Mechanical engineer John Norton ran through the details of the architecture, which Nvidia contributed to the Open Compute Project last year. He said that the cunning plan was to stop having to endlessly redesigning systems for picky hyperscaler customers who wanted slightly different mixes of CPUs, NICs or management gear.

The solution was MGX, a Lego-like set of building blocks with shared specs and interfaces. Customers can now tweak what they want without Nvidia having to rip up the entire design.

The racks themselves are meaty. Each GB200 system carries 300 chips across compute and switch trays, stitched together with NVLink running at 200 Gb/s per lane. One rack chews through 120 kilowatts of juice and can crank out around 1.4 exaflops. Individual trays pack two Grace CPUs and four Blackwell GPUs, all tied together via micro-MGX boards and liquid cooling.

To keep it all running, Nvidia redesigned the bus bar to carry up to 1,400 amps and stretch the rack footprint from 1,068 millimetres to 1,200 millimetres. Everything is liquid-cooled with quick-disconnect fittings, another OCP-standard touch.

Norton flagged that Nvidia is following OCP rack pitch standards to squeeze in more density than traditional 19-inch gear, while leaving modular zones at the front of trays so customers can swap in their own SSDs, NICs, or management hardware.

Both GB200 and GB300 are now shipping and already running inside hyperscale data centres. Nvidia said it plans to stick to an annual upgrade cadence, promising more density, more power draw, copper, and liquid pipes every year, with NVLink Fusion waiting in the wings as the next open interconnect trick.