Graph written 2026-09-03
Supply Chain
Supplier–customer relationships between named firms, and the layers with only one supplier.
Supplier–customer relationships between named firms, and the layers with only one supplier.
megawatts operating, of 7,875 recorded across 10 fleets; the rest is announced. CoreWeave has the largest built fleet at 703 MW.
facilities in which a rostered neocloud is the recorded AI-hardware owner, across 4 countries. 32 state a capacity. The remainder are drawn nowhere on the axis, because a missing figure is not a zero.
H100-equivalent accelerators, summed over the 32 facilities that state a count. That denominator is the figure’s own coverage limit: the total is a floor.
The neocloud fleet. One bar per company, length in megawatts of recorded data-center capacity, segmented by build status: solid where a site is recorded as operational, half-toned under construction, and outlined where it is announced. The role drawn is AI-hardware ownership, Epoch AI’s field for who owns the accelerators inside a data center, which Epoch states is not necessarily the facility’s owner or operator. For a neocloud that is the right role, since the business is owning hardware inside somebody else’s building: a bar here is a fleet of machines, whatever building houses them. Announced capacity is drawn as an outline and never summed into a delivered figure. Facilities that state no capacity appear nowhere on the axis and are counted separately in the row, because a missing megawatt figure is not a zero. The accelerator column has its own denominator for the same reason.
AI-focused clouds promise to widen compute access beyond the hyperscalers. Across 10 of them the substrate records 40 facilities in 4 countries where the company owns the AI hardware, and 1,280 of 7,875 recorded megawatts are operating today. The gap between those two numbers is the neocloud question: capacity announced at this scale is a claim on future supply, and the upstream chokepoints on Chokepoints constrain those claims.
Whether neoclouds diversify supply or inherit it is answered by their supplier sets: if every one of them buys from the same short list, a single upstream disruption reaches all of them at once. Scrutica records 57 relationship rows involving these companies, and 16 of them (28 per cent) are publishable. The other 41 arrive from a licensed database held under subscription and not redistributed.
A shared-supplier figure drawn from the publishable remnant alone would measure what happens to be publishable, so none is drawn. The counts above are corpus metadata, which is licensing-safe; the withheld rows inform the aggregate denominators on the dependency graph and are published nowhere. The public subset is enumerated below.
Bilateral relationships between these companies and their counterparties, on the public-provenance subset only. Read it as a sample of a larger graph; the ratio above is the sampling fraction.