leafcloud prints its rate card by the month and its hourly rates in a comparison table beside other providers'; we publish its own hourly row only, and only after checking each figure against the same flavor's monthly price over 730 hours and its card count. Its H100 names no form factor, so it is the neutral H100 and does not enter the H100 SXM index. The definition of the published figure did not change. | unchanged | |
| 1.28.14 | patch | Coverage: packet.ai joins the rate cards, read from its own static pricing page. Six figures on 2026-09-26 - L40S $0.92, RTX 4090 $0.39, A100 80GB $1.43 and B200 $6.99 an hour on its Dedicated plan (a whole card, single-tenant), RTX PRO 6000 $0.66 and B200 $3.75 on its Dynamic plan (a shared card at the same peak performance and VRAM, by packet.ai's own description). The plan is the row's region, the way RunPod's Secure and Community Cloud are; both plans are billed hourly and neither is pre-emptible, so both are on-demand. The cards whose plans read "Launching soon" - RTX 5090, H100 SXM, H200 - and the Clusters plan, which is quoted rather than priced, are not published, and the comparison table of other providers' rates on the same page is not read at all. The definition of the published figure did not change. | unchanged | |
| 1.28.13 | patch | Coverage: the NVIDIA GeForce RTX 5080 (16 GB) joins the catalogue with two sources - RunPod $0.59 an hour in its Secure Cloud and $0.39 in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.3356 an hour over 8 verified single-GPU offers on 2026-09-26, all reporting 16 GB. Both parsers had been reading the name and dropping it. RunPod's other GeForce cards - the RTX 4080, 4080 SUPER, 3080 and the rest - stay unread until a second source prices them. The definition of the published figure did not change. | unchanged | |
| 1.28.12 | patch | Coverage: the NVIDIA RTX PRO 5000 Blackwell (48 GB) joins the catalogue with two sources - RunPod $0.96 an hour in its Secure Cloud and $0.82 in its Community Cloud, read from RunPod's API as every RunPod row is, and the Vast.ai marketplace floor, $0.6434 an hour over 18 verified single-GPU offers on 2026-09-26. Both parsers had been reading the name and dropping it. Vast.ai sells a 72 GB edition under the same name, which is a different card with no id here, so its offers are kept out of this card's floor by a memory window. The definition of the published figure did not change. | unchanged | |
| 1.28.11 | patch | Coverage: Vast.ai is a third source for the NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB), whose floor was $1.3716 an hour over 9 verified single-GPU offers ($1.0030-$1.7370) on 2026-09-26, beside EmpirioLabs $1.80 and CloudRift $1.40 for a month. The marketplace calls it "RTX PRO 6000 Max-Q" and it is read as a book of its own: none of its offers reach the RTX PRO 6000 floor, which is taken over the Server and Workstation editions only. The definition of the published figure did not change. | unchanged | |
| 1.28.10 | patch | Coverage: the NVIDIA RTX 4000 SFF Ada (20 GB) joins the catalogue with two sources - Koyeb $0.50 an hour for its serverless instance, RunPod $0.44 in its Secure Cloud and $0.18 in its Community Cloud, read from RunPod's API as every RunPod row is. The SFF edition is a 70 W half-height card and is kept apart from the full-height RTX 4000 Ada listed by DigitalOcean, RunPod and Akamai: the two ids never mix, and neither card enters the other's figures. Both parsers had been reading this name and dropping it. The definition of the published figure did not change. | unchanged | |
| 1.28.9 | patch | Coverage: the NVIDIA RTX PRO 6000 Blackwell Max-Q (96 GB) joins the catalogue with two sources - EmpirioLabs $1.80 an hour on demand, CloudRift $1.40 an hour on a one-month commitment, its on-demand cell reading "Upon request". The Max-Q is the lower-power edition of the RTX PRO 6000 and is a card of its own here: its figures never enter the RTX PRO 6000's, and both parsers had been reading the name and dropping it. RunPod's Max-Q cells stay out, its secure price being the $0.50 placeholder the API returns in unrelated cells - below its own $1.64 community price. The definition of the published figure did not change. | unchanged | |
| 1.28.8 | patch | Coverage: the NVIDIA RTX PRO 4500 Blackwell (32 GB) joins the catalogue with three sources, a name three of our parsers already read and dropped - EmpirioLabs $1.40 an hour, Massed Compute $0.92, RunPod $0.72 in its Secure Cloud and $0.34 in its Community Cloud, the last read from RunPod's API as every RunPod row is. RunPod lists the card in two editions, plain and Server; they are the same card, as the RTX PRO 6000 pair already is, and the Server Edition's $0.50 community cell is the placeholder the API returns in unrelated cells, so it is not published. The definition of the published figure did not change. | unchanged | |
| 1.28.7 | patch | Coverage: Enverge (RTX Pro 6000 $1.95 and B300 $7.50 for a single-GPU machine) and EcoHash (RTX Pro 6000 $1.89 per GPU-hour) are read. Two candidates are listed as not listed, with the reason: BreezeHost (one card, out of stock, its on-demand rate only in the page title) and 1Legion (the page prints two different 12-month rates for the same card, so neither can be taken as its price). The definition of the published figure did not change. | unchanged | |
| 1.28.6 | patch | Coverage: six more providers from the survey of competitors' lists, each a public rate card. Lyceum (its on-demand VM table: H100 $2.79, H200 $4.29, B300 $7.99), Canopy Wave (H200 $4.00, H100 SXM $3.25 per GPU, from two-GPU machines), Atlas Cloud (H200 $3.50, H100 SXM $2.95), Nova Cloud (RTX 4090 $0.42, RTX 5090 $0.66, RTX PRO 6000 $1.56), Neysa (four cards marked "starts at", so shown as advertised floors: H100 SXM $4.39) and HexGrid (twelve cards, RTX A4000 $0.20 to B200 $7.50). Every H100 whose page names no form factor stays out of the H100 SXM figures. The definition of the published figure did not change. | unchanged | |
| 1.28.5 | patch | Coverage: Hugging Face Inference Endpoints are read - eight GPUs at the single-card size of a dedicated endpoint, on the cloud each row names (AWS today): T4 $0.50, L4 $0.80, L40S $1.80, A100 80GB $2.50, H100 $4.50, H200 $5.00, B200 $9.25, RTX PRO 6000 $2.75 an hour. An endpoint serves a model rather than giving a machine to log into, and the provider page says so. Its H100 names no form factor and stays out of the H100 SXM figures. The definition of the published figure did not change. | unchanged | |
| 1.28.4 | patch | Coverage: Daytona is read - eight GPUs of its sandbox pricing, on demand and preemptible (recorded as spot), each the GPU alone, since vCPU, memory and storage are billed on top: B300 and B200 $6.25, MI355X $5.99, H200 $4.54, H100 $3.95 on demand; H100 $2.27 preemptible. Its H100 names no form factor and stays out of the H100 SXM figures. The definition of the published figure did not change. | unchanged | |
| 1.28.3 | patch | Coverage: Civo is read - A100 40GB $1.09, A100 80GB $1.79, H100 SXM $2.99, H200 $3.49 and L40S $1.29 an hour on demand for one GPU with its CPU, memory and disk, and the same cards at 12- and 36-month commitments (H100 $2.69 and $2.49). Civo's 6- and 24-month terms have no market on this site and are not recorded. Its H100 is the SXM card its table names, so it enters the H100 SXM Price Index panel on the index's own rules. The definition of the published figure did not change. | unchanged | |
| 1.28.2 | patch | Coverage: EmpirioLabs is read - its GPU Cloud table of 21 cards, each billed per second at an hourly rate, among them H100 $2.60, A100 80GB $1.70, B300 $16.00. A row priced as an "8-GPU node" is kept as a node price with its count. Rows the page shows as "0 available" are published as listed; the site does not track stock. The GeForce RTX 3090 (24 GB) joins the catalogue with three sources: RunPod $0.22 community and $0.50 secure, Vast.ai's marketplace floor, EmpirioLabs $0.40. The definition of the published figure did not change. | unchanged | |
| 1.28.1 | patch | Coverage: Akamai Cloud (Linode) is read, from its documented keyless plan API - its pricing page refuses our client and is not read - and two cards join the catalogue, each with three sources. Quadro RTX 6000 (Turing, 24 GB): Seeweb EUR 0.29 ($0.33), Lambda $0.69, Linode $1.50. RTX 4000 Ada (20 GB): RunPod $0.20 community and $0.28 secure, Linode $0.52 for its smallest plan, DigitalOcean $0.76. The SFF edition of the RTX 4000 Ada is a different card and stays out. The definition of the published figure did not change. | unchanged | |
| 1.28.0 | minor | Two changes to what can be read, by the owner's decision. A card joins the catalogue when two providers publish its price verifiably, where three were needed before; its median, as for every card, needs two rate cards, so a card with two sources shows a thin but honest range. And a provider whose page draws its prices with JavaScript can now be read: a headless browser renders the page as a visitor's browser would, reads robots.txt first, identifies itself as our parser, and stops at any bot check or captcha - a page that asks is recorded as not listed, never worked around. No provider is read this way yet, and no published figure changed. | unchanged | |
| 1.27.0 | minor | A price read from a provider's API may be published although its public page does not print it, by the owner's decision - the row says "read from the provider's API" with the endpoint, as every RunPod, Vast.ai, AWS, Azure, Oracle, Cudo and Google Cloud row already does, so a reader checks it where it comes from. Until now a RunPod card was taken only if runpod.io/pricing printed it. First use: RunPod's RTX A4000 ($0.17 community, $0.25 secure) and H200 NVL ($3.79 secure). RunPod's API also returns exactly $0.50 in a dozen unrelated cells - the H200 NVL's community price beside $3.79, the RTX 2000 Ada's community price above its own secure one - so a $0.50 cell is published only where /pricing prints that price for that card and tier. The NVIDIA H200 NVL (141 GB, PCIe) joins the catalogue with three sources: Massed Compute $3.62, RunPod $3.79 and Vast.ai's floor $4.14. It is not the SXM H200 and never enters its figures. The definition of the published figure did not change. | unchanged | |
| 1.26.1 | patch | Azure's H200 spot quote is retired. Azure's price API lists a "Spot" meter for the H200 at exactly the on-demand price, in all 32 regions it prices the card - no discount, so no spot market to report; we had been publishing $10.60 per GPU-hour as spot beside the same $10.60 on demand. The rule is the one RunPod's identical spot rows were retired under on 2026-09-11: a spot price is published only where it differs from on-demand for the same machine in the same region. It also removes 19 of Azure's 20 GB200 spot meters, but not the one published (Korea South, $6.22, which does differ). No other published figure changed. | unchanged | |
| 1.26.0 | minor | Azure and Google Cloud rows carry the spread across the regions we read. Both are read in every region and the published row is the cheapest region per card and market, as before; the row now also says how many regions were read and what the dearest of them asks per GPU-hour, each region at its own cheapest offer - "cheapest of 38 regions read; dearest $7.61" for Azure's A100 80GB on demand ($3.67). Two new fields in prices.json and its CSV, regions_read and region_max_usd_per_gpu_hour, null for every other provider. The count is of regions read, not of the regions a provider has. Azure Government and Azure China regions are left out of both the spread and the pick; none had ever been the cheapest. No published figure changed. | unchanged | |
| 1.25.11 | patch | Coverage: AWS is read for the NVIDIA B300 - p6-b300.48xlarge at $142.416 an hour on demand in us-east-1, eight cards a node, $17.80 per GPU-hour. The instance sat unmapped in the same pricing JSON the other nine AWS rows come from; AWS prints one row for it, "8 Ultra" GPUs and 2,144 GB HBM3e, which is 268 GB a card - the B300 DataCrunch sells as "B300 SXM6 268GB". It is the dearest B300 we read today, above Oracle ($15.00) and Hyperstack ($7.40), as AWS is for the B200 ($14.24). Spot is unchanged: it needs signed API calls and is still not recorded. The definition of the published figure did not change. | unchanged | |
| 1.25.10 | patch | Coverage: Bentaus is read - AMD Instinct MI325X at $2.25 per GPU-hour on demand and $1.95 reserved, the two priced tiers of its card page. The MI325X now has three sources, beside DigitalOcean ($3.80) and TensorWave (from $2.25). Only AMD cards are mapped, as at TensorWave, so nothing Bentaus publishes can reach the H100 SXM Price Index. Its NVIDIA B300 page prices both tiers as a band ("$6.50 - $8.50 per GPU-hour") and is not read at all: one end of a band is not a rate. The reserved tier states no term, so the quote carries none. The definition of the published figure did not change. | unchanged | |
| 1.25.9 | patch | Coverage: Thunder Compute is read - RTX A6000 $0.35, L40 $0.79, A100 80GB $1.09 and H100 PCIe $3.20 an hour, each an instance at its smallest configuration (more CPU cores are an add-on), from its single-GPU price list; the 2- to 8-GPU tabs are exact multiples and are not read twice. Its H100 is the PCIe card, so it does not enter the H100 SXM Price Index. The definition of the published figure did not change. | unchanged | |
| 1.25.8 | patch | Coverage: fal is read - B300 $8.50, B200 $6.25, H200 $4.50, H100 $4.50 and RTX PRO 6000 $2.99 per GPU-hour, its serverless list price, each machine type with its CPU and memory. The "as low as" figures beside them are offered to custom deployments through sales and are not recorded. Its H100 names no form factor on the priced page, so it is the neutral H100 card and stays out of the H100 SXM Price Index. The definition of the published figure did not change. | unchanged | |
| 1.25.7 | patch | Coverage: Replicate is read - A100 80GB $5.04, H100 $5.49, L40S $3.51 and T4 $0.81 an hour, each an instance with its CPU and memory, from the self-serve hardware table. Its "Additional hardware" (H200, and 4- and 8-GPU machines) is sold by committed-spend contract and is not recorded. Its H100 names no form factor, so it is the neutral H100 card and stays out of the H100 SXM Price Index. The definition of the published figure did not change. | unchanged | |
| 1.25.6 | patch | Coverage: Azure's T4 (NCasT4_v3, $0.526 on demand) and A10 (NVadsA10_v5 from the 36-vCPU size up, the smallest carrying a whole card; the 1/6 to 1/2 sizes are fractions of a GPU and are not read), on demand, spot and 1- and 3-year reservations. Azure quotes are now never taken from an AT&T edge zone: the cheapest T4 was $0.343 in "ATT Dallas 1", an operator zone a customer does not simply deploy into, against $0.526 in every public US region. No quote published before came from one, so no figure changed. T4 now has six providers, A10 five. The definition of the published figure did not change. | unchanged | |
| 1.25.5 | patch | Coverage: the hyperscalers' inference cards. AWS - T4 (g4dn.xlarge), L4 (g6.xlarge), V100 (p3.2xlarge). Oracle - A10, V100, A100 40GB and RTX PRO 6000; its price list names three of them by an internal series, each matched to its card through the part number the price-list page prints beside the shape. Google Cloud - T4, V100 and L4 on demand and on 1- and 3-year commitments (the L4 was in the parser's map and never read); no spot rate for these three, because the page prints none and the catalog's T4 spot rate, $0.0315, fails the parser's plausibility floor. T4 now has five providers, A10 four. None of these is an H100, so the Price Index is untouched; the definition of the published figure did not change. | unchanged | |
| 1.25.4 | patch | Coverage: six cards the parser already read every hour and dropped as unmapped join the catalogue - NVIDIA A10 (Lambda, Modal, Cerebrium), T4 (Modal, Cerebrium, Baseten), A30 (E2E, Jarvislabs, Massed Compute, Seeweb), A40 (Denvr, E2E, RunPod), RTX A5000 (Massed Compute, Paperspace, RunPod, Vast.ai) and RTX A4000 (Hyperstack, Paperspace, Vast.ai). Two rates the RunPod API returns are not published because no RunPod page prints them: RTX A4000 (printed on its model page, not on /pricing, where every RunPod row links) and MI350X (no page at all). Baseten's A10G stays out: it is AWS's variant of the A10, not the A10. None of the six is an H100, so the Price Index is untouched; the definition of the published figure did not change. | unchanged | |
| 1.25.3 | patch | Coverage: TensorWave, an AMD-only cloud, is read - MI300X $1.71, MI325X $2.25, MI355X $2.95 per GPU-hour, each the "Starting at" figure its accelerator page prints for a whole node, so each is tagged "from" (advertised_from) like Sesterce and GMI Cloud. MI325X now has two providers, MI355X three. Voltage Park's and Latitude.sh's "Starting At" is still not recorded: that rule is about an H100 floor entering the Price Index, and TensorWave prices AMD only, which the parser enforces by mapping AMD ids alone. The definition of the published figure did not change. | unchanged | |
| 1.25.2 | patch | Coverage and one panel change. A new card, "NVIDIA H100 80GB (form factor not stated)", for sources that price an H100 without saying SXM or PCIe: GMI Cloud, Baseten, Cerebrium, E2E, Koyeb, Massed Compute, Northflank and Sesterce (whose row reads "SXM5 / PCIe"). Seven of them were read before and dropped; GMI Cloud was published as H100 SXM on the strength of a research note, not its page, which reads "NVIDIA H100 GPU". GMI Cloud therefore leaves the H100 SXM Price Index panel from the week of 2026-09-14; closed weeks keep its quote, and the base-100 index moves only on price changes of providers present in both weeks. The definition of the published figure did not change. | unchanged | |
| 1.25.1 | patch | Coverage: AMD cards on more price lists. MI300X is read at Azure (ND MI300X v5, $48.00 for an 8-GPU VM, $6.00 per GPU-hour, with its spot and 1- and 3-year rates), Oracle ($6.00 per GPU-hour) and Hot Aisle, an AMD-only cloud read for the first time ($2.99 per GPU-hour for its pay-as-you-go VMs; its bare-metal node, a one-month commitment priced above the VM, is not read). MI355X is tracked from today, priced by Oracle ($8.60) and by DigitalOcean spot ($4.50), whose spot cards are now read for the AMD cards; a spot card printed at the on-demand price (B300) is not published. The definition of the published figure did not change. | unchanged | |
| 1.25.0 | minor | The weekly Price Index series carries the middle half of each week: p25_usd_per_gpu_hour and p75_usd_per_gpu_hour in series[] of price-index.json, two new last columns of price-index.csv and a column of the weekly table on /price-index - the 25th and 75th percentile of the quotes in the week's receipt, the same quotes its median was taken over. Until now only the current week had it. By the owner's decision it was computed for every published week, closed weeks included, from the receipts those weeks already published (week of 2026-09-07: $3.12-$6.00; 2026-09-14: $2.89-$5.95). No published median moved; before adding the band, each week's receipt was checked to reproduce its published median. | unchanged | |
| 1.24.3 | patch | Coverage: the first AMD cards. AMD Instinct MI300X is read at Crusoe ($3.45), DigitalOcean GPU Droplets ($2.59, $1.91 for 12 months), RunPod Secure Cloud ($2.39) and Seeweb (EUR 2.35); MI325X at DigitalOcean ($3.80); MI350X at DigitalOcean ($6.16) and CloudRift ($3.65) - per GPU-hour, on-demand, each read from a page the parsers already opened and until now dropped as unmapped. RunPod's API also returns $0.50 for the MI300X on Community Cloud; no RunPod page prints that figure (the card page reads "from $2.39/hr"), so it is not published. MI355X is not tracked: its one source says "Contact sales". The Price Index is an H100 series and is untouched. The definition of the published figure did not change. | unchanged | |
| 1.24.2 | patch | Coverage: Vast.ai V100 16 GB is read - the marketplace lists 16 GB and 32 GB cards under one name, "Tesla V100", so only offers below 24 GB count. On this date the single-GPU floor was withheld (two verified offers, a floor needs three) and the multi-GPU quote was published ($0.48 for a 4-GPU machine, $0.12 per GPU). Marketplace floors stay out of every median. The definition of the published figure did not change. | unchanged | |
| 1.24.1 | patch | Coverage: Seeweb is read for the first time - listed in euro per GPU, converted at the ECB rate like Scaleway and OVHcloud: H100 SXM EUR 1.89, H200 EUR 2.60, A100 80GB EUR 0.99, L40S EUR 0.85, RTX PRO 6000 EUR 1.25, RTX A6000 EUR 0.74, L4 EUR 0.38 per GPU-hour on-demand, each with its 12-month rate. That the figure is per GPU was read from the page's own price script, which multiplies the rate by the GPU count. Spot is not read: the page promises up to 40% off while its script charges 40% of the price. EUR listings stay out of the Price Index. The definition of the published figure did not change. | unchanged | |
| 1.24.0 | minor | The risk tag the rate tables print is now documented and published as data: /methodology#risk says what it measures (whether capacity can be taken back mid-run, read off the market: spot and community high, serverless medium, on-demand and reserved low) and what it does not (provider reliability, stock, the cost of a commitment billed for idle hours), and prices.json and prices.csv carry it as risk. A new field; no published figure changed. | unchanged | |
| 1.23.2 | patch | Coverage: Northflank, Cerebrium and Baseten are read for the first time. Northflank: GPU-only per hour - L4 $0.80, A100 40GB and 80GB, RTX PRO 6000. Cerebrium: serverless per second, GPU only - L4 $0.80 per hour, A100 40GB and 80GB, L40S, RTX PRO 6000, H200, B200. Baseten: dedicated deployments per minute, instance - L4 $0.85 per hour, A100 80GB, B200. The L4 goes from seven rate cards to ten. All three list an H100 without its form factor, which is not recorded. Exoscale is recorded as not listed: its price file gives the same figures under EUR, CHF and USD. The definition of the published figure did not change. | unchanged | |
| 1.23.1 | patch | Coverage: Vast.ai lists the A100 under two names, "A100 SXM4" and "A100 PCIE"; both are now read as one book per memory size, as other providers' SXM and PCIe A100 already count as one card. The A100 80GB floor, withheld that morning because only two verified 80 GB SXM4 offers were rentable, is published again ($1.00 per GPU-hour single, on offers filtered to 80 GB by the marketplace). The definition of the published figure did not change. | unchanged | |
| 1.23.0 | minor | Beside the median rate card, from four rate cards: the middle half - the 25th and 75th percentile of the same rate cards, interpolated as percentile_cont, so the 50th is the median itself - on /spread, each card page, gpus.json (with the median and the rate-card count, new there too) and, for the index quotes, price-index.json. /spread now takes its rates from the same rate cards as every other card median: until this date it kept its own filter and counted community capacity (RunPod Community Cloud) as on-demand, which 1.18.0 had already removed everywhere else, so its cheapest rate and median could differ from the card pages. No other published figure changed. | changed | |
| 1.22.1 | patch | Coverage: the Vast.ai floor for the RTX PRO 6000 is read for the first time ($1.34 per GPU-hour single, $1.33 per GPU on an 8-GPU machine). The parser had asked the marketplace for "RTX PRO 6000", a name it does not use; it lists the card as "RTX PRO 6000 S" and "RTX PRO 6000 WS", read now as one book, while the lower-power Max-Q edition stays out. A check of every name the parser reads the same day found no other card sold under one name in two memory sizes besides the A100 and RTX 4090 already separated that morning. The definition of the published figure did not change. | unchanged | |
| 1.22.0 | minor | A new file: /changelog.xml, an Atom feed of providers' own rate changes - the newest 100 moves from changelog.json, one entry per move with a stable id, linked to the card page and the source. The marketplace floor (Vast.ai) is left out of the feed because it moves almost every hour; on this date the newest 100 moves were all Vast.ai floors. No published figure changed. | unchanged | |
| 1.21.3 | patch | Coverage: E2E Networks and Koyeb are read for the first time. E2E: on-demand per GPU-hour, before taxes - L4 $0.57, L40S, A100 80GB and 40GB, RTX PRO 6000, H200 and B200. Koyeb: serverless GPU instances with CPU, memory and disk included - L4 $0.70, RTX A6000, L40S, A100 80GB, RTX PRO 6000, H200 and B200. The L4 goes from five rate cards to seven. Both list an H100 without its form factor, which is not recorded. The definition of the published figure did not change. | unchanged | |
| 1.21.2 | patch | Coverage: Atlantic.Net and Denvr Dataworks are read for the first time. Atlantic.Net: H100 NVL ($3.94 per GPU-hour) and L40S ($1.67) servers, on-demand and on 1- and 3-year terms, the hourly figure as the page prints it (monthly price / 730). Denvr: on-demand GH200 ($3.87), H100 SXM ($2.45), A100 80GB ($1.45) and A100 40GB ($1.15) per GPU-hour. The GH200 gets a third rate card and the H100 NVL a fourth; Denvr's H100 SXM rate enters the Price Index from the next hourly scrape. The definition of the published figure did not change. | unchanged | |
| 1.21.1 | patch | Coverage: two Azure series the Retail Prices API already returned are read - NC H100 v5, one or two H100 NVL (94 GB) per VM ($6.98 per GPU-hour on-demand in eastus, spot and 1- and 3-year reservations), and NCv3, one to four V100 16 GB per VM ($3.06 per GPU-hour). The H100 NVL, until now on two providers' rate cards, gets a third. The definition of the published figure did not change. | unchanged | |
| 1.21.0 | minor | price-index.json carries the typical discount of /price-index#discounts as market_discounts: per term (spot, 1-year, 3-year) its status, median, range, pairs and providers, and every rate pair behind it with both rates, the difference in $ and % and the source link - withheld terms included, with their figures null and the reason. New fields; no published figure changed. | unchanged | |
| 1.20.0 | minor | A new aggregate figure on /price-index: the typical discount of spot and committed rates against the same provider's on-demand rate for the same card, region and price scope - the median of those per-quote differences across every card, with the number of rate pairs and providers behind it, and a term stated only with three or more providers. On this date: spot -50% (33 pairs, 5 providers), 1-year commitments -31% (9 pairs, 3 providers); 3-year commitments withheld, published by Azure and Google Cloud only. Rates with no stated term are not one term and are left out. The Price Index is unchanged. | unchanged | |
| 1.19.0 | minor | Two new cuts of the same quotes on card and provider pages. Every reserved and spot rate shows its difference, in dollars and per cent, from the same provider's on-demand rate for the same card, region and price scope (Azure H100 SXM 3-year $4.86: -$6.21, -56% against its $11.06); a rate with no such on-demand rate beside it shows none. Card pages state the cheapest and dearest on-demand quote priced as a whole 8-GPU node (H100 SXM: $29.23 Scaleway to $88.49 Azure per node-hour), and eight GPUs at the cheapest on-demand rate beside it. No published figure changed. | unchanged | |
| 1.18.1 | patch | Coverage: CloudRift and Beam are read for the first time. CloudRift: on-demand RTX 4090 ($0.39), RTX 5090 ($0.60), A100 80GB ($1.05) and V100 per GPU-hour, and a 1-month reserved rate for those and for L40S, RTX PRO 6000 and H200, which it sells reserved only. Beam: on-demand machines "from" a price (RTX 4090 $0.44, RTX 5090 $0.72, and B200, H200, H100 PCIe, A100 80GB, RTX PRO 6000, L40S, RTX A6000) and serverless GPU seconds for RTX 4090, RTX 5090 and H100 PCIe. The RTX 4090 and RTX 5090, until now on one provider's rate card, get a median rate card. Neither lists the H100 SXM, so the Price Index does not move. The definition of the published figure did not change. | unchanged | |
| 1.18.0 | minor | The median rate card takes on-demand rates proper, as "cheapest on-demand" has since 1.17.0: community capacity (RunPod Community Cloud) leaves it with marketplace floors, so the median and the count of providers listing lower on a provider page rest on the same providers. Where a provider had only community capacity for a card it drops out of that card's median. Moved on this date: A100 40GB $2.10 across 9 providers to $2.42 across 8; V100 $0.51 across 4 to $0.79 across 3; H100 NVL $2.75 to $3.05; B300 $7.73 to $7.85; B200 $6.69 to $6.79; RTX PRO 6000 $1.87 to $1.99; H200 $4.50 to $4.54; RTX 6000 Ada $0.94 to $0.96. The cheapest rate card moves with it (RTX 4090 $0.34 to $0.74, RTX 5090 $0.69 to $0.99, both RunPod Secure Cloud), and so does break-even at the median. H100 SXM ($3.90), A100 80GB, L40S and the Price Index are unchanged. | changed | |
| 1.17.0 | minor | "Cheapest on-demand" and "who is cheaper" rest on on-demand rates proper: community capacity (RunPod Community Cloud, on third-party hosts) and marketplace floors (Vast.ai) are listed as on-demand at the source but are shown in their own column on /cheapest and /compare, never decide a verdict on /compare/{a}-vs-{b}, and are no longer the cheapest rate on /buy-vs-rent. Card pages already labelled them community, high risk; the Lambda vs RunPod B200 verdict had been decided by a Community Cloud rate. Pair pages stay; a shared GPU resting on community capacity is shown and not counted. The median rate card and the Price Index are unchanged. | changed | |
| 1.16.0 | minor | Every retail price says what kind of price it is - a store's own price, a marketplace seller's listing, or a completed sale - and a retail median is computed inside one kind, never across. A card leads with the store median where three or more stores carry it, else the marketplace median, each named, with the other kind beside it. Three medians changed: A100 80GB $15,760 over 14 sellers of both kinds became $11,986 at 7 stores and $16,999 at 7 marketplace sellers; RTX 4090 $3,195 became $3,103 and $3,200; RTX 5090 $4,841 became $4,753 and $5,250. Break-even on /buy-vs-rent and the card pages moves with them. Rental figures, prices.json and the Price Index are untouched. | changed | |
| 1.15.0 | minor | A median rate card is published only with two or more providers. A card on one provider's rate card (RTX 4090 and RTX 5090 on this date, where most supply is the Vast.ai floor) shows that rate card by name and the cheapest marketplace floor beside it, and gets no break-even, no median column on /buy-vs-rent and no median comparison on /gpus - a median of one is a provider's price, not a market figure. | unchanged | |
| 1.14.0 | minor | One card median across the site, the median rate card - each provider once, at its cheapest on-demand rate, marketplace floors out - printed on card pages, /buy-vs-rent, the GPU comparisons and their JSON-LD, as the cards table and /spread already did. It replaces a median over every on-demand listing, which counted a provider once per region or tier and took marketplace floors in (H100 SXM: $3.85 to $3.90). The Price Index is unchanged. | changed | |
| 1.13.0 | minor | The weighting rule is stated where the median is printed - one quote per provider and region, so a provider that lists several regions enters once per region - and every week carries a check next to the index: the same quotes as the median of each provider's own median, with the largest provider's quote count and an HHI, in price-index.json as weighting_check and on /price-index. The index keeps its definition; no published figure changed. | unchanged | |
| 1.12.0 | minor | Every live quote carries its change over seven days: the listed figure per GPU against the same quote's read seven days before the newest scrape (within 48 hours of that moment), from our own reads only, in prices.json and its CSV as listed_7d_ago, read_7d_ago_at and change_7d_pct. A card's change is the geometric mean of its rate cards' own changes. New fields; no published figure changed. | unchanged | |
| 1.11.0 | minor | Vast.ai multi-GPU machines are read as a second quote per GPU model, region "Multi-GPU machines": the machine's own price with its GPU count, the per-GPU figure that price divided by the count, from the machine whose per-GPU ask is the median of the five cheapest, under the same guards. The single-GPU floor did not change and never mixes with it, because a share of a multi-GPU machine is not a machine you can rent alone. | unchanged | |
| 1.10.2 | patch | Coverage: Massed Compute is read for the first time - on-demand H100 SXM ($2.89), H100 NVL, A100 80GB, L40S, L40, RTX PRO 6000, RTX 6000 Ada and RTX A6000 from its VM pricing page. Headings that name no form factor (a bare H100 80GB) or a card we do not track as such (H200 NVL) are left out rather than guessed. Its H100 SXM rate enters the Price Index from the next hourly scrape. The definition of the published figure did not change. | unchanged | |
| 1.10.1 | patch | Coverage: Crusoe is read for the first time - on-demand H100 SXM ($3.90), H200, A100 80GB and L40S from its pricing table; GB200 and B200 are contact-sales there and are not recorded. Its H100 rate enters the Price Index from the next hourly scrape, as any provider's would. The definition of the published figure did not change. | unchanged | |
| 1.10.0 | minor | A marketplace floor withheld as too thin is recorded with the read that withheld it, and the quote is marked thin_market from that read on: it keeps its last floor with that floor's read time, leaves every computed figure at once instead of counting for up to 48 hours, and never ages into unconfirmed or stopped while the source is still being read. prices.json and its CSV gain withheld_at and withheld_reason. Which quotes count as current changed; no published definition did. | unchanged | |
| 1.9.0 | minor | The Price Index is also published cut by operator class: the same weekly median over the same quotes, split into hyperscalers (AWS, Azure, Google Cloud, Oracle) and specialist GPU clouds, each with its quotes and providers - by_operator_class in price-index.json, four columns in price-index.csv, a table on /price-index. Marketplaces have no value, because their floors never enter the index. Every provider carries its class (providers.operator_class, filled by hand). A new cut, not a new definition: the published value did not move. | unchanged | |
| 1.8.0 | minor | Each weekly receipt of the Price Index lists every H100 SXM on-demand quote read for that week, not only the ones that made the value: the quotes kept out carry included=false and, in a new excluded_by column, the rule that kept them out - marketplace_floor, eur_listing or not_confirmed_48h, the codes price-index.json publishes as exclusion_rules. The reproduce commands read the included rows, so no published figure moved. | unchanged | |
| 1.7.0 | minor | The home and card pages print one modelled figure for the first time: cost per million output tokens, a listed rate divided by a throughput NVIDIA published for that card, with the run that produced it (model, framework, precision, sequence, concurrency, date) beside it and a link to the benchmark. It is published per card and never as a ranking, because NVIDIA runs each card in its own configuration. Every other figure on the site is still read from a provider page. | unchanged | |
| 1.6.1 | patch | Coverage: the committed-use rates of Google Cloud and Azure are read for the first time, one-year and three-year, from the same endpoints as their on-demand prices. Azure publishes a reservation as the total for its term, so the hourly figure is that total divided by the hours in the term (8,760 per year). AWS still publishes only on-demand where we read it. No existing figure changed. | unchanged | |
| 1.6.0 | minor | Every quote says which market it is from: on-demand, spot, community (shared capacity - RunPod Community Cloud and the Vast.ai floor), serverless (billed by the second, no instance to rent - Modal), or a commitment with a stated term (Paperspace, one year) or without one (Nebius, Hyperstack, Together). Read from each source, never inferred from the price; Together was checked against its own page and sells on-demand clusters, not serverless. The Price Index still selects exactly as before, so no published figure moved. | unchanged | |
| 1.5.0 | minor | Every quote carries the name the provider gives what was priced - instance type, SKU, machine family or marketplace offer id (p5.48xlarge, HGX B200, offer 48635744) - in prices.json, its slices and prices.csv. Fifteen of twenty-one sources publish one; the rest list a single rate per card and the field stays empty rather than being derived. No published figure changed. | unchanged | |
| 1.4.0 | minor | Every published figure carries methodology_version: top level in price-index.json and prices.json and their slices, a column in prices.csv, price-index.csv, price-index-history.csv and in each weekly receipt. This dated log is published beside it. | unchanged | |
| 1.3.2 | patch | A Vast.ai marketplace floor is withheld when the window it came from cannot carry one: fewer than three verified single-GPU offers, or a median more than 3x the cheapest ask beside it. On 2026-09-11 that window published $53.34 per RTX 5090 GPU-hour. Marketplace floors never enter the index, so no published index week moved. | unchanged | |
| 1.3.1 | patch | Collection and coverage only: RunPod spot rows that repeated the on-demand price are no longer published (the API returned spot == on-demand for every GPU and tier), and Sesterce was added as the 21st provider. The definition of the published figure did not change. | unchanged | |
| 1.3.0 | minor | A quote not read again within 48 hours of the newest scrape stops counting as current: it is marked unconfirmed (stopped after 7 days), leaves every computed figure including the running index week, and carries last_confirmed_at and status in the files. Each published week ships its receipt, every live row says since when the figure has been listed, and a withheld week prints why. | unchanged | |
| 1.2.0 | minor | A chain-linked base-100 index is published beside the dollar median, so a change can be read without the panel moving it, and the reconstructed SkyPilot-catalog history is published as a separate, second-hand series that is never merged with our own. | unchanged | |
| 1.1.0 | minor | A week is published only when at least 8 providers contributed; weeks below that are withheld, never backfilled. Every quote declares its price_scope, marketplace floors and quotes listed in a currency other than USD are shown on the site but kept out of the index, with the listed price and the ECB rate beside the converted figure. | unchanged | |
| 1.0.0 | base | First published week of our own series: the Cloud GPU Price Index is the median across providers of the latest publicly listed on-demand NVIDIA H100 SXM rate, per GPU per hour, node prices divided by GPU count, each quote linked to the page it was read from. | unchanged |
The log starts on , the first published week of our own series; the entries up to September 30, 2026 were reconstructed from the repository while that history was six days long, and every change from here is recorded as it ships. The reconstructed history from the SkyPilot catalog is a separate, second-hand series and carries its own description on theindex page — it is not versioned here.
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 figure together with its methodology_version; version 1.30.17 is what the files carry today.
| Plain | GPU Economy, "Methodology changelog", September 29, 2026. https://gpueconomy.com/methodology/changelog |
|---|---|
| Markdown | [GPU Economy, "Methodology changelog" (September 29, 2026)](https://gpueconomy.com/methodology/changelog) |
| APA | GPU Economy. (2026-09-29). Methodology changelog. Retrieved from https://gpueconomy.com/methodology/changelog |
Data updated: