Where to rent an H100 SXM at the cheapest price

The H100 entered the cloud at hyperscaler prices. Azure's ND H100 v5 has listed at $98.32 an hour for eight GPUs, $12.29 per GPU, since December 2023⁶, and Silicon Data's series puts the hyperscaler median at about $7.76 per GPU-hour from August to December 2023, rising to $9.34 in the second half of 2024⁴¹. The fall came first from below. The neocloud wave of 2024 put a median of $2.99 on specialised clouds and $2.58 on marketplaces between June and December 2024, and marketplace rates slid to $2.29 in the first five months of 2025 and $1.95 in the second half of the year⁴¹. SemiAnalysis' index of one-year H100 contracts, compiled from monthly surveys of more than 100 market participants, bottomed at about $1.70 per GPU-hour in October 2025⁴².

The hyperscalers followed in June 2025, when AWS cut P5 on-demand prices by 44 percent from 1 June²⁵, which brought the p5.48xlarge to $55.04 an hour, $6.88 per GPU²⁴; Silicon Data's hyperscaler median dropped to $6.94 that month and to $6.26 in the second half of 2025⁴¹. That was the bottom. By March 2026 the SemiAnalysis contract index had climbed to $2.35, a rise of almost 40 percent in six months⁴², Lambda raised its 8x H100 SXM rate from $2.99 to $3.99 on 9 June 2026⁴³, and in October Nebius moved from $3.85 to $4.50 and AWS Capacity Blocks reached $5.97²⁷.

Today the typical on-demand hour is far below the 2023 hyperscaler price but above the 2025 trough: GetDeploying's median was $3.83 across 39 providers on 15 October 2026⁵, and Ornn's daily OCPI-H100 index, which moved between $2.49 and $3.21 over the previous month, settled at $3.13 that day⁴⁴. The cheapest hours did not rise with the median. A single Lium pod at $1.39³ and Vast.ai bids from $0.69¹ sit below even the 2025 marketplace median, which is the practical lesson of the chart: a cheap H100 hour is no longer a matter of waiting for the market to fall, but of choosing the seller, the terms and the risk you are willing to carry.

Six kinds of offer sit on the H100 SXM list, and each answers a different question. A managed on-demand cloud sells an instance at a fixed hourly price from its own fleet, with a stated amount of CPU, RAM and disk, and the thing to check is what the price leaves out: Runpod's $3.99 Secure Cloud pod includes 20 vCPU and 125 GB of RAM but not its volumes¹³, whereas Lambda's $3.99 per GPU on its 8x instance comes with 208 vCPU, 1,800 GiB of RAM and 22 TiB of SSD¹⁴. A peer-to-peer marketplace such as Vast.ai or Lium passes through a price set by an independent host, so two listings at the same price can be different machines in different countries with different disk and bandwidth charges¹³. Spot or interruptible capacity is the same hardware sold on the understanding that it can be taken back, at anything from $0.69 on Vast.ai to $2.08 at Verda¹⁸.

Reserved capacity is where the cheapest dependable hours are, but it is sold by the week, the month or the year, often by the whole node and often prepaid: Clore.ai's own dated snapshot quotes 8-GPU bare-metal contracts at $1.81 to $2.62 per GPU-hour with a 30-day minimum¹⁵, Together AI charges $3.19 for 91 to 180 days¹⁶ and DigitalOcean $3.26 on a 12-month contract¹⁷. Serverless platforms bill only while code runs, so Modal's $0.001097 per second, about $3.95 an hour for the GPU alone, can beat any always-on instance for a service that is busy a fraction of the day¹⁸. The large clouds, finally, sell the H100 SXM almost exclusively as 8-GPU instances, so their per-GPU price is the instance price divided by eight and one hour of it commits you to eight cards. The fair comparison is always like with like: a marketplace headline against another marketplace headline, and a managed all-in price against what the managed competitor charges once CPU, RAM, disk, IPv4 and traffic are added.

The captures below were taken on 16 Oct 2026 from each provider's own H100 or pricing page, so every figure can be read next to the conditions printed beside it.

On a managed, all-inclusive, per-second basis, GPUBank's $1.85 is the lowest H100 SXM price we could confirm on a provider's own page. The hour includes 20 vCPU, 125 GB of RAM, a 2 TB local NVMe disk, one public IPv4 and all inbound and outbound traffic, in machines of one to eight GPUs with NVLink at 900 GB/s, and the only other line a bill can carry is a persistent disk at $0.03 per GB-month⁴. The meter starts after the machine passes its health check, a failed launch is not billed, and a host failure during a job restarts the machine on a healthy host with the same disk and credits the interrupted session in full. Interruptible capacity costs $1.11 with five minutes of notice and a stop rather than a deletion, a week or a month reserved costs $1.48, and top-up bonuses bring the effective on-demand hour to about $1.54 with a $25 top-up, $1.37 with $100 and $1.16 with $1,000⁴. The card is offered in Paris, Frankfurt, Ashburn and Singapore, with a median of 48 seconds from launch to a machine that has passed its health check, and any Docker image or a ready template for vLLM, SGLang, Unsloth, Axolotl, ComfyUI or JupyterLab can be started on it²⁸.

It loses on headline price to every spot tier below $1.11, Vast.ai's $0.69 bid and Nebius' dynamic price from $0.79 among them¹², and to Lium's marketplace pods at $1.39 for one GPU or $1.69 per GPU on an 8-GPU node, as long as such a listing is there when you need it³. Its machines stop at eight GPUs and it lists no fabric between servers, so a job that must span 16 or 64 GPUs belongs on Lambda's 1-Click Clusters, Crusoe, Together AI or a hyperscaler, at two to four times the price per GPU-hour. Payment is a prepaid US dollar balance topped up in crypto from $25 to $5,000; card payments are not accepted for now, which rules GPUBank out for a buyer whose finance team can only pay by card or invoice, and the bonus only lowers the real price if the credit is actually spent, although the balance never expires⁴.

The four large clouds sell the H100 SXM by the 8-GPU instance, and their list prices are three to seven times those of the specialised clouds. AWS's p5.48xlarge costs $55.04 an hour on demand in us-east-1, $6.88 per GPU, with 192 vCPU and 2,048 GiB of RAM, and EBS storage and egress on top²⁴; that price dates from June 2025, when AWS cut P5 on-demand rates by 44 percent²⁵. Google Cloud's a3-highgpu-8g lists at $88.49 an hour in us-central1, $11.06 per GPU²⁶. Azure's Standard_ND96isr_H100_v5 costs $98.32 an hour for Linux in East US, $12.29 per GPU, a rate unchanged since December 2023⁶. Oracle's price list charges $10.00 per GPU-hour for the H100 shape, $80 per node, the same in every public region²⁷.

Commitments and spot change those figures sharply, which is how large customers actually pay. Azure's own price list puts its ND H100 v5 spot rate at $18.17 per VM-hour, $2.27 per GPU, and its one-year and three-year reservations at about $7.87 and $5.40 per GPU⁶. Google's a3-highgpu-8g spot price works out to $6.29 per GPU and its three-year committed-use price to about $4.86²⁶, AWS's three-year reserved p5.48xlarge to about $2.97 per GPU²⁴, and AWS Capacity Blocks, which reserve P5 capacity for a fixed window, cost $47.757 per instance-hour, $5.97 per accelerator, in every listed region, up from about $5.19 in US regions before the October revision, so a guaranteed P5 window now costs almost as much as the on-demand hour⁷. What these prices buy is integration with the rest of the cloud, compliance coverage, and large EFA, InfiniBand or RDMA clusters; for raw GPU-hours on one node they are rarely the cheap choice.

The H100 SXM is the right card for LoRA and full fine-tuning of models up to about 70B parameters on one 8-GPU node, for FP8 inference of 8B to 70B models, and for training that needs NVLink bandwidth between eight GPUs. It stops being the right card when the model and its KV cache no longer fit in 80 GB per GPU, which is where the 141 GB H200 or the 180 GB B200 saves both GPUs and time; when the job is small enough that HBM3 and NVLink sit idle, as with a 7B LoRA or speech transcription, which an A100 does for about half the price; and when a job spans many nodes, where the fabric between servers matters more than the card.

The cleanest training measurement is MLPerf Training's Llama 2 70B LoRA test, a fine-tune on the GovReport dataset at a sequence length of 8,192, timed to a fixed quality target. On one 8-GPU H100 node, NVIDIA's verified entry 4.1-0050 in MLPerf Training v4.1 finished in 27.93 minutes³¹, which is 3.72 GPU-hours, or $6.89 at GPUBank's on-demand rate. In v4.0 an 8-GPU H200 system finished in 24.7 minutes³², 3.29 GPU-hours or $8.20 at GPUBank's $2.49, and in v5.1 an 8-GPU HGX B200 system finished in 9.55 minutes³³, 1.27 GPU-hours or $5.28 at $4.15. The B200 is the cheapest per job despite costing more than twice as much per hour, and the H200 is the most expensive per job on this test, because its gain in speed is smaller than its gain in price. A real project takes longer than the benchmark, since MLPerf uses a tuned NVIDIA stack, a modest dataset and a single successful run.

For serving, the cleanest cross-card measurement is MLPerf Inference v5.0, where NVIDIA's DGX H100 (entry 5.0-0057) served Llama 2 70B at 31,106 tokens per second in the server scenario and 31,307 offline, against 33,072 and 34,988 for NVIDIA's 8-GPU H200 system and 98,443 and 98,858 for the DGX B200, with both Hopper systems in FP8, the B200 in FP4, and TensorRT 10.8 and CUDA 12.8 throughout³⁴. Divided by eight, the H100 makes about 14 million tokens an hour per GPU, $0.132 per million at GPUBank's $1.85, against $0.167 on the H200 at $2.49 and $0.094 on the B200 at $4.15. On a 70B model that fits on both Hopper cards, the H200's 6 percent more throughput does not pay for its 35 percent higher price, and the B200 is cheaper per token only when the stack runs in FP4.

For inference, NVIDIA's TensorRT-LLM performance overview, generated with trtllm-bench on the PyTorch backend at maximum load on synthetic data, reports 14,992 output tokens per second for Llama 3.1 8B in FP8 on a single H100 SXM at 1,000 input and 1,000 output tokens, 26,401 at 128 and 128, and 9,462 at 2,048 and 2,048³⁵. At $1.85 an hour, 14,992 tokens per second is about 54 million tokens an hour, or $0.034 per million output tokens. Llama 3.3 70B in FP8 on two H100s with tensor parallelism of 2 reaches 4,181 tokens per second at 1,000 and 1,000, or about $0.25 per million on two GPUs at $3.70 an hour³⁵. These are reference points under full batching, not what a single user sees, and a user who measured 15,897 tokens per second where the documentation for v0.20 expected 27,688 at 128 and 128 was told the gap came mainly from running a PCIe card³⁰.

Per hour the H100 SXM sits in the middle of GPUBank's datacenter cards, at $1.85 against $0.95 for the A100 SXM, $1.35 for the H100 PCIe, $1.65 for the MI300X, $2.49 for the H200 SXM and $4.15 for the B200 SXM⁴. Per job the order changes: on the MLPerf Llama 2 70B LoRA fine-tune above, the B200 is the cheapest run and the H200 the dearest³¹³²³³, so a team running large fine-tunes all month pays less on the more expensive card, and the MLPerf Inference v5.0 results give the same order for serving a 70B model: $0.094 per million tokens on the B200, $0.132 on the H100 and $0.167 on the H200³⁴. The H200 earns its premium in memory rather than speed: with 141 GB of HBM3e at 4.8 TB/s, a 70B model in FP8 fits on one GPU with room for a long KV cache, where the H100 needs two GPUs or a shorter context.

At the other end, the A100 SXM at $0.95 has the same 80 GB but HBM2e at 2.0 TB/s and no FP8, so it remains the cheaper tool for small LoRA runs, embedding jobs and inference of models that run in 16-bit anyway, and the H100 PCIe at $1.35 suits single-GPU inference that never uses NVLink. The MI300X at $1.65 offers 192 GB of HBM3 at 5.3 TB/s, more memory per dollar than any NVIDIA card here, for teams whose stack runs well on ROCm. We did not find measured results for the A100, the H100 PCIe or the MI300X on the same benchmarks and settings, so we do not compute a cost per job for them.

The realistic purchase unit for the H100 SXM is a whole 8-GPU HGX server, because the module only runs on an HGX baseboard. Mercatus, writing in July 2026, gives $250,000 to $320,000 for a new system with about $285,000 typical from the major OEMs³⁶, and CloudZero gives the same range in its October 2026 update³⁷. Used machines are cheaper and thinly traded: Rack Rate's index of 18 July 2026 showed a median of $250,000 for used, tested 8-GPU systems, in a range of $225,000 to $275,000, and $265,000 for refurbished ones with a 90-day warranty³⁸. Mercatus' own resale estimates for a system after 36 months run from $62,000 to $146,000 counting the GPUs alone³⁶, which shows how fast this value is falling as Blackwell ships.

For the arithmetic we take a new system at $285,000, worth 45 percent of that after two years, or $128,250, and add $2,000 a month for colocation, network, remote hands and spares, $48,000 over 24 months; all three are our assumptions, not quotes. The fixed cost is then $204,750 over 24 months, about $8,531 a month, and the money tied up in the server, at 8 percent a year on $285,000, adds about $1,900, for $10,431 a month. For power we take the roughly 10.2 kW peak per system in NVIDIA's DGX H100 specifications and SuperPOD design guide³⁹ at the US commercial average of 13.97 cents per kWh for January to July 2026⁴⁰, about $1.42 per node-hour under load. Renting the same eight GPUs at GPUBank costs $14.80 per node-hour on demand, so owning would need 10,431 / (14.80 - 1.42) = 780 node-hours a month to break even, more than the 730 a month contains.

Leave the cost of capital out, as a buyer paying cash that would otherwise earn nothing may, and the break-even with on demand falls to about 638 node-hours, 87 percent of the month; against GPUBank's reserved rate, $11.84 per node-hour, it is about 819 hours and against the effective $9.25 of $1,000 top-ups about 1,090, so a new server never wins against either. A used system at $225,000 that resells for $90,000 lowers the cash-only threshold to about 570 node-hours, but the resale assumption moves everything: every 10 points of resale value on a $285,000 system changes the monthly cost by about $1,190³⁸. Buy only if a team will keep an 8-GPU node busy more than nine hours in ten for two years against on-demand rates, already has power and rack space, and can carry the resale risk; everyone else pays less by renting, and least by renting reserved or prepaid capacity.