# Server OEM and System Integration Accelerator designers do not, for the most part, build servers. A distinct industry does: designing boards, assembling systems, building racks, validating them, and shipping them to facilities. Understanding who performs that work explains a good deal about how quickly capacity can be delivered, and why accelerator vendors have been buying into it. > [!note] The one-sentence version > Somebody has to turn chips into racks at volume, and that somebody is a manufacturing business with its own expertise and its own capacity limits. ## Who does what **Original equipment manufacturers** design and sell complete systems under their own brand, with their own support, firmware, management tooling, and service organization. Enterprises and many specialist operators buy this way, because the support relationship and the serviceable, documented platform have real value. **Original design manufacturers** design and build to a customer's specification, shipping systems the customer brands or deploys directly. The largest operators work this way, specifying systems to their own requirements, often to open hardware specifications, and buying at a price that reflects the absence of a brand and a support organization; see [[wiki/OCP Accelerator Module|OCP Accelerator Module]]. **Integrators** assemble, cable, cool, validate, and commission racks and clusters from components, whether those came from an equipment manufacturer or a design manufacturer; see [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]]. ## Why the accelerator vendors bought in Rack-scale systems changed the economics of this boundary. When the product was a card, a vendor could sell silicon and let others build systems. When the product is a seventy-two-accelerator liquid-cooled rack weighing seven thousand pounds, the system is the product, and its thermal, power, mechanical, and firmware design cannot be delegated without delegating the performance. One accelerator vendor accordingly acquired a systems manufacturer in 2025 for approximately four point nine billion dollars, having identified that systems capability, and the company acquired had already built the systems on which some large models were trained. The same logic produced the other vendor's move to selling complete racks. That shift matters beyond corporate strategy. It means delivery of compute capacity now depends on manufacturing and integration throughput rather than on chip supply alone, and integration capacity is itself something a country either has or imports; see [[wiki/Industrial Competence|Industrial Competence]] and [[wiki/Supply Chain Chokepoint|Supply Chain Chokepoint]]. ## Relationships - **builds:** [[wiki/Compute and Accelerator Systems|Compute and Accelerator Systems]] · [[wiki/Compute Tray|Compute Tray]] · [[wiki/Data Center Rack|Data Center Rack]] - **performs:** [[wiki/Rack and Cluster Integration|Rack and Cluster Integration]] - **builds to:** [[wiki/OCP Accelerator Module|OCP Accelerator Module]] - **serves:** [[wiki/Neocloud|Neocloud]] · [[wiki/Hyperscaler|Hyperscaler]] · [[wiki/Colocation|Colocation]] - **civic dimension:** [[wiki/Industrial Competence|Industrial Competence]] - **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 distinction between original equipment manufacturers selling branded supported systems, original design manufacturers building to customer specification, and integrators assembling and validating racks and clusters; the shift of the product unit from card to rack and the resulting inseparability of system design from performance. - **Established:** an accelerator vendor's 2025 acquisition of a systems manufacturer for approximately four point nine billion dollars, with that manufacturer having previously built systems used to train large models. - **Strongly indicated:** integration and manufacturing throughput rather than chip supply alone governing compute delivery; the structural logic is clear and comparative throughput data is not public. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; the 2025 systems-manufacturer acquisition, its value, and its prior role building systems on which large models were trained.