This website uses cookies

Read our Privacy policy and Terms of use for more information.

Last Updated: November 21, 2026

An Nvidia H100 costs about $2 to $11 per GPU-hour to rent on demand in November 2026, depending on the provider, and roughly $25,000 to $40,000 to buy new. Used H100 SXM cards have been selling for about $14,250, according to one appraisal firm, while the market rental benchmark sat at $3.11 an hour on November 19.

The H100 launched in October 2022 and is no longer Nvidia's newest chip, yet its rental price went up this fall, not down. Here is what it costs at the big clouds and the GPU specialists, checked against their own pricing pages, plus the monthly math and a simple rent-or-buy test.

Free download: The AI Stats Sheet 2026. 36 key AI numbers for business on four printable pages, with the source under every figure. Join 2,000+ business leaders who read AI Business Weekly and get it in your inbox right away.

In this guide

H100 Price at a Glance

The H100's price depends on whether you rent by the hour, rent a full eight-GPU server, or buy the card. These are the numbers that matter in November 2026.

Way to pay

Typical H100 price

Source

Market rental benchmark

$3.11 per GPU-hour (Nov 19, 2026)

Silicon Data H100 Rental Index

Cheapest listed rental

From $1.99 per GPU-hour (H100 PCIe, marketplace)

RunPod

GPU cloud, on-demand

$4.39 to $6.16 per GPU-hour (H100 SXM or HGX)

RunPod, Nebius, CoreWeave

Hyperscaler

$5.97 (AWS Capacity Blocks) to $11.06 (Google Cloud on-demand) per GPU-hour

AWS, Google Cloud

Buy new

About $25,000 to $40,000 per card

Silicon Data

Buy used

About $14,250 per H100 SXM (typical sale)

American Compute

Per-GPU figures for eight-GPU servers are our own math: the server's hourly price divided by eight.

How Much Does an H100 Cost per Hour?

Renting one H100 costs between about $2 and $11 per hour in November 2026. GPU marketplaces and specialist clouds sit at the low end, and Google Cloud's on-demand price sits at the top.

Provider

H100 option

Price per GPU-hour

Per GPU, 730 hours (1 month)

RunPod Community Cloud

H100 PCIe, 80 GB

From $1.99

$1,453

RunPod Community Cloud

H100 SXM, 80 GB

From $2.69

$1,964

RunPod Secure Cloud

H100 NVL, 94 GB

$3.19

$2,329

RunPod Secure Cloud

H100 SXM, 80 GB

$4.39

$3,205

Nebius

HGX H100, on-demand

$4.50

$3,285

AWS Capacity Blocks

p5.48xlarge (8x H100)

$5.97

$4,358

CoreWeave

HGX H100 (8 GPUs), on-demand

$6.16

$4,493

Google Cloud

A3 High (8x H100), on-demand

$11.06

$8,075

Sources: RunPod pricing, Nebius prices, AWS Capacity Blocks pricing, CoreWeave pricing and Google Cloud accelerator pricing, all checked November 20, 2026. RunPod's Community Cloud rates are "from" prices on a marketplace, so the rate you get can be higher.

The spread is wide because you are not buying the same thing. A single marketplace card is fine for testing a model. An eight-GPU server with fast networking, support and enterprise security is what a company needs for training or steady production work, and that costs more.

Discounts can cut the H100 price by more than half

Spot and committed prices change the picture. CoreWeave lists an eight-GPU H100 server at $49.24 an hour on demand and $19.71 on spot in North America, a 60% gap. Google Cloud's three-year committed price for A3 High is $38.86 an hour against $88.49 on demand, a 56% cut. Nebius lists H100 spot capacity "from $0.79" an hour, but says spot prices move with demand and capacity.

The catch is commitment. A three-year contract on a chip that has been on the market since 2022 is a real bet. Spot capacity can be taken back. Most businesses should price both, then decide how much interruption they can live with.

What Does an 8-GPU H100 Server Cost to Rent per Month?

Renting a full eight-GPU H100 server costs about $28,000 to $65,000 a month at the big providers in November 2026, running around the clock. Here is the math at 730 hours a month.

Provider and plan

Hourly (8 GPUs)

Monthly (730 hours)

Yearly (8,760 hours)

Google Cloud A3 High, 3-year commitment

$38.86

$28,371

$340,452

AWS p5.48xlarge, Capacity Block

$47.76

$34,863

$418,351

CoreWeave HGX H100, on-demand

$49.24

$35,945

$431,342

Google Cloud A3 High, 1-year commitment

$61.38

$44,810

$537,721

Google Cloud A3 High, on-demand

$88.49

$64,598

$775,172

Few teams run a server at 100% for a full month. If you only need it 40 hours a week, the CoreWeave on-demand bill drops to about $8,500 a month (173 hours). That is why hourly flexibility often beats a cheaper rate you cannot fully use.

Are H100 Rental Prices Going Up?

Yes, recently. The Silicon Data H100 Rental Index, which tracks neo-cloud providers and GPU marketplaces daily, moved from $2.86 per GPU-hour on October 21 to $3.11 on November 19, 2026, an 8.7% rise in a month, according to Silicon Data. For a chip that has been on the market since 2022, that is a sign demand for AI computing is still strong.

Rows of blade servers lit in blue inside a data center rack

Rental rates for the same servers moved up this fall. Photo: Getty Images / Unsplash+

The hyperscalers moved too, in both directions. In June 2025, AWS cut prices on its H100-based P5 instances by up to 45%. A year later the direction flipped: AWS raised Capacity Block prices for its Nvidia GPU instances, P5 included, by about 20% from July 1, 2026, according to AI Weekly, citing The Information. AWS says its current Capacity Block prices are scheduled to be updated next in January 2027.

The practical reality: do not assume older GPUs get cheaper every quarter. As long as companies are short of compute, an H100 is still useful, and providers price it that way.

How Much Does It Cost to Buy an H100?

A new H100 card has typically sold for $25,000 to $40,000, a range that held roughly flat from mid-2024 into early 2026, according to Silicon Data's resale analysis.

Close-up of a processor chip mounted on a dark circuit board

Buying the hardware outright trades a big upfront cost for lower hourly math. Photo: Getty Images / Unsplash+

Used cards are much cheaper. Silicon Data says used and refurbished H100s traded as high as $50,000 in mid-2024, then its resale chart shows prices sliding from the low $40,000s to the mid-$20,000s by late 2025. American Compute, which appraises GPU hardware, puts the typical realized sale for a used H100 SXM 80 GB at $14,250 and a used eight-GPU HGX H100 server at $195,000 (about $24,400 per GPU), based on its October 2026 report with data through August 2026.

The card is only part of the bill. A GPU needs a server, networking, power, cooling and a place to run it. The SXM version ships in HGX partner and Nvidia-Certified systems with four or eight GPUs, or in Nvidia's own eight-GPU DGX H100, according to its H100 product page. The PCIe-based H100 NVL fits standard servers with one to eight GPUs and comes with a five-year Nvidia AI Enterprise software subscription.

Rent vs buy: a quick break-even

Here is a simple test, using the $3.11 market rental rate. It ignores the server, power and staff, so treat it as a floor, not a full business case.

Purchase price

Rental hours to match it

Days at 24/7

$14,250 (used SXM)

4,582

191

$25,000 (new, low end)

8,039

335

$40,000 (new, high end)

12,862

536

If you would keep an H100 busy around the clock for more than a year or two, owning can win on paper. If your usage is spiky, or you lack a data center team, renting is simpler and the risk sits with the provider. American Compute's own residual estimates point to sharp value loss ahead, with conservative bands of 11% to 19% of original list price in 2027.

H100 vs A100 Price: Is the H100 Worth Double?

An H100 rents for about twice the price of an A100. RunPod's Secure Cloud lists the H100 SXM at $4.39 an hour against $2.19 for the A100 80 GB SXM, and CoreWeave charges $49.24 for an eight-GPU H100 server against $21.60 for an eight-GPU A100 server.

Nvidia says the H100 delivers up to 4x faster training than the prior generation on GPT-3 (175B) and up to 5x higher Llama 2 70B inference in H100 NVL servers versus A100 systems. Those are vendor claims, and your gains depend on the model and software. But if a job finishes in a third of the time at twice the hourly rate, the H100 is cheaper per finished job. For small models and light inference, the A100 can still be the better deal.

Spec

H100 SXM

H100 NVL

Memory

80 GB

94 GB

Memory bandwidth

3.35 TB/s

3.9 TB/s

FP8 Tensor Core (with sparsity)

3,958 teraFLOPS

3,341 teraFLOPS

Max power

Up to 700 W

350 to 400 W

Form factor

SXM (HGX or DGX systems)

PCIe, dual-slot, air-cooled

Specs from Nvidia's H100 page.

Is the H100 Still Worth It in 2026?

Yes, for many business workloads. The H100 is no longer Nvidia's top chip, but every provider we checked still offers it, and it costs less than newer parts. CoreWeave's eight-GPU H200 server costs $50.44 an hour and its B200 server $68.80, against $49.24 for the H100. Nebius lists the H200 at $5.40 and the B200 at $8.50 per GPU-hour, against $4.50 for the H100.

Pick the H100 when your model fits in 80 GB per GPU, you want the lowest price per hour for proven hardware, or you are fine-tuning and serving mid-sized models. Step up to the H200 for more memory (141 GB) for 2% to 20% more per hour at CoreWeave and Nebius, or to Blackwell when speed on the largest models matters more than the hourly rate. Our H200 price guide and B200 price guide run the same numbers for those chips.

Sameer's take: Sameer Khan thinks Nvidia's lead in AI chips is safe for years, thanks to CUDA and its scale.

How Should a Business Choose Where to Rent H100s?

Start with the workload, not the lowest sticker price. Here is a simple way to match the provider to the job:

  1. Testing or a short project: a marketplace or per-second billing (RunPod-style) keeps the bill small. Expect less support.

  2. Steady production work: a GPU cloud such as CoreWeave or Nebius, priced on demand first, then with a commitment once usage is predictable.

  3. Already deep in AWS or Google Cloud: the higher hourly rate can be worth it for security reviews, existing contracts and data that already lives there. Ask about Capacity Blocks or committed-use discounts.

  4. Check the whole bill: storage, data transfer and idle time can matter as much as the GPU rate.

For the wider picture on chip costs and competitors, see Nvidia's competitors and our TPU vs GPU guide.

The Bottom Line

An H100 costs about $3 an hour at the market rate, $4.39 to $6.16 at mainstream GPU clouds, and up to $11 at Google Cloud on demand. Buying one runs $25,000 to $40,000 new, or about $14,000 used, before the server around it.

For most businesses, renting is the right default: price it with spot and committed options, and only buy if you can keep the cards busy all year. Prices moved up this fall, and AWS plans its next Capacity Block update in January 2027, so check the provider pages again before you sign anything.

Related Articles

⭐ Get more AI Business Weekly in Google: add us as a preferred source on Google.

Written by Sameer Khan, with support from the AI Business Weekly Editorial Team. AI-assisted and fact-checked against the sources linked above.