GPU Market Watch
What this page is. A dated log of changes in the GPU cloud market that affect how you should budget: provider price announcements, new GPU generations reaching clouds, capacity shortages, contract-term changes and relevant policy news. Each entry links to the primary source. What it is not: a price index. We do not publish hourly rates, because they change faster than any page and differ by region and term; the provider directory links to where current prices live.
How to use these notes
Read the latest entry before you lock in a budget or sign a term commitment. Three questions matter: is the card you plan to use getting cheaper, scarcer or superseded; are providers changing contract terms in ways that shift risk to you; and is anything in the news likely to move demand in the next quarter. Then open the provider pages linked from each note and confirm the current figure yourself. Our guide to why GPU prices keep rising gives the longer background on supply, power and demand, and take-or-pay contracts explains the term-commitment trend this log tracks.
Week of October 13, 2026
Themes we are tracking. Three patterns from the past quarter continue and should shape any budget written this month.
- Term commitments are spreading downmarket. Reserved and committed-use pricing, once mainly for large buyers, is being offered to small teams with shorter terms. The discount is real; so is the obligation to pay for unused hours. Before accepting one, read what take-or-pay means for your budget and compare against a well-designed spot strategy, which gives a similar discount with a different risk.
- Previous-generation cards are becoming the value option for inference. As newer generations land in the big clouds, the cards they replace fall in price on specialist providers while remaining capable for serving most models at 8-bit precision. For serving workloads, benchmark one generation back before paying for the newest; our serving guide describes how.
- Regional price divergence is widening. The same card is increasingly priced differently across a provider’s regions as capacity concentrates near demand hubs. Training and batch jobs that can move regions are the beneficiaries; regions, availability and price explains how to design for that mobility.
Reading list this week. The official pricing and announcement pages listed in the provider directory are the primary sources; check the ones for the providers you use. Where a provider has announced a change, it appears on its own pricing or blog page before anywhere else.
Budget implication. If you are writing a quarterly GPU budget now: price inference at one generation back with a quantization step, price training on spot with an interruption allowance, and treat any term commitment as a decision to be made after the first month of measured utilisation, not before.
Week of October 19, 2026
What moved (12–18 October). Four developments this week bear on budgets; each links to the primary source, and none of them is a price we publish.
- NVIDIA opened its Vera Rubin rack and compute-tray designs to the Open Compute Project (OCP Global Summit, 13–14 October). The designs were contributed through the MGX architecture, with Google, Meta and Microsoft named among adopters, so next-generation racks can be built by many manufacturers rather than one. For buyers this shortens the path from Rubin announcement to Rubin capacity in the big clouds, and it is the clearest signal yet that Hopper- and Blackwell-class cards will keep sliding into the value tier for inference. Source: NVIDIA blog.
- Scaleway announced a new GPU zone in Copenhagen built on NVIDIA HGX B300 (ai-PULSE, 13 October). No opening date was given. It adds a European option for Blackwell Ultra-class capacity, which matters if your data has to stay in the EU; treat it as a 2027 planning item until the zone is listed on Scaleway’s availability page. Source: Scaleway newsroom.
- CoreWeave is entering India with 240 MW of capacity at AdaniConneX’s Navi Mumbai campus (announced 6 October; first phase expected mid-2028). Nothing to budget for this year, but it continues the pattern this log has tracked since the summer: specialist providers are building where power and demand are, and the same card will increasingly carry a different price in each region. Source: CoreWeave news.
- AWS states on its EC2 Capacity Blocks for ML pricing page that current prices are scheduled to be updated next in January 2027. Several third-party price trackers reported a revision to Capacity Block rates in early October; the official table remains the source of truth, and it still lists rates by Region. If Capacity Blocks are part of your plan, confirm the figure for your Region now and put a price review in your January calendar. Source: AWS Capacity Blocks pricing.
Budget implication. Nothing this week changes last week’s guidance: price inference one generation back, price training on spot with an interruption allowance, and keep term commitments short enough that a scheduled January repricing or a Rubin arrival cannot strand you. Add one item: if a provider has told you its next price update date, write it into the budget as a review date rather than treating the current figure as fixed.
How this log is maintained
A new entry is added each week with the date in its heading; entries are not edited after the following week’s entry is published, except to correct an error, in which case the correction is noted in place. The log keeps the most recent twelve weeks on this page. Entries describe changes and link to sources; they do not reproduce provider prices, which belong to the providers and go stale within days. If you see a market change we have missed, tell us through the contact page.