Availability & Utilization
Two lenses on GPU supply. Availability — of the genuine-capacity venues we probe, how many show launchable stock (X of Y), tracked over time so the slope reads demand tightening or easing. Utilization — of the capacity the one marketplace we can measure actually lists, how much is rented.
Availability over time
Rolling 7-day breadth — the share of genuine-capacity venues (X of Y) showing stock; falling = venues selling out. Dashed verticals mark a change in the venue panel. Cluster (8×) vs Fractional.
The panel
Each genuine-capacity source, one vote · 7-day availability
Excludes catalog-eligibility (Azure) and probabilistic-score (AWS) sources from the headline. Resellers that duplicate a directly-probed operator are deduped. Sources below the minimum observed-cycle threshold are held out until they accumulate probes.
Measured utilization one marketplace source
Vast.ai — the only marketplace that discloses rented / total. Rolling 7-day occupancy, denominator fixed to each window's median listed capacity.
Methodology
Each source's window rate is its share of capacity-probes that found stock, count-weighted within source. The headline is breadth — of the Y genuine-capacity venues that disclosed a signal, the share X/Y that found any stock — shown as Fractional (any shape) and Cluster (full 8-GPU node). Breadth, not a median, because availability is bimodal (venues clump dry or flush) and a median collapses to 0 or 1 on a bare majority; its slope over time is the demand read. Each rolling point uses the venues present in that window, and dashed verticals flag where the panel set changed, so added venues don't read as a demand move. Utilization is a separate measured depth metric — never blended into availability.