The NVIDIA H100 is still the default card for serious training and fast inference in 2026. Newer parts like the H200, B200 and B300 have arrived, but most teams that need real compute start with the same question: what does an H100 cost per hour, and is it cheaper to rent or to buy?
As of 28 October 2026 the H100 remains an active NVIDIA data-center GPU, still used across commercial and research workloads and still widely rentable. NVIDIA cites SemiAnalysis InferenceX benchmarks (as of April 2026) showing the H100 serving GPT-OSS-120B inference at about $0.09 per million tokens at 66 tokens per second per user, which helps explain why the card stays in demand even as newer silicon ships. Public rental prices, though, vary sharply by provider, region, form factor and whether the capacity is guaranteed, so a single “H100 price” can be misleading.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
The hourly price range
H100 prices now vary more by provider type than by the card itself. Across public listings in October 2026, on-demand rates run from roughly $2 to $12 per GPU-hour depending on the provider, region, form factor and instance size.
- Hyperscalers (the big three clouds) charge the most per GPU hour, often sell H100s only in eight-GPU instances, and give the best rates to customers who sign multi-year reservations. AWS, for example, offers H100 capacity mainly through its eight-GPU p5 family, and Google Cloud’s on-demand H100 is an eight-GPU SXM VM at $88.49 per hour, so its per-GPU rate assumes you use all eight.
- Specialist GPU clouds charge noticeably less on demand and usually let you rent a single GPU. Committed options exist too, such as Lambda’s clusters with terms ranging from two weeks to one year.
- Marketplaces, where many independent data centers list their hardware, are usually the cheapest, at roughly $2.40 per GPU-hour or less on some unreserved listings. Prices move with supply, so the cheapest card today may not be the cheapest one later.
At the low end of the market, an H100 80GB can be rented from about $1.60 per GPU hour on demand (Lium, as of 28 October 2026). The table below shows how on-demand rates compared across provider types in October 2026. The figures are not perfectly like-for-like: per-GPU prices drawn from eight-GPU instances assume every GPU is in use, and Azure’s single-GPU figure is for an H100 NVL card with 94 GB rather than a standard 80 GB H100.
| Provider and instance | Provider type | GPUs in instance | Price per GPU-hour | As reported |
|---|---|---|---|---|
| RunPod H100 SXM (Community Cloud) | Specialist marketplace | 1 | $2.69 | 26 Oct 2026 |
| Lambda H100 PCIe | Specialist cloud | 1 | $3.29 | 19 Oct 2026 |
| Lambda H100 SXM | Specialist cloud | 8 | $3.99 | 19 Oct 2026 |
| Lambda H100 SXM | Specialist cloud | 1 | $4.29 | 19 Oct 2026 |
| Azure NC40ads H100 v5 (NVL, 94 GB) | Hyperscaler | 1 | $6.98 | 19 Oct 2026 |
| Google Cloud a3-highgpu-1g | Hyperscaler | 1 | about $11.06 | Oct 2026 |
| Azure ND96isr H100 v5 | Hyperscaler | 8 | $12.29 (per-GPU equivalent) | 19 Oct 2026 |
Reports in late October 2026 described strong demand for rented NVIDIA compute and rising prices for older H100 capacity, so always check the live price on the day you rent. Rates still change week to week.
What the hourly price leaves out
The sticker price is only part of the bill. Before comparing, check:
- Billing granularity. Per-hour billing rounds a 10-minute test up to a full hour, while per-second or per-minute billing charges only for what you use. Terms differ by provider: AWS Capacity Blocks bill Linux, Ubuntu Pro and RHEL per second (SLES carries a one-hour minimum), and Lambda is reported to bill on-demand instances by the minute.
- Minimum commitment. Many cheap rates require a monthly or yearly term. If your workload is bursty, a lower reserved rate can cost more than on demand.
- Storage and egress. Moving large datasets and checkpoints in and out can add a real share to the bill, and a fixed monthly server can remove the egress meter entirely.
- Idle time. A GPU left running overnight costs the same as one doing useful work. Shutting it down when the job ends is the easiest saving there is.
- Interconnect and form factor. The H100 is listed as PCIe, SXM and NVL cards. SXM parts carry 80 GB of HBM3 and fast GPU-to-GPU NVLink for multi-GPU training, while the NVL listing is a 94 GB PCIe card. A cheaper PCIe card is fine for inference but can be slow for large training runs.
Interruptible or spot capacity is cheaper still, with some marketplace H100 PCIe offers reported around $0.35 per GPU-hour, but those instances can be reclaimed mid-job, so they suit fault-tolerant work rather than a long training run you cannot restart.
Where the H100 sits next to newer cards
The H100 is no longer NVIDIA’s newest data-center GPU, and part of its appeal now is price relative to the latest parts. The H200 keeps the same generation but widens memory to 141 GB of HBM3e. The Blackwell generation goes further: the B200 carries 192 GB of memory, and the newer B300 is rolling out in cloud deployments, with Google bringing Blackwell-based systems, including B300, into service during 2026. Those cards cost more per hour where they are listed; one specialist cloud’s B200 was reported at about $6.69 per GPU-hour, above its single-GPU H100 rates.
For many workloads the H100 remains the value pick: it is plentiful on specialist clouds and marketplaces, well supported by common frameworks, and cheap enough on demand that renting one to run a job is often the most economical option. Renting also lets you move up to an H200 or a Blackwell card later, without being locked into either generation.
When renting beats buying
A new H100 server is a large up-front cost before power, cooling, networking and staff time. One July 2026 estimate put H100 hardware at roughly $25,000 to $40,000 per card depending on the variant and reseller (that is a market estimate, not an official NVIDIA price). Renting shifts that to an hourly or monthly meter instead: one 2026 provider comparison put the cost of running a single H100 around the clock at about $2,621 to $7,738 a month across the clouds it tracked.
Buying makes sense when the hardware will be busy almost all the time for years. Renting wins when:
- usage is uneven: experiments, fine-tuning runs, or traffic spikes;
- you want to try newer cards (H200, B200) without being locked into the last generation;
- you need capacity this week, not after a hardware lead time;
- your team does not want to run a data center, with its power, networking, maintenance and financing costs.
A simple rule: work out how many hours a month the GPU would actually be busy. If it is well below full time, renting on demand is almost always cheaper once power and depreciation are counted. As a practical target in October 2026, a flexible single-GPU H100 runs roughly $2 to $4 per GPU-hour on specialist clouds and marketplaces, against about $7 to $12 per GPU-hour at the major hyperscalers. Marketplace prices can come with weaker availability guarantees, while hyperscaler prices generally buy more standardized infrastructure and service integration.
How to rent an H100 without overpaying
- Start on demand, measure real usage, and only commit once the pattern is stable.
- Rent the smallest setup that fits: one GPU for inference and most fine-tuning, eight GPUs with NVLink only for large training.
- Use per-second billing where available and stop instances as soon as a job finishes.
- Compare across provider types, not just within one cloud, because the spread from marketplace to hyperscaler is several times the price.
Per-second billing matters most for short jobs and frequent restarts, when per-hour rounding quietly inflates the bill. On Lium you can rent an H100 by the hour with per-second billing and no long-term commitment, starting from a single GPU, so a ten-minute test is charged as ten minutes rather than a full hour. Pairing that with the habit of shutting instances down the moment a job ends removes the two most common sources of waste.
Because supply and demand move the market week to week, the single most useful habit is to read today’s rate before you start rather than trusting a number from last month. Decide the form factor you need, size the instance to the job, and compare the on-demand price across at least one hyperscaler, one specialist cloud and one marketplace before you commit.
Prices quoted are on-demand “from” prices as read on 28 October 2026 and change with supply.
