H100 Rental Price per Hour: $2.69 to $12.29, Same Chip
The same NVIDIA H100 rents for anywhere from $2.69 to $12.29 per hour depending on the provider, a 4.6x spread. Over a 730-hour month that is about $1,964 versus $8,972 for identical silicon.

Last updated: November 2026
Key takeaways
- An open dataset of GPU rental prices reports H100 pricing from $2.69 to $12.29 per hour. The chip is the same, the price is not.
- Over a month that spread is about $7,008 per GPU, or about $84,000 a year.
- Cost per million tokens depends on price and utilization, so a cheap listing at low utilization can lose to a pricier one kept busy.
- The lazy choice is the expensive one: the biggest brand name is rarely the cheapest rate.
How much does an H100 cost per hour?
According to the open data-receipts GPU pricing dataset, discussed on Hacker News, the same H100 costs $2.69 per hour at one provider and $12.29 at another. Rates differ by provider type, region, commitment term and whether you accept interruptions. Calcaas tracks GPU cloud providers like Lambda and Nebius so you can compare current rates, but verify every quote on the provider's own page before signing.
What does the spread cost over a month and a year?
| | $2.69 per hour | $12.29 per hour | Difference | |---|---|---|---| | Per day (24h) | $64.56 | $294.96 | $230.40 | | Per month (730h) | $1,963.70 | $8,971.70 | $7,008.00 | | Per year (8,760h) | $23,564.40 | $107,660.40 | $84,096.00 |
If you run eight GPUs, the yearly gap is about $672,768 for the same hardware. That is a hiring plan.
What does it mean per million tokens?
Hourly price only becomes a unit cost when you add throughput. Say one H100 serves 1,000 tokens per second at full load, which is 3.6M tokens per hour. This throughput is hypothetical and varies a lot by model and batching.
| Utilization | Cost per 1M tokens at $2.69/h | Cost per 1M tokens at $12.29/h | |---|---|---| | 100% | $0.75 | $3.41 | | 50% | $1.49 | $6.83 |
Halving utilization doubles your unit cost, which is why a cheap-but-idle GPU can lose to an expensive busy one. We explain the full conversion in GPU hour price vs real LLM inference cost.
Why do the same GPUs have such different prices?
Common drivers include on-demand versus reserved terms, spot versus guaranteed capacity, region, bundled networking and storage, and minimum commitments. Our post on GPU cloud rate cards, reserved vs on-demand and utilization goes through each one, and the Oracle Cloud H100, H200 and B200 pricing breakdown shows what a single hyperscaler lists.
How should you shop for H100 hours?
- 1Get three quotes with the same term, region and interconnect.
- 2Convert each to cost per 1M tokens using your own measured throughput.
- 3Add storage, egress and support to the hourly price.
- 4Ask what happens at 50% utilization, because that is your real life.
- 5Compare against a per-token API before you commit: the GPU hour vs inference cost post has the break-even.
Source: the open data-receipts GPU pricing dataset and its Hacker News discussion (171 points, 87 comments). The month, year and per-token examples are our own calculations.
Takeaway: the cheapest H100 is not a hardware decision, it is a procurement decision. Turn an hourly rate into a per-token cost with the Calcaas LLM cost calculator.
Frequently asked questions
How much does an H100 cost per hour?
An open GPU pricing dataset reports H100 rental prices from $2.69 to $12.29 per hour depending on the provider.
How much does an H100 cost per month?
Over a 730-hour month, $2.69 per hour is about $1,964 and $12.29 per hour is about $8,972, a difference of about $7,008 per GPU.
Why is the same H100 so much cheaper at some providers?
Prices vary with commitment terms, spot or on-demand capacity, region, bundled services and provider type, even though the chip is identical.
How do I convert an H100 hourly price to cost per million tokens?
Divide the hourly price by tokens served per hour in millions. At a hypothetical 3.6M tokens per hour, $2.69 per hour is about $0.75 per million tokens at full utilization. Place the JSON-LD below in a <script type="application/ld+json"> tag in the page head.
More from the blog
LLM EconomicsOpenRouter vs Direct API Cost: The Real Math
A gateway like OpenRouter pays for itself when its fee is smaller than the engineering time and downtime it replaces. At a hypothetical 5% fee, a $10,000 monthly token bill adds $500 a month, so the question is whether failover, routing and one invoice are worth that to you.
LLM EconomicsGPT-6.1 Terra vs Gemini 3.6 Flash: Which to Route Where
Route by cost per task, not price per token. In an illustrative 80/20 split between a cheap and a premium model, a 50,000-conversation bot costs $236 a month instead of $750 on the premium model alone, about 68% less.
LLM EconomicsHugging Face Inference Providers Pricing: Open vs Closed Cost
Hugging Face adds no markup on Inference Providers requests, so you pay the provider's price. Open-weight models average $0.28 input and $1.62 output per 1M tokens, versus $1.27 and $6.36 for closed models, about 4x less.
Pricing math, in your inbox.
One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.