# OCP Accelerator Module An accelerator module standard specifies the physical and electrical envelope an accelerator must fit: its dimensions, its mounting, its power delivery, its cooling interface, and its connector pinout. With such a standard, a system builder designs one baseboard and several vendors' accelerators drop into it. Without one, every accelerator requires a bespoke system, and the system becomes an extension of one vendor's product line. > [!note] The one-sentence version > Agree on the socket, and the market gets competition; keep the socket private, and the market gets a supplier. ## Why open hardware standards matter here The accelerator business has a structural tendency toward lock-in, because a device's value depends on the fabric, software, and system around it. Open specifications counteract that by defining the interfaces at which substitution becomes possible. The Open Compute Project, a consortium through which operators and manufacturers publish hardware designs as open specifications, has become the principal venue, and its work now covers the module, the baseboard, the rack, the power shelf and busbar, and the liquid-cooling interfaces. Current rack-scale systems are explicitly built to the consortium's wide-rack specification, with four-accelerator trays conforming to it. The commercial logic deserves to be stated plainly, because it is the reason a buyer should care about a mechanical drawing. A published socket lets a second vendor build a compatible part; a second vendor creates price competition and supply resilience; and supply resilience is strategically meaningful when the alternative is a single source with a committed order book. For a nation building compute capacity at scale, open hardware standards are an industrial policy instrument that costs nothing and requires no legislation. ## What the specifications cover - **Mechanical envelope:** module dimensions, mounting pattern, and keep-out zones. - **Power delivery:** input voltage and current, now extending to the several-hundred-volt direct-current rack bus; see [[wiki/High-Voltage Direct Current Distribution|High-Voltage Direct Current Distribution]]. - **Thermal interface:** cold plate mounting, clamping load, and coolant connection geometry, which is what permits a third party to build the plate; see [[wiki/Cold Plate|Cold Plate]]. - **Host and fabric connectors:** the pinout carrying host attachment and accelerator-to-accelerator fabric links; see [[wiki/PCIe|PCIe]] and [[wiki/UALink|UALink]]. - **Baseboard topology:** how modules on one board interconnect, which determines the scale-up domain a tray provides. - **Management:** the interfaces through which the module is inventoried, monitored, and updated; see [[wiki/Baseboard Management Controller|Baseboard Management Controller]]. ## Where openness stops A standard socket does not produce a standard ecosystem. Software remains the decisive differentiator: a module that fits the baseboard still requires compilers, kernels, collective libraries, and framework support before anyone can use it productively; see [[wiki/CUDA|CUDA]] and [[wiki/ROCm|ROCm]]. Scale-up fabrics have likewise been proprietary at the leading edge, which is the gap open consortium efforts aim at; see [[wiki/UALink|UALink]] and [[wiki/Ultra Ethernet|Ultra Ethernet]]. The honest summary is that hardware interfaces have been substantially opened while software and fabric remain the contested ground, and that the contest is being run through consortia with real membership and shipping products. ## Relationships - **standardizes:** [[wiki/Accelerator|Accelerator]] · [[wiki/Compute Tray|Compute Tray]] · [[wiki/Data Center Rack|Data Center Rack]] - **specifies interfaces to:** [[wiki/Cold Plate|Cold Plate]] · [[wiki/High-Voltage Direct Current Distribution|High-Voltage Direct Current Distribution]] · [[wiki/PCIe|PCIe]] - **complemented by:** [[wiki/UALink|UALink]] · [[wiki/Ultra Ethernet|Ultra Ethernet]] - **limited by:** [[wiki/CUDA|CUDA]] · [[wiki/ROCm|ROCm]] software ecosystems - **strategic dimension:** [[wiki/Supply Chain Chokepoint|Supply Chain Chokepoint]] · [[wiki/National Compute Capacity|National Compute Capacity]] - **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:** the Open Compute Project as the consortium publishing open hardware specifications covering accelerator modules, baseboards, racks, power shelves, busbars, and liquid-cooling interfaces; current rack-scale systems built to the consortium's wide-rack specification with four-accelerator trays. - **Established:** software ecosystems and scale-up fabrics remaining the principal differentiators despite open mechanical and electrical interfaces. - **Strongly indicated:** open socket standards broadening accelerator supply and improving buyer position; the mechanism is clear and measured market effects are still accruing. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; Open Compute Project Open Rack Wide alignment and four-accelerator tray design. - Open Compute Project published specifications, as reviewed October 2026.