GPU marketplaces
Vast.ai, SF Compute, the GPU exchanges and the decentralised compute networks. 28 providers on EveryNeocloud so far.
GPU marketplaces match buyers of compute with many sellers instead of owning the hardware themselves. Some, such as Vast.ai and SF Compute, let data centres and owners list spare GPUs at prices that change with demand. Others are decentralised networks, such as Akash, Render and io.net, that pay suppliers in crypto tokens. Routers for model APIs, such as OpenRouter, apply the same idea to inference.
How the market works
GPU time is perishable: an idle hour cannot be sold later. Marketplaces let owners sell spare hours and let buyers bid for them, often at a discount to fixed prices. AWS introduced spot instances, spare capacity sold at a floating price that can be taken back at short notice, in 2009. Rackspace runs an auction for servers and GPUs called Rackspace Spot. SF Compute sells short, reserved blocks of large clusters that buyers can resell.
Trade-offs
Marketplace GPUs can be much cheaper, but quality varies. Machines may sit in different places with slow links between them, so they suit single-server jobs, tuning and inference better than large training runs. Decentralised networks add questions about security and about whether token rewards, rather than real demand, drive supply. Several firms now publish GPU price indices drawn from these markets, giving the industry a view of rental prices.
Famous marketplace providers
Further reading
Photo credits on this page
The small pictures above come from Wikimedia Commons under free licences. Some are cropped to fit.
- OpenRouter: Wikideas1, CC0, source
- Amazon Elastic Compute Cloud: Raysonho @ Open Grid Scheduler / Grid Engine, CC BY-SA 3.0, source