| Host | Advertised H100 / hour | Median performance (all GPU models) | Real cost per H100-hour | Jobs completed | Billing |
|---|---|---|---|---|---|
| Supabase | $1.39 | 55% | $2.53 (lab figure) | 7/20 | Per hour |
| BriefGPU | $1.49 | 94% (81–100% by host) | About $1.59 (our calculation) | 18/20 | Per second |
| GPURento | $2.10 | 99% | About $2.12 (our calculation) | 19/20 | Per minute |
| GPURentHub | $2.40 | 100% | $2.40 (our calculation) | 20/20 | Per second |
| RunPod | Not recorded | 98% | $2.91 (lab figure) | 14/20 | Whole started hours |
| Vast.ai | Not recorded | 97% | Not recorded | 13/20 | Instance + storage |
How the real cost is worked out
A slower GPU does less work per paid hour, so the lab compares hosts on the price of the work, not the advertised hour: the hourly price divided by the measured performance. Two figures in the table are the lab’s own: $2.53 for Supabase and $2.91 for RunPod. The other three are our calculation from the same test data, with two limits. The performance used is each host’s median across the four GPU models tested (RTX 5090, L40S, A100 80 GB, H100 80 GB), because an H100-only figure was not recorded. And no rerun cost is added.
Reruns are left out of our three figures. GPURentHub completed 20/20 jobs, so it had none to redo. GPURento lost one job to a maintenance window announced only on its dashboard, and BriefGPU lost two when unverified hosts rebooted mid-job; the cost of redoing those jobs was not recorded. The lab’s Supabase figure matches its price divided by its 55% median, with nothing added for failed jobs, although only 7/20 cleared the bar. Only GPURentHub has a public pricing page cited in the record, reviewed on September 25, 2026; the other prices are those the lab recorded during the test.
Supabase: the lowest listed price, the slowest cards
Among the prices recorded, Supabase lists the cheapest H100 in the test at $1.39 an hour. On the instances, nvidia-smi showed H100 PCIe cards capped at 250 W where "H100 SXM" was advertised, and the A100s and part of the L40S were throttled as well; only the RTX 5090s ran at full speed. Its median performance across the test was 55%, only 7/20 jobs reached the 90% bar, and it scored 6.2/10, last of the six rated hosts. No identity was requested. Measured on the work done, its $1.39 hour became $2.53.
How to check the GPU you rented
The lab ran the same four checks on every instance. First, nvidia-smi, which shows the exact card and its power limit: that is how the PCIe cards capped at 250 W appeared at Supabase. Second, a five-minute benchmark against the GPU’s reference performance; a job only counted as successful at 90% or more. Third, a search for files left by a previous client; at GPURentHub, GPURento and BriefGPU, none were found. Last, a witness file on the "persistent" storage: at Vast.ai, data on that storage was gone after 5 of 9 reboots.
Run these checks in the first minutes, while stopping still costs little: GPURentHub and BriefGPU bill by the second, GPURento by the minute. At RunPod, which bills whole started hours, the hour you start is charged in full. If the card, the power limit or the benchmark does not match what was advertised, stop the instance before you start the real job, and keep checkpoints somewhere other than the rented machine.
The billing unit changes the bill
GPURentHub and BriefGPU bill by the second, GPURento by the minute and Supabase by the hour. RunPod bills whole started hours, interruptions included, and its instances were preempted with 30 seconds’ notice: none of its four twelve-hour jobs finished, the longest ran 5 h 40 min, and only 14/20 jobs met the bar. That is how its real H100 work reached $2.91 an hour despite a 98% median performance. Vast.ai bills the instance plus storage, and the storage kept being billed while the instance was stopped, at a rate absent from the price grid but present on the invoice.
Frequently asked questions
What is the cheapest H100 per hour without KYC?
By advertised price, Supabase at $1.39 an hour, the lowest of the four H100 prices recorded in our 4-month test. By the cost of the work done, BriefGPU: $1.49 listed and a 94% median performance, about $1.59 per full-speed H100-hour on our calculation, billed per second. Supabase’s 55% median put its real cost at $2.53.
Was Supabase’s "H100 SXM" really an SXM card?
Not on the instances we rented: nvidia-smi showed H100 PCIe cards capped at 250 W, and Supabase’s median performance across the test was 55% of reference. Only 7/20 jobs reached the 90% bar; only its RTX 5090s ran at full speed.
Why did H100 work on RunPod cost $2.91 an hour?
Because RunPod bills whole started hours, interruptions included, and its instances were preempted with 30 seconds’ notice. None of our four twelve-hour jobs there finished; the longest ran 5 h 40 min. Its advertised H100 price was not recorded, so $2.91 is the lab’s figure for work actually completed, the highest real-cost figure recorded in the test.
Does BriefGPU’s performance depend on the host?
Yes. BriefGPU is a marketplace where thousands of hosts rent out their GPUs, and measured performance ranged from 81% to 100% depending on the host, for a 94% median. Its two failed jobs came from unverified hosts whose machines rebooted mid-job; with the "verified hosts" filter, none recurred. Launch is slower, 240 s median, because the host first downloads your image.
How can I check that a rented GPU is really an H100?
Run nvidia-smi as soon as the instance starts: it shows the exact model and the power limit. Then run a short benchmark and compare it with the card’s reference performance; our test counted a job as successful only at 90% or more. That is how the lab found PCIe cards capped at 250 W where Supabase advertised "H100 SXM".
Do no-KYC GPU hosts ask for ID to rent an H100?
Not the six hosts rated in our 4-month test: none requested identity verification, even for H100s held at 100% for hours. Modal did, outside the ranking: during its first twelve-hour H100 job it suspended the account for "suspected mining", asked for an identity document, a selfie and a project description, and held $85.00.