H100-first v2 methodology

How Compute Index turns live H100 prices into a market signal

The primary index is not a generic GPU average. It is a H100-focused market heat score built from live or availability-verified spot prices, provider-level medians, source confidence, and freshness gates.

What the score means

The 0-100 score measures H100 market heat. Lower scores indicate a buyer-friendlier market with softer prices or better supply. Higher scores indicate seller-friendlier conditions with rising prices, tighter supply, wider provider spreads, or weaker source health.

0-25
Buyer's market
26-50
Neutral-buyer
51-75
Neutral-seller
76-100
Seller's market

Component weights

H100 momentum
Price movement in the H100 provider-weighted spot tape.
30%
H100 supply tightness
Live provider and offer depth, stockout pressure, and availability signals.
25%
Provider spread
Dispersion between provider-level H100 medians after normalizing per GPU-hour.
15%
H100 volatility
Short-window movement in the H100 spot price history.
15%
Source health
Source confidence, freshness, and provider data quality.
10%
Stockout pressure
Unavailable or reference-only H100 pressure separated from live market rows.
5%

What enters the spot index

  • GPU class must resolve to H100-family inventory.
  • Price must be normalized to USD per GPU-hour.
  • Offer must be live or availability-verified for the primary spot tape.
  • Rows must pass source-quality gates and freshness windows.
  • Provider medians are computed before the global aggregate to reduce marketplace row-count bias.

What is excluded

  • Hardcoded fallback rows.
  • Unknown or unmatched GPU aliases.
  • Stale rows outside provider freshness windows.
  • Official list prices without verified current capacity.
  • Quote-required or contact-sales rows without public spot availability.

Freshness windows

Freshness depends on the source type. Fast-moving marketplace APIs should refresh much more often than official catalog pages.

Marketplace APIs
stale after roughly 30 minutes
Catalog APIs
stale after roughly 24 hours
Pricing pages
reference after roughly 7 days
Fallback rows
always excluded from market aggregates

Current limitations

The index is optimized for public H100 spot-market visibility, not every private enterprise contract. Some providers expose only official list prices, contact-sales pricing, or limited capacity signals. Those rows are useful context, but they are visually and methodologically separated from the primary live spot index.

ACI AI

Split View Active

Research while you chat.

People Also Ask

Free-- questions left