Wednesday at 10 p.m., I nearly burned through my experiment budget
When I was first building an AI tool, I got stuck on a pretty basic question: the same H100 GPU could cost more than four times as much per hour depending on the platform. I was afraid of picking the wrong one, so I reviewed the pricing pages of the usual cloud services three times. The more I checked, the more anxious I got—the GPU model looked identical, but the bill after clicking the button was a different story.
This isn’t only a technical buying decision. People running side projects, working independently, or leading teams of one to five often test AI features on client projects. A small price difference is tolerable. If the runtime estimate is wrong, though, dozens of dollars can disappear. For us, understanding the real cost matters more than chasing the “strongest GPU.”
It’s the same hardware, with very different prices
In the GPU pricing data collected by Data Receipts, hourly H100 prices ranged from about $2.69 to $12.29. The difference usually comes from more than the provider: cloud plans, minimum rental durations, whether the instance is dedicated, region, storage and network fees, and the platform’s discount structure.
A friend of mine, Lin Yue, summarizes voice recordings for clients. In March 2025, she ran the same batch of recordings in two stages in Shanghai: first, she used a cheaper instance for formatting; then, she switched to a pricier one for the critical task. Afterward, she told me, “I didn’t used to be bad at choosing. I just assumed every H100 was the same.” Her case shows what a pricing table is actually useful for. It doesn’t hand you the one right answer—it shows which step belongs on the cheap instance and which one shouldn’t be cut.
I’ve made the opposite mistake too: I saw the lowest price, started a long-running instance immediately, and before the job had finished, the cost was already close to the expensive option. Since then, I run a five-minute sample first, estimate from the measured runtime, and the comparison feels much more natural.
Reproducing this pricing table takes neither much money nor much time
Reproduction cost: the basic version is $0. If I want less hassle, I can subscribe to the repository’s data feed for about $10 per month, or use the public sample for free. Time: the first price comparison can be finished in 15 minutes; if I want to record my actual billing results, I’ll allow another 30–60 minutes for a test run. Technical barrier: I only need to understand the hourly rate and estimated task duration—no server configuration skills required. First step: open Data Receipts’ GPU Pricing page, find H100, put the hourly prices from different platforms into one table, then enter my estimated runtime.
I calculate three numbers first: what the full task costs on the cheap instance, what it costs on the expensive one, and what each costs after storage, downloads, and traffic are included. If the difference is under $10, I don’t fuss over it and choose whichever option is easier. If the gap is meaningful, I split the test run from the production run. If I’m not ready to test now, that’s fine too. When I already have a reliable provider, logging actual runtimes for a week is often more worthwhile than switching platforms.
Choose for the stage I’m at—no need to do everything at once
If I’m just getting started: I’d use public pricing tables to set a budget instead of rushing into a long-term plan. I’d pick one small task, confirm the output quality, and then decide whether spending five minutes on an expensive instance is worthwhile. I’d cap the initial budget at $20 so “just a quick test” doesn’t become a recurring bill.
If I already have one or two clients: I’d split each client project into a test run and a production run, recording the actual duration of each stage. Lin Yue’s approach works well: steps that tolerate a little error go on the cheaper instance, while the final result moves to a more stable, reasonably priced option. The goal isn’t the lowest possible rate; it’s less time lost to repeated rework.
If I’m scaling up: I’d compare more than the hourly GPU price. I’d factor in the total cost of batch concurrency, queue times, storage and traffic, and how easily I can exit. Saving $2 while a job waits two days can hurt client delivery. Once usage grows, it’s worth having a finance or technical partner review the table with me and negotiate enterprise discounts.
Not everyone needs this tool. If I’m only running an occasional small test, sticking with my current platform may be easier. But if I’m already paying repeated “tuition fees” for cloud services, I can take three real tasks tonight and calculate their costs. That usually shows whether I’m overpaying.