The capital expenditure cycle behind generative AI runs through a six-layer industrial supply chain. Each layer below lists the public companies operating in it, with live quotes and the Veridion Score. This is identification — not investment advice.
Each layer enables the layer above. Compute at the apex. Power at the foundation. Every public company actually doing the work.
GPUs, custom accelerators, and the silicon performing the matrix multiplication at the core of every training and inference workload. Designers, not foundries.
Integrators that take silicon and assemble it into rack-scale systems, plus the thermal management and power distribution layer enabling 100+ kW per rack.
High-bandwidth memory (HBM), NAND flash, enterprise SSDs, and storage arrays. Memory bandwidth is the documented bottleneck for large-model training and inference at scale.
Ethernet and InfiniBand switching, 400G/800G optical transceivers, DSPs, and the back-end fabric required to connect thousands of accelerators into a single training cluster.
Specialty GPU cloud operators outside the three hyperscalers. Business model: take delivery of accelerators at scale and lease compute capacity to AI labs, enterprises, and sovereign training programs.
Independent power producers, integrated utilities, gas turbine OEMs, and grid-scale storage. Hyperscaler capacity additions are constrained by interconnect availability and dispatchable generation.
Constituent lists are reviewed quarterly against revenue mix, segment disclosures, and capex commentary. Quotes are sourced from licensed market data; the Veridion Score is computed from six published factors. Inclusion is not a recommendation.