Renting a single H100 80GB costs anywhere from $1.85 to $98.32 per GPU-hour depending on the provider - a 53.1× spread for the same silicon. On a week-long fine-tuning job that's the difference between a $311 bill and a $16,518 one.
CloudMart tracks GPU pricing across 12 providers, and every price in this article is live. Here's where H100 rental prices actually stand, who's cheapest at each tier, and which provider fits which workload.
H100 80GB: single-GPU prices
| Provider | On-demand /hr | Spot /hr | Notes |
|---|---|---|---|
| Lium | $1.85 | - | Marketplace, priced per GPU-hour by node, billed by the second |
| Crusoe | $3.90 | - | Flat rate, no interruption risk |
| RunPod | $3.99 | $1.99 | Per-second billing, Secure + Community tiers |
| Lambda Labs | $4.89 | - | ML-focused tooling, flat rate |
| CoreWeave | $6.16 | $2.46 | InfiniBand interconnect, built for clusters |
| Microsoft Azure | $98.32 | $18.17 | ND96isr H100 v5, enterprise integration |
Prices are per GPU-hour, pulled from the CloudMart GPU marketplace each time this page loads and sorted by price - they're current as you read this, not as of publication. Marketplace providers move continuously through the day.
The independents have compressed hard. A year ago the cheap tier sat around $2.70-$3.00; today 1 providers are at or under $3.00 on demand, and the floor is $1.85. The top of the market has not followed - Microsoft Azure is asking 53.1× what the cheapest provider charges for the same card.
Marketplaces vs flat-rate providers
Two of the providers above work differently from the rest, and the distinction matters more than the headline number.
Lium and Vast.ai are marketplaces. You rent a specific node from a specific host, and the price depends which node you get. All Lium prices here are per GPU-hour; their H100s currently run $1.85-$5.00/GPU-hr depending on availability, and the low end is the cheapest node online right now rather than a rate card - it moves through the day as capacity is rented and released. Billing is by the second with no contract or minimum, and they expose a CLI, SDK, API and an MCP server, which makes Lium unusually easy to drive from an agent rather than a dashboard.
Crusoe, Lambda, OVH and CoreWeave are flat-rate. You pay the posted price, the hardware is uniform, and nothing gets reclaimed. Crusoe at $3.90/GPU-hr is the notable one here - flat pricing with no interruption risk at all.
8× H100 nodes (distributed training)
| Provider | 8× H100 /hr | Per GPU | Spot /hr |
|---|---|---|---|
| RunPod | $21.52 | $2.69 | $15.92 |
| Crusoe | $31.20 | $3.90 | - |
| Lambda Labs | $36.72 | $4.59 | - |
| CoreWeave | $49.24 | $6.16 | $19.71 |
| Google Cloud | $88.49 | $11.06 | - |
| AWS (p5) | $98.32 | $12.29 | $57.76 |
Look at the bottom of that table: AWS spot works out to $7.22 per GPU-hour - more than RunPod's on-demand price, and 1.9× Crusoe's. If you are not already locked into AWS, hyperscaler GPU pricing is difficult to justify at any tier.
A caveat on the cheap end: at cluster scale, interconnect matters as much as the hourly rate. CoreWeave's premium buys InfiniBand with near-linear scaling across nodes - for serious distributed pre-training, a cheaper cluster that scales at 70% efficiency can cost more per useful FLOP. For single-node jobs, which covers almost all fine-tuning, the cheap providers are simply cheaper.
Don't need an H100? A100 80GB is the value play
For most fine-tuning and inference workloads, an A100 80GB delivers ~60-70% of H100 training throughput at roughly half the price:
| Provider | A100 80GB /hr | Spot /hr |
|---|---|---|
| Paperspace | $1.15 | - |
| RunPod (SXM) | $1.79 | $1.14 |
| Crusoe | $2.42 | - |
| CoreWeave | $2.70 | $1.21 |
| Lambda Labs | $3.19 | - |
And if you're doing inference on a quantized 7B-13B model, you may not need datacenter GPUs at all: an RTX 4090 24GB runs varies/hr on Vast.ai and $0.89/hr on RunPod. Lium lists RTX PRO 6000 Blackwell - 96GB of VRAM - from $1.49 per GPU-hour, which is a lot of memory for the money if your model is memory-bound rather than compute-bound.
Newer silicon: H200, B200, B300
Blackwell has reached the independents, and it is no longer quote-only. Lium currently lists H200 141GB from $3.80, B200 180GB from $3.70, and B300 288GB from $4.00 - all per GPU-hour, billed by the second. These are marketplace rates and move with availability; the live listing carries current figures.
Whether they're worth it depends entirely on whether you're memory-bound. For a model that spills out of 80GB and forces you into multi-GPU sharding, compare a single 141GB H200 at $3.80 against 2× H100 at $3.70. That's close enough that the simpler single-card setup usually wins on both throughput and engineering time. For anything that already fits on one H100, the extra VRAM is dead weight - buy the cheaper card.
What a real job costs
- LoRA fine-tune, 7B model, ~6 hours on 1× A100 80GB: $7-$11 on Paperspace or RunPod. This is the workload most people actually have.
- Full fine-tune, 13B model, ~3 days on 1× H100: ~$133 on Lium, ~$281 on Crusoe, ~$287 on RunPod on-demand, ~$7,079 on Azure.
- Week of 8× H100 distributed training: ~$3,615 on RunPod, ~$3,615 on RunPod, ~$16,518 on AWS. At this scale, negotiate - listed prices are a starting point.
Which provider should you pick?
- Cheapest H100 on the board right now: Lium at $1.85/GPU-hour. Marketplace pricing, so check what's actually available before you plan around it.
- Cheapest with no surprises: Runcrate at $3.58 flat - no reclaim risk, no node lottery.
- Iterating on fine-tunes: RunPod - per-second billing and pre-built PyTorch templates make start/stop cycles cheap.
- Agent-driven or scripted workloads: Lium's CLI/SDK/MCP surface, or RunPod's API. Both are built to be driven by code rather than a console.
- 8+ GPU distributed training: CoreWeave - the InfiniBand premium pays for itself at scale.
- Already on AWS/GCP/Azure: The integration tax is real. So is a 53× price difference against the cheapest provider. Move the training job out if you possibly can.
Not sure what GPU your job needs?
Tell the Planner your model size and workload - it picks the GPU, provider, and estimates the job cost.
Plan my training run →Disclosure: Lium is a paid partner of CloudMart and paid a one-time fee for inclusion in this article and the GPU marketplace. Their prices come from their public pricing feed, polled every 15 minutes because marketplace prices move through the day; other providers are refreshed twice weekly. Every table in this article is sorted by live price at the moment you load it, so no provider's position - Lium's included - is fixed by hand. No other provider in this article has a commercial relationship with us.
Related: GPU spot pricing explained - when interruptible capacity actually saves money, and LLM API pricing compared if you just want to call a model rather than host one.