GPU rental price index
Series 1 · GPU rental prices · snapshot 16 Oct 2026
The hourly price of renting one GPU, model by model: the cheapest published offer, the typical rate, the most expensive list price and, where a provider publishes one, the 12-month reserved rate. All values are US dollars per GPU-hour.
Key facts
- As of 16 Oct 2026, renting an NVIDIA H100 80GB costs $3.49 per GPU-hour at the median on-demand rate, with published offers from $1.30 to $12.29.
- As of 16 Oct 2026, the median on-demand price is $4.40 per GPU-hour for the H200 141GB, $6.25 for the B200 and $1.75 for the A100 80GB.
- A published 12-month reserved H100 rate of $3.26 per GPU-hour is −7% against the on-demand median (16 Oct 2026).
- The H100 median moved +79% between 1 Oct 2025 ($1.95) and 16 Oct 2026.
- The Compute Census GPU index covers 10 models and 29 providers with a recorded price (snapshot 16 Oct 2026).
Figure 1.1On-demand price range by model, USD per GPU-hour, 16 Oct 2026
§1The index
Table 1.1On-demand and reserved GPU rental prices, USD per GPU-hour
| Model | VRAM | Floor | Median | Ceiling | Reserved 12 mo | vs previous snapshot | Median history | Providers |
|---|---|---|---|---|---|---|---|---|
| GB200NVIDIA | 186 GB | $8.00 | $13.25 | $27.04 | $10.58 | 0%since 2026-10-14 | 5 | |
| MI355XAMD | 288 GB | $5.99 | $7.00 | $8.60 | $4.25 | 0%since 2026-10-09 | 6 | |
| B200NVIDIA | 180 GB | $3.69 | $6.25 | $16.11 | not published | 0%since 2026-10-14 | 9 | |
| H200NVIDIA | 141 GB | $2.09 | $4.40 | $13.78 | $3.40 | 0%since 2026-10-14 | 5 | |
| H100NVIDIA | 80 GB | $1.30 | $3.49 | $12.29 | $3.26 | +6.1%since 2026-10-14 | 8 | |
| MI300XAMD | 192 GB | $1.71 | $2.99 | $12.00 | $1.91 | 0%since 2026-10-09 | 7 | |
| A100NVIDIA | 80 GB | $0.53 | $1.75 | $5.07 | $1.59 | 0%since 2026-10-09 | 5 | |
| L40SNVIDIA | 48 GB | $0.38 | $1.53 | $3.77 | $1.09 | 0%since 2026-10-14 | 7 | |
| RTX 5090NVIDIA | 32 GB | $0.25 | $0.67 | $1.09 | not published | 0%since 2026-10-14 | 2 | |
| RTX 4090NVIDIA | 24 GB | $0.13 | $0.42 | $0.74 | $0.27 | 0%since 2026-10-09 | 4 |
Snapshot 2026-10-16. "vs previous snapshot" compares the current median with the latest history point dated before this snapshot. Sparklines: dashed grey = points before 1 Aug 2026 seeded from published indices; solid = collected by the tracker; red square = latest. Linked models have their own page.
§2What the snapshot says
The GB200 sits at the top with a median of $13.25 per GPU-hour; the RTX 4090 is the cheapest card in the index at $0.42. Between them the order follows memory and generation more than vendor: AMD's MI355X prices next to NVIDIA's Blackwell parts, and the older MI300X competes with the H100 class rather than with B200.
Dispersion matters as much as the median. The L40S ceiling is 9.9 times its floor: the same chip is offered by Lium.io at $0.38 and by AWS at $3.77. The tightest market is the MI355X at 1.4×, where only a handful of clouds offer the part at all.
Reserved capacity is not uniformly cheaper. The deepest published 12-month discount against the on-demand median is on the MI355X (−39%); the shallowest is on the H100 (−7%), where several on-demand offers already sit below the published term rate.
The most recent median move in the history was the H100 on 16 Oct 2026: $3.29 → $3.49 (+6.1%).
§3Floor and ceiling providers
Table 1.2Who sets the cheapest and the most expensive recorded price, USD per GPU-hour
| Model | Cheapest recorded | Price | Highest on-demand | Price | Quotes as of |
|---|---|---|---|---|---|
| GB200 | GMI CloudOn-demand | $8.00 | Microsoft AzureND128isr GB200 v6, $108.16/hr for 4 GPUs | $27.04 | 2026-09-14 |
| MI355X | DaytonaOn-demand | $5.99 | Oracle Cloudbare metal | $8.60 | 2026-09-18 |
| B200 | DeepInfraOn-demand | $3.69 | Google Clouda4-highgpu-8g | $16.11 | 2026-09-28 |
| H200 | BeamServerless | $2.09 | Microsoft AzureND96isr H200 v5, $110.24/hr for 8 GPUs | $13.78 | 2026-09-21 |
| H100 | Lium.ioMarketplace | $1.30 | Microsoft AzureND96isr H100 v5, $98.32/hr for 8 GPUs | $12.29 | 2026-10-16 |
| MI300X | TensorWaveSales quote | ~$1.71 | Microsoft AzureND96isr MI300X v5, $48-96/hr for 8 GPUs depending on region | $6.00–$12.00 | 2026-09-25 |
| A100 | Vast.aiMarketplace | ~$0.53 | Google Clouda2-ultragpu | ~$5.07 | 2026-09-25 |
| L40S | Lium.ioMarketplace | $0.38 | AWSus-east-1, 8-GPU configuration | $3.77 | 2026-09-21 |
| RTX 5090 | Salad CloudMarketplace | ~$0.25 | Beamfrom | ~$0.72 | 2026-09-21 |
| RTX 4090 | Vast.aiMarketplace | $0.13–$0.14 | RunPodSecure Cloud | $0.74 | 2026-09-18 |
Provider quotes are the individual prices recorded in the tracker's per-model notes, dated by when that note was last written. "~" marks a price the source states approximately.
§4Price history, model by model
Each panel uses its own vertical scale so the shape of each market is visible; compare levels with Table 1.1, not across panels.
GB200
MI355X
B200
H200
H100
MI300X
A100
L40S
RTX 5090
RTX 4090
Methodology in brief
- Prices are published list prices: provider price pages, instance-pricing databases and comparison trackers. No private quotes, no estimates.
- Multi-GPU instances are divided by their GPU count, so every figure is per single GPU-hour.
- Spot, preemptible and interruptible tiers are recorded but excluded from the on-demand floor.
- History before 1 Aug 2026 is seeded from published indices (Silicon Data, SemiAnalysis, Introl); later points are snapshots taken by the tracker.
- A model gets its own page only with at least 3 distinct provider prices. RTX 5090 (2) falls short and stays in this table only.
Sources: every provider price page, tracker and seed index behind this table is listed on Sources.
Need GPUs at a firm price?
Index prices are list prices for small volumes. For clusters, a reserved block or hardware you own, send the specification — we come back with supplier pricing and lead times.