H100 GPU price per hour: 8 clouds ranked, Sep 2026
Ranked by lowest listed on-demand H100 price per GPU-hour, checked 28 September 2026, RunPod leads at $2.89 (H100 PCIe) and CoreWeave is last at $6.16 ($49.24 per 8-GPU node, divided by eight).
RunPod lists the lowest on-demand H100 price of the eight GPU clouds we ranked: $2.89 per GPU-hour for an 80 GB H100 PCIe, on its pricing page checked 28 September 2026.1 CoreWeave sits at the other end. Its 8-GPU HGX H100 node lists at $49.24 an hour, or $6.16 per GPU-hour.2 That’s a gap of about 2.1 times for one chip family, and the variant, billing unit and host bundle explain much of it.
TL;DR
- Ranked by the lowest listed on-demand H100 price per GPU-hour, checked 28 September 2026, RunPod ($2.89) and Lambda ($3.29) lead. Both prices are for H100 PCIe.1,3
- On SXM-class H100s, RunPod still leads at $3.49, ahead of Nebius at $3.85 and Crusoe at $3.90.1,4,5
- The middle of the market is tight. Five of the eight list between $3.85 and $4.41, and the median of all eight is $3.925.
- Interruption and commitment cut the price further. Nebius lists preemptible H100s from $0.79 per GPU-hour, Together AI lists $1.99, and DigitalOcean lists a 12-month reserved rate of $3.26.4,6,7
- Vast.ai’s marketplace median for H100 SXM was $2.16 an hour, below every fixed on-demand SXM price in the ranking.8
- Nebius lists $4.50 from 1 October 2026. At that rate it falls from third to seventh.4
- The hourly rate is one input to cost per task. Our explainer on what controls AI inference cost covers what sits beyond the GPU line.
How this ranking was built
Glamdring Research selected eight GPU clouds that rent H100 capacity by the hour: RunPod, Lambda, Nebius, Crusoe, Modal, Together AI, DigitalOcean and CoreWeave. We captured each company’s own pricing page on 28 September 2026, plus the Vast.ai pricing page as a marketplace reference.
- Source: each provider’s public pricing page, captured 28 September 2026.
- Field: the lowest listed on-demand H100 price per GPU-hour. Any H100 variant counts: PCIe, SXM, HGX or NVL.
- Unit: US dollars per GPU per hour, as each page prints it. Lambda lists its prices “plus applicable sales tax/VAT/GST”.3 DigitalOcean says it applies tax in some countries.7
- Conversions: CoreWeave prices a whole 8-GPU node, so its node price is divided by eight. Modal prices per second, so its rate is multiplied by 3,600. No other figure is changed.
- Row count: nine pages captured, eight providers listing an on-demand H100, eight ranked. Vast.ai is shown beside the ranking, not in it, because hosts set its prices on a marketplace.8
- Not ranked: spot, preemptible and reserved rates. They have their own table below.
These are company statements about company prices, which is how our research methodology treats them. The ranking belongs to our model infrastructure coverage.
Which cloud rents the H100 cheapest on demand?
RunPod does, at $2.89 per GPU-hour for H100 PCIe, checked 28 September 2026.1 Lambda is second at $3.29 for a 1-GPU H100 PCIe instance.3 CoreWeave is last at $6.16, the per-GPU share of its $49.24 8-GPU node.2
| Rank | Provider | H100 variant priced | On-demand (US$ per GPU-hour) | How the page lists it |
|---|---|---|---|---|
| 1 | RunPod | H100 PCIe, 80 GB | 2.89 | $2.89/hr per GPU |
| 2 | Lambda | H100 PCIe, 80 GB, 1-GPU instance | 3.29 | $3.29 per GPU-hour, plus tax |
| 3 | Nebius | HGX H100 | 3.85 | $3.85 per GPU-hour; $4.50 from 1 October 2026 |
| 4 | Crusoe | HGX H100, 80 GB | 3.90 | $3.90/GPU-hr |
| 5 | Modal | H100 SXM5 | 3.95 | $0.001097 per second, × 3,600 |
| 6 | Together AI | HGX H100, GPU Clusters | 3.99 | $3.99 per GPU-hour |
| 7 | DigitalOcean | HGX H100 | 4.41 | $4.41/GPU/hour |
| 8 | CoreWeave | HGX H100, 8-GPU node | 6.16 | $49.24 per node-hour, ÷ 8 |
Each row comes from that provider’s own pricing page, checked 28 September 2026.1,3,4,5,9,6,7,2
The hourly figure buys a different host at each provider. RunPod’s H100 PCIe comes with 16 vCPUs and 188 GB of RAM.1 Lambda’s 1-GPU H100 PCIe instance carries 26 vCPUs, 225 GiB of RAM and a 1 TiB SSD.3 DigitalOcean’s H100 Droplet lists 20 vCPUs, 240 GiB of memory and a 5 TiB NVMe scratch disk, in 1-GPU or 8-GPU sizes.7 CoreWeave’s node price covers eight GPUs, 128 vCPUs, 2,048 GB of RAM and 61.44 TB of local storage.2 So the lower rows aren’t simply dearer copies of the top row.
Over a 30-day month of 720 hours, RunPod’s rate comes to $2,080.80 per GPU and CoreWeave’s to $4,435.20.
What is the cheapest H100 SXM or HGX price?
RunPod’s H100 SXM, at $3.49 per GPU-hour, is the cheapest SXM-class listing among the eight, checked 28 September 2026.1 Nebius’s HGX H100 follows at $3.85, then Crusoe’s HGX H100 at $3.90 and Modal’s H100 SXM5 at $3.95.4,5,9 The distinction matters because the top two ranked prices are both PCIe cards.
| Provider | SXM-class H100 listing | On-demand (US$ per GPU-hour) |
|---|---|---|
| RunPod | H100 SXM, 80 GB | 3.49 |
| Nebius | HGX H100 | 3.85 |
| Crusoe | HGX H100, 80 GB | 3.90 |
| Modal | H100 SXM5 | 3.95 |
| Lambda | H100 SXM, 8-GPU instance | 3.99 |
| Together AI | HGX H100, GPU Clusters | 3.99 |
| DigitalOcean | HGX H100 | 4.41 |
| CoreWeave | HGX H100, 8-GPU node | 6.16 |
NVIDIA’s specification separates the parts. The H100 SXM carries 80 GB of memory at 3.35 TB/s, draws up to 700 W and links GPUs over 900 GB/s NVLink. The PCIe-based H100 NVL carries 94 GB at 3.9 TB/s, draws 350 to 400 W and uses 600 GB/s NVLink.10 NVIDIA lists HGX H100 systems with 4 or 8 GPUs as the server option for the SXM part, so an HGX listing is an SXM listing.10 Its specification table has no column for the 80 GB H100 PCIe that RunPod and Lambda price lowest.
Lambda’s SXM price rises as the instance shrinks: $3.99 per GPU on the 8-GPU instance, $4.09 on four GPUs, $4.19 on two and $4.29 on one.3 RunPod also lists the 94 GB H100 NVL at $3.19.1
How were node and per-second prices converted?
Two of the eight don’t print a per-GPU-hour figure, so we converted them. CoreWeave lists its HGX H100 as an 8-GPU node at $49.24 an hour, and $49.24 ÷ 8 = $6.155, shown as $6.16.2 Modal lists the H100 SXM5 at $0.001097 per second, and $0.001097 × 3,600 = $3.9492, shown as $3.95.9
Both conversions agree with the companies’ own figures. CoreWeave’s table also carries an inference single-GPU price for the HGX H100 of $6.16 an hour.2 Modal’s serverless worked example on the same page uses $3.95 per GPU-hour.9
Billing granularity is a separate question from price. Modal says customers pay only for the compute time they use.9 DigitalOcean bills GPU Droplets per second with a 5-minute minimum, and it keeps billing a powered-off Droplet until it is destroyed.7 RunPod’s pricing page offers per-hour and per-second views.1 A per-second bill saves money only when the workload releases the GPU.
What do spot and reserved H100 prices cost?
Five of the eight publish a spot, preemptible or reserved H100 figure. The rest send those terms to sales or don’t list them. Checked 28 September 2026:
| Provider | Spot or preemptible (US$ per GPU-hour) | Reserved or committed (US$ per GPU-hour) | On-demand (US$ per GPU-hour) |
|---|---|---|---|
| Nebius | from 0.79 | Up to 35% below on-demand for multi-month clusters; no H100 figure | 3.85 |
| Together AI | 1.99 | 3.69 (7–30 days), 3.45 (31–90), 3.19 (91–180); 181+ days on request | 3.99 |
| CoreWeave | 2.46 in North America ($19.71 node ÷ 8); 2.44 in Europe ($19.51 ÷ 8) | No H100 figure listed | 6.16 |
| DigitalOcean | No H100 spot plan listed | 3.26 on a 12-month contract | 4.41 |
| Lambda | Not listed | 1-Click Clusters, 2 weeks to 1 year: 6.16 (16 GPUs), 5.85 (64), 5.54 (256) | 3.29 |
| Crusoe | Contact sales | Contact sales | 3.90 |
| RunPod | Not listed for H100 | Contact sales | 2.89 |
| Modal | Not listed | Volume-based discounts on Enterprise | 3.95 |
Sources: each provider’s pricing page, checked 28 September 2026.4,6,2,7,3,5,1,9
Spot carries the deepest discount and the least certainty. Nebius says its spot prices are dynamic, can change as often as every 15 minutes, and top out one cent below the on-demand rate.4 Together AI’s $1.99 preemptible H100 is about half its $3.99 on-demand rate.6 CoreWeave’s spot node costs about 40% of its on-demand node.2
Commitment discounts are smaller. DigitalOcean’s 12-month rate is $1.15, or 26%, below its on-demand price.7 Together AI’s 91–180 day rate is 20% below on-demand.6 Lambda runs the other way. Its 1-Click Cluster rates, for 16 to 256 H100s, sit above its $3.99 8-GPU SXM instance, and the pricing page doesn’t state what the premium covers.3
How does the Vast.ai marketplace median compare?
Vast.ai’s median H100 SXM listing was $2.16 an hour on 28 September 2026. That’s $1.33 below RunPod’s $3.49, the cheapest fixed on-demand SXM price in the ranking.8,1 Vast.ai’s medians for H100 PCIe ($2.53) and H100 NVL ($2.76) also sit below RunPod’s matching listings of $2.89 and $3.19.8
Vast.ai says its prices are “set by supply and demand across 40+ data centers”, and it offers on-demand, interruptible and reserved rental.8 Its lowest H100 SXM offer was $1.73 an hour.8 A median of host-set offers is a different figure from a list price. Hosts differ in hardware, reliability and location. The page also offers 30-day and 90-day views, and the page text we checked doesn’t say which window the median uses.8 That is why Vast.ai sits beside the ranking rather than inside it.
Why do other published H100 prices disagree?
Most disagreement comes from the variant, the date and the unit. Jarvislabs, which isn’t among the eight, published $2.69 per GPU-hour for H100 SXM, checked 8 September 2026.11 The same guide quotes Lambda at $4.29, the price of Lambda’s 1-GPU H100 SXM instance.11 This ranking uses Lambda’s $3.29 1-GPU H100 PCIe instance, and Lambda’s page lists both.3 Jarvislabs also announces new H100 rates from 5 October 2026.11
Purchase prices answer a different question. One new H100 80 GB PCIe card was listed on eBay at US$47,520 or best offer, shipping from Shenzhen, with the listing last updated 27 September 2026.12 One seller’s asking price is a company statement, not a market price. Jarvislabs’s buy-versus-rent example uses $30,000 per GPU of installed cost and calls it “a budgeting assumption, not a current market quote.”11
What changes in October 2026?
Nebius has published its next H100 rate: $4.50 per GPU-hour from 1 October 2026, up $0.65, or 17%, from $3.85.4 At $4.50, Nebius moves from third to seventh, just above CoreWeave. The rest of the order holds unless another page changes.
Together AI’s Dedicated Inference H100 lists at $5.49, with a $3.99 promotion valid until 30 September 2026.6 The $3.99 ranked here is a different line: the GPU Clusters on-demand rate, which the page doesn’t mark as a promotion.6
Two other pages were repriced recently. DigitalOcean’s new GPU pricing took effect on 1 August 2026, and RunPod dates its page 27 September 2026.7,1 This is Glamdring’s first H100 price ranking, so there’s no earlier release to compare.
What the ranking cannot tell you
The ranking covers eight selected clouds, not the whole market. Jarvislabs’s $2.69 SXM price from 8 September would rank first if it still held on 28 September.11 We didn’t check its pricing page for this ranking.
It doesn’t measure availability. A listed price states what a provider charges, not whether an H100 is free to launch in your region today. Jarvislabs makes the same point about its own comparison: a snapshot of listed rates is “not a guarantee of available capacity”.11
It doesn’t measure cost per task. Host CPU, memory and storage differ by provider, and the variant changes memory bandwidth and interconnect.10 Storage, network and idle time add to the GPU line. Run one representative job on the two or three cheapest options that fit it, and compare the cost of the finished job. Confirm the variant, region and tax in the current quote.
One ranked figure carries a label question. RunPod’s page has Community Cloud and Secure Cloud views, and the page text we checked doesn’t label which view carries the $2.89.1
New price rankings in this field go out in our free email briefing.
Frequently asked questions
How much does an H100 GPU cluster cost?
Among pages that price a full 8-GPU H100 configuration on demand, an 8-GPU node runs from $31.92 an hour at Lambda to $49.24 at CoreWeave, checked 28 September 2026. Lambda’s figure is eight GPUs at its $3.99 SXM rate.3 DigitalOcean’s 8-GPU Droplet works out to $35.28 an hour at $4.41 per GPU.7 CoreWeave lists its HGX H100 node at $49.24.2 Larger clusters cost more per GPU at Lambda: its 1-Click Clusters list $6.16 per GPU-hour for 16 GPUs, or $98.56 an hour.3 Together AI lists H100 GPU Clusters at $3.99 per GPU-hour on demand, and RunPod sends H100 cluster buyers to sales.6,1
How much does it cost to rent an H100 for a month?
One H100 running for 720 hours, a 30-day month, costs $2,080.80 at RunPod’s $2.89 on-demand rate and $4,435.20 at CoreWeave’s $6.16, checked 28 September 2026.1,2 Mid-table rates land near $2,800: $2,808 at Crusoe’s $3.90 and $2,844 at Modal’s $3.95.5,9 DigitalOcean’s 12-month reserved rate of $3.26 comes to $2,347.20 for the same hours.7 These figures cover the GPU line only, before storage, network and tax.
Does the hourly H100 price include tax?
Not always. Lambda lists its prices “plus applicable sales tax/VAT/GST”.3 DigitalOcean says it is required by law to apply taxes in some countries.7 All eight pages price in US dollars, and a US-dollar price doesn’t mean the GPU sits in the United States, as Jarvislabs notes in its own guide.11 Confirm the tax line and the deployment region in the quote.
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.
