Teardowns
A full rack, walked from the cabinet inward, with every line mapped to a part category and the companies reported on it. Use these as worked examples of the catalog assembled into one machine — the same part and company pages you'd reach from Parts orCompanies, just read in the order a technician would open the rack rather than the order the taxonomy lists them.
Three different bets on how to wire an AI cluster together: NVL72 is Nvidia's own proprietary, 72-GPU rack; Helios is an open-standard reference design (OCP Open Rack Wide, UALink, Ultra Ethernet) that multiple system builders implement independently; TPU7x isn't GPU-based at all — it's Google's custom ASIC, scaled to 9,216 chips with a reconfigurable optical circuit switch in place of an electrical one. Company placements marked with an asterisk on any page come from partner lists or named trade/analyst reporting, not a company's own disclosure — see each teardown's own note for its exact sourcing. See the specs side by side on thecomparison page.
Nvidia's liquid-cooled, 72-GPU rack — the reference design most of the current AI build-out is standardized on.
AMD's open, OCP-standard answer to NVL72 — 72 MI455X GPUs on a double-wide Open Rack Wide chassis, built by AMD, HPE, Celestica and others rather than one vendor alone.
Google's custom-ASIC alternative to a GPU rack entirely — 9,216 chips per superpod, wired with a reconfigurable optical circuit switch instead of NVSwitch- or UALink-style electrical switches.