Vast.ai

Vast.ai is a GPU rental marketplace with market-set hourly prices across 20,000+ GPUs in 40+ data centers. Users deploy instances, serverless endpoints or clusters from a console, CLI, Python SDK or REST API, with per-second billing and a $5 minimum start.

Vast.ai is a GPU rental marketplace and cloud platform for teams that need compute for AI and machine learning work. Instead of a fixed price list, rates are set by supply and demand across a pool of more than 20,000 GPUs in over 40 data centers, covering 68+ GPU types. Consumer cards such as the RTX 3090, 4090 and 5090 sit alongside data center hardware like the H100, H200 and B200, and the homepage shows live "from" and median hourly prices that refresh hourly. The workflow is built around code. A user adds credit (the minimum is $5), takes an API key from the console, searches offers by model, VRAM, price and availability, and launches an instance. The same steps can be done in the web console, with a command-line tool, with the vastai Python SDK, or through a REST API. A sample on the homepage searches for an eight-GPU H100 SXM offer and launches a vLLM container image on it. Billing is per second, and the site positions the API as the way automated agents can procure and tune their own compute. There are three deployment modes. GPU Cloud gives on-demand instances with full control. Serverless turns models into endpoints with automatic benchmarking across GPU types and scaling to zero, so charges accrue only for compute time. Clusters offer dedicated multi-node setups with InfiniBand networking for large training runs. A model library supplies pre-configured templates for popular open-source models, and use-case pages cover fine-tuning, image and video generation, transcription, rendering, batch processing and GPU programming. Vast.ai also has a hosting side, where owners of hardware and data centers can list capacity and estimate earnings. The company states it is SOC 2 certified and lists customers such as Bosch, IBM, Brave and Inria. Compared with hyperscaler GPU instances, the marketplace model tends to suit cost-sensitive developers, researchers and startups who are comfortable choosing hardware themselves, while managed AI platforms suit teams that want fewer infrastructure decisions.