Paid links may earn Compute Screener a referral fee. They never affect prices, weights or the order.
About NVIDIA H200
The NVIDIA H200 rental price today is $5.10 per GPU-hour, down 0.3% since yesterday, up 2.3% over the past week and up 12.3% over the past month. At that rate one H200 costs about $3,720 for a month of 730 hours. Over the past year it has ranged from $3.57 to $5.12. The index is the price of one benchmark contract: H200 SXM 141GB, on-demand, 8-GPU node, US hardware, CPU-neutral. It is standardized across 11 cloud providers and marketplaces, from Civo at $3.45 to CoreWeave at $6.31, and is updated every 2 to 3 hours.
The H200 is a memory upgrade of the H100. NVIDIA announced it on November 13, 2023 as the first GPU with HBM3e memory: 141 GB at 4.8 TB/s, against the H100's 80 GB at 3.35 TB/s, on the same Hopper compute engine. It fits the same HGX servers as the H100, and its extra memory makes it a common choice for serving large models: OpenAI adopted Azure's H200 machines in October 2024.
Specifications
| Spec | H200 SXM |
|---|
| Architecture | Hopper |
| Memory | 141 GB HBM3e, 4.8 TB/s |
| BF16 / FP16 tensor | about 990 TFLOPS dense, 1,979 with sparsity, the same as the H100 |
| FP8 tensor | about 1,979 TFLOPS dense, 3,958 with sparsity |
| NVLink | 900 GB/s |
| Power | up to 700 W |
| Other versions | H200 NVL, a PCIe card of up to 600 W |
| Typical server | HGX H200, 4 or 8 GPUs |
Source: NVIDIA H200, accessed October 24, 2026.
Where the H200 is rented
CoreWeave said it was the first cloud to offer the H200, on August 28, 2024. AWS followed with P5e on September 9, 2024, at first only as reserved Capacity Blocks, then Azure ND H200 v5 on October 2, AWS P5en on December 2 and Google Cloud A3 Ultra at the end of 2024. GetDeploying tracked 50 H200 providers on October 23, 2026, of which 33 posted on-demand prices and 8 offered custom contracts only.
What this price measures
The index is the posted on-demand price of one benchmark contract, an H200 SXM 141GB in an 8-GPU node of US hardware, rented by the hour from GPU clouds and marketplaces; prices for other node sizes are converted to that contract, and prices from other regions are counted without a regional adjustment for now. It is not a hyperscaler, reserved or spot price.
The gap between those is large for the H200. On October 23, 2026, it was listed at $4.29 per GPU-hour at Crusoe, $6.30 on demand and $2.62 spot at CoreWeave, $7.91 on demand at AWS, $10.00 at Oracle, and $13.78 on demand or $6.86 for a 36-month reservation at Azure.
What moves H200 rental prices
- Memory-bound inference. Serving large models and long contexts needs memory more than compute. AWS said in September 2024 that one P5e machine with H200s could serve Llama 3.1 405B at up to 1.72 times the throughput of two P5 machines with H100s.
- Exports to China. The US allowed H200 sales to China in December 2025 on condition of a 25% payment to the government, and in May 2026, Reuters reported, cleared sales to 10 Chinese companies. Chips sold to China are not rented out in the US market.
- Next-generation supply. Blackwell capacity coming online through August and September 2026 was already booked by April, and in March it was "increasingly impossible to find any H100s, H200s or B200 rental capacity for any term".
- Hyperscaler pricing. AWS cut on-demand prices of P5en machines with H200s by 25% from June 1, 2025.
- Memory supply. High-bandwidth memory is reported to be the binding constraint on GPU manufacturing, with hyperscalers buying production years ahead, RunPod wrote in September 2026.
Price history
GetDeploying's public weekly H200 series starts in October 2025 with 21 providers; its on-demand provider median rose from $3.50 to $5.00, about 43%, over the year to the week of October 19, 2026, while its provider count grew to 35. On October 23, 2026 the median posted on-demand price was $4.68 per GPU-hour across providers and configurations, against $3.85 for the H100. Hardware estimates in January 2026 put an 8-GPU H200 board at about $308,000 to $315,000, against about $215,000 for an H100 board. This index's history before October 23, 2026 is GetDeploying's weekly H200 series, rescaled to join the daily index.
Futures and benchmarks
No H200 futures contract has been filed. ICE and Ornn named the H200 among planned contracts in May 2026, but ICE's September 29, 2026 notice lists only H100 and B200 contracts, and CME's filed contracts are also for H100 and B200. See the H100 and B200 pages.
FAQ
How does the H200 compare with the H100 on price? On GetDeploying's October 23, 2026 medians of $4.68 against $3.85, an H200 cost about 22% more per hour than an H100 for the same compute, 76% more memory and 43% more memory bandwidth.
What is the H200 NVL? A PCIe card version for standard servers, up to 600 W, linked in pairs or fours by NVLink bridges. This index measures the SXM version in 8-GPU servers.
How often is this index updated? Every 2 to 3 hours, with one value per day in US Central time that is final when the day closes.
Sources
Text reviewed Oct 24, 2026; the figures in it are as of that date. The figures at the top of this page update every 2 to 3 hours. Not investment advice.
