Which cloud provider is cheapest for an H100?
We track 15 providers publishing a price for H100 80G: the median is ¥18.83/GPU/hour, the low is ¥11.09 (jarvislabs), the high is ¥32.40 (nebius) — a 2.9x spread. Half of all providers sit between ¥16.20 and ¥23.04; start there, then look at the tails.
Prices: provider-published list prices via rentgpu.link (CC BY 4.0) plus operator self-reported quotes · amounts in CNY as published (no FX conversion) · as of 2026-11-05 · same caliber as /providers and /v1/market/models
The three cheapest and the three dearest
| # | Provider | Price / GPU / h | vs median | Billing | Region | Last |
|---|---|---|---|---|---|---|
| 1 | jarvislabs | ¥11.09 | -41% | spot | 亚太 | 2026-11-02 |
| 2 | ppio | ¥11.88 | -37% | spot | 中国 | 2026-11-02 |
| 3 | latitude | ¥12.10 | -36% | on-demand | 北美 | 2026-11-02 |
| ⋯ | ||||||
| 14 | crusoe | ¥28.08 | +49% | on-demand | 北美 | 2026-11-02 |
| 15 | lambda | ¥29.45 | +56% | on-demand | 北美 | 2026-11-02 |
| 16 | nebius | ¥32.40 | +72% | on-demand | 欧洲 | 2026-11-02 |
Check the billing mode: 2 of the 3 lowest quotes are spot instances, interruptible at any time. Spot and on-demand prices should not share the same "cheapest" label — which is exactly why this page gives the full table instead of a single name.
A ⚑ means the quote sits outside 0.5–2x of the median, usually a unit or SKU mapping error (a whole-node price read as per-GPU is the most common). We do not delete these, we flag them — "cheapest" is the phrase most likely to be quoted wholesale, and the one most easily carried off by an upstream mistake.
All 16 priced quotes
| # | Provider | Price / GPU / h | vs median | Billing | Region | Last |
|---|---|---|---|---|---|---|
| 1 | jarvislabs | ¥11.09 | -41% | spot | 亚太 | 2026-11-02 |
| 2 | ppio | ¥11.88 | -37% | spot | 中国 | 2026-11-02 |
| 3 | latitude | ¥12.10 | -36% | on-demand | 北美 | 2026-11-02 |
| 4 | runpod | ¥14.33 | -24% | spot | 北美 | 2026-11-02 |
| 5 | tensordock | ¥16.20 | -14% | on-demand | 北美 | 2026-11-01 |
| 6 | vast | ¥16.82 | -11% | spot | 北美 | 2026-11-05 |
| 7 | verda | ¥17.43 | -7% | spot | 欧洲 | 2026-11-02 |
| 8 | hyperstack | ¥18.00 | -4% | on-demand | 欧洲 | 2026-11-02 |
| 9 | massed | ¥19.66 | +4% | on-demand | 北美 | 2026-11-02 |
| 10 | gcore | ¥21.29 | +13% | on-demand | 欧洲 | 2026-11-02 |
| 11 | gongji | ¥22.50 | +19% | on-demand | 中国 | 2026-11-02 |
| 12 | thunder | ¥23.04 | +22% | on-demand | 北美 | 2026-11-02 |
| 13 | jarvislabs | ¥25.13 | +33% | on-demand | 亚太 | 2026-11-02 |
| 14 | crusoe | ¥28.08 | +49% | on-demand | 北美 | 2026-11-02 |
| 15 | lambda | ¥29.45 | +56% | on-demand | 北美 | 2026-11-02 |
| 16 | nebius | ¥32.40 | +72% | on-demand | 欧洲 | 2026-11-02 |
The full table is here rather than a top three because "who is fourth" is the question people actually ask next — and because only a complete table deserves to be called citable. Another 28 providers are tracked as quoting H100 without a public price (quote-only) and are not listed.
Three sources of cheap — and what each costs you
| Type | Why cheaper | What it costs you |
|---|---|---|
| Idle-capacity markets | Consumer cards on home bandwidth, supply floats with the market and sellers go offline at will | Reclaimable at any time; no SLA; multi-GPU interconnect and egress are unpredictable |
| Regional pools | Lower power and bandwidth costs, heavy competition, no cross-border links | Slow from abroad; compliance and invoicing need separate confirmation |
| Owned-fleet clouds | Volume purchasing and own data centres — pricier but stable | Higher unit price; usually requires committed spend |
What this answers, and what it does not
② Cannot: be the basis of an order. Quotes exclude storage, egress, redundancy and SLA, and do not guarantee current stock.
③ Cannot: be added to device-level listings (per-machine snapshots on vast.ai / Clore.ai) — different subject.
④ Cannot: treat billing modes as one market. On-demand, spot and reserved are supposed to differ.