Cloud GPU providers ranked by H100 price, 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). All eight clouds rent H100s.
RunPod lists the cheapest on-demand H100 of the eight GPU clouds we checked on 28 September 2026: $2.89 per GPU-hour for an 80 GB H100 PCIe.1 Lambda is second at $3.29, also for an H100 PCIe.2 All eight rent H100s. CoreWeave lists the highest rate, $49.24 an hour for an 8-GPU HGX H100 node, which is $6.16 per GPU-hour.3
That’s a gap of about 2.1 times for one chip family. The form factor, the bundle and the billing unit explain much of it.
TL;DR
- Ranked by lowest listed on-demand H100 price per GPU-hour, checked 28 September 2026: RunPod $2.89, Lambda $3.29, Nebius $3.85, Crusoe $3.90, Modal $3.95, Together AI $3.99, DigitalOcean $4.41, CoreWeave $6.16.
- The top two prices are for H100 PCIe. Among the other H100 listings, RunPod’s H100 SXM at $3.49 is still the cheapest.1
- Billing units differ. Modal and DigitalOcean bill by the second, Lambda by the minute, Crusoe and Together AI by the hour, and CoreWeave prices a whole 8-GPU node.4,5,6,7,8,3
- Commitment cuts the rate. Together AI lists $3.19 for 91 to 180 days and DigitalOcean $3.26 on a 12-month contract. The other six give a maximum discount, a cluster rate or a sales contact instead of a reserved H100 price.8,5
- The order moves on 1 October 2026, when Nebius lists $4.50 and drops from third to seventh.9
How this ranking was built
We ranked one field: the lowest on-demand H100 price each provider lists on its own public pricing page, in US dollars per GPU-hour. The pages were captured on 28 September 2026. Our methodology page sets out the research standard these figures follow.
- Scope. Any H100 variant counts: PCIe, SXM, NVL or HGX. Each provider’s cheapest H100 listing is its ranked price.
- Conversions. CoreWeave prices an 8-GPU node, so its node-hour rate is divided by eight. Modal prices per second, so its rate is multiplied by 3,600. No other figure is changed.
- Exclusions. Spot, preemptible and reserved prices are reported beside the ranking, not in it. Vast.ai is a marketplace of host-set offers, so its median is shown for reference and not ranked.10
- Row count. Nine pricing pages captured. Eight list an on-demand H100 price. Eight ranked.
These are company-published list prices. They establish what each provider says it charges. They don’t establish what a buyer pays after a quote, a discount or a stock-out.
Which cloud GPU providers rent H100s, and which is cheapest?
All eight rent H100s on demand, and RunPod is cheapest at $2.89 per GPU-hour. More GPU and serving-cost analysis sits in our model infrastructure research.
| Rank | Provider | H100 listing ranked | On-demand price (USD per GPU-hour, as listed) | Price as the provider lists it |
|---|---|---|---|---|
| 1 | RunPod | H100 PCIe 80 GB | 2.89 | $2.89/hr, Secure Cloud1,11 |
| 2 | Lambda | H100 PCIe 80 GB, single-GPU instance | 3.29 | $3.29 per GPU-hour2 |
| 3 | Nebius | HGX H100 | 3.85 | $3.85 per GPU-hour; $4.50 from 1 October 20269 |
| 4 | Crusoe | HGX H100 80 GB | 3.90 | $3.90/GPU-hr7 |
| 5 | Modal | H100 SXM5 | 3.95 | $0.001097 per second4 |
| 6 | Together AI | HGX H100, GPU Clusters | 3.99 | $3.99 per GPU per hour8 |
| 7 | DigitalOcean | HGX H100 | 4.41 | $4.41/GPU/hour5 |
| 8 | CoreWeave | HGX H100, 8-GPU node, North America | 6.16 | $49.24 per node-hour3 |
Five of the eight fall between $3.85 and $4.41. The spread at the edges is wider: $2.89 at the bottom and $6.16 at the top.
RunPod’s pricing page offers Community Cloud and Secure Cloud views. RunPod’s own guide gives $2.89 as a Secure Cloud starting rate.11 Lambda and Nebius state their prices exclude tax, and DigitalOcean says it applies tax in some countries.2,9,5 The other five pages don’t state tax status in the captured text.
Which GPUs does each provider list?
Every provider lists more than the H100, and all eight also list at least one Blackwell B200 or B300 system. The lists below are what each pricing page names, including models priced only through sales.
| Provider | H100 variants listed | Other GPUs on the pricing page |
|---|---|---|
| RunPod | PCIe 80 GB, SXM 80 GB, NVL 94 GB | B300, B200, H200, A100 PCIe and SXM, RTX Pro 6000, L40S, L40, RTX 6000 Ada, A40, RTX A6000, RTX A5000, RTX 5090, RTX 4090, RTX 3090, L41 |
| Lambda | SXM 80 GB, PCIe 80 GB | B200 SXM6, GH200, A100 SXM and PCIe, A10, A6000, Tesla V100, Quadro RTX 60002 |
| Nebius | HGX H100 | GB300 NVL72, GB200 NVL72, HGX B300, HGX B200, HGX H200, RTX PRO 6000, L40S9 |
| Crusoe | HGX H100 80 GB | GB200 NVL72, B200, H200, A100 SXM and PCIe 80 GB, L40S, AMD MI355X, AMD MI300X7 |
| Modal | H100 SXM5 | B300, B200, H200 SXM, RTX PRO 6000, A100 80 GB and 40 GB, L40S, A10, L4, T44 |
| Together AI | HGX H100 | GB200 NVL72, GB300 NVL72, HGX B200, HGX B300, HGX H2008 |
| DigitalOcean | HGX H100 | HGX B300, HGX H200, RTX 4000 Ada, RTX 6000 Ada, L40S, AMD MI355X, MI350X, MI325X, MI300X5 |
| CoreWeave | HGX H100 | GB300 NVL72, GB200 NVL72, HGX B300, HGX B200, HGX H200, GH200, RTX PRO 6000 Blackwell Server Edition, L40S, L40, A1003 |
The breadth splits into two groups. RunPod, Lambda and Modal run long lists that reach down to cards such as the RTX 4090, Tesla V100 and T4.1,2,4 Together AI’s GPU Clusters list only H100, H200 and Blackwell systems.8 Nebius adds RTX PRO 6000 and L40S options, and CoreWeave adds those plus A100, L40 and GH200 nodes.9,3 Crusoe and DigitalOcean are the two that list AMD Instinct accelerators beside NVIDIA’s.7,5
A listed GPU isn’t always a priced GPU. Crusoe lists its GB200, B200 and MI355X under “Contact sales”, and CoreWeave does the same for its GB300 NVL72.7,3
How does each provider bill for GPU time?
Six of the eight state a price per GPU-hour, Modal states a price per second and CoreWeave states a price per 8-GPU node-hour. The metering behind the headline differs again.
| Provider | Price unit on the page | Metering stated | Smallest H100 unit |
|---|---|---|---|
| RunPod | per GPU-hour, with a per-second view | per second11 | not stated |
| Lambda | per GPU-hour | by the minute6 | 1 GPU (PCIe and SXM) |
| Nebius | per GPU-hour | not stated in the capture | not stated |
| Crusoe | per GPU-hour | by the hour7 | not stated |
| Modal | per GPU-second | per second; CPU and memory billed separately4 | not stated |
| Together AI | per GPU-hour | hourly8 | cluster, size not stated |
| DigitalOcean | per GPU-hour | per second, 5-minute minimum5 | 1 or 8 GPUs |
| CoreWeave | per 8-GPU node-hour | not stated in the capture | 8-GPU node3 |
The unit matters most for short jobs. Per-second and per-minute metering charge a 20-minute test for about 20 minutes, and DigitalOcean’s 5-minute minimum only bites on runs shorter than that.5 Crusoe and Together AI describe their GPU pricing in hourly terms, and the captured pages don’t state a shorter increment.7,8
Modal prices the GPU on its own. It charges CPU at $0.0000131 per physical core per second and memory at $0.00000222 per GiB per second on top.4 RunPod, Lambda, Nebius, DigitalOcean and CoreWeave list vCPUs and memory beside each GPU price. DigitalOcean adds one condition for idle machines: a powered-off GPU Droplet is still billed until it is destroyed.5
For how a GPU-hour rate turns into the cost of serving a model, see our explainer on what controls AI inference cost.
How much less do reserved and spot H100s cost?
Commitment and interruptible capacity both cut the on-demand rate. Only Together AI and DigitalOcean publish an exact reserved H100 price; the rest give a maximum discount, a cluster rate or a sales contact.
| Provider | On-demand H100 (USD per GPU-hour) | Reserved or committed | Spot or preemptible |
|---|---|---|---|
| RunPod | 2.89 | H100 SXM reserved clusters, 1 to 12+ months: contact sales1 | not listed for H100 |
| Lambda | 3.29 | reserved capacity: contact sales; H100 1-Click Clusters $6.16 to $5.54 for 2 weeks to 1 year2 | not listed |
| Nebius | 3.85 | up to 35% less for multi-month cluster reservations9 | from $0.799 |
| Crusoe | 3.90 | contact sales for “our lowest rates”7 | contact sales7 |
| Modal | 3.95 | committed spend through the AWS and GCP marketplaces; Enterprise volume discounts4 | not listed |
| Together AI | 3.99 | $3.69 (7 to 30 days), $3.45 (31 to 90 days), $3.19 (91 to 180 days)8 | $1.998 |
| DigitalOcean | 4.41 | $3.26 on a 12-month contract5 | no H100 spot plan listed |
| CoreWeave | 6.16 | up to 60% off on-demand for committed usage3 | $19.71 per node-hour, $2.46 per GPU-hour3 |
Three published figures can be read directly. Together AI’s 91-to-180-day rate is 20% below its on-demand rate. DigitalOcean’s 12-month rate is 26% below its on-demand rate. CoreWeave’s North American spot node rate is 60% below its on-demand node rate.8,5,3
Lambda’s cluster table runs the other way. Its 16-GPU H100 cluster lists at $6.16 per GPU-hour for 2 weeks to 1 year, above every one of its self-serve H100 instance rates.2 That rate buys a production cluster for a fixed term; Lambda describes these clusters as running from 16 to 2,000+ GPUs.2
Spot prices move. Nebius says its spot prices “may vary during the active use, including as frequently as every 15 minutes”.9 DigitalOcean says spot Droplets “can be reclaimed at any time”.5
Why do H100 prices differ so much between providers?
The cheapest two listings are for the PCIe card, and the rest are SXM, NVL or HGX listings with different CPU, memory and storage attached. Compare like for like before comparing prices.
Set the two PCIe listings aside and RunPod still leads, with its H100 SXM at $3.49.1 Nebius follows at $3.85, Crusoe at $3.90 and Modal at $3.95.9,7,4 Together AI and Lambda’s 8-GPU SXM instance both list $3.99.8,2 RunPod also lists a 94 GB H100 NVL at $3.19.1
The bundle varies with the price:
- RunPod’s H100 PCIe comes with 16 vCPUs and 188 GB of RAM.1
- Lambda’s H100 PCIe comes with 26 vCPUs, 225 GiB of RAM and a 1 TiB SSD.2
- Nebius’s HGX H100 comes with 16 vCPUs and 200 GB of RAM.9
- DigitalOcean’s H100 Droplet comes with 20 vCPUs, 240 GiB of memory, a 720 GiB boot disk and a 5 TiB scratch disk.5
Instance size changes the rate at Lambda. Its H100 SXM falls from $4.29 per GPU-hour on the single-GPU instance to $3.99 on the 8-GPU instance.2
Modal’s base rate carries conditions. Its page lists region selection at 1.15 to 1.75 times base prices and non-preemptible execution at 3 times base prices.4 On the H100, three times the base rate is $11.85 per GPU-hour.
The decision changes when the workload can’t be interrupted or needs a fixed region. The ranked price holds for the default terms each page shows. Confirm this in the current quote.
What changes the ranking after 28 September 2026?
Nebius’s listed price change on 1 October 2026 moves it from third to seventh. Its HGX H100 goes from $3.85 to $4.50 per GPU-hour, which puts it above DigitalOcean’s $4.41 and below CoreWeave’s $6.16.9
Three other dated items bear on the table:
- RunPod’s pricing page is marked “Updated September 27, 2026”, one day before capture.1
- DigitalOcean’s current prices took effect on 1 August 2026.5
- Together AI’s Dedicated Inference H100 carries a $3.99 promotional rate, down from $5.49, “valid until 09/30/26”. Its GPU Clusters H100 rate of $3.99, the one ranked here, carries no promotion label.8
The table is refreshed when any of the nine pricing pages changes.
What the ranking cannot tell you
List prices say nothing about stock. This ranking doesn’t record whether an H100 was available to launch at the time of capture.
The ranking doesn’t measure performance per dollar. It prices the rental hour, not the work done in it, and the PCIe, SXM and HGX listings aren’t the same machine.
It covers eight GPU clouds, not AWS, Microsoft Azure or Google Cloud. Google Cloud lists the H100 among its Compute Engine GPUs and bills per second, but its H100 rate wasn’t captured for this ranking.12 AWS and Azure weren’t checked at all.
Marketplace prices sit outside it. Vast.ai showed a median of $2.16 per hour for H100 SXM listings, with offers from $1.73.10 Those are prices set by individual hosts, so the median moves with supply rather than a published rate card.
Frequently asked questions
Which cloud providers offer H100 GPUs?
All eight providers in this ranking list H100s: RunPod, Lambda, Nebius, Crusoe, Modal, Together AI, DigitalOcean and CoreWeave. Google Cloud’s Compute Engine also lists the H100.12 Vast.ai offers H100 SXM, H100 PCIe and H100 NVL through host-set marketplace listings.10
Which cloud GPU provider is best for H100s?
On listed on-demand price, RunPod is cheapest at $2.89 per GPU-hour for H100 PCIe and $3.49 for H100 SXM.1 For a published multi-month rate without a sales call, Together AI lists $3.19 per GPU-hour for 91 to 180 days.8 For whole 8-GPU nodes, CoreWeave prices by the node.3 The right choice depends on the variant, term and billing unit the workload needs.
Can I get a cloud GPU for free?
Modal’s Starter plan includes $30 a month of free compute, and its Team plan $100 a month.4 Modal also offers academics up to $10k in compute credits.4 Nebius advertises an AI Builder Program with “$400+ in credits and discounts”.9 None of the eight pages lists a free on-demand H100.
Do cloud GPU prices include tax?
Not on the pages that say. Lambda lists its prices “plus applicable sales tax/VAT/GST”.2 Nebius shows prices “without any applicable taxes, including VAT”.9 DigitalOcean says it applies taxes in some countries as required by law.5 All eight quote in US dollars.
New GPU cloud rankings go out in our free research email. Subscribe to our research.
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.
