Buyer Guide

Rent an H100: Prices, Providers, and What to Watch

By Tyler Babin · FlopSource Research

Published Nov 04, 2026

The H100 is still the default rental GPU for serious training and high-throughput inference, and it is also the easiest GPU to overpay for. The same card, in the same week, rents for $1.09 per GPU-hour as an interruptible Spot VM on Google Cloud and $5.38 per GPU-hour on demand on the same platform (FastGPU, Nov 04, 2026). This guide walks through what the live market actually charges, which of the three H100 variants you are really being offered, what the FlopSource directory records say, and the traps that turn a cheap headline rate into an expensive run.

What an H100 costs right now

As of November 4, 2026, the median on-demand price across 46 providers with a priced H100 configuration is $3.61 per GPU-hour, and the cheapest verified in-stock on-demand offer is $1.79 (GetDeploying, Nov 04, 2026). Reserved capacity medians $3.23 and Spot medians $2.39 in the same tracker. At the median on-demand rate, one H100 run for a 720-hour month costs about $2,599. Prices are drifting up, not down: the same median is up about 8% over the past 90 days and about 13% over the past 12 months (GetDeploying, Nov 04, 2026).

The hyperscaler spread is the widest part of the market. Google Cloud charges $5.38 per GPU-hour on demand for an H100 in its cheapest region, 3.0x the $1.79 market floor, while its Spot rate of $1.09 undercuts every on-demand offer in the market (FastGPU, Nov 04, 2026). The catch is structural: Google sells H100 machines on demand only as whole eight-GPU VMs, so the smallest uninterrupted Google bill is eight times the per-GPU figure, $43.04 per hour, and a Spot VM can be stopped at any time (FastGPU, Nov 04, 2026).

The FlopSource directory tells the same story from its own records, with one honest caveat: most H100 directory records were last verified on June 08, 2026, so treat them as a dated snapshot, not live pricing. The directory lists 29 providers offering the H100. Across the 28 with a published representative rate, the median is $2.25 per GPU-hour, from Vast.ai at $1.53 and Hetzner at $1.55 at the low end to Microsoft Azure at $12.29 at the high end (FlopSource Directory dataset, records last verified Jun 08, 2026). Named mid-market records in that snapshot include Lambda Labs at $2.39, CoreWeave at $2.65, RunPod at $3.19, Google Cloud at $3.95, and AWS at $4.10, all last verified Jun 08, 2026. Directory rates are each provider's representative GPU-hour rate, not a per-H100 quote for your exact configuration; confirm the live price with the provider before you commit. Browse the live directory at directory.flopsource.com.

First question: which H100 are you actually renting?

"H100" on a rental listing can mean three different boards, and they do not perform alike. Per NVIDIA's own product specifications, the H100 SXM carries 80 GB of HBM3 at 3.35 TB/s, 3,958 TFLOPS of FP8 Tensor Core throughput with sparsity, NVLink at 900 GB/s, and a 700 W power envelope. The H100 NVL carries 94 GB of HBM3 at 3.9 TB/s and 3,341 FP8 TFLOPS, with NVLink at 600 GB/s for a bridged two-GPU pair (NVIDIA, H100 product page, accessed Nov 04, 2026). The H100 PCIe card carries 80 GB of slower HBM2e at about 2 TB/s, 3,026 FP8 TFLOPS with sparsity, and a 350 W envelope, and it does not support NVLink (RunPod, NVIDIA H100 guide, accessed Nov 04, 2026).

• Multi-GPU training (4 or more GPUs, tightly coupled): rent SXM. NVLink at 900 GB/s per GPU is what keeps gradient exchange from becoming the bottleneck; the PCIe card has no NVLink at all (NVIDIA, accessed Nov 04, 2026; RunPod, accessed Nov 04, 2026).
• Single-GPU fine-tuning and most inference: PCIe is the honest buy. You keep the same 80 GB of memory and the Transformer Engine at a lower hourly rate and half the power draw (RunPod, accessed Nov 04, 2026).
• Memory-bound inference on a large model: look at the NVL. Its 94 GB holds more weights and KV cache per card than the 80 GB variants (NVIDIA, accessed Nov 04, 2026).

Do not judge the variant by its hourly rate alone. In an independent rental test of all three variants serving the same model (Qwen3-8B in BF16 on vLLM, priced at RunPod Secure Cloud rates on the day), the SXM was the most expensive card per hour and the cheapest per output token: $0.31 per million output tokens on SXM, against $0.43 on PCIe and $0.57 on NVL (TokenWatt, Oct 2026). Cost per token at your quality bar is the number that matters; our guide to the inference economy explains why.

What to watch: seven traps

• The listing does not name the variant. "H100 80GB" can be SXM or PCIe, and the performance gap between them is large (memory bandwidth 3.35 TB/s vs about 2 TB/s; NVIDIA and RunPod, accessed Nov 04, 2026). Ask, in writing, before you pay.
• Power limits set below spec. In the same rental test, every NVL card drawn was capped at 310 W, below NVIDIA's 350 to 400 W configurable range for the card, and the cards reported that cap as their default (TokenWatt, Oct 2026). Run a two-minute health check on any new card: sustained clocks and memory bandwidth under load tell you what you actually got. Our GPU cloud benchmark guide has the full checklist.
• Whole-node billing. A per-GPU rate can hide an eight-GPU minimum. Google Cloud sells its H100 machines only as whole eight-GPU VMs on demand, so the real entry price is $43.04 per hour, not $5.38 (FastGPU, Nov 04, 2026).
• Spot is a different product, not a discount. Google's $1.09 Spot H100 can be stopped at any time (FastGPU, Nov 04, 2026). Spot pays only for checkpointed, restartable work. The break-even math for on-demand vs reserved vs spot is in our pricing-model guide.
• Availability wording. In the FlopSource directory snapshot, H100 availability ranges from "Immediate" to "2-4 weeks" (Hetzner) and "Contact for availability" (Lambda Labs, Vultr), all last verified Jun 08, 2026 (FlopSource Directory dataset). A low rate you cannot provision for three weeks is not a low rate for a deadline.
• The rate is not the run cost. Storage, egress, failed runs, and idle time sit on top of the GPU-hour. Our breakdown of the true cost of training and the GPU-hours estimator will size the real bill before you book.
• Stale comparison pages, including ours. H100 prices moved about 8% in 90 days (GetDeploying, Nov 04, 2026). Any price page without a date on it, including a directory record from June, is a starting point for a conversation with the provider, not a quote.

A worked budget

Take a concrete job: an eight-GPU training run that needs two weeks of wall-clock time. At the live median on-demand rate of $3.61 per GPU-hour (GetDeploying, Nov 04, 2026), the GPU line alone is 8 GPUs x 24 hours x 14 days x $3.61, which is $9,703.68 (FlopSource arithmetic from the GetDeploying median, Nov 04, 2026). The same run at the cheapest verified in-stock on-demand rate of $1.79 would be $4,810.08, and at Google Cloud on-demand, billed as a whole eight-GPU VM at $43.04 per hour, it would be $14,458.88 (rates: GetDeploying and FastGPU, Nov 04, 2026; arithmetic by FlopSource). That spread, on identical silicon, is why the variant question and the billing-model question come before the provider question. If the job can checkpoint and restart, a Spot or reserved mix cuts it further; if it cannot, availability and SLA matter more than the last $0.20 per hour. Shortlist providers with our provider checklist, then compare the finalists in the FlopSource Directory. If your workload is drifting toward larger memory, the H100 vs H200 vs MI300X comparison shows when stepping up a tier beats renting more H100s.

Sources

• NVIDIA, "H100 GPU" product specifications (SXM: 80 GB HBM3, 3.35 TB/s, FP8 3,958 TFLOPS with sparsity, NVLink 900 GB/s; NVL: 94 GB HBM3, 3.9 TB/s, FP8 3,341 TFLOPS, NVLink 600 GB/s) · accessed Nov 04, 2026 · nvidia.com/en-in/data-center/h100
• RunPod, "NVIDIA H100 on Runpod: price, specs and SXM vs PCIe" (PCIe variant: 80 GB HBM2e, about 2 TB/s, FP8 3,026 TFLOPS with sparsity, 350 W, no NVLink; SXM 700 W) · accessed Nov 04, 2026 · runpod.io/articles/guides/nvidia-h100
• GetDeploying, "Nvidia H100 Rental Prices: Compare 62 Providers" (median on-demand $3.61 across 46 providers, cheapest in-stock on-demand $1.79 at Lium, reserved median $3.23, spot median $2.39, monthly estimate $2,599 at the median, median up about 8% in 90 days and 13% in 12 months) · Nov 04, 2026 · getdeploying.com/gpus/nvidia-h100
• FastGPU, "Google Cloud GPU Pricing in 2026: On-Demand, Spot and DWS" (Google Cloud H100 $5.38 on demand, $1.09 Spot, cheapest region; market floor $1.79 at Vast.ai; H100 sold only as whole eight-GPU VMs on demand, $43.04 per hour minimum; Spot VMs can be stopped at any time) · Nov 04, 2026 · fastgpu.co/blog/gcp-gpu-pricing-2026
• TokenWatt, "H100 PCIe vs SXM vs NVL: Which H100 to Rent for LLMs" (measured serving cost per million output tokens: SXM $0.31, PCIe $0.43, NVL $0.57, Qwen3-8B BF16 on vLLM at RunPod Secure Cloud prices on the day; NVL cards found power-capped at 310 W against a 350-400 W configurable range) · Oct 2026 · tokenwatt.io/blog/h100-sxm-vs-pcie-vs-nvl-llm-inference.html
• FlopSource Directory dataset, H100 provider records (29 providers offering H100; 28 with representative rates, median $2.25, low Vast.ai $1.53, high Microsoft Azure $12.29; Lambda Labs $2.39, CoreWeave $2.65, RunPod $3.19, Google Cloud $3.95, AWS $4.10) · records last verified Jun 08, 2026 · directory.flopsource.com