H200 GPU price per hour by provider, September 2026
Crusoe lists the cheapest on-demand Nvidia H200 at $4.29 per GPU-hour (US dollars, before tax) on 28 September 2026, with DigitalOcean, Nebius, Modal and RunPod within 30 cents and Together AI ($5.99) and CoreWeave ($6.30 per GPU from an eight-GPU node) a tier higher.
Crusoe rents the cheapest on-demand Nvidia H200 of seven GPU clouds that publish an on-demand H200 price, at $4.29 per GPU-hour on its own pricing page.1 The ranking uses each cloud’s lowest listed on-demand H200 price per GPU-hour, checked 28 September 2026, in US dollars before tax. CoreWeave comes last at $6.30, derived from its $50.44 hourly price for an eight-GPU node.2 Five of the seven span 30 cents, from $4.29 to $4.59.3 Nebius moves first: it lists $5.40 from 1 October 2026, up from $4.50 today.4
TL;DR
- Crusoe lists the lowest on-demand H200 at $4.29 per GPU-hour. DigitalOcean ($4.47), Nebius ($4.50), Modal ($4.54) and RunPod ($4.59) follow within 30 cents.1,5,4,6,3
- Together AI ($5.99) and CoreWeave ($6.30) form a higher tier. An eight-GPU node run for 720 hours costs $36,316.80 at CoreWeave’s list price and $24,710.40 at Crusoe’s.7,2,1
- Nebius raises its listed H200 price by 20.0% to $5.40 on 1 October 2026. The rise drops it from third to fifth.4
- At the same provider, the H200 costs 6 cents to $2.00 more per GPU-hour than the H100. The median premium across the seven is 59 cents, or 14.9%.5,7,6
- Per gigabyte of GPU memory, the H200 is cheaper than the H100 at all seven providers. The premium pays back only when a workload uses the extra memory.8
- Spot and reserved H200s cost far less but carry different terms at each provider. Nebius lists preemptible H200s from $0.79 and DigitalOcean a 12-month reserved rate of $3.40.4,5
How this ranking was built
The ranking belongs to our model infrastructure research and follows the Glamdring research method. The source is each GPU cloud’s own public pricing page, captured on 28 September 2026. The ranked field is the lowest listed on-demand price for one Nvidia H200 with 141 GB of memory, per GPU per hour. Where a page shows an on-demand GPU-hour figure (Nebius labels it On-demand, GPU-hour), the figure is used unchanged.4
Two prices needed arithmetic. Modal lists the H200 SXM at $0.001261 per second, which is $4.54 per hour at 3,600 seconds.6 CoreWeave’s North America table prices the HGX H200 as an eight-GPU node at $50.44 an hour. Divided by eight, that is $6.305 per GPU-hour, shown as $6.30 in the table.2
Nine GPU clouds were checked. Lambda lists B200, H100, A100 and GH200 instances but no H200, so it drops out.9 Vast.ai publishes prices that hosts set through supply and demand, so its median stays outside the ranking as a reference point.10 Seven providers remain. Spot, preemptible and reserved prices sit in their own section because each provider attaches different terms to them.
What does each GPU cloud charge for an H200?
Crusoe charges the least at $4.29 per GPU-hour and CoreWeave the most at $6.30, a difference of $2.01.1,2
| Rank | GPU cloud | H200 as listed | On-demand price (US$ per GPU-hour, before tax) | Basis |
|---|---|---|---|---|
| 1 | Crusoe | HGX H200, 141 GB | 4.29 | Listed per GPU-hour |
| 2 | DigitalOcean | HGX H200, 141 GB | 4.47 | Listed per GPU-hour; 1 or 8 GPUs per Droplet |
| 3 | Nebius | HGX H200 | 4.50 | Listed per GPU-hour; $5.40 from 1 October 2026 |
| 4 | Modal | H200 SXM | 4.54 | $0.001261 per second x 3,600 |
| 5 | RunPod | H200 SXM, 141 GB | 4.59 | Listed per hour (Pods) |
| 6 | Together AI | HGX H200, GPU Clusters | 5.99 | Listed per GPU-hour |
| 7 | CoreWeave | HGX H200, 8-GPU node, 141 GB | 6.30 | $50.44 per node-hour / 8 = $6.305 |
Sources for each row: Crusoe1, DigitalOcean5, Nebius4, Modal6, RunPod3, Together AI7 and CoreWeave2.
The top five sit so close that the hourly price rarely decides between them. The package around the GPU does. DigitalOcean lets a Droplet hold one or eight GPUs and bills per second with a five-minute minimum.5 RunPod’s H200 Pod comes with 24 vCPUs and 276 GB of system memory.3 Modal bills CPU cores and memory on top of the GPU, at $0.0000131 per core-second and $0.00000222 per GiB-second.6
The higher tier sells a different unit. CoreWeave’s main table prices whole eight-GPU nodes. It also shows a single-GPU inference price of $6.31 an hour for the H200.2 Together AI’s $5.99 is its GPU Clusters rate. Its Dedicated Inference table gives no H200 price and asks buyers to contact the company.7
The gap compounds at node scale. Eight GPUs for 720 hours is 5,760 GPU-hours. At Crusoe that costs $24,710.40. At CoreWeave the same node-month costs $36,316.80, or $11,606.40 more, before storage, networking and tax.1,2
Which H200 prices are changing?
Nebius is the only ranked provider with a dated H200 price change on its page. It lists $4.50 per GPU-hour today and $5.40 under the column GPU-hour (Effective October 1, 2026).4 The rise is 20.0%. For an eight-GPU node at 720 hours, it adds $5,184 a month at list price. After the change, Nebius ranks fifth, behind RunPod and ahead of Together AI.
The same Nebius column raises other GPUs too. The H100 goes from $3.85 to $4.50, up 16.9%. The HGX B200 goes from $7.15 to $8.50 and the HGX B300 from $7.85 to $9.50.4
DigitalOcean’s page states that new GPU pricing took effect on 1 August 2026 and that listed prices are subject to change. The captured page does not show the earlier rates.5 RunPod’s pricing page carries an “Updated September 27, 2026” stamp without naming a change.3
Together AI’s H100 promotion ends on 30 September 2026. Its Dedicated Inference table shows the HGX H100 at $3.99 against a standard $5.49 until that date.7 Its GPU Clusters H100 rate of $3.99, used in the H100 comparison below, carries no end date.
Thunder Compute published an H200 price table last reviewed on 18 September 2026. It lists the same prices as this ranking for Crusoe, DigitalOcean, Nebius, Modal and RunPod.11 Its CoreWeave figure is $6.31, normalised from an eight-GPU node, which matches CoreWeave’s $50.44 node price rounded up. The six listed prices were the same on both dates.
How much more does an H200 cost than an H100?
At the same provider, the H200 costs between 6 cents and $2.00 more per GPU-hour than the H100. The median premium is 59 cents, or 14.9%.5,7,6
| GPU cloud | H200 (US$/GPU-hour) | H100 (US$/GPU-hour) | Premium (US$) | Premium (%) | H200 per GB-hour (US$) | H100 per GB-hour (US$) |
|---|---|---|---|---|---|---|
| DigitalOcean | 4.47 | 4.41 | 0.06 | 1.4 | 0.0317 | 0.0551 |
| CoreWeave | 6.305 | 6.155 | 0.15 | 2.4 | 0.0447 | 0.0769 |
| Crusoe | 4.29 | 3.90 | 0.39 | 10.0 | 0.0304 | 0.0488 |
| Modal | 4.54 | 3.95 | 0.59 | 14.9 | 0.0322 | 0.0494 |
| Nebius | 4.50 | 3.85 | 0.65 | 16.9 | 0.0319 | 0.0481 |
| RunPod | 4.59 | 3.49 | 1.10 | 31.5 | 0.0326 | 0.0436 |
| Together AI | 5.99 | 3.99 | 2.00 | 50.1 | 0.0425 | 0.0499 |
H100 prices come from the same pages: Crusoe1, DigitalOcean5, Nebius4, Modal’s H100 SXM5 at $0.001097 per second6, RunPod’s H100 SXM3, Together AI’s GPU Clusters7 and CoreWeave’s eight-GPU node at $49.24 an hour2.
The premium depends on the provider far more than on the chip. DigitalOcean and CoreWeave charge almost the same for either GPU. RunPod and Together AI charge 31.5% and 50.1% more for the H200. Across providers, the cheapest H200 (Crusoe, $4.29) costs 80 cents or 22.9% more than the cheapest H100 SXM (RunPod, $3.49).1,3
Memory explains the price. NVIDIA lists the H200 with 141 GB of HBM3e at 4.8 TB/s, which it describes as nearly double the H100’s capacity with 1.4 times the memory bandwidth.8 Crusoe, DigitalOcean, RunPod and CoreWeave list their H100s at 80 GB.1,5,3,2 Divided by memory, the H200 is the cheaper chip at every provider. At Crusoe it costs $0.0304 per GB-hour against $0.0488 for the H100. The per-GB figures assume 80 GB for the H100 at Nebius, Modal and Together AI, whose pages don’t state it.
The decision changes when the model fits in 80 GB. Thunder Compute notes that the two chips share the same GH100 die, so the H200’s advantage appears where memory is the limit.11 NVIDIA claims up to twice the H100’s inference speed on Llama 2, a company figure tied to its own test settings.8 For how the hourly GPU price flows into a serving budget, read our explainer on what controls AI inference cost.
How much cheaper are spot and reserved H200s?
Interruptible and committed H200s cost well below on-demand, but each provider sets its own terms, so these prices stay out of the ranking.
- Nebius lists preemptible H200s from $0.79 per GPU-hour. It also offers up to 35% off on-demand rates for large clusters reserved for several months.4
- Together AI lists a preemptible H200 at $2.99. Its reserved H200 falls from $4.99 for 7 to 30 days to $3.99 for 91 to 180 days.7
- CoreWeave lists H200 spot capacity at $20.93 per eight-GPU node-hour in North America, which works out to $2.62 per GPU-hour.2
- DigitalOcean lists a 12-month reserved H200 at $3.40 per GPU-hour. Its spot plans on the captured page cover the B300 and AMD Instinct GPUs only.5
- Crusoe routes H200 spot pricing through its sales team.1
The reserved prices put a floor under negotiation. DigitalOcean’s $3.40 for a year is 23.9% below its own on-demand rate of $4.47.5
Why do other published H200 prices disagree?
Other trackers measure different things: marketplace medians, the single lowest offer or node prices rounded to one GPU.
Vast.ai’s own page shows H200 offers from $2.63 an hour with a median of $4.30, set by hosts across its marketplace.10 Thunder Compute reports Vast.ai at $3.96, the median of six verified US and Canadian hosts priced with 100 GB of storage.11 gpus.io’s search listing reports H200s from $2.63 per GPU-hour with a median of $4.40 across 21 configurations.12 The $2.63 floor matches Vast.ai’s lowest offer on the same day.
Variant matters. Vast.ai lists the H200 NVL separately at a median of $4.00 an hour.10 The NVL is NVIDIA’s PCIe card, rated up to 600 W against up to 700 W for the SXM version.8 Check which variant a quoted price covers.
The hyperscalers sit far higher. Thunder Compute’s 18 September table lists AWS at $7.91 per GPU-hour, normalised from an eight-GPU node in its cheapest US region.11 It lists Oracle Cloud at $10.00, Azure at $10.60 and Google Cloud at $10.85.11 Those figures are Thunder Compute’s. This ranking did not check the hyperscalers’ own pages.
What the ranking cannot tell you
A list price is a statement from the company that sets it. It does not show whether capacity exists at that price on a given day. Crusoe sends buyers to sales for guaranteed capacity, and Together AI for H200 reservations beyond 180 days.1,7
The ranked figure also leaves out most of the bill. DigitalOcean keeps charging for a powered-off GPU Droplet until it is destroyed.5 Modal charges CPU and memory separately from the GPU.6 Storage, network egress and local tax vary by provider and by buyer.
Region matters at CoreWeave, whose figure comes from its North America table.2 Negotiated and committed prices sit outside the ranking, and so does performance per dollar. The AWS, Azure, Google Cloud and Oracle rates above come from a third-party table, not from their own pages.
Confirm the price in the provider’s console on the day you commit. Use one real workload to test whether it needs 141 GB before paying the H200 premium.
Frequently asked questions
How much does it cost to rent an H200 GPU per hour?
On 28 September 2026, listed on-demand prices for one H200 ran from $4.29 per GPU-hour at Crusoe to $6.30 at CoreWeave, in US dollars before tax.1,2 Marketplace offers on Vast.ai started at $2.63 with a median of $4.30.10 Interruptible capacity went lower, with Nebius listing preemptible H200s from $0.79.4
Why is the H200 so expensive?
The H200 carries 141 GB of HBM3e memory at 4.8 TB/s, which NVIDIA describes as nearly double the H100’s capacity.8 Cloud prices reflect that memory. Measured per gigabyte, an H200 rents for less than an H100 at all seven providers checked. Crusoe charges $0.0304 per GB-hour for the H200 and $0.0488 for the H100.1
Is the H200 the most powerful GPU?
No. Thunder Compute describes the B200 as NVIDIA’s Blackwell successor to the H200.11 GPU clouds price the newer chips higher. RunPod lists the B200 at $6.79 an hour and the B300 at $7.89, against $4.59 for the H200.3 Nebius lists the HGX B200 at $7.15 and the HGX B300 at $7.85 per GPU-hour.4
Does Lambda rent the H200?
Not on its public price list as of 28 September 2026. Lambda lists B200, H100, A100 and GH200 instances.9 The GH200 it lists at $2.29 per GPU-hour carries 96 GB of GPU memory and is a different product from the H200.9
Sources checked
Company-owned pages establish what a company says. They do not prove a market conclusion. Each source is dated so readers can judge each claim.
