GPU price & market log

GPU rental price changes, market milestones and corrections, with dated evidence for each entry.

117 logged entries · 9 archive months · Source archives through .

This week on the desk

Week ending Nov 22, 2026 on the desk: 7 curve sessions archived, 1 venue-set change behind a class aggregate, 7 provider price changes and 1 correction logged. Thin markets this session — no reliable move. Thin markets this session — no reliable curve shape for B200, H200, H100, A100 and RTX 5090 (each has a thin or interpolated point among its anchor, Nov ’26 and Aug ’27 ladders).

Composed from the week’s logged events and the canonical curve notes · week ending 2026-11-22

Showing the 40 newest of 117 entries. Filters apply to these entries; open a month for its complete record. Corrections use the date they were recorded, which can be later than the source archives.

Event date range

Inclusive event dates, which can differ from evidence or archive snapshot dates.

Showing 40 of 40 events on this page

  • Price rise Verda raised 7 posted GPU prices by 2.5% to 4.5% in the week of Nov 16, 2026Verda · A100 80GB, B200, B300, H100, H200, L40S, V100

    In the week of Nov 16, 2026, Verda changed these posted prices in steps of less than 5% each. First and last price per GPU-hour and net change: A100 80GB, 1-GPU instance 1A100.22V: $2.386 → $2.446, +2.5% in 4 changes between Nov 16 and Nov 20, 2026; B200, 1-GPU instance 1B200.30V: $9.508 → $9.796, +3.0% in 3 changes between Nov 16 and Nov 19, 2026; B300, 1-GPU instance 1B300.30V: $12.41 → $12.92, +4.1% in 4 changes between Nov 16 and Nov 20, 2026; H100, 1-GPU instance 1H100.80S.30V: $5.241 → $5.453, +4.0% in 4 changes between Nov 16 and Nov 20, 2026; H200, 1-GPU instance 1H200.141S.44V: $6.701 → $6.973, +4.1% in 4 changes between Nov 16 and Nov 20, 2026; L40S, 1-GPU instance 1L40S.20V: $2.018 → $2.079, +3.0% in 3 changes between Nov 16 and Nov 19, 2026; V100, 1-GPU instance 1V100.4V: $0.242 → $0.253, +4.5% in 5 changes between Nov 11 and Nov 20, 2026. One includes smaller changes from before that week, which had not yet added up to 2%. Posted list prices, not transactions.

    Source: https://api.verda.com/v1/instance-types

    Permalink for Verda raised 7 posted GPU prices by 2.5% to 4.5% in the week of Nov 16, 2026

  • Venue change B300 typical price: Verda added between the Oct 20 and Nov 16, 2026 quote snapshotsB300 · typical price

    The venue set changed between the Oct 20 and Nov 16, 2026 quote snapshots: Verda entered at $12.41. The archived typical price moved $9.50 → $10.50 (+10.5%), which is not a like-for-like price change. Like-for-like, the 3 venues quoted in both snapshots moved 0.0% at the median ($9.50 → $9.50).

    Derived from archived quote days 2026-10-20 · 2026-11-16

    Permalink for B300 typical price: Verda added between the Oct 20 and Nov 16, 2026 quote snapshots

  • Correction Correction: Compare pages, badges and the reference-price file kept the Oct 20 quote tableGPU quotes

    Was: From Nov 2 to Nov 16, 2026 the production site kept building its quote table from the copy collected on Oct 20, 2026 (51 rows), although the weekly refreshes of Nov 2, 9 and 16 had each published a newer one. The compare pages, the chip badges and social cards, the reference figures in /reference-prices.json, the quote table in /llms-full.txt and the quote snapshot in /status.json showed the older table: the H100 typical price read $3.99 (median of 4 specialist clouds, observed Oct 14 to Oct 20) and Lambda's H100 rate read $3.99. Corrected to: The production build was set to read the published weekly quote export, and the site was rebuilt on Nov 16, 2026 from the table collected that day: 62 rows, with Verda, Jarvislabs and Hot Aisle among them. The H100 typical price read $4.25 (median of 6 specialist clouds, observed Nov 16) and Lambda's H100 rate read $4.59 on those surfaces. Where: /compare/ pair pages (titles, descriptions and reference figures); /badge/ chip and site badges and the chip social cards; /reference-prices.json reference figures and citations; /llms-full.txt quote and series tables; /status.json quoteSnapshot. Found by a review of published pages and data.

    Evidence: Quote snapshot stamp and row count · Reference prices · Lambda pricing · Full record

    Permalink for Correction: Compare pages, badges and the reference-price file kept the Oct 20 quote table

  • Price rise Lambda raised its posted A100 80GB price +14.3% to $3.19 per GPU-hour between Nov 2 and Nov 15, 2026A100 80GB · Lambda · 8-GPU instance

    Lambda's posted price for this A100 80GB offer was $2.79 per GPU-hour when last recorded on Nov 2, 2026 and $3.19 when next recorded on Nov 15, 2026 (+14.3%). It was not recorded on the 12 days in between, so the change took effect at some point in that window. The offer: 8-GPU instance; NVIDIA HGX A100 80GB; 240 vCPUs, 1800 GiB RAM and 19.5 TiB SSD storage per instance; published per GPU-hour. Region: Provider regions vary. Lambda renamed the product from “NVIDIA A100 SXM 80GB” to “NVIDIA HGX A100 80GB”. It is the same 8-GPU instance: 240 vCPUs, 1800 GiB RAM and 19.5 TiB SSD storage. The earlier price was read from lambda.ai, the later one from lambda.com. A posted list price, not a transaction.

    Source: https://lambda.com/pricing

    Permalink for Lambda raised its posted A100 80GB price +14.3% to $3.19 per GPU-hour between Nov 2 and Nov 15, 2026

  • Price rise Lambda raised its posted B200 price +14.9% to $7.69 per GPU-hour between Nov 2 and Nov 15, 2026B200 · Lambda · 8-GPU instance

    Lambda's posted price for this B200 offer was $6.69 per GPU-hour when last recorded on Nov 2, 2026 and $7.69 when next recorded on Nov 15, 2026 (+14.9%). It was not recorded on the 12 days in between, so the change took effect at some point in that window. The offer: 8-GPU instance; NVIDIA HGX B200 180GB; 208 vCPUs, 2900 GiB RAM and 22 TiB SSD storage per instance; published per GPU-hour. Region: Provider regions vary. Lambda renamed the product from “NVIDIA B200 SXM6 180GB” to “NVIDIA HGX B200 180GB”. It is the same 8-GPU instance: 208 vCPUs, 2900 GiB RAM and 22 TiB SSD storage. The earlier price was read from lambda.ai, the later one from lambda.com. A posted list price, not a transaction.

    Source: https://lambda.com/pricing

    Permalink for Lambda raised its posted B200 price +14.9% to $7.69 per GPU-hour between Nov 2 and Nov 15, 2026

  • Price rise Together AI raised its posted B200 price +13.4% to $9.29 per GPU-hour between Nov 10 and Nov 15, 2026B200 · Together AI · NVIDIA HGX B200 GPU Cluster

    Together AI's posted price for this B200 offer was $8.19 per GPU-hour when last recorded on Nov 10, 2026 and $9.29 when next recorded on Nov 15, 2026 (+13.4%). It was not recorded on the 4 days in between, so the change took effect at some point in that window. The offer: NVIDIA HGX B200 GPU Cluster; published per GPU-hour. Region: Provider regions vary. A posted list price, not a transaction.

    Source: https://www.together.ai/pricing

    Permalink for Together AI raised its posted B200 price +13.4% to $9.29 per GPU-hour between Nov 10 and Nov 15, 2026

  • Price rise Together AI raised its posted B300 price +15.0% to $11.49 per GPU-hour between Nov 10 and Nov 15, 2026B300 · Together AI · NVIDIA HGX B300 GPU Cluster

    Together AI's posted price for this B300 offer was $9.99 per GPU-hour when last recorded on Nov 10, 2026 and $11.49 when next recorded on Nov 15, 2026 (+15.0%). It was not recorded on the 4 days in between, so the change took effect at some point in that window. The offer: NVIDIA HGX B300 GPU Cluster; published per GPU-hour. Region: Provider regions vary. A posted list price, not a transaction.

    Source: https://www.together.ai/pricing

    Permalink for Together AI raised its posted B300 price +15.0% to $11.49 per GPU-hour between Nov 10 and Nov 15, 2026

  • Price rise Lambda raised its posted H100 price +15.0% to $4.59 per GPU-hour between Nov 2 and Nov 15, 2026H100 · Lambda · 8-GPU instance

    Lambda's posted price for this H100 offer was $3.99 per GPU-hour when last recorded on Nov 2, 2026 and $4.59 when next recorded on Nov 15, 2026 (+15.0%). It was not recorded on the 12 days in between, so the change took effect at some point in that window. The offer: 8-GPU instance; NVIDIA HGX H100 80GB; 208 vCPUs, 1800 GiB RAM and 22 TiB SSD storage per instance; published per GPU-hour. Region: Provider regions vary. Lambda renamed the product from “NVIDIA H100 SXM 80GB” to “NVIDIA HGX H100 80GB”. It is the same 8-GPU instance: 208 vCPUs, 1800 GiB RAM and 22 TiB SSD storage. The earlier price was read from lambda.ai, the later one from lambda.com. A posted list price, not a transaction.

    Source: https://lambda.com/pricing

    Permalink for Lambda raised its posted H100 price +15.0% to $4.59 per GPU-hour between Nov 2 and Nov 15, 2026

  • Price rise Verda raised 6 posted GPU prices by 3.0% to 6.2% in the week of Nov 9, 2026Verda · A100 80GB, B200, B300, H100, H200, L40S

    In the week of Nov 9, 2026, Verda changed these posted prices in steps of less than 5% each. First and last price per GPU-hour and net change: A100 80GB, 1-GPU instance 1A100.22V: $2.316 → $2.386, +3.0% in 4 changes between Nov 11 and Nov 15, 2026; B200, 1-GPU instance 1B200.30V: $9.138 → $9.508, +4.0% in 4 changes between Nov 11 and Nov 15, 2026; B300, 1-GPU instance 1B300.30V: $11.69 → $12.41, +6.2% in 6 changes between Nov 9 and Nov 15, 2026; H100, 1-GPU instance 1H100.80S.30V: $4.938 → $5.241, +6.1% in 6 changes between Nov 9 and Nov 15, 2026; H200, 1-GPU instance 1H200.141S.44V: $6.44 → $6.701, +4.1% in 4 changes between Nov 11 and Nov 15, 2026; L40S, 1-GPU instance 1L40S.20V: $1.901 → $2.018, +6.2% in 6 changes between Nov 2 and Nov 15, 2026. One includes smaller changes from before that week, which had not yet added up to 2%. Posted list prices, not transactions.

    Source: https://api.verda.com/v1/instance-types

    Permalink for Verda raised 6 posted GPU prices by 3.0% to 6.2% in the week of Nov 9, 2026

  • Curve sessions 7 forward-curve sessions archived · week of Nov 9, 2026Forward board

    Archived sessions 2026-11-09, 2026-11-10, 2026-11-11, 2026-11-12, 2026-11-13, 2026-11-14 and 2026-11-15 — each at its dated permalink; notable moves are logged separately.

    Derived from archived curve sessions 2026-11-09 · 2026-11-10 · 2026-11-11 · 2026-11-12 · 2026-11-13 · 2026-11-14 · 2026-11-15 · Data: Kalshi (KalshiEX LLC)

    Permalink for 7 forward-curve sessions archived · week of Nov 9, 2026

  • Price rise Verda raised 6 posted GPU prices by 2.5% to 3.0% in the week of Nov 2, 2026Verda · A100 80GB, B200, B300, H100, H200, V100

    In the week of Nov 2, 2026, Verda changed these posted prices in steps of less than 5% each. First and last price per GPU-hour and net change: A100 80GB, 1-GPU instance 1A100.22V: $2.248 → $2.316, +3.0% in 3 changes between Nov 2 and Nov 7, 2026; B200, 1-GPU instance 1B200.30V: $8.869 → $9.138, +3.0% in 3 changes between Nov 2 and Nov 7, 2026; B300, 1-GPU instance 1B300.30V: $11.35 → $11.69, +3.0% in 3 changes between Nov 2 and Nov 7, 2026; H100, 1-GPU instance 1H100.80S.30V: $4.793 → $4.938, +3.0% in 3 changes between Nov 2 and Nov 7, 2026; H200, 1-GPU instance 1H200.141S.44V: $6.25 → $6.44, +3.0% in 3 changes between Nov 2 and Nov 7, 2026; V100, 1-GPU instance 1V100.4V: $0.236 → $0.242, +2.5% in 3 changes between Nov 2 and Nov 7, 2026. Posted list prices, not transactions.

    Source: https://api.verda.com/v1/instance-types

    Permalink for Verda raised 6 posted GPU prices by 2.5% to 3.0% in the week of Nov 2, 2026

  • Price rise Verda raised 6 posted GPU prices by 4.6% to 5.1% in the week of Oct 26, 2026Verda · A100 80GB, B200, B300, H100, H200, L40S

    In the week of Oct 26, 2026, Verda changed these posted prices in steps of less than 5% each. First and last price per GPU-hour and net change: A100 80GB, 1-GPU instance 1A100.22V: $2.149 → $2.248, +4.6% in 5 changes between Oct 26 and Oct 31, 2026; B200, 1-GPU instance 1B200.30V: $8.439 → $8.869, +5.1% in 5 changes between Oct 26 and Oct 31, 2026; B300, 1-GPU instance 1B300.30V: $10.80 → $11.35, +5.1% in 5 changes between Oct 26 and Oct 31, 2026; H100, 1-GPU instance 1H100.80S.30V: $4.56 → $4.793, +5.1% in 5 changes between Oct 26 and Oct 31, 2026; H200, 1-GPU instance 1H200.141S.44V: $5.947 → $6.25, +5.1% in 5 changes between Oct 26 and Oct 31, 2026; L40S, 1-GPU instance 1L40S.20V: $1.809 → $1.901, +5.1% in 5 changes between Oct 26 and Oct 31, 2026. Posted list prices, not transactions.

    Source: https://api.verda.com/v1/instance-types

    Permalink for Verda raised 6 posted GPU prices by 4.6% to 5.1% in the week of Oct 26, 2026

  • Curve sessions 6 forward-curve sessions archived · week of Oct 26, 2026Forward board

    Archived sessions 2026-10-26, 2026-10-27, 2026-10-28, 2026-10-29, 2026-10-30 and 2026-10-31 — each at its dated permalink; notable moves are logged separately.

    Derived from archived curve sessions 2026-10-26 · 2026-10-27 · 2026-10-28 · 2026-10-29 · 2026-10-30 · 2026-10-31 · Data: Kalshi (KalshiEX LLC)

    Permalink for 6 forward-curve sessions archived · week of Oct 26, 2026

  • Venue change MI300X typical price: Hot Aisle added, Vultr removed between the Oct 11 and Oct 25, 2026 quote snapshotsMI300X · typical price

    The venue set changed between the Oct 11 and Oct 25, 2026 quote snapshots: Hot Aisle entered at $2.99 and Vultr ($3.99) left. The archived typical price moved $3.99 → $2.99 (−25.1%), which is not a like-for-like price change. No venue was quoted in both quote snapshots, so there is no like-for-like move.

    Derived from archived quote days 2026-10-11 · 2026-10-25

    Permalink for MI300X typical price: Hot Aisle added, Vultr removed between the Oct 11 and Oct 25, 2026 quote snapshots

  • Curve sessions 7 forward-curve sessions archived · week of Oct 19, 2026Forward board

    Archived sessions 2026-10-19, 2026-10-20, 2026-10-21, 2026-10-22, 2026-10-23, 2026-10-24 and 2026-10-25 — each at its dated permalink; notable moves are logged separately.

    Derived from archived curve sessions 2026-10-19 · 2026-10-20 · 2026-10-21 · 2026-10-22 · 2026-10-23 · 2026-10-24 · 2026-10-25 · Data: Kalshi (KalshiEX LLC)

    Permalink for 7 forward-curve sessions archived · week of Oct 19, 2026

  • Correction Correction: Grok estimates omitted documented batch and long-context cache ratesToken prices

    Was: Grok 4.7, 4.6 and 4.5 had no recorded batch rates, and no Grok row had a combined batch cache-hit rate or long-context discounts. The estimator used its disclosed full-rate fallback for missing rates and said batch and caching were not stacked. The earlier reproduced Grok 4.7 workload of 100 million input tokens, no output, 100% batch and 100% cache therefore estimated $200. Corrected to: A scoped reading of xAI's pricing page recorded the stated 20% batch discount for Grok 4.7, 4.6, 4.5 and 4.3, including cached tokens and long context. The estimator used only these explicit combined and tier-specific rates, giving $40 for that short-context Grok 4.7 workload. Grok Build 0.1 remained outside the batch discount list. All other models kept their observations and missing-discount fallbacks. Where: /tools/ API estimator; /ai-model-api-prices/ and Grok model details; /watchlist/ API estimates; /token-prices.json and /token-prices.csv. Found by a review of published pages and data.

    Evidence: xAI pricing and exact-model batch discounts · Recorded rates · Full record

    Permalink for Correction: Grok estimates omitted documented batch and long-context cache rates

  • Correction Correction: HUMAIN’s Kingdom-wide capacity target was classified as one campusBuildout

    Was: The HUMAIN AI factories row lacked a multi-site marker, so its up-to-500 MW target was classified as one Riyadh campus. Its retained NVIDIA claim and ledger note described a program across factories in Saudi Arabia. Corrected to: The row was explicitly classified as a multi-site program. Campus counts and single-campus rankings excluded it, while the program comparison retained its up-to-500 MW target, original claims, dates and stable identifier. No capacity figure, milestone, source publication date or dataset review date was changed. Where: /buildout campus counts, entry-type filters and comparisons; /buildout/humain-riyadh/ and buildout ledger exports. Found by a review of published pages and data.

    Evidence: NVIDIA: HUMAIN AI factories in Saudi Arabia · HUMAIN program’s retained capacity claims and sources · Full record

    Permalink for Correction: HUMAIN’s Kingdom-wide capacity target was classified as one campus

  • Correction Correction: A headline capex multiple compared guidance with actuals on different accounting basesFinancials

    Was: The Infrastructure financial summary presented combined 2026 guidance as approximately 1.8 times 2025 actuals. Amazon's retained guidance was net cash capex while its historical figures were gross purchases, and Microsoft's retained guidance included a lease reclassification. Detailed notes described the differences, but the headline multiple did not qualify its comparison. Corrected to: The headline multiple was removed. Individual company guidance retained its recorded ranges or point estimates, reporting periods, source dates and accounting bases beside the figures. Historical values were preserved, with basis changes disclosed rather than interpreted as like-for-like spending growth. No financial observation or dataset review date was changed. Where: /buildout financial summary, guidance comparison and chart explanations. Found by a review of published pages and data.

    Evidence: Company guidance, reporting periods and accounting basis · Full record

    Permalink for Correction: A headline capex multiple compared guidance with actuals on different accounting bases

  • Correction Correction: Hardware and rental FAQs selected older quotes than the current comparisonGPU quotes

    Was: Hardware and the rental comparison used different quote selectors, producing inconsistent MI300X and MI355X headlines. Some cheapest-provider FAQ answers also selected the older reviewed snapshot without identifying the differing price selection. The separate retired-provider correction records why the comparison’s then-selected Vultr offers were ineligible. Corrected to: Hardware rental economics, rental summaries, specification strips, cheapest-provider FAQs and their structured data were aligned to the comparison’s current eligible offer selection and original dates. Historical survey statements retained explicit dates. No provider observation or immutable price history was rewritten. Where: /hardware/ rental economics; /gpu-rental-prices/ detail summaries, FAQs and structured data. Found by a review of published pages and data.

    Evidence: Current rental offers · Hardware rental economics · Full record

    Permalink for Correction: Hardware and rental FAQs selected older quotes than the current comparison

  • Correction Correction: Custom GPU comparisons did not follow the rental index’s current price selectionGPU quotes

    Was: The custom comparison page derived rental prices only from the reviewed quote snapshot while the main GPU index and Watchlist could select newer eligible daily observations. A selected accelerator could therefore show a different reference price, coverage and observation date without identifying the differing data selection. Corrected to: The custom comparison’s reference, specialist and big-cloud prices were prepared from the GPU index’s exact current profiles, including their latest eligible provider observations, mixed daily and reviewed source dates, retained older prices and comparable historical changes. Deferred history was verified against every selected full profile before plotting. Reviewed marketplace quotes remained separate. No recorded source price, observation date or specification review date was rewritten. Where: /compare/ current rental prices, source dates and inline history; /chart-data/compare.json prepared comparison catalog. Found by a review of published pages and data.

    Evidence: GPU rental index’s current prices and dated coverage · Custom accelerator comparison · Full record

    Permalink for Correction: Custom GPU comparisons did not follow the rental index’s current price selection

  • Correction Correction: Homepage rental headlines used an older price selection than the comparisonGPU quotes

    Was: The homepage hero selected rental headlines from the reviewed quote snapshot while its rental teaser and the core comparison could select newer daily observations. On October 24 the H100 floor carried an October 14 observation in the hero and an October 23 observation in the teaser. Corrected to: The duplicated reviewed headline figures were replaced with a comparison entry point. Homepage offer cards were prepared from the core comparison’s exact current offers, including original dates, freshness, provider coverage and recorded instance terms. No source observation or historical price was rewritten. Where: / homepage rental headlines and offer cards. Found by a review of published pages and data.

    Evidence: Homepage current offer cards · Current rental comparison and original observations · Full record

    Permalink for Correction: Homepage rental headlines used an older price selection than the comparison

  • Correction Correction: Retired Vultr plans returned as the MI300X and MI355X typical pricesGPU quotes

    Was: Current rental prices accepted a provider missing from the reviewed quotes when its daily observation was timed after the review. The Oct 14, 2026 review that retired Vultr's non-deployable plans carries a date without a time, so Vultr's observations from 09:38 that day counted as newer. Its $3.99 MI300X and $2.59 MI355X plans became the only providers behind those chips' typical prices, on the watchlist, the chip pages and the GPU rental overview. Corrected to: A review recorded as a bare date was treated as covering its whole day, so a provider absent from it needed a collection on a later day to count as current. The Vultr plans left the current prices, and MI300X and MI355X used their remaining reviewed provider quotes. Exactly timed reviews kept the exact comparison. Where: MI300X and MI355X typical prices on /gpu-rental-prices and the chip pages; /watchlist GPU rows. Found by a review of published pages and data.

    Evidence: MI300X rental prices · MI355X rental prices · Full record

    Permalink for Correction: Retired Vultr plans returned as the MI300X and MI355X typical prices

  • Correction Correction: Watchlist reference prices did not follow the main rental index’s current data selectionGPU quotes

    Was: The Watchlist prepared GPU reference prices from the reviewed quote snapshot while the main GPU index could select newer eligible daily observations. The same GPU could therefore show a different reference price and date without identifying the differing source selection. API price review dates also appeared under a generic Observed label. Corrected to: Watchlist reference prices, coverage, observation dates and comparable changes were prepared from the main rental index’s exact selected profile. Eligible daily collections and dated reviewed fallbacks were distinguished. API review dates were labeled Reviewed. No source observation, recorded price, review date or historical billing condition was rewritten. Where: /watchlist reference prices, source labels and evidence dates; /chart-data/watchlist.json prepared public catalog. Found by a review of published pages and data.

    Evidence: Current GPU rental reference prices and dated coverage · Watchlist current profiles and source clocks · Full record

    Permalink for Correction: Watchlist reference prices did not follow the main rental index’s current data selection

  • Correction Correction: Two launch months and three purchase-price estimates lacked supportHardware specs

    Was: The B300's first availability read Jul 2025, the month the first GB300 NVL72 rack shipped, and the MI325X's read Dec 2024, when AMD had only said it would ship. The V100, A100 and MI300X purchase-price estimates ($13,000, $12,000 and $15,000) rested on no reputable report, and the Gaudi 3 row carried no estimate flag although its launch month is unconfirmed. Corrected to: The B300 moved to Sep 2025, when Supermicro began volume shipments of HGX B300 systems, and the MI325X to Feb 2025, when Vultr became the first cloud to offer it; both stayed estimates. The V100 and MI300X estimates were removed, the A100's became the roughly $10,000 CNBC reported in February 2023, and the Gaudi 3 row was flagged as an estimate. Where: /hardware spec table; llms-full.txt accelerator specs; /accelerators.json and /accelerators.csv (launch, est). Found by a review of published pages and data.

    Evidence: Supermicro: volume shipments of NVIDIA Blackwell Ultra systems (Sep 11, 2025) · Vultr: first cloud to offer AMD Instinct MI325X (Feb 18, 2025) · CNBC: Nvidia's A100 is the $10,000 chip powering the race for A.I. · Full record

    Permalink for Correction: Two launch months and three purchase-price estimates lacked support

  • Correction Correction: API estimates applied cache discounts to unsupported batch trafficToken prices

    Was: For Grok 4.7, 100 million monthly input tokens, zero output, 100% batch and 100% cache produced a $50 estimate despite the stated standard-rate fallback implying $200. The estimator also ranked workloads beyond recorded model context limits. Corrected to: Requested batch traffic retained its allocation when no batch rate was recorded and used the standard rate without an overlapping cache discount. The reproduced Grok workload estimated $200. Models exceeding their recorded context limit were excluded from priced rankings, while unknown limits remained explicitly unknown. Where: /tools/ API bill estimator; /ai-model-api-prices/ workload estimates. Found by a review of published pages and data. Corrected again on Oct 25, 2026: Grok estimates omitted documented batch and long-context cache rates.

    Evidence: API bill estimator and eligibility rules · Recorded model rates and limits · Full record

    Permalink for Correction: API estimates applied cache discounts to unsupported batch traffic

  • Correction Correction: An unsupported date was given for Claude Sonnet 5's standard priceToken prices

    Was: The token-price table said Claude Sonnet 5's introductory $2 and $10 per million input and output tokens were made its standard price on Aug 10, 2026. Anthropic's pricing page states the change but no date for it. Corrected to: The date was removed. The row kept the $2 and $10 standard price and the note that the rise to $3 and $15 planned for Sep 1, 2026 would not occur, both of which Anthropic's pricing page states. Where: /ai-model-api-prices/ table and Claude Sonnet 5 model page; /token-prices.json and /token-prices.csv; llms-full.txt token prices. Found by a review of published pages and data.

    Evidence: Anthropic API pricing · Full record

    Permalink for Correction: An unsupported date was given for Claude Sonnet 5's standard price

  • Correction Correction: Gemini 3.6 Flash's scheduled price rise was missing from the token-price tableToken prices

    Was: The token-price table, reviewed on Oct 11, 2026, listed Gemini 3.6 Flash at $0.75 per million input tokens and $3.75 per million output tokens with no note, although Google's pricing page scheduled the same change for it as for Gemini 3.8 Flash. Corrected to: The row was given Google's stated terms: $0.75 and $3.75 through Dec 31, 2026, then $1.50 and $7.50 from Jan 1, 2027, in its Terms and in the payloads' temporary-price fields. Where: /ai-model-api-prices/ table and Gemini 3.6 Flash model page; /token-prices.json and /token-prices.csv; llms-full.txt token prices. Found by a review of published pages and data.

    Evidence: Gemini Developer API pricing · Full record

    Permalink for Correction: Gemini 3.6 Flash's scheduled price rise was missing from the token-price table

  • Correction Correction: Homepage API savings were stated as general discounts instead of model-specific termsToken prices

    Was: The homepage displayed roughly 50% batch and 90% cached-input discounts and said production pays below list. Those constants did not represent every model’s recorded terms or establish realized customer bills. Corrected to: The general savings claim was removed. The homepage showed current model examples with separate input and output list prices, review dates, and each model’s recorded cache, batch and context terms. Missing prices remained explicitly unrecorded; no provider price was changed. Where: / homepage API price preview. Found by a review of published pages and data.

    Evidence: Current model-level prices and billing terms · Model pricing dataset · Full record

    Permalink for Correction: Homepage API savings were stated as general discounts instead of model-specific terms

  • Correction Correction: Older provider-history rows inherited later configuration descriptionsGPU quotes

    Was: Provider history grouped observations by GPU, provider, SKU, region and billing basis without including the configuration description. A later description replaced the metadata for the whole series. Nebius H100, H200, B200 and B300 observations dated October 11 and 14, 2026 consequently carried their October 20 descriptions instead of their original recorded descriptions. Corrected to: The immutable observations were re-exported with separate configuration identities, preserving each observation's recorded description. The metadata repair did not change its recorded price or observation date. New observations include a configuration fingerprint in their comparison identity, and current-quote fallbacks match the actual configuration and evidence before reusing a historical point. Lambda instance resources and Hyperstack's explicitly labeled maximum resource limits were retained in the new collector's configurations. Where: /price-history.json and /price-history.csv provider series; Per-GPU history payloads and provider-history charts. Found by a review of published pages and data.

    Evidence: GPU price history with recorded provider configurations · Nebius GPU pricing documentation · Full record

    Permalink for Correction: Older provider-history rows inherited later configuration descriptions

  • Correction Correction: Full-power electricity estimates were described as guaranteed cost floorsHardware specs

    Was: The Hardware page labeled chip TDP multiplied by assumed facility overhead and a tracked state's industrial electricity rate as a lower bound on rental cost, displaying a greater-than-or-equal sign. Its explanation assumed full-TDP draw for the rented hour, although actual chip power depends on the workload. Corrected to: The calculation was relabeled as an approximate electricity-cost scenario at full published chip power, with ordinary rounding and no bound inequality. The explanation retained the stated power, facility-overhead and electricity-rate inputs and their data date, and identified workload dependence and omitted host-server power and capital costs. It made no claim to measured consumption or a guaranteed minimum rental cost. Where: /hardware electricity-cost scenario and explanatory research. Found by a review of published pages and data.

    Evidence: Full-power electricity scenario and stated assumptions · Full record

    Permalink for Correction: Full-power electricity estimates were described as guaranteed cost floors

  • Market structure CFTC — Comment period on compute derivatives closesCFTC

    The comment period on the CFTC’s request for comment closes. The CFTC’s NYMEX review extension and ICE’s OCPI listing notice both point to this request. Status after: comment closed. Source: CFTC letter to CME Group (comment period closing Oct 20, 2026).

    Source: https://www.cftc.gov/filings/documents/2026/orgdcmnymexcompcontr260921.pdf

    Permalink for CFTC — Comment period on compute derivatives closes

  • Correction Correction: Prices from five GPU clouds whose terms bar reuse, and three aggregator estimates, were publishedGPU quotes

    Was: The site published 22 hand-surveyed list prices from CoreWeave (6), Crusoe (5), RunPod (8), Salad (2) and Voltage Park (1), the price ledger's earlier observations of all five and 59 archived CoreWeave and Crusoe list prices, although each provider's own terms bar that use: CoreWeave's Terms of Use (July 8, 2024) bar displaying site content in connection with products or services that are not CoreWeave's; Crusoe's Website Terms of Use (version 1.0, Aug 12, 2024) bar copying, reproducing, publishing or displaying its content without written permission; Salad's Terms of Service (Sep 29, 2026) bar using or displaying any element of the site without written consent; RunPod's Terms of Service (Mar 24, 2026) bar systematic retrieval to compile a collection or database; Voltage Park's Terms of Service (Feb 5, 2025) bar collecting data from its platform by any automatic or manual process without written permission. The quote tables also listed three secondary estimates observed on Jul 6, 2026 beside providers' own prices, two of them under a provider's name: the RunPod MI300X price of $1.99 and the TensorWave MI355X price of $2.95 were GetDeploying's figures, and the V100 market average of $0.99 was AIMultiple's; the ledger kept 13 GetDeploying-sourced estimates in all. The quarterly price-history charts drew a typical-price line of midpoints compiled from specialist list prices that included CoreWeave's, Crusoe's and RunPod's, and the homepage H100 insight read a 2023–25 price collapse from it. Corrected to: Every value from the seven sources was taken off the site's pages, payloads and downloads: the 25 rows, 67 price-ledger observations and the 38 class-aggregate points they could have fed, the 59 archived list prices, their venues in the archived quote snapshots, and the quarterly typical-price line with the figures read from it (its flat-quarter streak and level, the H100, H200 and B200 calendar y/y changes, the H100 collapse years and the B200 “+20% y/y” annotation); the quarterly charts kept the big-cloud list price. The H100 cheapest on-demand price became Hyperstack's $3.20, and the /hardware price per PFLOP-hour ranking was led by B200 instead of H100. The L40S cheapest on-demand price and typical price became Nebius's $1.55, a single venue's “from” price for its smallest preset, and A100 80GB's typical price, left with two specialist venues, became their $1.35–$2.79 range. GB200 NVL72 and MI300X lost their specialist quotes, so their cheapest on-demand prices became Oracle OCI's $16.00 and $6.00 and both left the price per PFLOP-hour ranking, as did RTX 4090 and RTX 5090, which lost every on-demand and marketplace quote and were given Beam's serverless per-second rates ($0.6912 and $1.0908 per GPU-hour, GPU only). The MI300X and MI355X all-market ranges lost their estimated low ends, which left MI355X with Oracle OCI's $8.60 alone, and V100 kept the three providers' own prices it already headlined. ACT.GB200-NVL72, ACT.RTX-4090 and ACT.RTX-5090 .SPEC-MED@primary-on-demand-v1 and ACT.RTX-4090 and ACT.RTX-5090 .MKT-FLOOR@marketplace-access-v1 were marked withdrawn, never to be reused; of the .SPEC-MED series of B300, B200, H200, H100, A100 80GB, MI300X and L40S, which stayed citable, those of H200, A100 80GB and L40S kept no ledger point at the withdrawal, only the current quote table's typical price; the re-survey's observations of Oct 20, 2026 gave H200's and L40S's a point that day (H200 from Nebius's $5.40 and Together AI's $5.99, L40S from Nebius's $1.55 alone) and A100 80GB's none. The CoreWeave, Crusoe and RunPod provider pages were redirected to the provider matrix, the RTX 4090 and RTX 5090 class-band embeds, badges and class-band chart tables kept their addresses with the withdrawal's notice instead of data (their price-history embeds kept theirs, drawing each chip's serverless prices from Oct 20, 2026), three desk-log entries computed from those prices became withdrawal notices under the same permalinks, B200's typical-price venue change of Oct 11, 2026 kept its permalink with its like-for-like move recomputed without CoreWeave's and RunPod's prices (+16.9% became +7.5%), and the closed evidence-digest day of Oct 11, 2026 got a new root without three of their digests. Earlier corrections records keep the figures they quote. Where: H100, H200, B200, B300, GB200 NVL72, A100 80GB, MI300X, MI355X, L40S, RTX 4090, RTX 5090 and V100 quote tables, ranges, chip-page titles and withdrawal notices on /gpu-cloud and the chip pages; the TPU v5p chip page's price history and withdrawal notice; the chip pages' social cards (og:image); /gpu-cloud/providers/ matrix and its other billing models; the CoreWeave, Crusoe and RunPod provider pages, redirected to the matrix and out of the sitemap; each other provider page's rank of its prices among its chip's quotes; /hardware price per PFLOP-hour ranking and the accelerator table's rental prices; /gpu-cloud/archived-prices/ and the chip pages' archived-price notes; homepage H100 quarterly chart, brief, lede and insight; the Board tiles of B300, B200, GB200 NVL72, H200, H100, A100 80GB, MI300X, L40S, RTX 5090 and RTX 4090, the market tape, and the GPU-cloud section's provider count and H100 price; the A100 80GB, H200 and B200 quarterly histories; /compare/ reference prices of a100-80gb-vs-h100, b200-vs-b300, h100-vs-b200, h100-vs-h200, h200-vs-b200 and l40s-vs-a100-80gb; /compare/h100-vs-mi300x/ and /compare/rtx-4090-vs-rtx-5090/ (noindex, out of the sitemap); /tools quote check; the H100, H200, B200, B300, GB200 NVL72, A100 80GB, MI300X and L40S badges, the badge gallery (/data/badges/) and the site badge's provider count; RTX 4090 and RTX 5090 class-band and price-history embeds, badges and class-band chart tables; GB200 NVL72 and MI300X class-band embeds, the GB200 NVL72 price-history embed and the embed gallery (/data/embeds/); the Board's citations (/chart-data/reference-citations.json); the core-review status on every page, whose manual GPU quotes kept inputs from Jul 6, 2026; four entries on /log and /log/2026-10/, /feed.xml and the H200, RTX 4090, RTX 5090 and B200 chip feeds; /price-history.json, the chip shards and /price-history.csv; /archive/quotes/ dated permalinks, the /archive/ index's aggregate counts and the class-band charts; /quotes.json, /quotes.csv, /reference-prices.json and /status.json; llms-full.txt quote rows and citable series; /evidence, /evidence/refs.csv, /evidence/digests.json and /evidence/digests/; the licence and attribution lines of the payloads, /data and /openapi.json, which no longer credit RunPod or Voltage Park; /terms third-party notices and /about. Found by a review of published pages and data.

    Evidence: CoreWeave Terms of Use (July 8, 2024), CoreWeave Sites’ Content · Crusoe Website Terms of Use (version 1.0, August 12, 2024), Intellectual Property · SaladCloud Terms of Service (September 29, 2026), 6. Use Restrictions · RunPod Terms of Service (March 24, 2026), §9 · Voltage Park Terms of Service (February 5, 2025), §5.2(h) · Methodology: GPU rental prices (providers’ own prices only) · The withdrawn citable series (withdrawnSeries) · Archived list prices, without the withdrawn sources · Full record

    Permalink for Correction: Prices from five GPU clouds whose terms bar reuse, and three aggregator estimates, were published

How entries and price moves are recorded

Automated entries are re-derived from the archives on every build. Each entry keeps its dated evidence and permanent month-page link. Market milestones after the source archive clock are marked Scheduled; corrections and recorded methodology updates use their recording dates.

Quote events diff the daily class-aggregate archive; curve events diff the forward-session archive under its movement-compatibility rules (same instrument, series, event, and methodology). Thresholds and semantics are versioned in code, and thin archives simply emit fewer events — cut/rise and floor-low entries activate as the daily record grows. Badges name each price move’s direction in words and colour it by what it means for a buyer: green for a cut or a new low, red for a rise.

A price change compares one provider’s posted price for the same offer on two days it was recorded; if the days between were not recorded, the entry says “between” the two dates. A renamed offer is followed through the rename when the instance is the same; another instance size or region is another offer. A change of 5% or more is one entry. Smaller ones share one entry per provider per week, which lists each offer once its price has moved 2% since it was last logged. Posted list prices, not transactions; a posted price does not confirm available capacity.

Market and quote events RSSFull correction records