Compute Exchange; Silicon Data

Compute Exchange connects enterprises buying GPU resources with infrastructure providers and hardware sellers. Carmen Li is CEO of both Compute Exchange and Silicon Data, two distinct businesses serving AI compute markets: Silicon Data develops data, benchmarks and derivatives, while Compute Exchange operates the marketplace for transactions. Compute Exchange’s co-founders include Don Wilson and Suna Said.

Compute Exchange’s Hardware Market covers used and refurbished GPUs, while Reserved GPU Rental aggregates committed cloud capacity. Its Forward Contracts let buyers agree on pricing and delivery for future compute needs, such as scheduled training runs or inference deployments. Buyers specify their requirements, compare provider quotes and service terms, and contract directly with their chosen provider. Compute Exchange facilitates procurement and does not itself operate the compute infrastructure.

Token Forwards let businesses commit in advance to inference volumes for open-weight models over terms of up to six months. Compute Exchange normalizes provider quotes for comparison; buyers lock a committed rate and draw against their balance as demand arrives. Settlement schedules and unused-balance rollover terms vary by provider. GPU rental customers run their own models and manage the infrastructure, while token-forward providers operate the infrastructure and deliver the inference.

Silicon Data’s Silicon Indices provide daily benchmarks for GPU rental prices. Its methodology converts rental records to an hourly rate per GPU, adjusts for rental type, geography, CPU platform and GPU variant, and combines provider-level prices in a weighted average. Separate H100 and A100 hyperscaler readings distinguish those markets from the neocloud and GPU-marketplace benchmarks. These adjustments let buyers compare costs without mistaking a change in the mix of rental configurations for a price movement. The company’s May 2025 product announcement also describes SiliconMark, which measures GPU performance degradation through workload benchmarking and supplies maintenance insights for GPU fleets.