When a single rack pushes past 250 kilowatts, no amount of fan speed can save you. At the OCP Global Summit in mid-October, NVIDIA announced it would open the rack and compute-tray designs of its next flagship platform, Vera Rubin. The updated MGX 1.2 spec introduces rack-level blind-mate liquid cooling — compute trays dock and coolant lines connect automatically. A CLS deep-dive on October 17 put it plainly: liquid cooling is shifting from "optional" to "mandatory," opening a market worth hundreds of billions of yuan.

NVIDIA writes liquid cooling into the platform spec

MGX 1.2 goes beyond blind-mate cooling: interoperable rack-level interconnects, 5,000-amp liquid-cooled busbars for high-power AI systems, configurable I/O bays and fanless liquid-cooled trays — supporting deployments up to 250kW per rack. Companion tooling turns raw AI hardware into production clusters "in hours instead of days." NVIDIA also co-authored the Open Data Center spec v0.7 with AMD, Google and Meta, a unified blueprint for power, cooling and layout; Foxconn's Visionbay unveiled its "K-1" AI-factory design combining 800VDC architecture with Vertiv thermal management.

Key numbers on the cooling race

· Max rack support in MGX 1.2: 250kW

· Cluster bring-up time: days → hours

· NVIDIA × Cisco USFP optics: 200G→400G per lane with doubled lanes — 4x bandwidth density, ~1.3x energy efficiency

· A-share landscape: 16 vendors all-in on cold-plate R&D, with widening gaps between progress and revenue

Why now: density crossed the tipping point

Rubin-class platforms keep raising rack density, and the twin ceilings of air cooling and facility power have turned liquid cooling from nice-to-have into unavoidable. Domestic vendors are all in: Lenovo and other leading AI-server makers are building supernodes, platforms and liquid cooling simultaneously; at a recent industrial expo, cooling gear was billed as the "unsung hero" of AI compute. Analysts expect steep penetration gains as domestic supernodes and intelligent-computing centers come online in volume.

"A downward cost curve is moving liquid cooling from technically feasible to economically feasible." — industry insider

Cool business, hot competition

The liquid-cooling chain spans cold plates, CDUs, manifolds, dielectric fluids and site retrofits, drawing both veteran thermal players and newcomers. Huxiu's review notes that while all 16 A-share vendors are pouring in R&D, "progress is stratifying and market realization is diverging" — orders will concentrate among those with proven mass delivery. For buyers, the question is no longer whether to go liquid, but whose system ships fastest, leaks least and saves most. The second half of the compute race is about keeping the heat down.

(Information synthesized from CLS, 36Kr, Huxiu and other public reports)