# Compute Tray The tray is the unit a technician actually handles. It slides out of the rack carrying processors, accelerators, memory, its share of the cooling loop, and its power connection, and it slides back in connected. In a rack-scale system the tray replaced the server as the thing that is bought, installed, serviced, and replaced, and that substitution is one of the quiet structural changes of this hardware generation. > [!note] The one-sentence version > Not a server in a box, but a drawer of accelerators that arrives plumbed and wired and leaves the same way. ## Why it exists A conventional server is a self-contained computer: its own chassis, power supplies, fans, and network connections, able to work alone on a shelf. That independence costs volume, power conversion, and air paths at exactly the moment when volume, power, and cooling have become the binding constraints. Disaggregating the server pushes power conversion into a shared shelf, cooling into a shared loop, and switching into shared trays, leaving the compute tray to carry only what computes. The result packs more accelerators into a rack and removes several hundred discrete power supplies and several thousand fans from it. ## What a tray contains Processors and accelerators on a shared board, coupled over a coherent interconnect; see [[wiki/Host CPU|Host CPU]]. System memory. [[wiki/Cold Plate|Cold plates]] over every high-power device, with internal tubing terminating in blind-mate quick disconnects at the rear. A power connection that mates to the rack busbar, taking direct current rather than converting its own; see [[wiki/High-Voltage Direct Current Distribution|High-Voltage Direct Current Distribution]]. Accelerator fabric connections to the switch trays. Local flash for the operating environment. And a [[wiki/Baseboard Management Controller|management controller]] with its own network path, so the tray can be inspected and controlled whether or not its operating system is healthy. A current rack-scale design places four accelerators and one processor on a tray and installs eighteen such trays in a rack alongside six switch trays, reaching seventy-two accelerators and eighteen processors in one frame. ## Blind-mate service, and why it matters The engineering achievement in a modern tray is that liquid cooling did not cost serviceability. Fluid, power, and data connections are made by pushing the tray home: self-sealing couplings mate without spilling, the power contacts engage the busbar, and the fabric connectors seat. Removal reverses it, with no draining, bleeding, or hose work. A technician services a liquid-cooled tray with roughly the same motions used for an air-cooled server, which is what allowed liquid cooling to be adopted at fleet scale rather than remaining a specialist practice. Trays arrive from integration already plumbed, wired, and tested; see [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]]. The practical consequence is that the liquid path has become part of the computer rather than part of the building, which is a deeper change than it first appears: the hardware vendor now owns the thermal design down to the channel geometry, and the facility supplies water at a temperature and flow rather than cooling the equipment itself. ## Relationships - **contains:** [[wiki/Host CPU|Host CPU]] · [[wiki/Accelerator|Accelerator]] · [[wiki/Cold Plate|Cold Plate]] · [[wiki/Baseboard Management Controller|Baseboard Management Controller]] - **installed in:** [[wiki/Data Center Rack|Data Center Rack]] · [[wiki/Rack-Scale Computing|Rack-Scale Computing]] - **powered by:** [[wiki/High-Voltage Direct Current Distribution|High-Voltage Direct Current Distribution]] · [[wiki/Busbar Busway|Busbar Busway]] - **cooled by:** [[wiki/Direct-to-Chip Liquid Cooling|Direct-to-Chip Liquid Cooling]] · [[wiki/Coolant Distribution Unit|Coolant Distribution Unit]] - **standardized through:** [[wiki/OCP Accelerator Module|OCP Accelerator Module]] - **assembled by:** [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]] · [[wiki/Server OEM and System Integration|Server OEM and System Integration]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** tray disaggregation moving power conversion, cooling, and switching to shared rack resources; blind-mate self-sealing fluid, power, and data connections permitting service without draining; factory-integrated plumbing and wiring; management controllers with independent network paths. - **Established:** a current rack-scale design with four accelerators and one processor per tray, eighteen compute trays and six switch trays per rack, totalling seventy-two accelerators and eighteen processors. - **Strongly indicated:** blind-mate service as the enabling condition for fleet-scale liquid cooling adoption; vendor and operator practice supports it and comparative service-time data is not public. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; tray and switch counts, accelerator and processor counts per rack, four accelerators under copper cold plates per processor. - Open Compute Project rack and tray specifications, as reviewed October 2026.