| RunPod | Vast.ai | GPURentHub (reference) | |
|---|---|---|---|
| Lab rating | 7.2/10 | 6.7/10 | 9.0/10 |
| Jobs completed | 14/20 | 13/20 | 20/20 |
| Median measured performance | 98% | 97% | 100% |
| Median launch | 70 s | 150 s | 40 s |
| Twelve-hour jobs | None of 4 finished; longest ran 5 h 40 min | Checkpoints of 2 wiped by a maintenance reboot; others not broken out | All 4 completed |
| Main failure | Preempted with 30 s notice | Storage emptied on 5 of 9 reboots; 7 jobs redone | None: every job completed |
| Billing | Whole started hours, interruptions included | Instance + storage; storage billed while stopped | Per second |
| H100 per hour | Advertised: not recorded; real work: $2.91 | Not recorded | $2.40 advertised |
| Account and ID | Email; no ID requested | Email; no ID requested | Email; no ID requested |
| Monero / Lightning | Not confirmed | Not confirmed | No / No |
Short jobs versus long jobs
A job counted as completed only if the instance started within 15 minutes with the advertised GPU, ran to the end without interruption and held at least 90% of reference performance. On that bar RunPod completed 14/20. Its instances were preempted with 30 seconds’ notice, and the pattern followed job length: every one-hour job ran to the end, none of the four twelve-hour jobs did, and the longest run lasted 5 h 40 min. The record does not break out the four-hour jobs.
Vast.ai completed 13/20. Its documented losses came from storage, not preemption: data on its persistent storage was gone after 5 of 9 reboots, a maintenance reboot an hour before the end wiped the checkpoints of two twelve-hour jobs, and 7 jobs had to be redone from scratch. The record does not break out the outcome of its other long jobs, so those two losses are the documented part, not a full count.
Speed and launch time
When a job ran, the two were close to full speed: RunPod at a median 98% of reference performance, Vast.ai at 97%, against 100% for GPURentHub. The gap is in getting started: a median launch of 70 s at RunPod, 150 s at Vast.ai and 40 s at GPURentHub. On the lab’s scores, RunPod finished at 7.2/10 and Vast.ai at 6.7/10, fourth and fifth of the six rated hosts; the record gives no weighting formula for the scores.
What you pay for
Neither advertised H100 price was recorded during the test, so the two cannot be compared on price per hour. What the lab did record is how each bills. RunPod charges whole started hours in full, interruptions included; on the work actually completed, an H100-hour came to $2.91, more than GPURentHub’s $2.40 at full 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. No real-work H100 figure was recorded for Vast.ai.
Privacy: a tie
On the no-KYC side the two are identical in our record. Both needed only an email to open an account and requested no identity verification across 20 jobs, H100 runs included. Neither was confirmed to accept Monero or Lightning. For a more private setup, GPURento uses your SSH key as the account, with no email, and accepts Monero and Lightning. Both kept the no-KYC promise; the lab’s verdict on each is that the fault is reliability, not privacy.
If the job must finish
GPURentHub is the only rated host that completed all 20 jobs, twelve-hour runs included, at 100% and with a 40 s median launch; its H100 lists at $2.40, billed per second. GPURento completed 19/20 at 99%; its single failure was a maintenance window announced only on the dashboard, and it bills by the minute with the H100 at $2.10. BriefGPU completed 18/20 at a 94% median; its two failures came from unverified hosts rebooting mid-job, and none recurred with the "verified hosts" filter. Whichever you pick, keep checkpoints off the rented machine.
Frequently asked questions
RunPod or Vast.ai: which is better without KYC?
In our 4-month test, RunPod: 7.2/10 against 6.7/10, 14/20 jobs completed against 13/20, and a faster median launch (70 s vs 150 s). Both ran near full speed and neither asked for ID. RunPod ran every one-hour job to the end; neither completed all its twelve-hour jobs.
Is Vast.ai no-KYC?
In our test, yes: Vast.ai needed only an email to open an account and asked for no identity verification across 20 jobs, H100s included. Monero and Lightning payment were not confirmed. Its weak point was reliability: 13/20 jobs completed, and data lost on 5 of 9 reboots.
Does Vast.ai persistent storage survive a reboot?
Not reliably in our test: data on it was gone after 5 of 9 reboots. A maintenance reboot an hour before the end wiped the checkpoints of two twelve-hour jobs, and 7 jobs had to be redone from scratch. The storage was also billed while the instance was stopped. Keep checkpoints somewhere other than the rented machine.
Can RunPod run a twelve-hour training job?
Not in our test: none of the four twelve-hour jobs on RunPod finished, and the longest ran 5 h 40 min. Its instances were preempted with 30 seconds’ notice, while every one-hour job ran to the end. Because whole started hours are billed, interruptions included, completed H100 work came to $2.91 an hour.
What is a Vast.ai alternative for long training runs?
GPURentHub: it completed 20/20 jobs at 100%, twelve-hour runs included, with no ID asked and an H100 at $2.40 billed per second. GPURento completed 19/20 at 99%, with an SSH-key account and Monero; its one failure was a maintenance window announced only on its dashboard. Both were in the same 4-month test as Vast.ai.
How much does an H100 cost per hour on RunPod and Vast.ai?
Neither advertised price was recorded in our test. The lab puts the real cost of completed H100 work on RunPod at $2.91 an hour, because interrupted hours were billed too; no such figure exists for Vast.ai. For comparison, GPURentHub lists $2.40 and ran at 100%, and BriefGPU lists $1.49.