The same AI chip costs $2.69 or $12.29 an hour. Price the task instead

Rent one Nvidia H100 GPU for a month and the bill can be $1,964 or $8,972. Same chip, same hours. The only difference is where you rent it. That is the finding of a price comparison published by Data Receipts, which normalised on-demand prices across AWS, Azure and smaller GPU clouds.

The short version

• The same H100 GPU costs from $2.69 to $12.29 an hour on demand: a 4.57× gap, according to the analysis.
• Part of the gap buys real things: fast networking for multi-server training, service guarantees and compliance. Part of it you may not need.
• Most businesses should not rent GPUs at all. Ask instead what each AI task costs to run, and who carries that cost.

What the numbers say

The analysis compares identical GPUs only, the H100 SXM with 80 GB of memory, at each provider's cheapest route. RunPod Community lists $2.69 per GPU-hour, RunPod Secure $3.99, AWS $6.88 and Azure $12.29. Over a 730-hour month, that is $1,964 at the bottom and $8,972 at the top.

The author is careful about what the gap means. Azure's high-end option includes very fast InfiniBand networking between GPUs, which matters when you train across many servers. RunPod Community is third-party hardware without a data-centre service guarantee. The figures are on-demand list prices in one region each; reserved and spot pricing would narrow the gap. And a cheap GPU you cannot actually get is not a price.

The lesson is not "rent the cheap one"

If you fine-tune a model on a single machine, you may be paying for networking you never use. If you train across 64 machines, the cheap option is not an option. The right price depends on the job.

For most companies, the bigger lesson sits one level up: you probably don't need a GPU of your own. Customer service assistants, document processing and voice agents usually run on AI services billed per use, where the provider carries the hardware.

Price the task, not the hardware

Here is an illustrative comparison for a support line. Say you are weighing a self-hosted model on one rented H100 against an AI service billed per use, for 15,000 calls a month.

Keeping that GPU running all month costs between about $1,964 and $8,972 at the prices above, before engineering time, monitoring and updates, and whether you take 1,500 calls or 15,000. With per-use billing, the cost follows your volume: a quiet month costs less, a busy month costs more, and nothing runs at night for nobody.

Self-hosting starts to make sense with large, steady volume, strict data-location rules or a team that already runs this kind of infrastructure. Below that, cost per handled task is the number to compare, not cost per GPU-hour.

Questions to ask any AI supplier

• What does one task cost at our volume: one answer, one document, one call?
• What happens to cost at twice the volume, and at half?
• Where does our data go, and in which country is it processed?
• Can we switch models or providers without rebuilding everything?

That last question is how we design AI voice agents with Holol-Call, SmartHolol's AI-powered inbound and outbound call center solution: the conversation is about cost per call handled and data location, not about which GPU sits underneath. You can see the wider approach on our AI agents page.

A 4.57× gap is a warning, not a bargain

When the same hardware can cost four and a half times more depending on the label, the useful habit is simple: price every AI project per task, and make the supplier show you the maths.

Sources

• Data Receipts: 109 — What an H100 actually costs per hour (Hacker News discussion)