Rentals · H100 pricing · with video · Updated October 2026
The Same H100 Costs $1.49 or $7 an Hour
Same chip, wildly different bills. An H100 rents for anywhere from about $1.49 to $7 an hour depending on who you rent it from. Here's what that means for real jobs, using our own measured H100 speeds, and the hidden costs that decide which provider is actually cheaper.
Watch: The Same H100 Costs $1.49 or $7 an Hour
One chip, eleven prices. When I lined up H100 listings across providers for the video, the cheapest was about $1.49 an hour and the most expensive about $7. Same chip. That's up to 4.7 times more for the exact same hardware, depending on where you click 'rent'.
The hourly number is the one everyone compares, and it's the wrong one. What matters is what the whole job costs, and that depends on speed, on the extras each platform charges, and on whether the machine actually works when you need it.
What real jobs cost on an H100 at each price
| Where | Rate | 1,000 FLUX.1-schnell images | 1M tokens, one user | 1M tokens, 32 users batched | 8 hrs a day for a month |
|---|---|---|---|---|---|
| Low end of the market (from the video) | $1.49/hr | $0.43 | $1.69 | $0.11 | $357.60 |
| Vast.ai, H100 SXM | $2.14/hr | $0.61 | $2.43 | $0.16 | $512.64 |
| RunPod, H100 SXM | $2.69/hr | $0.77 | $3.06 | $0.21 | $645.60 |
| Big-cloud on-demand (from the video) | $7.00/hr | $2.00 | $7.96 | $0.54 | $1,680.00 |
Our measured H100 SXM speeds: FLUX.1-schnell at 58.36 images/min, Qwen3 8B at 244.22 tok/s for one user and Qwen2.5 7B at 3,601 tok/s with 32 requests in flight. Vast.ai and RunPod rates are the cheapest we tracked on October 5, 2026; the $1.49 and $7 rows are the extremes from the video. Bandwidth, storage and startup time are not included.
Vast.ai: cheapest on paper. Vast.ai is a marketplace: most of the GPUs belong to independent hosts, and each host sets its own price for bandwidth, storage and CPU on top of the hourly rate. For text generation and data work, where you move very little data, it's the best deal going. A million tokens for one user costs $2.43 at today's cheapest Vast H100.
Once you start generating high-res images and video, the bandwidth bill a host can set starts to matter. I've personally burned more credits on supposedly cheaper Vast GPUs than on RunPod, between bandwidth charges and hosts whose internet dropped halfway through downloading a 50GB model. You pay by the hour whether the download is moving or not.
RunPod and Modal: you pay for reliability. RunPod sets one price per card and runs the hardware itself, so the bill is predictable, and the bandwidth is good enough that big model downloads don't eat your hours. Modal lets you download models on a cheap CPU machine and only then attach a GPU, which is how I run most of the GPU Battle benchmarks. For heavy image and video batches, that predictability is often cheaper in the end than the lowest sticker price.
The big clouds, AWS and Azure, are the expensive end. They make sense for companies that need specific regions, compliance or long-term reserved contracts. For one person running batch jobs, there's no reason to pay $7 an hour for a chip you can rent for a third of that.
When a smaller card is the better buy. For a model that fits in 32GB, you don't need an H100 at all. An RTX 5090 runs Qwen3 8B at 243.91 tok/s on our bench, about the same as the H100, and rents for $0.39 an hour. That's $0.44 per million tokens against $2.43 on the cheapest H100. On Qwen3 32B it's $1.52 vs $8.01.
The H100 earns its price when you need what consumer cards don't have: 80GB of memory, data-center bandwidth and hosts that stay up. Check whether a listing is PCIe or SXM before comparing prices, since they are different cards, and price the job, not the hour.
Our verdict
The same H100 rents for $1.49 to about $7 an hour, but the hourly rate isn't the bill. Vast.ai is cheapest for text and data work, RunPod is more predictable for heavy image and video batches, and the big clouds only make sense for compliance or reserved contracts. Batching cuts the cost per token far more than switching providers, and for models that fit in 32GB a rented RTX 5090 at $0.39/hr beats any H100 per token.
FAQ
- How much does it cost to rent an H100?
- On October 5, 2026 the cheapest H100 SXM we tracked was $2.14/hr on Vast.ai and $2.69/hr on RunPod. Across providers the range runs from about $1.49 to $7 an hour.
- Why is Vast.ai cheaper than RunPod?
- Vast.ai is a marketplace where independent hosts rent out their own GPUs and set their own prices, including bandwidth and storage. RunPod runs its hardware at a fixed price. Vast is usually cheaper per hour; for bandwidth-heavy jobs the extras can close the gap.
- How much does it cost to generate 1,000 images on an H100?
- With FLUX.1-schnell at our measured 58.36 images/min, 1,000 images take about 17 minutes: $0.43 at $1.49/hr, $0.61 at today's cheapest Vast.ai rate and $2.00 at $7/hr.
- Is an H100 worth it for running a small LLM?
- Usually not. On our bench an RTX 5090 runs Qwen3 8B at 243.91 tok/s, about the same as an H100, for $0.39/hr instead of $2.14/hr. The H100 is worth it when you need its 80GB of memory or are serving many users at once.
How we test
Job costs are hourly rate x hours of work, where hours come from our own measured H100 SXM speeds: FLUX.1-schnell (images/min), Qwen3 8B single-stream (tok/s) and Qwen2.5 7B served with 32 concurrent requests via vLLM (tok/s). Vast.ai and RunPod rates are the cheapest we tracked on October 5, 2026 (the rentals page has live numbers); the $1.49 and $7 extremes are the listings compared in the GPU Battle video 'The Same H100 Costs $1.49 or $7 an Hour'. Bandwidth, storage and startup time vary by provider and are not included.