H100 Rental Prices Compared: What One H100 Hour Actually Costs Across Providers
Pricing guide · Updated November 11, 2026 · Every rate below was read from the provider's official pricing page on November 11, 2026
How to read these prices. Each figure comes from the provider's own public pricing page, checked on November 11, 2026 — not from third-party roundups. GPU cloud prices change often and vary by region and availability, so treat this as a dated snapshot and always confirm the live rate before budgeting or committing.
On this page
The short answer
On November 11, 2026, a single H100 hour costs anywhere from $1.99 (RunPod Community Cloud, PCIe, marketplace-style capacity) to $6.16 (CoreWeave, priced as part of a full 8-GPU HGX node) — a spread of more than 3x for what is nominally the same chip. Self-serve on-demand SXM rates cluster tightly: RunPod Secure Cloud at $3.99/hr, Nebius at $4.50/hr, and Lambda at $4.59/hr. Where you land in that range depends less on the GPU than on how the provider sells it: single pod vs. full node, fixed list price vs. marketplace, on-demand vs. interruptible.
The H100 price table
All rates are per GPU per hour, US dollars, as published on each provider's pricing page on November 11, 2026. CoreWeave sells H100s only as 8-GPU nodes, so its node price and the per-GPU equivalent are both shown.
| Provider | H100 offering | On-demand $/GPU/hr | Spot / interruptible | How it is sold |
|---|---|---|---|---|
| RunPod | H100 PCIe 80GB (Secure Cloud) | $2.89 | Community Cloud from $1.99 | Self-serve pods, billed per second |
| RunPod | H100 NVL 94GB | $3.19 | — | Self-serve pods, billed per second |
| RunPod | H100 SXM 80GB (Secure Cloud) | $3.99 | Community Cloud from $2.69 | Self-serve pods, billed per second |
| Nebius | HGX H100 80GB | $4.50 | Preemptible from $0.79 (dynamic spot pricing) | Per-GPU instances, self-serve console |
| Lambda | HGX H100 80GB | $4.59* | — (reserved capacity via sales) | Self-serve instances, per-GPU rate |
| CoreWeave | HGX H100 80GB (8-GPU node) | $6.16 (node: $49.24/hr) | Spot node $19.71/hr ($2.46/GPU) | Full nodes and clusters, North America list |
* Lambda's pricing page lists $4.59/GPU/hr with a note that the posted rates take effect November 16, 2026; until then, current rates are shown after login. Sources: runpod.io/pricing, lambda.ai/pricing, nebius.com/prices, coreweave.com/pricing — all accessed November 11, 2026.
Why the same GPU costs different amounts
The variant matters. "H100" covers three products in the table above. The PCIe card is the cheapest way in and is fine for single-GPU work. The SXM card carries NVLink, which only pays off when several GPUs train together and need to exchange data fast. The NVL variant trades a little interconnect for more memory (94GB vs 80GB) — useful for serving large models. Comparing a PCIe price against an SXM price without noticing is the most common way these tables mislead.
The sales model matters more. RunPod's two tiers show the split clearly: Secure Cloud is a fixed list price on managed capacity, while Community Cloud is marketplace-style capacity that starts lower ($1.99/hr for PCIe) but comes and goes. CoreWeave sits at the other end — it does not sell single GPUs at all. Its $49.24/hr buys a complete 8-GPU HGX node with high-bandwidth networking, so the $6.16/GPU figure is a unit conversion, not a product you can buy one of.
Interruptible capacity is a different market. Nebius lists preemptible H100 capacity from $0.79/GPU/hr, and CoreWeave's spot node works out to $2.46/GPU/hr — both far below on-demand, both able to disappear when the cloud needs the capacity back. For checkpointable batch work that trade is often worth it; our spot instance guide covers how to budget for interruptions.
What a full month of H100 costs
Hourly rates hide the real number. Run one H100 continuously for a 730-hour month at on-demand rates and the table above becomes:
| Option | $/GPU/hr | One GPU, 24/7 for a month (730 hrs) |
|---|---|---|
| RunPod H100 PCIe (Community Cloud, from) | $1.99 | ~$1,453 |
| RunPod H100 PCIe (Secure Cloud) | $2.89 | ~$2,110 |
| RunPod H100 SXM (Secure Cloud) | $3.99 | ~$2,913 |
| Nebius HGX H100 | $4.50 | ~$3,285 |
| Lambda HGX H100 | $4.59 | ~$3,351 |
| CoreWeave HGX H100 (per-GPU equivalent) | $6.16 | ~$4,497 (node: ~$35,945) |
The lesson is not "pick the top row." It is that at 24/7 utilization, the gap between the cheapest and most expensive on-demand option is roughly $3,000 per GPU per month — and that nobody should run steady workloads on pure on-demand rates without first pricing a reservation. Committed-use and reserved pricing exists precisely for this case; see how GPU cloud pricing works for the on-demand vs. spot vs. reserved mechanics, and run your own hours through the GPU Cost Calculator.
Which provider fits which job
- Cheapest single GPU for experiments and fine-tuning: RunPod. The PCIe card at $2.89/hr on Secure Cloud (or from $1.99 on Community Cloud when capacity is there) with per-second billing is the lowest-friction way to rent an H100 today.
- Multi-GPU training on a budget: RunPod SXM at $3.99/hr or Nebius at $4.50/hr. Both are self-serve with NVLink-class HGX hardware; Nebius adds a genuinely cheap preemptible tier for the parts of the job that can be interrupted.
- Predictable self-serve HGX capacity: Lambda at $4.59/GPU/hr. The rate is published, the product is a straightforward per-GPU instance, and reserved clusters are available when you outgrow on-demand. (For a deeper two-way comparison, see CoreWeave vs Lambda Labs.)
- Full-node, production-scale training: CoreWeave. At $49.24/hr for an 8-GPU HGX H100 node you are buying the complete machine — networking, storage, and support included — which is the right comparison unit for serious multi-node runs, not the per-GPU conversion.
- Fault-tolerant batch work: spot and preemptible tiers. Nebius preemptible from $0.79/GPU/hr and CoreWeave spot nodes at $19.71/hr reward jobs that checkpoint well — and punish jobs that do not.
The fine print worth knowing
- Lambda's rate is dated. The $4.59/GPU/hr on Lambda's pricing page is posted as taking effect November 16, 2026; the page directs logged-in users to current pricing until then. We list the posted rate because it is the only public figure, and flag the date so you can verify.
- "From" prices move. RunPod Community Cloud and Nebius preemptible rates are floors, not fixtures — both vary with available capacity. Nebius states explicitly that its spot prices are dynamic and can change as often as every 15 minutes.
- Clusters are a fourth RunPod price. RunPod's Instant Clusters list H100 SXM at $4.79/GPU/hr — above its own $3.99 pod rate, because managed multi-node networking is bundled in. Compare cluster prices against cluster prices.
- Storage and egress sit outside the GPU hour. RunPod, for example, charges separately for container and network storage. Our guide to reading a GPU cloud invoice walks through every line beyond GPU-hours.
- Not every cloud publishes an H100 price. The hyperscalers sell H100 capacity inside bundled instance families with region-dependent pricing, which is why this comparison sticks to providers with a public per-GPU rate. Our provider directory tracks where to check each cloud's current numbers.
Run the numbers for your workload: multiply these rates by your real GPU-hours in our GPU Cost Calculator — and always verify the live rate on the provider's pricing page before you commit.
Frequently asked questions
What is the cheapest way to rent an H100?
As of November 11, 2026, the lowest published H100 rate is RunPod's Community Cloud, where an H100 PCIe starts at $1.99/hr — availability varies because it is marketplace-style capacity. The cheapest fixed list rate is RunPod Secure Cloud's H100 PCIe at $2.89/hr, billed per second.
How much does it cost to rent an H100 for a month?
Running one H100 around the clock (730 hours) at on-demand rates costs roughly $1,453/month at RunPod Community Cloud PCIe rates, $2,913/month at RunPod Secure Cloud SXM rates, about $3,285/month on Nebius, about $3,351/month on Lambda, and about $4,497/month at CoreWeave's per-GPU on-demand rate. For month-long steady work, reserved or committed rates usually beat all of these.
Why is CoreWeave's H100 price higher per GPU?
CoreWeave sells H100s as full 8-GPU HGX nodes ($49.24/hr on demand), not single GPUs. The $6.16/GPU/hr figure is that node price divided by eight, and it buys a complete node with high-bandwidth networking and enterprise support — a different product from a self-serve single-GPU pod.
Does it matter whether an H100 is SXM, PCIe, or NVL?
For a single GPU, mostly no: PCIe cards are cheaper and fine for fine-tuning, inference, and prototyping. SXM matters when several GPUs train together, because NVLink gives them a much faster interconnect. The NVL variant adds memory (94GB vs 80GB), which helps large-model inference.
How current are these H100 prices?
Every rate in this guide was read from the provider's official pricing page on November 11, 2026. GPU cloud prices change frequently — Lambda's listed rate, for example, is posted as effective November 16, 2026 — so always confirm the live rate on the provider's pricing page before budgeting.