At the Open Compute Project Global Summit in San Jose, Hon Hai Precision Industry put its newest hardware on the floor: a liquid-cooled cabinet holding 72 Nvidia Rubin graphics processors and 36 Vera central processors, with high-speed interconnects, power distribution and cooling plumbed together and delivered as one unit.
The rack, displayed by Foxconn subsidiary Ingrasys, is the group’s clearest statement yet of where it wants to sit in the AI supply chain. Foxconn built a business assembling servers to other companies’ specifications. It now wants to sell the cabinet itself, including the parts inside it and the systems that keep it cool and powered, to buyers building AI data centers by the gigawatt.
Nvidia helped set the stage. The chipmaker said it contributed rack and compute tray designs for its Vera Rubin platform to the Open Compute Project through its MGX architecture, and released an updated 1.2 specification that adds blind-mate liquid cooling connectors, native 19-inch equipment support and rack designs carrying up to 250 kilowatts. Nvidia also put forward a design for a separate cabinet filled with Vera CPUs, aimed at the orchestration and sandboxing work that AI agents generate.
Foxconn’s exhibition followed that outline. Ingrasys showed the Vera Rubin NVL72 platform and said it supports the full MGX-based Vera Rubin line. It displayed a Vera CPU rack compute tray that the company said can carry up to 256 Vera CPUs per cabinet and support 22,500 sandboxes for agentic workloads, alongside a rack built for Groq’s LPX accelerators and an HGX Rubin NVL8 system for model training. Foxconn put the inference efficiency of the new platform at up to 10 times that of Nvidia’s previous Blackwell generation, a figure the company supplied and has not published test conditions for.
FIT, the group’s interconnect unit, filled in the plumbing. Its display included a 480-kilowatt liquid-cooled busbar, high-current AC whips rated to 400 amps, quick disconnects that meet the project’s specifications, a 1.6-terabit optical module line and a denser pluggable optical platform rated at 12.8 terabits, all aimed at the bandwidth, power and heat that a 250-kilowatt cabinet produces. High-speed signaling runs on PAM4 and supports PCIe 6.0 and 7.0.
The strategic shift is in the integration. “The rack is becoming the unit of sale,” an executive at a rival contract manufacturer said, asking not to be named to speak about a competitor. “Whoever assembles it takes the warranty, the thermal risk and the schedule risk. That is a different business from screwing boards into a chassis.”
Margin follows that risk in both directions. Rack-scale integration raises the revenue Foxconn can book per order and pulls in content from its connector, power and cooling businesses. It also puts the company on the hook when a liquid coupling weeps onto a tray of accelerators, and it forces it to coordinate a bill of materials that Nvidia largely defines and a schedule that Nvidia largely sets.
The event drew a crowd of companies making the same move. Wiwynn and Wistron were named among Nvidia’s first pilot build partners for the Vera Rubin NVL72 platform, carrying out engineering validation for the design. Pegatron showed a rack that integrates the same 72 processors and 36 CPUs, and Giga Computing displayed a full platform with figures of its own, including up to 20.7 terabytes of high-bandwidth memory per rack. Taiwanese contract manufacturers are competing for slots in the same build plans.
Nvidia’s decisions at the summit shaped what those companies can differentiate on. By contributing the rack and tray designs to an open project and standardizing the liquid cooling connections, the chipmaker turns the cabinet architecture into a common template. What remains contested is execution: how fast a manufacturer can ramp a liquid-cooled line without defects, how tightly it holds the tolerance of blind-mate connectors that must align sight unseen inside a closed rack, and how much of the hardware it can supply itself.
Foxconn’s answer is vertical integration, from connectors and cold plates up to the finished cabinet. Executives have said automated production lines are part of the pitch, allowing volume manufacturing without a matching rise in defects. The company has also deepened work with Intel on rack-scale systems for agentic AI applications, according to people familiar with the arrangement, an effort that runs parallel to its Nvidia line.
The contract manufacturing business has never been a high-margin one, and the shift to racks does not change the arithmetic of assembling someone else’s design. What it changes is the size of the order and the number of internal suppliers that can claim a piece of it. Foxconn’s bet is that the group’s components business grows faster when the final product is a cabinet the group can price as a whole.
Watch, then, whether the content Foxconn supplies per rack rises in the next generation, and whether a design contributed to an open project ends up making cabinets easier to build. Both would show up in margins rather than in press releases, and neither is visible yet.
