There's no agreed way to value a GPU running inference—and compute futures now settle on these indexes

AccBalanced · x · 2026-09-05

There is no single agreed-upon way to measure the value of a GPU running AI inference, and that's becoming a bigger problem because these measurements are starting to determine the price of compute.

Once a futures contract settles against a compute index, the methodology stops being a technical footnote: which GPUs count, which regions count, rental duration, utilization, and which transactions get filtered out all determine the number money changes hands against.

The market is already splitting around different definitions of compute: SemiAnalysis publishes InferenceMAX and its successor InferenceX, measuring inference performance across hardware, serving stacks, workloads, latency targets and cost assumptions. NVIDIA has worked closely with the project and regularly cites Blackwell's gains on those benchmarks.

Original post →

There's no agreed way to value a GPU running inference—and compute futures now settle on these indexes

More from Venture