NVIDIA H100 cloud prices

The cheapest H100 SXM hour on 16 October 2026 was an interruptible bid on Vast.ai at $0.687 per GPU-hour, on an unverified host in California that pauses your instance if someone outbids you 8. Nebius' dynamic spot price starts at $0.79 in eu-north1 14. The cheapest hour that cannot be taken back was a single-GPU Lium marketplace pod at $1.39, with 8-GPU nodes from $1.69 per GPU, although listings come and go during the day 6. Among managed clouds that fix the machine and put CPU, RAM, local disk, an IPv4 address and traffic in the price, the lowest we confirmed on a provider's own page is GPUBank at $1.85, billed by the second, or about $1.16 effective with a $1,000 top-up 1. CUDO Compute, still shown at $1.79 by trackers, now quotes on request 25.

An H100 SXM costs between about $0.69 and $12.29 per GPU-hour in October 2026, depending on who sells it and on what terms. GetDeploying's median on-demand price was $3.83 across 39 providers on 15 October 10, and Ornn's daily OCPI-H100 index settled at $3.13 the same day 36. The low end is spot and marketplace capacity: Vast.ai bids from $0.69, Nebius spot from $0.79 and Lium pods from $1.39 8 14 6. Managed clouds run from GPUBank's all-in $1.85 through Runpod at $2.69 and $3.99 and Lambda at $3.99 to $4.29 up to Nebius and Crusoe at $4.50 1 11 12. The hyperscalers sit at the top, selling 8-GPU instances: AWS at $6.88 per GPU, Oracle at $10.00, Google Cloud at $11.06 and Azure at $12.29 28 33 31 32.

The H100 SXM is the fastest of the three and the one sold on 8-GPU HGX baseboards, with 80 GB of HBM3 at 3.35 TB/s, NVLink at 900 GB/s and a power limit of up to 700 W 3. The H100 PCIe is a standard add-in card with 80 GB of slower HBM2e at about 2 TB/s and a much lower power limit, and it rents for less: $1.35 at GPUBank and $1.99 or $2.89 on Runpod 1 11. The H100 NVL is a 94 GB PCIe card with 3.9 TB/s of bandwidth and NVLink at 600 GB/s between pairs, listed at $2.59 and $3.19 on Runpod. Many price lists label all three 'H100 80GB', and GetDeploying finds that only about half of providers state the form factor 10, so check the variant on the provider's page or in nvidia-smi before comparing prices.

A marketplace H100 is the same silicon but not the same product. On Vast.ai, Lium, Clore.ai or Runpod Community Cloud, an independent host sets the price, the CPU, RAM and disk, the maximum rental length and often the storage and bandwidth fees, and the listing ends if that host goes offline 9. Lium shows one H100 host with a reduced power limit and reliability scores from 64 to 96 percent on its nodes 6, and Vast.ai shows a reliability score on every offer 8. Marketplaces are the right choice for short, checkpointed or exploratory work where the lowest hourly rate matters most and a lost session costs little. A managed cloud is the safer choice for production inference, long jobs without frequent checkpoints, and anything where a lost night of compute costs more than the difference in price.

H100 SXM spot prices on 16 October 2026 started at $0.687 on a Vast.ai interruptible bid 8 and $0.79 on Nebius, whose spot price has been dynamic since 8 October and can change every 15 minutes 14. GPUBank's interruptible tier costs $1.11 with five minutes of notice 1, Together AI's preemptible clusters $1.99 17, Verda $2.08 20 and Azure's ND H100 v5 spot about $2.27 per GPU 32. Spot is worth it for training and batch jobs that write checkpoints every 15 to 30 minutes and can resume on another machine, because a saving of 40 to 80 percent outweighs the occasional lost interval. It is not worth it for a service that must answer requests, or for jobs whose state is expensive to rebuild, where the hours lost to a reclaim quickly eat the discount.

You can rent a single H100 SXM from most specialised clouds, but Azure and Oracle sell their H100 SXM instances only by eight GPUs, and the AWS instance most buyers know, the p5.48xlarge, is also an 8-GPU unit, although Akash's September snapshot lists a single-GPU p5.4xlarge at about $6.88 32 33 28 27. GPUBank, Runpod, Lambda, Verda, JarvisLabs, DigitalOcean, Nebius, Vast.ai and Lium all rent a single GPU with a share of the host's CPU and RAM, at prices from $1.39 on a Lium pod to $4.41 on DigitalOcean 6 22. Lambda charges more per GPU on small instances, $4.29 on 1x against $3.99 on 8x 12, and some marketplaces price whole nodes lower per GPU than single pods. One H100 is enough to fine-tune with LoRA or serve models up to about 30B parameters in FP8; a 70B model in FP8 needs two GPUs for its weights and a useful cache, and full fine-tuning of 70B needs the whole 8-GPU NVLink node.

The hyperscalers charge $6.88 to $12.29 per H100 SXM GPU-hour on demand because they sell 8-GPU instances with very large CPU and RAM allocations, cluster networking, compliance coverage and integration with the rest of their cloud, and because their large customers pay through commitments rather than list prices 28 31 32. AWS's p5.48xlarge lists at $55.04 an hour, Google Cloud's a3-highgpu-8g at $88.49 and Azure's ND96isr H100 v5 at $98.32. Commitments narrow the gap: Azure's three-year reservation works out to about $5.40 per GPU 32, Google's three-year committed use to about $4.86 31, AWS's three-year reserved instance to about $2.97 28, and AWS Capacity Blocks cost $5.97 30. Egress, billed separately on AWS and Google Cloud, adds to the bill of any inference service.

A LoRA fine-tune of Llama 2 70B as defined by MLPerf Training, on GovReport at a sequence length of 8,192 to a fixed quality target, took 27.93 minutes on one 8-GPU H100 SXM node in NVIDIA's verified v4.1 entry 37. That is 3.72 GPU-hours, or about $6.89 at GPUBank's $1.85 on-demand rate 1. A real project takes longer, because MLPerf uses a tuned NVIDIA stack, a modest dataset and a single successful run, while real work adds data preparation, evaluation and repeated attempts; multiplying the benchmark by 5 to 20 for a first real run puts a 70B LoRA project in the range of tens to a few hundred dollars of H100 time. On a B200 node the same benchmark costs about $5.28 at GPUBank's prices, and on an H200 node about $8.20 39 38.

Per job, the B200 often costs less than the H100 despite its higher hourly price, and the H200 usually costs more. At GPUBank's prices, the MLPerf Llama 2 70B LoRA fine-tune costs $6.89 on an 8-GPU H100 node at $1.85 per GPU, $8.20 on an H200 node at $2.49 and $5.28 on an HGX B200 node at $4.15, because the B200 finished in 9.55 minutes against 27.93 for the H100 and 24.7 for the H200 37 38 39. The H200's advantage is memory: 141 GB of HBM3e lets a 70B model in FP8 run on one GPU with a large KV cache, which matters for long contexts and many concurrent users. The A100 at $0.95 remains the cheapest tool for small LoRA jobs that do not use FP8 1.

Buying an 8-GPU HGX H100 server is cheaper than renting only for a node kept busy almost all the time for two years, and then only if the money tied up in it costs you nothing. A new system costs about $285,000 42 43; at 45 percent resale after 24 months, $2,000 a month for colocation and maintenance, 8 percent a year on the capital and power at the US commercial average, owning costs about $10,431 a month plus $1.42 per node-hour, or about $11,470 at full use. Renting the same eight GPUs at GPUBank costs $10,804 a month on demand, $8,643 reserved by the month and about $6,753 at the effective price of $1,000 top-ups 1 5. Without the cost of capital, owning beats on demand above about 638 node-hours, 87 percent of the month, and never beats the reservation or the top-up rate. Used systems sold for a median of $250,000 in July 2026 and refurbished ones with a 90-day warranty for $265,000 44, but the resale value of an H100 server two years from now is the largest uncertainty in the sum.

H100 rental prices fell from 2023 to late 2025 and have been rising since. Silicon Data's series puts the hyperscaler median near $7.76 per GPU-hour in late 2023 and $9.34 in late 2024, and the marketplace median at $1.95 in the second half of 2025 34; AWS cut P5 on-demand prices by 44 percent in June 2025 29, and SemiAnalysis' one-year contract index bottomed at about $1.70 in October 2025 35. In 2026 that index rose to $2.35 by March, Lambda raised its 8x H100 SXM rate from $2.99 to $3.99 in June 13, Nebius went from $3.85 to $4.50 on 1 October 14 and AWS Capacity Blocks reached $5.97 30. Cheap hours still exist, on marketplaces, spot tiers and smaller clouds, but they come from choosing the seller rather than from a general fall.

A stopped H100 instance usually costs its storage, and on some offers much more. GPUBank charges only the persistent disk at $0.03 per GB-month, so 500 GB kept for a month costs $15 1. Nebius stops charging compute when a VM is stopped but keeps billing its disks 14. On Vast.ai the instance's storage keeps billing for as long as the instance exists, at a rate each host sets, $0.20 per GB-month on the cheapest H100 SXM host on 16 October and up to about $0.66 on others 8 9. Runpod bills a stopped pod's volume disk at $0.20 per GB-month and network volumes at $0.07 11. On bare-metal servers rented by the node, the whole machine usually keeps billing until it is released, so check which lines keep running before you stop rather than delete, and move large datasets to a cheap volume if the GPU will sit idle for days.

Yes, at a few providers. GPUBank is paid only from a prepaid US dollar balance topped up in crypto, with Bitcoin, Ethereum, USDT, USDC, Monero and more than a dozen other coins accepted for top-ups of $25 to $5,000, and its bonus of 20 to 60 percent brings the effective H100 SXM hour to about $1.54 with $25, $1.37 with $100 and $1.16 with $1,000 1. The balance never expires, every account and team member can have a monthly spend cap, and cards are not accepted for now. Several marketplaces, including Vast.ai, Lium and Clore.ai, also work from prepaid credit 9 6 19, while the large clouds and most managed providers bill a card or an invoice. A top-up bonus is a discount only on credit you actually spend, so size it to the hours you expect to use.