The Reference Rate for AI Compute
 
Early access · Invitation only
Live Market Data

AI Compute
Reference Rates

Transparent, independently calculated reference rates for GPU compute. Used by enterprises to benchmark spot pricing and anchor procurement negotiations.

ICX H100S
H100 SXM5 · NVLink · 80GB
$
Live
ICX H100P
H100 PCIe · 80GB
$
Live
ICX A100
A100 SXM4 · 80GB
Coming Soon
ICX H200
H200 SXM5 · 141GB
Coming Soon
ICX L40S
L40S · 48GB
Coming Soon
ICX H100 SXM5 Reference Rate
$
per GPU-hr · Settlement · Live
24h Change
7-Day Range
Providers
ICX H100 SXM5 — 90-Day Rate History
$3.40$3.00$2.60$2.20$1.80
Current ICX Rate
7-Day Low
7-Day High
Panel Providers
Programmatic Access to ICX Rates
Embed ICX reference rates directly into your procurement systems, trading infrastructure, or financial models. REST API with JSON responses, available on request.
REST API

Current Rates

Query the live ICX rate for any tracked GPU type. Returns rate, panel size, and computation timestamp.

Historical

Rate History

Access the full time-series of ICX rates. 90-day default window, extendable for subscribers.

Webhooks

Rate Alerts

Subscribe to threshold alerts — receive a callback when ICX crosses a configured level. Coming soon.

View API Documentation →
How ICX Calculates Each Rate
Transparent. Repeatable. Independent of hardware cost models. Each index uses the same methodology applied to its own provider panel.
01

Collect

Automated scrapers run hourly, fetching executable on-demand pricing from a diversified panel of US-based cloud providers. Multi-GPU clusters normalized to per-GPU equivalent.

02

Normalize

All prices are converted to $/SCU-hour (Standard Compute Unit) using the ICX SCU formula: SCU = CNF × P-tier × U × Q. The A100 80GB is the baseline (1.0×). H100 SXM5 = 3.3× (locked, MLPerf-validated). P-tier captures interconnect quality. U and Q adjust for availability and service reliability.

03

Filter & Clean

Only SCU v1 eligible providers enter the panel: AWS, Azure, GCP, CoreWeave, Atlas Cloud. Consumer/community-grade providers (RunPod, Vast.ai, TensorDock, Lambda Labs) are excluded per P-05 eligibility findings. Minimum panel enforced — rate not published on thin data. Bottom and top 15% of NDPs trimmed.

04

Publish

Rates updated hourly from live scrapes. Each computation uses only currently-available providers, so the rate reflects executable spot prices — not stale listings. Each GPU type is stored as an independent time series.

05

Price ≠ Cost

ICX reflects transaction prices, not hardware ASP or production expense. Commodity benchmarks measure markets, not drilling costs. The rate anchors procurement — it is not a cost model.

06

Governance

Published methodology. Documented update process. Minimum data requirements enforced in code. Variant separation documented and applied consistently. The remedy for data risk is governance, not complexity.

Benchmark First. Then Everything Else.
Derivatives cannot precede credibility. ICX follows the established path: measure → adopt → structure → financialize.
Phase 1 — Now

Measurement

Publish consistent hourly reference rates across GPU variants. Establish volatility patterns, seasonal shifts, and supply tightness in neutral instruments.

Adoption

Enterprises benchmark vendor quotes to ICX. Procurement teams embed ICX in forecasts. The signal: "ICX plus spread" enters the market vocabulary.

Structure

Term expectations, forward curves, and realized volatility measures derived from observed spot data. Analytic layers, not separate businesses.

Financialization

Bilateral swaps. Cash-settled futures. Only after the benchmark is sufficiently trusted and liquid. Trust is the prerequisite, not the product.

ICX measures the market
so the market can mature.

Get the daily rate delivered to your inbox. Access the API. Reference ICX in your next vendor negotiation.