The Compute Tightness Index is 54.0/100 as of 2026-09-27, remaining in the Balanced regime after dropping 3.7 pts over the past month as AI infrastructure capacity loosens. Public rental listings show H100, H200 and B200 spot GPUs pricing at 26% to 41% below on-demand, a spread that reflects slack and pulls the blended gauge down from the acute-tightness highs of its 28–80 range since 2025-09-01. It is the input-side companion to the Token Demand Index (the output — tokens consumed).
← looser · 45 balanced · 60 tight · tighter →
As of 2026-09-28 the blended index sits at 55.1 (Balanced) and has been loosening over the past 30 days (-1.6 pts). Listing-weighted across H100, H200 and B200 by on-demand market depth.
⤓ Download full daily series (CSV)| GPU | On-demand $/hr | Spot $/hr | Spot discount | Listings | CTI | 30d |
|---|---|---|---|---|---|---|
| B200 | $6.49 | $4.78 | 26% | 65 | 68.0 | ▲ +4.5 |
| H200 | $4.50 | $2.62 | 42% | 121 | 51.7 | ▼ -14.9 |
| H100 | $3.38 | $2.00 | 41% | 218 | 53.1 | ▲ +4 |
Spot discount = how far below on-demand the spot rate trades. A shrinking discount (rising CTI) means spare capacity is drying up; a widening discount (falling CTI) means idle GPUs are being dumped on spot.
Independent reads of the same market, not in the blended score. Datacenter clouds: provider list prices and whether each listed region can be provisioned right now, from a multi-cloud catalog (12% of all listed regions across every GPU type were provisionable at the snapshot, so a GPU's deviation from that norm is the signal, not the absolute share). Marketplace: a large GPU marketplace's lowest on-demand price and its own stock grade, for its datacenter-hosted ("secure") tier and its individually-hosted ("community") tier. As of 2026-09-28.
| GPU | Datacenter clouds | Marketplace · secure | Marketplace · community | |||||
|---|---|---|---|---|---|---|---|---|
| Median list $/hr | p25–p75 | Provisionable / listed | vs catalog | $/hr | Stock | $/hr | Stock | |
| B200 | $6.52 | $6.33–$6.76 | 0 / 13 | +12 pts | $6.79 | Low | — | — |
| H200 | $4.47 | $4.00–$4.52 | 0 / 24 | +12 pts | $4.59 | Low | $3.59 | Low |
| H100 | $2.60 | $2.36–$3.59 | 4 / 65 | +5 pts | $3.49 | Low | — | — |
For each GPU on each day: CTI = 100 × ( 0.8 · spot_avg/on_demand_avg + 0.2 · unavailable/(unavailable + on_demand_listings) ). The scarcity term is the share of listed capacity that has gone dark (clamped to 1). The blended index is the listing-weighted mean across B200, H200 and H100 (weighted by on-demand listing count, a proxy for market depth). Inputs are a daily aggregate of public GPU rental listings across cloud providers — on-demand, reserved and spot medians plus availability — one snapshot per day.
Reading it: <45 loose · 45–60 balanced · 60–80 tight · >80 acute scarcity. The spot-premium term dominates day to day; the availability term only bites when providers actually go dark, so it spikes the index during genuine capacity crunches. Young series (data since 2025-09-01); thresholds will calibrate as history accrues. Full daily history is available as CSV.
This is the input side of the AI-compute economy. For the output side, see the Token Demand Index (tokens consumed + the effective price of inference). Want the underlying rows or to query from an agent? Grab the CSV, or use the Bargo MCP endpoint (get_spot_pricing). See also the research notes and Congress trades.