Louer Nvidia Tesla V100S.

No provider has listed a Nvidia Tesla V100S in the last 24 hours, so there is no current price. Listings for older or less common cards come and go; the price history below shows when it was last available.

With 32 GB it runs 7–8B models at FP16, 13–14B at INT8 and about 32B at 4-bit. A 70B model needs a 48 GB card or two 32 GB cards. It can QLoRA fine-tune 7–14B models.

Marketplaces such as Vast.ai and Clore.ai list machines from independent hosts who set their own prices, so competition pushes rates down but reliability varies by host. Operator clouds such as Lambda and the hyperscalers charge more for consistent hardware, support and networking. Prices here are per GPU, so they compare directly.

Neither. Renting a GPU is on-demand: the rate is per hour the card is allocated to you — whether it is running your job or sitting idle — and most providers bill per second and stop billing the moment you stop the instance. There is no weekly or monthly plan to commit to (the monthly figure we show is simply 730 × the hourly rate), and nothing is metered by usage. Usage-based pricing — per-token rates with limits like TPM or concurrency — is how hosted model APIs charge, which is a different product we track on our AI-model pages, not here.

We collect listings from each provider's public API or price page — every 5–10 minutes for marketplaces, up to daily for operator clouds — and show the median hourly price per provider from the last 24 hours, so one unusually cheap listing doesn't set the headline. Details are on the methodology page.