How far apart are GPU prices across providers?

Top 14 models by priced-offer count. The provider is shown next to the price; where the provider count reads "18/25", the denominator is every provider we track quoting that model and the numerator is those that actually publish a price. Spread = highest / lowest. The same model can differ several-fold between providers, which is why "what does an H100 cost per hour" has no single answer — only "which provider, which form factor, which billing mode".

"No public price" is not a data gap — those providers are quote-only by business model. They are listed because the usual procurement mistake is assuming a provider is expensive when there was never a public price to compare. "Priced" counts platform × model × billing mode, so a model offered both on-demand and spot counts twice and the number can exceed "Models".

One more thing that gets misread: a provider with few models here does not have few machines. This table counts only model-level public quotes; per-machine listings are a separate caliber — 5,618 real machines are currently listed on vast.ai / Clore.ai and are counted per device, not here. vast shows only 4 models precisely because it publishes per-machine offers instead of model quotes. Both numbers are useful; they just cannot be added.

Only models with 2+ priced offers on both sides. A single offer is not a median. Domestic prices typically bundle local bandwidth and compliance; overseas prices usually exclude connectivity back to mainland China — comparing sticker prices alone misses that.