A reference price for the listed GPU rental market.
Daily fixings for spot, on‑demand, and reserved rentals on public venue APIs across 12 venues. Complementary to survey‑based contract indices. Open data. Open methodology. Published at 00:30 UTC.
Renting a GPU now costs between $0.78 and $5.88 per hour depending on the chip and where you rent it. GPU Markets publishes the daily reference price across every major rental venue — so anyone building on, investing in, or reporting on AI infrastructure can see what compute actually costs today, and how that price has moved.
FOR NEOCLOUD CFOs·FOR INFRASTRUCTURE ANALYSTS·FOR JOURNALISTS·FOR RESEARCHERS
Struck 2026‑04‑18 · 00:30 UTC · 8 series · 1 off-strike (◦ marker)
| Series | Today | Δ 1d | Δ 7d | Δ 30d | Obs | Venues | 30d |
|---|---|---|---|---|---|---|---|
| $2.1432 | −0.78% | −0.99% | −7.72% | 2,847 | 9/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$2.1687·+1.19%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·4 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·1 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $2.8714 | −0.14% | −0.71% | −0.72% | 1,204 | 8/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$2.9012·+1.04%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·3 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·1 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $1.9208 | +0.73% | −0.20% | −1.69% | 412 | 6/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$1.9315·+0.56%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·2 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·1 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $3.4117 | −0.71% | −2.18% | −2.14% | 1,893 | 7/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$3.4523·+1.19%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·3 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·1 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $5.8843 | −0.15% | +0.53% | −11.63% | 1,147 | 6/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$5.9712·+1.48%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·3 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. | |||||||
| |||||||
| $1.1284 | −0.89% | +0.42% | −0.41% | 3,281 | 11/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$1.1198·−0.76%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·5 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·3 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $0.7812 | −0.46% | +1.76% | +1.80% | 1,624 | 9/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·$0.7934·+1.56%
vs fix2 T1 Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. ·4 T2 Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. ·2 T3 Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. | |||||||
| |||||||
| $2.9408 | −0.17% | −1.92% | −4.12% | 318 | 4/12 | ||
CW Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. ·suppressed · insufficient tier coverage (all 4 surviving venues in T2) | |||||||
| |||||||
- GPUM.H100.SXM.SPOT · Hyperbolic observation excluded from today’s fix: residual > 2.8 MAD vs. the surviving-venue median. Venue remains Eligible; rejection applies to this fix only. ↩
- GPUM.MI300X.SPOT · Off-strike backfill. TensorDock API returned 5xx on the 00:30 UTC poll; this row was struck at 06:45 UTC from the same day’s observations once the venue became reachable. Δ columns compare to the 00:30 UTC fix from 2026‑04‑17. ↩
Fixings quoted as USD per GPU‑hour, rounded to four decimal places. Δ columns show percentage change against the previous fix, the fix seven days ago, and the fix thirty days ago. Obs is the number of observations contributing to today's estimator after outlier rejection; Venues is eligible venues over total venues covered.
Interactive price chart. Hover or tap to inspect a daily fixing. When focused, use the left and right arrow keys to move through the series.
› Constituent venues for H100 SXM Spot 9 eligible · 1 outlier · 1 cross-check
| Venue | Price | Δ 7d | Weight | Contributed |
|---|---|---|---|---|
| Vast.ai | $1.8704 | −4.20% | 18.3% | Yes |
| RunPod | $2.0412 | −2.80% | 16.1% | Yes |
| Lambda Labs | $2.2900 | −1.10% | 14.7% | Yes |
| Prime Intellect | $1.9816 | −3.50% | 15.2% | Yes |
| Hyperstack | $2.1100 | −2.20% | 13.8% | Yes |
| TensorDock | $2.0144 | −3.10% | 11.9% | Yes |
| DataCrunch | $2.2500 | −1.80% | 9.0% | Yes |
| Hyperbolic | $2.6100 | +0.40% | — | No (outlier) |
| Shadeform | $2.0900 | −2.50% | — | Cross‑check |
Weights proportional to inverse MAD distance from the trimmed-mean estimator after outlier rejection. Hyperbolic excluded today: >2.5 MAD above the cluster. Full rejection log in methodology.
Struck 2026‑04‑18 · 00:30 UTC · 5 series
| Series | Today | Δ 1d | Δ 7d | Δ 30d | Obs | Venues | 30d |
|---|---|---|---|---|---|---|---|
| $0.4318 | −0.48% | −0.35% | +2.37% | 4,128 | 10/12 | ||
| |||||||
| $0.8942 | +0.75% | −0.74% | −6.17% | 982 | 5/12 | ||
| |||||||
| $0.2684 | −0.45% | −0.04% | −1.43% | 2,156 | 8/12 | ||
| |||||||
| $0.1247 | −0.40% | −0.80% | −1.27% | 1,893 | 9/12 | ||
| |||||||
| $1.4820 | +0.66% | +0.74% | −2.39% | 204 | 3/12 | ||
| |||||||
Inference‑tier series are priced the same way as training series, but include consumer and workstation cards (RTX 4090, RTX 5090) and low‑power accelerators (L4, T4) that dominate high‑throughput serving workloads.
| SPOT | Per-hour rate on a marketplace or community cloud. Pay-as-you-go, cheapest, no commitment, supply varies by hour. |
| OD | On-demand rate from a dedicated provider. Higher price than spot, guaranteed supply, cancel anytime. |
| R1Y | One-year reserved contract. Locked in for 12 months at a discounted rate in exchange for committed capacity. |
| CW | Capacity-weighted companion research series. Weighted median of the same surviving observations, using venue-capacity tiers rather than equal weights. Not part of the headline fix family. |
| T1 | Capacity tier 1. Venues with more than 10,000 GPUs. Weight 3 in the capacity-weighted companion. |
| T2 | Capacity tier 2. Venues with 1,000 to 10,000 GPUs. Weight 2 in the capacity-weighted companion. |
| T3 | Capacity tier 3. Venues with fewer than 1,000 GPUs. Weight 1 in the capacity-weighted companion. |
The GPU rental market in 2026 looks the way the oil market looked in the 1970s, before Platts began publishing Dated Brent: enormous, economically consequential, and almost entirely opaque. A neocloud CFO modelling next quarter's revenue, an analyst writing a sector note, or a journalist reporting on AI capex has no shared reference for what an H100‑hour costs today — only a thicket of venue‑specific dashboards, sales‑team quotes, and screenshots passed around in Slack.
GPU Markets is an attempt to fix that with the most boring possible tool: a daily statistical release. Every day at 00:30 UTC we publish a single number per series — the trimmed‑mean rate across every venue we can observe, with the full input set, the outliers we rejected, and the code that produced the fix all committed to a public repository. The methodology is the product. If you disagree with a fix, you can rerun it.
The point is not to compete with the venues, predict prices, or recommend trades. It is to give the GPU economy the same kind of public, auditable price reference that every other commodity market already has — so that the conversations being had above this market (capacity planning, hedging, journalism, policy) can rest on a shared set of numbers instead of vibes.
Each daily fix is a 20% trimmed mean of every observed venue rate for the series, with observations more than 2.5 median absolute deviations from the cluster rejected as outliers. Reference clouds (AWS, GCP, Azure) are reported alongside but excluded from the estimator. No smoothing, no carry‑forward, no proprietary adjustments — if a number is in the fix, you can find the observation that produced it.
| Venue | Access | Eligible | Pipeline since | n · 30d |
|---|---|---|---|---|
| Vast.ai | REST API | Yes | 2026‑02‑14 | 3,184 |
| RunPod | GraphQL API | Yes | 2026‑02‑14 | 2,947 |
| Lambda Labs | HTML | Yes | 2026‑02‑14 | 1,862 |
| Prime Intellect | HTML + CLI | Yes | 2026‑02‑17 | 1,421 |
| Hyperstack | REST API | Yes | 2026‑02‑17 | 1,308 |
| TensorDock | REST API | Yes | 2026‑02‑18 | 1,154 |
| DataCrunch | REST API | Yes | 2026‑02‑18 | 982 |
| Hyperbolic | REST API | Yes | 2026‑02‑21 | 847 |
| Shadeform | Aggregator | Cross‑check | 2026‑02‑21 | 612 |
| AWS (p5 family) | Official Pricing API | Reference | 2026‑02‑14 | — |
| GCP (A3 family) | Cloud Billing API | Reference | 2026‑02‑14 | — |
| Azure (ND v5) | Retail Prices API | Reference | 2026‑02‑14 | — |
Eligibility is a property of the observation pipeline, not a quality judgement. Hyperscaler list prices are posted alongside as a Reference series but excluded from the trimmed‑mean estimator — their n · 30d column therefore shows —. The Pipeline since column records the date live observation began for each venue; the hero chart's 180‑day series extends earlier via backfill from venue‑published rate‑card archives and public snapshots, so chart history predates pipeline start by design. The n · 30d column is the count of individual observations from each venue that survived outlier rejection over the last thirty strike windows, aggregated across all series — a coarse depth signal: venues with higher n have more bearing on the fix, even when two sit in the same Eligible tier. Per‑venue terms of service, scrape cadence, and the latest eligibility review date are available on the venues directory.
Three integration patterns for a GPU compute derivatives market
Six regulated, crypto‑native, and auction‑based venues for GPU‑compute derivatives are in development. Each has the same structural dependency: an independent, auditable, low‑friction reference price to settle against. Three integration modes — cash‑settled futures, perpetual funding oracles, and listed‑minus‑transacted basis trading — cover the landscape.
- 2026‑04‑11 Decomposing the B200 curve — where did the demand go?
- 2026‑04‑04 The Lambda–Vast spread is structural, not arbitrage.
- 2026‑03‑28 Constructing a forward curve from reserved‑rate differentials.
- 2026‑03‑21 Implied GPU residual value: H100 vs H200.
- 2026‑03‑14 Could a neocloud have hedged the 2024 crash? A backtest.
SemiAnalysis H100 1Y contract series now cited as external validation signal
The lead research note at /research/b200‑curve‑decomposition/ now references SemiAnalysis's contract‑side index to discriminate between demand‑absorption and demand‑exhaustion hypotheses on the listed‑side decline. Listed spot down 28% / 6mo while contract 1Y up ~38% over same window — forces the absorption reading.
- 2026‑04‑19 Methodology update — fix surface defined as 'listed advertised rates' [ MARKET ]
- 2026‑04‑17 Datavault AI launches edge GPU sites in NY/Philadelphia; targets 1,000 U.S. sites by end‑2026 [ PRESS · Electronics Media ]
- 2026‑04‑13 CoreWeave surges on analyst upgrade and AI contract momentum [ PRESS · The Motley Fool ]
- 2026‑04‑09 CoreWeave takes as much financial engineering as datacenter design [ PRESS · The Next Platform ]