An NVIDIA H100 rents for $3.84 per GPU-hour on average across 27 cloud providers, an H200 for $4.43 across 20, and a B200 for $6.39 across 12, as tracked by the Mercatus GPU Index on September 13, 2026. Those averages hide the number that actually matters to a buyer: the same H100 lists at $1.79 an hour from one provider and $10.00 from another. On an 8-GPU node running around the clock, that is the difference between $10,500 and $58,400 a month for identical hardware.
This page is the static reference for that market: the current average and range for each data-center GPU, the hourly rate at every provider the index tracks for H100, H200, and B200, what a node costs per month at each end of the range, and how a GPU-hour translates into the price of a token. The live version, updated every five minutes, is the Mercatus GPU Index.
GPU rental prices at a glance
Volume-weighted average on-demand price per GPU-hour, with the low and high across tracked providers. Prices in USD as of September 13, 2026.
| GPU | VRAM | Average $/GPU-hr | Range | Providers | 30-day change |
|---|---|---|---|---|---|
| B300 | 288GB HBM3e | $8.89 | $7.50 to $24.22 | 7 | +5.3% |
| B200 | 192GB HBM3e | $6.39 | $3.74 to $16.14 | 12 | 0.0% |
| AMD MI300X | 192GB HBM3 | $4.98 | $2.39 to $6.94 | 6 | -0.1% |
| H200 | 141GB HBM3e | $4.43 | $2.25 to $11.71 | 20 | -0.8% |
| H100 | 80GB HBM3 | $3.84 | $1.79 to $12.52 | 27 | -0.4% |
| RTX PRO 6000 | 96GB GDDR7 | $2.07 | $0.99 to $5.64 | 11 | +1.4% |
| A100 80GB | 80GB HBM2e | $2.05 | $0.82 to $4.89 | 18 | +0.4% |
| L40S | 48GB GDDR6 | $1.41 | $0.58 to $4.24 | 16 | 0.0% |
Three things stand out.
The H100 is the reference price, and it has been rising. At $3.84 it sits 6.1 percent above where it was 90 days ago ($3.62), with the period low on June 27 and the high of $3.87 on August 15. A GPU that launched in 2022 got more expensive to rent over the summer of 2026, which is not what a depreciation schedule would predict. Demand for inference capacity, not the age of the silicon, is setting the rate. The H100 GPU cost page covers what that hardware costs to buy outright.
The H200 premium over H100 is 15 percent for 76 percent more memory. At $4.43 against $3.84, the H200 is the cheapest HBM per dollar in the data-center tier: about 32GB per dollar-hour versus 21GB on the H100. For memory-bound inference, which is most of it, the H200 is the better deal on paper, and its 7-day move of -2.4 percent suggests providers are pricing it to fill.
Blackwell carries a 66 percent premium over Hopper. B200 at $6.39 is 1.66x the H100 rate. Whether that is worth it depends entirely on throughput per dollar on your workload, which is the calculation in GPU ROI. The B300 at $8.89, up 5.3 percent in 30 days, is the only data-center GPU in the table moving meaningfully in either direction.
H100 rental price by provider
Lowest listed on-demand price per GPU-hour at each provider, sorted cheapest first, with the configurations offered. Two of the 27 tracked providers (AWS and Azure) are in the index but not in the per-provider snapshot below.
| Provider | $/GPU-hr | GPU counts offered |
|---|---|---|
| Vast.ai | $1.47 to $3.00 | 1, 2, 4, 8 |
| Latitude.sh | $1.66 to $2.24 | 1, 4, 8 |
| CUDO Compute | $1.79 | 1 |
| Horizon Compute | $1.95 | 8 |
| Voltage Park | $1.99 | 1, 2, 4, 8 |
| Yotta Labs | $2.15 to $3.60 | 1, 2, 4, 8 |
| Shadeform | $2.48 | 1, 2, 4, 8 |
| Denvr Cloud | $2.50 | 8 |
| Gcore Cloud | $2.50 | 8 |
| Hyperstack | $2.50 to $3.20 | 1, 2, 4, 8 |
| Massed Compute | $2.73 to $3.14 | 1, 2, 4, 8 |
| Sesterce | $2.75 to $6.59 | 1, 2 |
| Lyceum Technology | $2.79 | 1 |
| Scaleway | $2.87 to $3.31 | 1, 2, 4 |
| RunPod | $2.89 to $3.49 | 1 |
| CanopyWave | $3.25 | 2, 4, 8 |
| Verda Cloud | $3.25 | 1, 2, 4, 8 |
| Lambda Labs | $3.29 to $4.29 | 1, 2, 4, 8 |
| Oblivus | $3.50 to $4.31 | 1, 2, 4, 8 |
| Crusoe Cloud | $3.90 | 1 |
| Enverge | $4.00 | 1 |
| DigitalOcean | $4.41 | 1, 8 |
| Paperspace | $5.99 | 1 |
| Google Cloud | $9.80 to $15.67 | 1 |
| Oracle Cloud Infrastructure | $10.00 | 8 |
The shape of this table is the story of the GPU cloud market. Nineteen providers sit between $1.47 and $3.50. Then there is a gap, and the hyperscalers sit at $9.80 and $10.00, 2.5x the volume-weighted average and 5x the specialist floor. Vast.ai’s $1.47 is a spot and community marketplace rate and sits below the index’s own $1.79 floor, which applies outlier filtering.
The gap is not a mistake. Hyperscaler pricing bundles enterprise support, compliance, and integration with the rest of the platform. A buyer who needs those things is paying for them. A buyer who only needs the GPU is paying for them anyway if they rent from a hyperscaler, and the market has responded by building a tier of specialist providers whose entire business is the $2 to $3.50 row. Why the same hardware carries such different prices is covered in more depth in Why token prices differ, since the same forces set both markets.
H200 rental price by provider
Same source and date. Eighteen of 20 tracked providers listed in the snapshot.
| Provider | $/GPU-hr | GPU counts offered |
|---|---|---|
| Vast.ai | $1.98 to $4.00 | 1, 8 |
| Amaya Cloud | $2.25 | 8 |
| Shadeform | $2.25 to $4.63 | 1, 8 |
| BoostRun | $2.45 | 8 |
| Horizon Compute | $3.00 | 8 |
| Yotta Labs | $3.50 to $4.41 | 1, 2, 4, 8 |
| Massed Compute | $3.62 to $3.82 | 1, 2, 4, 8 |
| RunPod | $3.79 to $4.59 | 1 |
| Hyperstack | $3.99 | 8 |
| CanopyWave | $4.00 | 8 |
| Verda Cloud | $4.00 | 1, 2, 4, 8 |
| Crusoe Cloud | $4.29 | 1 |
| Lyceum Technology | $4.29 | 1 |
| Sesterce | $4.40 | 8 |
| DigitalOcean | $4.47 | 1, 8 |
| Enverge | $5.50 | 1 |
| Google Cloud | $9.31 to $11.17 | 1 |
| Oracle Cloud Infrastructure | $10.00 | 8 |
The cheapest H200 nodes, at Amaya and BoostRun, are priced below the average H100. An 8xH200 node from Amaya at $2.25 costs $13,140 a month; the same node at Oracle costs $58,400. The H200 floor is where the price-sensitive inference buyer should be looking today, since 141GB of HBM3e at $2.25 an hour is the best memory-per-dollar in the entire index. Purchase-side context, including what an 8-GPU H200 server costs to own, is in H200 server price.
B200 rental price by provider
Eleven of 12 tracked providers listed in the snapshot.
| Provider | $/GPU-hr | GPU counts offered |
|---|---|---|
| BoostRun | $3.74 | 8 |
| Shadeform | $4.51 | 8 |
| Yotta Labs | $5.80 to $6.69 | 1, 2, 4, 8 |
| Verda Cloud | $6.11 | 1, 2, 4, 8 |
| Lyceum Technology | $6.49 | 1 |
| Vast.ai | $6.63 to $7.88 | 1, 8 |
| Lambda Labs | $6.69 to $6.99 | 1, 2, 8 |
| RunPod | $6.79 | 1 |
| Nebius AI Cloud | $7.15 | 1 |
| Sesterce | $7.17 to $7.69 | 1 |
| Oracle Cloud Infrastructure | $14.00 | 8 |
Blackwell is a younger market and the table shows it: the specialist tier clusters tightly between $5.80 and $7.69 rather than spreading from $1.50 to $3.50 the way the H100 tier does. BoostRun lists a full 8-GPU B200 node at $3.74, below the average H200 rate, and Shadeform at $4.51; everyone else is above $5.80. The B200 index range runs to $16.14, above Oracle’s $14.00, which reflects a hyperscaler listing not in the snapshot. The B200 server price page covers the purchase side.
What an 8-GPU node costs per month
Eight GPUs, on demand, running 730 hours a month, at the volume-weighted average and at each end of the provider table. Reserved and committed-use pricing runs lower than these figures; spot runs lower still and can be interrupted.
| Node | At cheapest 8-GPU provider | At index average | At most expensive provider |
|---|---|---|---|
| 8x H100 | $11,388 (Horizon Compute, $1.95) | $22,426 | $58,400 (Oracle, $10.00) |
| 8x H200 | $13,140 (Amaya Cloud, $2.25) | $25,871 | $58,400 (Oracle, $10.00) |
| 8x B200 | $21,842 (BoostRun, $3.74) | $37,318 | $81,760 (Oracle, $14.00) |
Two conclusions.
Provider choice moves the bill more than GPU choice does. The gap between the cheapest and most expensive H100 provider ($47,000 a month on one node) is larger than the gap between an average H100 node and an average B200 node ($15,000). A team choosing between Hopper and Blackwell is making a smaller financial decision than a team choosing between a specialist and a hyperscaler.
The rent-or-buy line sits inside this range. At $22,426 a month, an average-priced 8xH100 node costs roughly $269,000 a year to rent. At $11,388, it costs $137,000. Whether either number beats owning depends on utilization, financing, and what the hardware is worth in three years, which is the framework in Should you buy or rent GPUs and the resale math in H100 resale value. What this table adds is that the answer changes depending on which rental price you compare against. A buy-vs-rent model built on hyperscaler rates makes owning look inevitable; one built on the specialist floor makes it a close call.
From GPU-hour to token price
The hourly rental rate is the floor under every token price in the market. A provider selling inference has to earn back the GPU-hour in tokens, so the rate card on a model like DeepSeek V4.1 Flash is really a bet about throughput.
The arithmetic is simple. At the index-average H100 rate, an 8-GPU node costs $30.72 an hour. To break even selling output tokens at $0.60 per million (DeepSeek’s official V4.1 Flash off-peak rate), that node has to produce 51.2 million output tokens an hour, about 14,200 tokens a second sustained. At $1.98 per million (the V4 Pro rate), the requirement drops to 15.5 million tokens an hour, or 4,300 a second. At $6.00 per million, a rate typical of closed frontier models, 1,400 a second is enough.
That is why the cheapest hosts of open-weight models are the ones renting at the bottom of the tables above, or running owned hardware at high utilization, and why hosts on hyperscaler rates cannot match them on price. It is also why the H100’s summer price rise matters beyond the GPU market: every dollar added to the hourly rate has to come out of somewhere, and on a model priced at $0.60 per million output tokens there is not much room. The full chain from hardware cost to model price is worked through in Cost per token: how infrastructure becomes model pricing, and the utilization side in GPU utilization.
How to read the rental market
Price the specialist tier, then decide if you need the hyperscaler. The $2 to $3.50 H100 row is where the volume is. Start there and add the hyperscaler premium only for a reason you can name.
Watch the H200 floor. At $2.25 to $2.45 for 8-GPU nodes, it is the best memory-per-dollar available and it is priced below the H100 average. That gap will not last if the buyers notice.
Treat single-GPU prices and 8-GPU prices as different markets. Several providers list only one or the other. A $1.79 single H100 at CUDO does not mean an 8-GPU node is available at that rate.
Recheck monthly, or use the live index. The H100 average moved 6 percent in 90 days and the B300 moved 5 percent in 30. The Mercatus GPU Index refreshes every five minutes and carries the 90-day series for every GPU in this table.
Frequently asked questions
How much does it cost to rent an H100 per hour? As of September 13, 2026, the volume-weighted average across 27 providers is $3.84 per GPU-hour. Specialist providers list from $1.79 (CUDO Compute, single GPU) and $1.95 (Horizon Compute, 8-GPU node). Google Cloud and Oracle list at $9.80 and $10.00. Spot and community marketplaces go lower, from $1.47 on Vast.ai.
How much does it cost to rent an H200 per hour? $4.43 average across 20 providers. 8-GPU nodes list from $2.25 (Amaya Cloud) and $2.45 (BoostRun). Hyperscalers list at $9.31 to $11.17.
How much does it cost to rent a B200 per hour? $6.39 average across 12 providers. 8-GPU nodes from $3.74 (BoostRun) and $4.51 (Shadeform); most specialist providers sit between $5.80 and $7.69; Oracle lists at $14.00.
What does an 8xH100 node cost per month? About $22,400 at the index average, $11,400 at the cheapest 8-GPU provider, and $58,400 at the most expensive, for 730 hours of on-demand use.
Why do GPU rental prices vary so much between providers? Hyperscalers bundle support, compliance, and platform integration into the rate. Specialist providers sell the GPU and little else. Marketplaces like Vast.ai resell community and spot capacity. The hardware is identical; the rate reflects what surrounds it and how much of the provider’s fleet is idle.
Are H100 rental prices going down? Not over the last 90 days. The index average rose from $3.62 to $3.84, up 6.1 percent, peaking at $3.87 on August 15. The 30-day move is -0.4 percent. H200 is down 2.4 percent over 7 days.
Is it cheaper to rent or buy an H100? It depends on utilization and which rental rate you compare against. At the specialist floor, renting an 8-GPU node costs about $137,000 a year; at the index average, $269,000. The purchase-side numbers and the payback math are in GPU ROI and H100 GPU cost.
Methodology
All prices are on-demand rates in USD per GPU-hour from the Mercatus GPU Index and its per-GPU pages for H100, H200, and B200, as of September 13, 2026. The index collects prices every five minutes from 50+ providers via API and web scraping, normalizes them to per-GPU-hour, and computes a volume-weighted average with outlier filtering. Provider tables show the lowest listed price and the GPU counts offered; the index range for each GPU may differ from the provider table because the range applies availability and outlier filters and includes providers not in the first-page snapshot. Monthly node costs assume 8 GPUs at 730 hours. Token break-even figures use DeepSeek’s official V4.1 Flash and V4 Pro off-peak output rates as of September 10, 2026, and describe required throughput, not measured throughput. Prices change continuously; the live index is the current source.