Compute Exchange Plans Token Forward Contracts for AI Model Costs

Compute Exchange is preparing token forward contracts that would let enterprises lock in AI inference token prices for up to six months, The Information reported. The plan follows the company's July launch of a marketplace for used and refurbished NVIDIA GPUs, which it announced in a press release.

The token contracts would differ from GPU rental agreements, which usually set prices for compute capacity by the hour. Compute Exchange plans to offer pricing tied to tokens used in AI model inference, with customers selecting models from one of six inference providers signed up with the platform.

Customers would submit requests that include the model they want, target pricing, throughput rates, and other specifications. Providers would compete for the request, and the customer would commit to using the purchased token volume for the contract term. Compute Exchange would charge a 4% fee.

Compute Exchange is also creating standardized token units to compare pricing proposals and model performance. Its hardware marketplace supports requests for used and refurbished GPU infrastructure, including H100 and A100 systems, and uses SiliconMark for independent GPU verification.