Cloud GPU rental price tracker (GRPI)

As of 2026-08-22, the GPU Rental Price Index is 114.6, or +14.6% from its 2026-07-05 base. GRPI tracks like-for-like changes in cloud rental $/GPU-hour. Each GPU family is compared only across providers that list it on both days, so a new provider cannot create a false price move.

Scope: compute rental offers only. Consumer resale prices and game-streaming subscriptions use different products and data sources, so they are not included. Use price historyfor model-level series and price movers for measured weekly changes.

Thin-ledger note. The index is 49 days old. Base 100 is anchored to 2026-07-05; early links are measured over a small provider set and the trend line fills in as daily snapshots accumulate. Every point is real, like-for-like, and never backfilled.

The index over time

Chain-linked level per snapshot date. The dashed line marks the base of 100.

A negative day change means the like-for-like cloud GPU rental price fell that day. "Link families" is how many GPU families had a like-for-like comparison that day; "Link providers" is the distinct providers that were like-for-like in at least one of those families (the union of the per-family intersections). Both grow as the ledger widens: a provider joins a family's link one day after it first lists that family.

Per-family pricing today

Live $/GPU-hour by GPU family in the latest snapshot (2026-08-22). Cohort median is a median of per-provider medians (equal provider weight), so a high-listing provider does not drown a newcomer.

Spread = priciest on-demand listing divided by the cheapest. Spot discount = how far the cheapest interruptible (spot / community) rate sits below the cheapest on-demand rate, shown only where positive.

The hyperscaler premium

What the big clouds charge for the same GPU versus the neocloud median. Posted on-demand prices, per GPU-hour.

Caveat (n=2). The hyperscaler set today is only AWS and Azure, so the hyperscaler median is a median over at most two providers per family (often one when only one lists a given SKU). GCP is not yet a source. Treat the premium as directional, not a market-wide average. Full detail on the methodology page.

How the index is built

The GPU Rental Price Index is chain-linked, the standard way real price indices absorb basket and composition change without spurious jumps in the level.

• Base 100 on 2026-07-05. The earliest snapshot date is the level anchor. The base fixes the level, never a fixed set of providers.
• A running product of daily links. Index(base) = 100, and Index(today) = Index(yesterday) times today's link. Each link is the price change from the previous snapshot to this one.
• Like-for-like, per family. The provider intersection is taken independently for each GPU family: a family's link uses only the providers that list that family with a firm offer in both consecutive snapshots. A provider that first appears today, or that stays live overall but is missing one family on either day, is not in that family's intersection, so it contributes nothing to today's move; it enters a family's ratio one day later, once it lists that family on two consecutive snapshots and a genuine like-for-like change can be measured.
• Equal family weight. Per GPU family we take the cohort median (a median of per-provider medians) over that family's like-for-like provider set on both days, giving one family ratio, then average those ratios with equal weight across families. No single family or high-listing provider dominates.
• Composition change never moves the level. By construction, a provider or family joining the ledger, or a provider dropping and re-adding a single family, cannot move the index on that day. This is why the 2026-07-05 base carrying only two providers does not pin the index to a two-provider universe: the like-for-like universe grows as the ledger widens.
• Append-only ledger. Every daily snapshot is kept forever and never rewritten. History cannot be backfilled, which is exactly why the index is credible.

Every number here is fetched from a provider source then computed by the formula above. Nothing is typed in by hand, and nothing stale is presented as live. Full collection detail on the methodology page. Changed how a number here is computed? It is logged in the methodology page's Revisions section, dated, with the magnitude of the change.

Cite the index

Referencing GRPI in an article, paper, model card or dashboard? Cite the snapshot you retrieved (August 2026):

APA

GPU Rental Prices. (2026). GPU Rental Price Index (GRPI) [Data set]. gpurentalprices.com. Retrieved 2026-08-22, from https://gpurentalprices.com/gpu-price-index

BibTeX

@misc{grpi_2026,
  title  = {GPU Rental Price Index (GRPI)},
  author = {{GPU Rental Prices}},
  year   = {2026},
  howpublished = {\url{https://gpurentalprices.com/gpu-price-index}},
  note   = {Chain-linked, base 100 = 2026-07-05. Accessed 2026-08-22}
}

HTML

<a href="https://gpurentalprices.com/gpu-price-index">GPU Rental Price Index (GRPI)</a>. GPU Rental Prices, 2026. Accessed 2026-08-22.

Today's snapshot is free under CC BY 4.0 with a visible link to gpurentalprices.com. The full historical ledger is licensed: see Data & API. The underlying daily ledger is archived with a citable DOI: 10.5281/zenodo.21435394.

Follow GPU pricing through Google? Add gpurentalprices.com as a preferred source and Google will surface our results more often when they are relevant.

Keep exploring

Price history · Price movers · All-time records · Data & API · Methodology