GPU Rental Price Index (GRPI)
One number for the cloud GPU rental market: a chain-linked index of $/GPU-hour prices, set to 100 on 2026-07-05 and moved only by like-for-like price changes. Composition change never moves the level.
Thin-ledger note. The index is 4 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.
| Date | Level | Day change | Link families | Link providers |
|---|---|---|---|---|
| 2026-07-05 | 100.0 | base | – | – |
| 2026-07-06 | 104.3 | +4.27% | 18 | 2 |
| 2026-07-07 | 104.2 | -0.10% | 19 | 11 |
| 2026-07-08 | 104.0 | -0.19% | 19 | 18 |
A negative day change means the like-for-like cloud GPU rental price fell that day. "Link families / providers" is the size of the universe that day's move was measured over: it grows as the ledger widens (a new provider joins the link one day after it first appears).
Per-family pricing today
Live $/GPU-hour by GPU family in the latest snapshot (2026-07-08). Cohort median is a median of per-provider medians (equal provider weight), so a high-listing provider does not drown a newcomer.
| GPU family | Providers | Cohort median | Cheapest | Priciest | Spread | Spot discount |
|---|---|---|---|---|---|---|
| RTX A4000 | 2 | $0.20 | $0.15 | $0.25 | 1.7x | – |
| RTX A5000 | 2 | $0.35 | $0.27 | $0.44 | 1.6x | 41% |
| RTX 3090 | 1 | $0.46 | $0.46 | $0.46 | 1.0x | 52% |
| RTX A6000 | 5 | $0.57 | $0.49 | $1.09 | 2.2x | 33% |
| RTX 4090 | 2 | $0.64 | $0.58 | $0.69 | 1.2x | 41% |
| L4 | 4 | $0.72 | $0.39 | $1.67 | 4.3x | 44% |
| Tesla V100 | 3 | $0.77 | $0.17 | $0.79 | 4.6x | – |
| RTX 6000 Ada | 3 | $0.79 | $0.77 | $1.04 | 1.4x | 4% |
| RTX 5090 | 2 | $0.83 | $0.68 | $0.99 | 1.5x | – |
| L40 | 4 | $0.93 | $0.82 | $1.25 | 1.5x | 16% |
| A40 | 2 | $1.08 | $0.44 | $1.71 | 3.9x | 20% |
| L40S | 10 | $1.59 | $0.88 | $3.77 | 4.3x | 10% |
| RTX PRO 6000 | 7 | $1.89 | $1.80 | $2.50 | 1.4x | 9% |
| GH200 | 4 | $2.14 | $1.88 | $6.50 | 3.5x | – |
| A100 | 13 | $2.15 | $0.89 | $5.12 | 5.8x | 24% |
| MI300X | 2 | $2.82 | $2.19 | $3.45 | 1.6x | – |
| H100 | 16 | $3.12 | $1.99 | $12.29 | 6.2x | 29% |
| H200 | 12 | $4.19 | $3.62 | $10.60 | 2.9x | 1% |
| B200 | 9 | $6.69 | $3.50 | $9.36 | 2.7x | – |
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.
Hyperscaler median $11.06/hr vs neocloud median $2.99/hr.
Hyperscaler median $4.10/hr vs neocloud median $1.54/hr.
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 only. Each link is measured only over providers present in both consecutive snapshots. A provider that first appears today is absent from the previous snapshot, so it contributes nothing to today's move; it enters the index one day later, when 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) 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 cannot move the index on its first 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.
Cite the index
Referencing GRPI in an article, paper, model card or dashboard? Cite the snapshot you retrieved (July 2026):
APA
GPU Rental Prices. (2026). GPU Rental Price Index (GRPI) [Data set]. gpurentalprices.com. Retrieved 2026-07-08, from https://gpurentalprices.com/gpu-price-indexBibTeX
@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 2026-07-08}
}HTML
<a href="https://gpurentalprices.com/gpu-price-index">GPU Rental Price Index (GRPI)</a>. GPU Rental Prices, 2026. Accessed 2026-07-08.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.
Keep exploring
Price history · Price movers · All-time records · Data & API · Methodology