The Same H100 Costs $2.69 or $12.29 an Hour — a 4.6x Gap

Imagine buying the same iPhone in two stores, and in one of them it costs 4.6 times more. With GPU rental, this is not a metaphor. According to a GitHub compilation that earned 110 points and 66 comments on Hacker News, the same NVIDIA H100 is rented for $2.69 per hour by some providers and for $12.29 by others. Same card, same chip, yet the bill differs almost fivefold.

What Is Behind the Numbers

Let's convert hours into money. A month has about 730 hours. At $2.69, one card costs roughly $1,960; at $12.29, it costs $8,970. The difference is about $7,000 per month for a single GPU. For a typical node of eight H100s, the gap comes to around $56,000 every month. For a team that trains a model or runs inference around the clock, this is no rounding error but the salaries of several engineers.

An important caveat: we rely on the extreme ends of the range from the headline, not on a full price list. Which providers sit at each end, and under what terms, you need to check in the repository itself.

Why the Price Swings So Much

An hourly cloud price guarantees almost nothing on its own. The final amount is usually affected by the payment model (on-demand, reserved for months, or spot, which can be taken away), the contents of the package (CPU, memory, disk, network, egress traffic), the format (a single card or a whole node with fast GPU-to-GPU interconnect), reliability with support, and a region with available capacity.

Large clouds often charge a premium for their ecosystem: ready-made services, access management, compliance, familiar billing. Specialized GPU providers and marketplaces are frequently cheaper, but may differ in stability and support quality. So you should compare not the price per card but the price per useful work.

What to Do in Practice

First: count the cost of the task, not the hour. If a cheap card sits idle in a queue or has a slow network, the savings evaporate. Second: run a short benchmark on your own workload with two or three providers, for example a one-hour training run or an inference load test, and divide the costs by the number of tokens or training steps. Third: separate your workloads. Experiments and batch jobs that survive interruption can go to cheap or spot hardware. Production with strict uptime requirements is better kept where you pay for guarantees.

Fourth: plan for migration in advance. Containers, a reproducible environment, and checkpoints in neutral storage turn switching providers from a quarterly project into a matter of a few days.

For a product manager, the takeaway is even simpler: the unit economics of an AI feature depend heavily on where it runs. If the margin holds only at $12 per hour while the market offers $2.69, then either the infrastructure was chosen poorly, or you are consciously paying for something specific. In the latter case, the reason is worth writing down explicitly.

Limitations

GPU prices change quickly, and a snapshot on GitHub is a cross-section at a particular date. A low price is not always available: there are queues, limits, and minimum rental terms. Debates in such HN discussions usually concern exactly these nuances, so check current rates yourself before migrating.

The bottom line: do not accept the default price. An hour of H100 can cost 4.6 times more with the same transistors, and comparing providers remains the cheapest optimization you can make this quarter.