The Sorso View Compute Index · SVCI
The benchmark for the cost of compute.
Published weekly. The Sorso View Compute Index reads the on-demand public rental rate of AI compute across twenty-seven providers in three tiers, normalized to per-GPU-hour, constructed into one composite value. Every data point sourced. Every weight disclosed.
Issue 004 · Now live
The latest SVCI level is 103.8. The index held steady from Issue 003. Fifteen providers published comparable public rates, twelve were captured and excluded, and the Tier 2 median remained $3.85. Read Issue 004 or browse the archive.
Free reader tools
Get the next SVCI issue when it publishes.
Receive the benchmark instead of hunting for it. The public issue stays open; the subscription delivers the Monday Compute Price Brief, the market reading that explains it, and buyer tools as they launch.
What the Index measures.
The on-demand, publicly-quoted, single-instance rental rate for the H100 SXM 80GB configuration, captured in the pre-publication window across the twenty-seven-provider panel. Normalized to per-GPU-per-hour. Currency is United States dollars; non-USD rates are reported alongside their USD equivalents. Source URL, capture window, and evidence record logged for every price.
This is the leverage layer. Every committed-use contract is negotiated against it. Every procurement memo references it. When the leverage layer moves, every buyer's position shifts in the same direction within the week, regardless of what any specific contract says. The Index reads that signal.
Read the full methodology →
What the Index does not measure.
Private contract pricing, which is confidential, customer-specific, and often materially different from public list. Large committed-use customers may pay materially below list, especially in multi-year capacity agreements. The Index does not see those numbers. Editorial covers them where they enter the public record through filings and disclosures.
Reserved and committed-use pricing. Bundled enterprise pricing combined with storage, networking, and credits. Operators who do not publish on-demand rates, including Nscale, IREN, and Fluidstack, are acknowledged in editorial coverage but excluded from the Index calculation.
The line between what is measured and what is not is published in the same place as the number itself. Stating limits plainly is not a weakness of the methodology. It is the methodology.