Cloud prices, history and price index

Cloud prices shows what renting each GPU costs right now, across providers, and how that price has moved. The same data prices every result on the site: the $/hr and $/1M tokens of the calculator, TCO, Optimize and fine-tuning come from it.

Where the prices come from

We collect prices ourselves, once a day, directly from each provider's public price list or pricing API — no aggregators and no affiliate feeds. The providers range from hyperscalers (AWS, Azure, Google Cloud, Oracle Cloud) to neoclouds such as RunPod, Lambda, CoreWeave, Nebius, Crusoe, Vast.ai, DataCrunch and Scaleway; a few only expose prices behind an API key and are collected when a key is configured. 14 providers have prices from the last seven days — the provider count shown across the site. Each adapter maps the provider's instance names to catalog GPUs, so "8× H100 80GB SXM" and "p5.48xlarge" land on the same row.

• Per GPU-hour. Multi-GPU instances are divided by their GPU count; where a provider offers several sizes of the same GPU, the cheapest per-GPU offer is shown.
• Three pricing types — on-demand, spot and community (marketplace hosts) — filtered separately. Market figures, history and every cost estimate use on-demand only.
• Currency. Prices listed in euros are converted to US dollars at the ECB reference rate of the collection day.
• What is excluded — regional variation, committed-use and enterprise discounts, reservations, storage, egress and support. Confirm on the provider's page; every row links to its source.

Reading the prices page

The header states how many providers and GPUs the latest collection covers and how old it is. The table lists every offer with provider, GPU, pricing type, $/GPU-hr and the day it was last seen; filter by GPU (search or popular picks), provider and pricing type, and sort by price or provider. A row whose price has not been seen for more than three days is marked stale in amber — the provider may have dropped it, or a collection failed.

The aside summarises the market for the selected GPU, or across the board when none is selected: providers today, the median with its 30-day change, the neocloud and hyperscaler medians and the 90-day low. Best deals now shows market cards — the cheapest on-demand offer against the median, the median's 30-day change and the 90-day low — for the leading GPUs; a card opens that GPU's history. If you visited before, a since your last visit banner lists what moved; the snapshot it compares against is kept in your browser only.

Price history

Opening a GPU (/cloud-pricing?gpu=<slug>) shows its daily history over 30, 90 or 365 days: the market band from cheapest to dearest provider with the median line, one step line per provider coloured cheapest first, and a dot on every change of at least 1 %, listed below the chart. The segment switch shows all providers, neoclouds or hyperscalers only, and a chip per provider greys it out. A provider's last list price is carried across days without a collection, so a gap does not draw as a drop.

In the calculator and TCO, 90-day cost history applies the same history to a configuration: the median price × the GPU count, day by day, as the cost of running the chosen setup.

How prices enter the estimates

Each tool prices a GPU at its current market median: per provider the cheapest on-demand offer on the latest collection day, then the median across providers. When no provider lists the GPU, a catalog estimate is used and the result says "catalog estimate"; when there is neither, the configuration has no cost and is ranked without one. The open calculator row shows which source applied, how many providers contributed and the date.

Price index, price pages and feeds

• Price index — every tracked GPU with its provider count, cheapest and median $/hr and 30-day change on one server-rendered page.
• Price pages /price/<gpu> — per GPU: the provider table, a 30-day chart and short answers. A page exists when at least three providers list an on-demand price from the last 14 days, so a thin market does not produce a misleading page.
• Atom feeds — /feeds/market.xml for market-wide changes and /feeds/gpu/<slug>.xml per GPU. An entry appears only on a day the median changed or a new 90-day low was set; a low is claimed only after 30 days of history. The feeds contain no personal data and need no account.
• Snapshot image — a 1200 × 630 PNG of today's market for posts and slides, from the share card of the prices page.