Research
The measurement problem
A lender financing a power plant can read a meter. A lender financing a data center cannot. There is no standardized way to measure whether GPU infrastructure is performing as underwritten.
Different hardware, different workloads, different facilities — and no common unit to compare them. Lenders price this uncertainty into spreads. Borrowers pay for it in cost of capital.
Our approach
The Standard Compute Unit normalizes heterogeneous GPU fleets into a single, comparable metric. It accounts for hardware differences, workload mix, multi-GPU scaling overhead, and energy draw — so a lender can assess collateral performance the same way across any facility.
The specification is formally verified — machine-checked, not self-assessed. Calibration is anchored to industry-standard benchmarks governed by MLCommons, not proprietary data.
Published work
arXiv:2406.19261
Commodification of Compute
Symmetric Research · June 2024
Mathematical foundations for standardized compute measurement.
Read the paper →