Compute Exchange has launched a secondary marketplace for Nvidia data-center GPUs on Friday, expanding beyond reserved compute capacity as organizations seek lower-cost hardware for production AI workloads.
Launched in January 2025 and co-founded by DRW’s Don Wilson and Suna Said of Woodside AI, Compute Exchange originally focused on reserved GPU capacity across a network of more than 100 providers. It also offers forward contracts that allow buyers to specify a GPU model, quantity, contract duration, start date and target price for future capacity.
The newest product compliments sister company, Silicon Data, a data firm working on real-time GPU pricing. Silicon Data is also helmed by Compute Exchange CEO Carmen Li.
“Not every workload requires the newest generation of GPUs,” Carmen Li, CEO of Compute Exchange and founder of Silicon Data, said in a statement to Blockspace. “For many organizations, used and refurbished hardware offers the fastest and most economical path to expanding AI infrastructure. We see the secondary GPU market as a natural step toward making AI infrastructure more efficient, transparent, and accessible.”
Compute Exchange acts as an intermediary and does not take ownership of the hardware. Buyers or sellers submit their requirements, receive introductions to verified counterparties and complete transactions directly, with the platform supporting procurement and logistics.
Hardware costs
The hardware marketplace connects enterprises, cloud providers, AI startups and infrastructure operators with suppliers of used, refurbished and OEM-surplus equipment. Available categories include Nvidia H100, H200, B200, A100, L40S and V100 GPUs.
Compute Exchange said demand is concentrated around established H100 and A100 systems rather than the latest Blackwell accelerators. Requests have ranged from hundreds to tens of thousands of GPUs, indicating that buyers are considering secondary equipment for deployments beyond small experimental clusters, the firm said in a statement.
Spot prices for computer hardware typically derive from perceived usefulness of the product, in this case the value of the compute hours for large language models, rendering, or other GPU work. Data from DI Metrics shows steady spot prices for most units over the past 6 months, with a noticeable increase in spot price for B200 units.
Used GPUs on the platform are generally sold as-is with supplier documentation, while refurbished units have been inspected, tested and recertified. Compute Exchange said used equipment can sell for as much as 40% to 60% below its original manufacturer suggested retail price, while refurbished hardware usually includes a limited warranty and commands higher prices.
