Market brief · 2026-11-08
GPU cloud market brief
Where rental prices for cloud GPUs stand today, written for a budget conversation: price ranges, what flexibility costs, and what a cluster runs per month. Every figure is computed from the price table; nothing is estimated by hand.
Based on 144 prices from 13 providers: AWS, OCI, Akamai, DataCrunch, OVHcloud, DigitalOcean, Nebius, Google Cloud, Together, Koyeb, Jarvislabs, Latitude.sh, Hot Aisle. Fetched 2026-11-08 13:40 UTC to 2026-11-08 18:45 UTC. Other GPU clouds are not covered yet; see who is missing and why. Listed prices only: they do not show whether capacity is available.
The short version
- H100 SXM on demand: $3.49 to $11.06 per GPU-hour across 8 providers; the middle price is $4.72. An always-on 8-GPU server at the middle price runs $27,558 a month.
- Spot: the cheapest interruptible H100 SXM is 67% below the middle on-demand price, for jobs that can restart from a checkpoint.
- Commit or stay flexible: for H100 SXM, a 1-year commitment beats on-demand once the GPUs are busy more than 69% of the time; a 3-year commitment, more than 43% (median across machines priced both ways).
- Most 16-bit compute per dollar on demand: MI300X at DigitalOcean, $1.98 per dense PFLOP-hour.
Price ranges per GPU
US dollars per GPU-hour. Low, middle (median) and high on-demand price across providers, with the lowest spot and reserved prices. Each provider counts once, at its lowest price for that GPU.
| GPU | Memory | On-demand low | Middle | High | Providers | Spot low | 1-yr low | 3-yr low |
|---|---|---|---|---|---|---|---|---|
| B300 | 270 GB | $8.00 Latitude.sh | $10.84 | $17.80 | 6 | $4.99 Together | $7.94 | — |
| GB300 | 279 GB | $13.44 DataCrunch | $15.72 | $18.00 | 2 | $6.72 DataCrunch | — | — |
| B200 | 180 GB | $5.50 Koyeb | $8.82 | $14.24 | 6 | $4.09 Together | $11.12 | $7.09 |
| GB200 | 186 GB | $16.00 OCI | $16.00 | $16.00 | 1 | — | — | — |
| H200 SXM | 141 GB | $3.00 Koyeb | $5.99 | $10.60 | 9 | $2.29 Jarvislabs | $3.40 | $4.65 |
| H100 SXM | 80 GB | $3.49 Jarvislabs | $4.72 | $11.06 | 8 | $1.54 Jarvislabs | $3.26 | $2.97 |
| A100 80GB | 80 GB | $1.49 Jarvislabs | $2.87 | $5.07 | 6 | $0.890 Jarvislabs | $2.18 | $1.46 |
| A100 40GB | 40 GB | $0.890 Jarvislabs | $2.20 | $3.48 | 4 | $0.790 Jarvislabs | $1.74 | $1.17 |
| L40S | 48 GB | $1.20 Koyeb | $1.80 | $3.50 | 7 | $0.970 DataCrunch | $1.17 | $0.804 |
| L4 | 24 GB | $0.440 Jarvislabs | $0.707 | $1.00 | 5 | $0.290 Jarvislabs | $0.445 | $0.318 |
| A10 | 24 GB | $2.00 OCI | $2.00 | $2.00 | 1 | — | — | — |
| A10G | 24 GB | $1.01 AWS | $1.01 | $1.01 | 1 | — | $0.634 | $0.435 |
| T4 | 16 GB | $0.526 AWS | $0.526 | $0.526 | 1 | — | $0.331 | $0.227 |
| V100 | 32 GB | $0.242 DataCrunch | $1.65 | $3.06 | 2 | $0.121 DataCrunch | $2.40 | $1.30 |
| RTX PRO 6000 Blackwell | 96 GB | $1.80 Nebius | $2.86 | $4.50 | 8 | $1.15 Jarvislabs | $3.10 | $1.98 |
| RTX 6000 Ada | 48 GB | $1.54 DataCrunch | $1.56 | $1.57 | 2 | $0.771 DataCrunch | — | — |
| RTX A6000 | 48 GB | $0.750 Koyeb | $0.801 | $0.852 | 2 | $0.426 DataCrunch | — | — |
| RTX 4000 Ada | 20 GB | $0.520 Akamai | $0.640 | $0.760 | 2 | — | — | — |
| Quadro RTX 6000 | 24 GB | $1.50 Akamai | $1.50 | $1.50 | 1 | — | — | — |
| MI300X | 192 GB | $2.59 DigitalOcean | $2.99 | $6.00 | 3 | — | $1.91 | — |
| MI325X | 256 GB | $3.80 DigitalOcean | $3.80 | $3.80 | 1 | — | $2.88 | — |
| MI355X | 288 GB | $8.60 OCI | $8.60 | $8.60 | 1 | $4.50 DigitalOcean | — | — |
What an 8-GPU server costs per month
Always on: 8 GPUs × 730 hours. Commitments bill every hour of the term whether or not the GPUs are busy; spot can be interrupted.
When a commitment pays off
A commitment is billed for every hour. It beats on-demand once the GPUs are busy more than this share of the time: the committed price divided by the on-demand price for the same machine and region, from providers that publish both.
| GPU | 1-yr: busy more than | 1-yr range | 3-yr: busy more than | 3-yr range | Machines compared |
|---|---|---|---|---|---|
| H200 SXM | 73% | 69%–76% | 44% | 44%–44% | 3 |
| H100 SXM | 69% | 69%–74% | 43% | 43%–44% | 13 |
| A100 80GB | 63% | 63%–83% | 43% | 43%–69% | 10 |
| A100 40GB | 63% | 63%–65% | 43% | 35%–53% | 22 |
| L40S | 63% | 63%–63% | 43% | 43%–43% | 64 |
| L4 | 65% | 63%–65% | 46% | 45%–46% | 116 |
| A10G | 63% | 63%–63% | 43% | 43%–43% | 96 |
| T4 | 66% | 63%–68% | 46% | 43%–51% | 84 |
| V100 | 62% | 62%–63% | 33% | 33%–41% | 8 |
| RTX PRO 6000 Blackwell | 69% | 69%–69% | 44% | 44%–44% | 8 |
| MI300X | 74% | 74%–74% | — | — | 1 |
| MI325X | 76% | 76%–76% | — | — | 1 |
Compute and memory per dollar
Lowest on-demand price divided by the GPU’s dense 16-bit throughput (from the manufacturer’s datasheet) and by its memory. Real throughput depends on the workload; use this to compare classes, not to forecast run times.
| GPU | Lowest on-demand | At | $ per dense PFLOP-hour | $ per GB of memory-hour |
|---|---|---|---|---|
| V100 | $0.242 | DataCrunch | $1.94 | $0.0076 |
| MI300X | $2.59 | DigitalOcean | $1.98 | $0.0135 |
| B200 | $5.50 | Koyeb | $2.44 | $0.0306 |
| A100 40GB | $0.890 | Jarvislabs | $2.85 | $0.0223 |
| MI325X | $3.80 | DigitalOcean | $2.91 | $0.0148 |
| H200 SXM | $3.00 | Koyeb | $3.03 | $0.0213 |
| L40S | $1.20 | Koyeb | $3.31 | $0.0250 |
| MI355X | $8.60 | OCI | $3.42 | $0.0299 |
| H100 SXM | $3.49 | Jarvislabs | $3.53 | $0.0436 |
| B300 | $8.00 | Latitude.sh | $3.56 | $0.0296 |
| L4 | $0.440 | Jarvislabs | $3.64 | $0.0183 |
| A100 80GB | $1.49 | Jarvislabs | $4.78 | $0.0186 |
| RTX A6000 | $0.750 | Koyeb | $4.84 | $0.0156 |
| GB300 | $13.44 | DataCrunch | $5.38 | $0.0482 |
| GB200 | $16.00 | OCI | $6.40 | $0.0860 |
| T4 | $0.526 | AWS | $8.09 | $0.0329 |
| Quadro RTX 6000 | $1.50 | Akamai | $11.49 | $0.0625 |
| A10 | $2.00 | OCI | $16.00 | $0.0833 |
Price moves
Daily history started on 2026-11-08. Price moves appear here from the second day of data, and week-on-week trends after a week.
Reading this brief
- Prices are list prices from each provider’s public price list, per GPU-hour, in US dollars. Negotiated discounts, taxes, storage and network charges are not included.
- A listed price does not mean capacity is free right now, especially for the newest GPUs.
- Reserved means a 1- or 3-year commitment with no upfront payment where the provider offers it (Google: resource-based committed use discounts). How each figure is computed.