Best GPU rental without KYC: lab tests and reviews

A flawless record and the widest catalogue

GPURentHub leads the ranking with 9.0/10. It completed 20/20 jobs at 100% of reference performance, with instances ready in 40 s. It is the only service where the lab launched a twelve-hour job before sleeping and did not think about it again. Across about 4,000 GPUs in 15 models and 11 countries, there was never a shortage on the four test models.

No KYC was requested, even for H100s, and no "business verification" appeared. On every disk the lab searched for a previous client’s files and found nothing. Support was human and round-the-clock. The H100 is the priciest of the podium at $2.40, billed by the second; top-ups take BTC, ETH, USDT and USDC, with no Lightning and no Monero.

The most private setup: your SSH key is your account

GPURento takes second place with 8.9/10, completing 19/20 jobs at 99%, launched in 90 s. Its no-KYC is the most radical here: your SSH public key is your account, with no email and no password, and support tickets are signed with the same key. Disks are encrypted and the datacentres sit in privacy-minded countries, including Iceland and Switzerland.

It is the only podium service to accept Monero, alongside Lightning (credited in about one second) and BTC. The H100 is $2.10, billed by the minute. The catalogue is the limit — about 800 GPUs in 6 models across 4 countries — and the single failure was a maintenance window announced only on the dashboard, since the site holds no email to reach us.

The biggest, cheapest marketplace

BriefGPU scores 8.9/10: a marketplace where thousands of hosts rent out their GPUs, so it has the biggest catalogue and the lowest prices — about 12,000 GPUs in 40 models across 60 countries, with the H100 at $1.49, billed by the second. It completed 18/20 jobs at a median 94% of reference performance (81% to 100% depending on the host), launched in 240 s because the host first downloads your image.

No KYC was requested; an email and a Lightning (or ten other coins) top-up are enough. The two failures came from unverified hosts whose machine rebooted mid-job; with the "verified hosts" filter, none recurred. The lab never found a previous client’s files, but isolation depends on each host, and support is slower (about twelve hours).

Fast, but long jobs do not survive

RunPod scores 7.2/10. It never asked for an identity and ran at 98% when a job ran, launched in 70 s. But its "dedicated, guaranteed uninterrupted" instances are in fact preemptible, cut with 30 seconds’ notice whenever someone pays more. The result depends entirely on job length: every one-hour job finished, but none of the twelve-hour jobs did — the longest lasted five hours and forty minutes.

Started hours are billed in full, interruptions included, so the work actually done came to about $2.91 per H100-hour — dearer than GPURentHub. Only 14/20 of the twenty jobs met our full-completion bar. The fault is reliability, not privacy.

"Persistent" storage that was not

Vast.ai scores 6.7/10. It asked for no identity and measured 97% when data survived, launched in 150 s. But the "persistent storage included" vanished on reboot five times out of nine; a maintenance reboot an hour from the end wiped the checkpoints of two twelve-hour jobs, and seven jobs had to be redone from scratch. Only 13/20 of twenty met our bar.

That storage is billed even when the instance is stopped, at a rate absent from the price grid but present on the invoice. Storage that clears on reboot is temporary storage; calling it "persistent" is mislabelling. Again, the fault is reliability, not privacy.

Cheapest per hour, throttled in practice

Supabase scores 6.2/10 and is the cheapest per hour at $1.39 for an H100 — but the benchmark fell to 55% of reference. nvidia-smi explained it: the advertised "H100 SXM" are H100 PCIe, further capped at 250 W, and the A100s and part of the L40S were throttled too. Only the RTX 5090s ran at full speed, so just 7/20 of twenty jobs cleared the 90% bar.

The real price is the price of the work done: at $1.39 and 55% performance, an hour of full-speed H100 compute costs about $2.53 once the work is done — well above BriefGPU, though still below RunPod, whose interrupted jobs bring it to about $2.91. No identity was demanded; the fault is throttled hardware, not privacy.