The same NVIDIA H100 GPU rents for $1.79 an hour in one place and $11 an hour in another. That spread, roughly sixfold for functionally identical silicon, is the defining fact of the GPU rental market in late 2026. List price is not a property of the chip. It is a property of the provider tier, the commitment term, and the fine print.
This post pins down what H100s actually cost right now, provider by provider, using each vendor's own pricing page where possible, and explains why the price curve bent back upward this year after two years of collapse.

The curve bent back up
From launch in 2022 through 2024, H100 rentals routinely ran $7 to $10 per GPU-hour. By late 2025 the same card rented for $2 to $4 on marketplace and specialist clouds, one of the fastest price corrections in infrastructure history, driven by overbuilding, ramped NVIDIA shipments, and a flood of new entrants (Silicon Data).
Then the decline stopped and reversed. Silicon Data's neocloud H100 rental index went from $2.00 per GPU-hour on 9 December 2025 to $2.20 by early January 2026 (Silicon Data), and printed $2.83 to $2.97 through October and early November 2026 (Silicon Data H100 index). That is roughly a 45% rise from the December 2025 baseline. On the contract side, SemiAnalysis tracked 1-year H100 rental pricing up almost 40%, from $1.70 per GPU-hour in October 2025 to $2.35 by March 2026, and attributes the reversal to inference demand for open-weight models outpacing the expected Blackwell-driven decay of Hopper pricing (SemiAnalysis).
So the 2026 buyer is not shopping a falling market. Capacity that looks cheap against 2024 is expensive against 2025, and waiting for prices to crater again has not been a profitable strategy this year.
Where the market sits today

Aggregate trackers put the median on-demand H100 SXM rate at $3.49 per GPU-hour across roughly 25 to 28 providers, with the cheapest listed rates at $1.79 on Cudo Compute and $1.80 at HyperAI, and spot H100s from about $1.42 (GPU Economy price index, GPU Rental Prices, Deploying).
First-party list prices, checked against each provider's own page in November 2026:
- RunPod (pricing, updated 9 November 2026): H100 PCIe from $1.99/hr on Community Cloud and $2.89/hr on Secure Cloud. H100 SXM from $2.69/hr Community, $3.99/hr Secure. H200 from $3.59/hr Community, $5.29/hr Secure. Billed per second, no minimum commitment.
- Lambda (pricing): H100 SXM at $3.99/GPU-hr on 8-GPU instances, rising to $4.29 on a single-GPU instance; H100 PCIe at $3.29 on 1-GPU instances. B200 from $6.69/GPU-hr. GH200 at $2.29. Billed by the minute, and Lambda advertises no egress fees.
- CoreWeave (pricing): HGX H100 on-demand at $49.24 per 8-GPU node-hour, which normalizes to $6.16 per GPU-hour, with spot at $19.71 per node-hour (about $2.46 per GPU). HGX B200 on-demand at $68.80 per node-hour ($8.60 per GPU), spot $34.11. There is no single-GPU tier; the minimum purchase is the whole node.
- Together AI (pricing): H100 clusters at $3.99/GPU-hr on-demand, with a preemptible tier at $1.99 and H200 at $5.99. Reserved H100 terms drop to $3.69 for 3-30 days, $3.45 for 31-90 days, and $3.19 for 91-180 days.
- Hyperscalers as the baseline: AWS p5.48xlarge normalizes to about $6.88/GPU-hr in us-east-1, Google Cloud a3 nodes to about $9.80, and Azure H100 v5 instances to $12.29 in eastus (Thunder Compute roundup, Economize).
The fine print moves more money than the sticker
List price per GPU-hour is the least important number on these pages for most buyers. Four asterisks do more work:
- Minimum purchase unit. CoreWeave sells H100 capacity only as an 8-GPU node. A job that needs one GPU-hour still buys eight. At $6.16 normalized per GPU, it is mid-market on the sticker and the most expensive option in this list for a single-GPU fine-tune, because you pay for the fabric whether you use it or not.
- Tiered pricing by instance size. Lambda charges $3.99 per GPU on 8x instances and $4.29 on 1x. Same card, 7.5% premium for not wanting eight of them.
- Community versus secure tiers. RunPod's $1.99 H100 PCIe lives on Community Cloud, hosted by third parties with no availability guarantee. The same workload on Secure Cloud, which carries an SLA, costs $2.89. Marketplace capacity is cheap until the host you depend on disappears mid-run.
- Interruption risk as a discount. The cheapest rates anywhere are preemptible or spot: Together's $1.99 H100 and CoreWeave's $2.46-per-GPU node spot exist because someone else's higher-priority job can reclaim the hardware at any moment. Fault-tolerant batch work should use them. Anything holding state should not.
Billing granularity matters too for spiky serving workloads. RunPod and Lambda bill per second and per minute respectively, and Modal's per-second serverless model means an endpoint that runs four hours a day pays for four hours, while an hourly rental bills 24 (Modal and per-second comparison). An always-on H100 at $3.49/hr runs about $2,550 a month idle or not; the same card used at 15% utilization costs a fraction of that on per-second infrastructure.
What to buy for what you're doing
For single-GPU fine-tuning and experiments: RunPod Community or marketplace hosts at $1.80 to $2.69, accepting availability risk, or Lambda H100 PCIe at $3.29 if you want a supportable vendor. The premium for certainty is $0.60 to $1.50 per GPU-hour, which is noise at single-GPU scale.
For multi-node training: CoreWeave and Lambda 1-Click Clusters, where InfiniBand fabric and node-scale availability are what you are actually paying for. If the run survives interruption, spot and preemptible tiers at $1.99 to $2.46 cut the bill in half against on-demand.
For inference: the honest comparison is no longer GPU-hour rental versus rental. Per-token serverless APIs (Together, RunPod Serverless, Modal) absorb utilization risk that you would otherwise carry yourself on an hourly rental, and the rental math only wins once you can keep GPUs meaningfully busy, which most serving workloads below a few thousand requests per minute cannot.
Bottom line
The floor of the market is now genuinely cheap: under $2 per H100-hour if you accept spot semantics or unvetted hosts. The ceiling at the hyperscalers remains about 2.5 to 4 times higher for the same silicon (GPU Economy). Most teams overpay not by choosing the wrong provider but by paying node minimums, secure-tier and always-on rates for workloads that can absorb interruption. And the trend line is the caution: H100 rentals rose roughly 45% from December 2025 to late October 2026 (Silicon Data). Locking a reserved term at $3.19 per GPU-hour is, for now, cheaper than waiting.
Sources
- Silicon Data, H100 Rental Price Over Time (2023-2025): https://www.silicondata.com/blog/h100-rental-price-over-time
- Silicon Data, H100 Price Spike: https://www.silicondata.com/blog/h100-price-spike
- Silicon Data, H100 Rental Price Index: https://www.silicondata.com/products/silicon-index/h100
- SemiAnalysis, The Great GPU Shortage - Rental Capacity: https://newsletter.semianalysis.com/p/the-great-gpu-shortage-rental-capacity
- RunPod GPU Cloud Pricing: https://www.runpod.io/pricing
- Lambda GPU Cloud Pricing: https://lambda.ai/service/gpu-cloud
- CoreWeave Cloud Pricing: https://www.coreweave.com/pricing
- Together AI Pricing: https://www.together.ai/pricing
- GPU Economy, Cloud GPU Price Index: https://gpueconomy.com
- GPU Rental Prices comparison: https://gpurentalprices.com
- Deploying, Cheapest Cloud GPUs: https://getdeploying.com/guides/cheapest-gpu-cloud
- Thunder Compute, NVIDIA H100 Pricing: https://www.thundercompute.com/blog/nvidia-h100-pricing
- Economize, Cloud GPU Pricing 2026: https://www.economize.cloud/resources/gpu/pricing
- Deepak Gupta, Top 7 GPU Cloud and AI Compute Providers 2026: https://guptadeepak.com/tools/top-7-gpu-cloud-ai-compute-providers-2026
