H100, H200 and B200 rental prices per GPU-hour from 13 providers, checked 7 November 2026

On 7 November 2026, an on-demand H100 SXM cost $2.69 to $6.88 per GPU-hour on the providers' own pages (RunPod Community Cloud to AWS us-east-1), an H200 $3.59 to $7.91, and a B200 $5.98 to $14.24. The middle of the market, Together AI, Nebius and Lambda, charged about $4 to $4.90 for an H100 and $7.70 to $8.50 for a B200. Prices are rising: Silicon Data's H100 neo-cloud index read $3.04, against $2.00 on 9 December 2025.

Per GPU-hour, USD, on demand. AWS figures are 8-GPU instance prices divided by eight. Details and sources in the next table.

How much does an H100, H200 or B200 cost per hour?

The spread for one GPU model is about 2.5x between the cheapest neocloud and AWS. I read each row from the provider's own page or price feed on 7 November 2026. Where a provider sells only 8-GPU instances, I divided by eight.

All checked 7 November 2026, USD per GPU-hour, on demand, excluding tax where the provider says so. Lambda's page states that the listed prices take effect on 16 November 2026 and that current prices show after login. AWS prices come from the on-demand price feed behind its pricing page (published 7 November 2026, Linux); Frankfurt p5 was $68.80 per instance, $8.60 per GPU. Vast.ai figures are my count of on-demand offers from its public search API, base GPU price only. Fireworks rates include its serving stack, so they are not raw compute. CoreWeave lists the same H100, H200 and B200 instance prices for Europe.

What does the extra money buy on an H200 or B200?

Memory and bandwidth. Per terabyte per second of bandwidth the three cost almost the same, $0.75 to $0.80 an hour, while the H200 is about a quarter cheaper per gigabyte. NVIDIA lists the H100 SXM at 80 GB and 3.35 TB/s, the H200 at 141 GB and 4.8 TB/s, and the DGX B200 at 1,440 GB and 64 TB/s across eight GPUs, which is 180 GB and 8 TB/s each. At RunPod's Community prices, by my division:

GPUwerk publishes this page and rents the last two rows; they are included as reference points, at GPUwerk's own published rates. RTX 5090 bandwidth from Tom's Hardware's review spec table; Spark from GPUwerk's spec page.

The H200 is the cheapest memory per hour among the three datacenter parts, and the cheapest single GPU here that holds about 100 GB of FP8 weights. The B200 buys compute as well as memory: NVIDIA's DGX B200 figure of 72 PFLOPS sparse FP8 for eight GPUs is 9 PFLOPS per GPU, against 3,958 TFLOPS sparse for an H200 SXM, and Hopper has no FP4 at all. For prefill-heavy or batch serving, the B200's higher hourly price can still win per token; the B300 vs B200 vs H200 comparison takes that further.

On-demand or reserved: how much does a commitment save?

Between 2% and about 33% at the providers that publish term prices, and some cluster products cost more per GPU than single instances. The published ladders:

• Together AI: H100 $3.99 on demand, then $3.69 (3 to 30 days), $3.45 (31 to 90), $3.19 (91 to 180). H200 $5.99 down to $3.99. B200 $8.19 down to $6.79. Preemptible: H100 $1.99, H200 $2.99, B200 $4.09.
• Hyperstack: reserved from $1.75 for its H100 PCIe, $2.72 for H100 SXM, $2.79 for H200 and $5.10 for B200.
• CloudRift: on-demand H100 and H200 are "upon request"; reserved H100 is $1.70 for one month and $1.33 for three, H200 $2.50 and $2.38.
• Verda: 2% off for one month, 8% for a year, 25% for two years.
• Lambda 1-Click Clusters (2 weeks to 1 year): H100 $5.54 to $6.16 and B200 $8.87 to $9.86 per GPU, above its own single-instance rates, so that commitment buys a multi-node cluster rather than a discount.

Monthly rental outside the US exists too. LeaderGPU lists an 8x H100 NVLink server at EUR 1,488 per card per month and a single H200 at EUR 2,987 a month, about EUR 2.04 and EUR 4.09 an hour over 730 hours by my division.

Why the same GPU costs $2.69 at one provider and $6.88 at another

Three things drive it: what else is in the box, the minimum unit you can rent, and what you pay around the GPU. AWS's p5.48xlarge comes with 192 vCPUs, 2 TiB of RAM, 8 x 3.84 TB of NVMe and 3,200 Gbps of EFA networking per the P5 instance page; you rent all eight GPUs or the single-GPU p5.4xlarge. CoreWeave's HGX instances and OVHcloud's H200 instance come with eight GPUs each. Marketplace offers on Vast vary by host, and only 6 of the 20 H100 SXM offers I counted came from verified hosts.

Around the GPU, costs vary less than the hourly rates do. CoreWeave, Nebius networking and Hyperstack list egress as free; Hyperstack charges about $0.10 per TB-hour for block storage; Nebius bills shared filesystems at $0.08 per GiB-month; RunPod charges $0.10 per GB-month for running volumes. Billing granularity runs from per second (RunPod, AWS with a 60-second minimum) to per minute (Hyperstack, JarvisLabs).

Are GPU rental prices going up in 2026?

Yes, by every series I could check. Silicon Data publishes daily rental indices from neocloud listings. Its own pages give these readings:

USD per GPU-hour. Each index page's FAQ quotes an older "current price" ($2.53, $3.10 and $5.45) that contradicts its own headline and daily table; I used the headline and table, which agree with each other.

By my arithmetic the H100 index is up 52% since 9 December 2025, the H200 index 31% since 4 May, and the B200 index 15% in the four weeks from 9 October. List prices are following: GPUwerk's DGX Spark vs H100 page recorded Lambda's H100 SXM at $3.99 to $4.29 on 11 October 2026, and Lambda's page now lists $4.59 to $4.89 from 16 November, a 14 to 15% rise. The memory side of this is in the 2026 memory shortage article.

When is a DGX Spark or an RTX 5090 the cheaper fit?

When the model fits and one person or a small team uses it; not when you need throughput or training. The test is memory first, then bandwidth. Single-stream decode speed is roughly bandwidth divided by the bytes of active weights, which is why GPUwerk measured Llama 3.3 70B at 6.0 tok/s on one Spark (vLLM 0.27.1, AWQ, 2 September 2026, logs on the benchmarks page) while the same arithmetic gives an H100 a ceiling around 90 tok/s.

• Model under about 30 GB, one user: an RTX 5090 at $0.89 an hour has 53% of an H100 SXM's bandwidth for 33% of RunPod's cheapest H100 SXM rate. Rent the 5090; RTX 5090 cloud pricing lists other providers.
• Sparse model of 30B class, small team: one Spark. GPUwerk measured Qwen3-Coder 30B-A3B at 80.9 tok/s single-stream and 3,235 tok/s total throughput at 256 concurrent requests.
• Weights of 80 to 128 GB, one user: one Spark at $0.79 against two H100s at $5.38 or more, or an H200 at $3.59 or more. The Spark is slower per token; the DGX Spark vs H200 page shows by how much.
• Dense 70B for many users, any training with NVLink, or FP8 batch serving: H100, H200 or B200. A Spark's 70B throughput peaked at 485 tok/s total (prompt plus generated, 162 generated) at 128 concurrent in GPUwerk's run and fell past that.

My recommendation: for evaluation and single-user work, rent by the hour on the cheapest tier that does not preempt you (GPUwerk's hourly rates are on its pricing page), and do not reserve until a week of on-demand use proves the load is steady. For an 8-GPU job that runs for months, a published reservation ladder like Together's or a quote for a whole server, which is what GPUwerk's 8x H100 and 8x B200 page offers, beats any hourly rate in this table. The older A100 80 GB, at $1.19 to $2.32 an hour on demand, is compared with a Spark in DGX Spark vs A100.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Pricing, RunPod
• GPU cloud pricing, Lambda
• Pricing, CoreWeave
• Prices, Nebius
• GPU pricing, Hyperstack
• Pricing, Verda
• Pricing, JarvisLabs
• Pricing, Together AI
• Pricing, Fireworks AI
• Amazon EC2 On-Demand Pricing, AWS
• Amazon EC2 P5 instances, AWS
• Public Cloud prices, OVHcloud
• GPU rental prices per hour, CloudRift
• Offer search API (public bundles endpoint), Vast.ai
• H100 Rental Price Index, Silicon Data, 2026-11-07
• H200 Rental Price Index, Silicon Data, 2026-11-07
• B200 Rental Price Index, Silicon Data, 2026-11-07
• H200 vs H100 rental prices, May to July 2026, Silicon Data, 2026-08-04
• H100 price spike, Silicon Data, 2026-01-09
• B200 rental price, August 2026 update, Silicon Data, 2026-09-24
• NVIDIA H100 Tensor Core GPU, NVIDIA
• NVIDIA H200, NVIDIA
• NVIDIA DGX B200, NVIDIA
• GPU servers for rent, LeaderGPU
• Nvidia GeForce RTX 5090 Founders Edition review, Tom's Hardware