How far apart can the same GPU be priced across providers?
Across 14 model + VRAM combinations, the median ratio of highest to lowest price is 3.0x; the widest is NVIDIA V100 · 16G at 6.6x (4 quotes). 7 models exceed 3x. Put differently: a price quoted without naming the provider carries an error bar of more than 100%.
Caliber: only model + VRAM combinations with 4+ providers publishing a price (thinner samples are noise) · latest observation per provider per billing mode · as of 2026-11-05
Distribution: 7 of 14 models span more than 3x
| Spread band | Models | Share | |
|---|---|---|---|
| < 1.5× | 1 | 7% | |
| 1.5–2× | 2 | 14% | |
| 2–3× | 4 | 29% | |
| 3–5× | 5 | 36% | |
| ≥ 5× | 2 | 14% |
"Spread" = max ÷ min across all published quotes for that model. It is not evidence of gouging; it measures how many supply tiers exist in that market.
The 12 widest spreads
| # | Model | Providers | Low | Median | High | Spread |
|---|---|---|---|---|---|---|
| 1 | NVIDIA V100 · 16G | 4 | ¥0.86 | ¥1.58 | ¥5.69 | 6.6× |
| 2 | NVIDIA RTX3090 · 24G | 13 | ¥0.65 | ¥1.28 | ¥3.30 | 5.1× |
| 3 | NVIDIA RTX4090 · 24G | 17 | ¥0.87 | ¥1.92 | ¥3.60 | 4.1× |
| 4 | NVIDIA RTX5090 · 32G | 13 | ¥1.53 | ¥3.25 | ¥5.50 | 3.6× |
| 5 | NVIDIA A6000 · 48G | 8 | ¥2.20 | ¥2.93 | ¥7.85 | 3.6× |
| 6 | NVIDIA H20 · 96G | 4 | ¥3.40 | ¥9.00 | ¥12.00 | 3.5× |
| 7 | NVIDIA A100 · 80G | 15 | ¥6.41 | ¥10.67 | ¥20.09 | 3.1× |
| 8 | NVIDIA H100 · 80G | 15 | ¥11.09 | ¥18.83 | ¥32.40 | 2.9× |
| 9 | NVIDIA L40S · 48G | 10 | ¥4.54 | ¥6.95 | ¥12.03 | 2.6× |
| 10 | NVIDIA H200 · 141G | 9 | ¥16.49 | ¥27.39 | ¥38.88 | 2.4× |
| 11 | NVIDIA RTXPRO6000 · 96G | 7 | ¥7.35 | ¥12.98 | ¥15.77 | 2.1× |
| 12 | NVIDIA B200 · 180G | 6 | ¥32.25 | ¥43.13 | ¥61.20 | 1.9× |
Wide spreads concentrate in two groups: consumer cards quoted both by idle-capacity marketplaces and by enterprise clouds (4090 / 3090), and high-end cards heavily resold (H100). The first differ on "is ops included", the second on "who is reselling". The thinnest rows (right at the 4-provider threshold) have their spread amplified by a single quote — treat as a lead, not a conclusion; the provider count column is there so you can judge.
Where the spread comes from, and how to use it
Four sources：(1) different supply structures (owned fleet vs idle capacity vs end-of-life stock); (2) different billing modes (on-demand / spot / reserved); (3) different form factors and interconnect (SXM has NVLink, PCIe does not); (4) resale markup — the same machine can appear at two prices on two platforms.
How to use it：Treat the spread as a map for sourcing and negotiation: first fix the tier you need (SLA or not, NVLink or not, cross-border or not), then look at the spread *within* that tier — it usually collapses from 3.0x to under 1.5x. Whatever remains is what is actually worth negotiating.
Two limits of this measurement
② Caliber limit: these are list prices, not transaction prices, and quotes from different times still carry a time component. Taking the latest observation per provider per billing mode compresses that, but does not remove it.