LG signs a pact to cool and power Microsoft's next 26 GW
A group-level supply pact signed in Redmond and a pair of hardware qualifications handed down by Nvidia landed on the same day — and both point at the same bottleneck: the parts of an AI data center that are not chips.
LG Group will supply the cooling, power and IT infrastructure for Microsoft's global AI data centers under a group-wide strategic partnership the two companies signed at Microsoft's Redmond headquarters on September 21. LG chairman Koo Kwang-mo and Microsoft CEO Satya Nadella met for the first time, with LG's chief operating officer and the CEOs of LG Electronics, LG Uplus, LG CNS and LG Science Park in the room — a deliberate "One LG" signal that the conglomerate is bidding as one supplier rather than five. The near-term ask is concrete: large-capacity cooling for Microsoft's key strategic sites. The medium-term ambition is bigger, extending to prefabricated data centers where power, cooling and racks are assembled in a factory and trucked to site — the construction model the industry needs if a gigawatt campus is to be built in months rather than years.
Why it matters: Microsoft has said it intends to grow from roughly 12 GW of data center capacity today to more than 38 GW by 2032 — the plan we covered on September 10 — and none of that exists unless somebody can deliver chillers, coolant loops and battery storage at industrial scale. LG is answering as a hardware group that already makes the compressors, motors, inverters and magnetic bearings inside its own chillers, and whose LG CNS designs and operates the buildings. That vertical integration is the pitch: a single vendor for the thermal stack from the cold plate on the GPU to the chiller rejecting heat outside the hall.
The rest of the deal is softer and worth reading as intent, not contract. LG will roll Microsoft's Copilot and its 365 E7 Frontier Suite out to office staff worldwide and redesign security, cost and operations around AI rather than bolting tools onto old workflows. On physical AI, the two will use Microsoft's cloud to train models on manufacturing-floor data and feed the results back into production — LG's version of a data flywheel across manufacturing, logistics and mobility. No order values, volumes or delivery dates were disclosed, and LG's own framing is that the two sides will "explore" supply opportunities. A partnership announcement in this market is a place in the queue, not a purchase order.
LG Electronics' largest coolant distribution unit has qualified under Nvidia's DSX Ready program, and LG Energy Solution's battery storage cleared the same bar — putting both LG arms inside Nvidia's reference design for AI factories. LG Electronics says its 2.6MW model met the additional requirements needed for the qualification, after its 600kW and 1MW units, and joins LiquidStack and Vertiv in the cooling category; LG Energy Solution's lithium-iron-phosphate battery system joins Hitachi Energy and Tesla in energy storage, built around modular 2.5MW / 5.1MWh blocks that smooth AI load swings and hold voltage steady. Nvidia launched the program on September 21 — we flagged it in yesterday's Today in AI — and is explicit that qualifying a component is not the same as guaranteeing a site works.
The numbers explain the urgency. Nvidia's own cited research puts liquid cooling penetration for AI chips rising from 14 percent in 2024 to roughly 60 percent by 2027, because racks dense enough to train frontier models cannot be cooled with air. Both LG qualifications are product-level facts rather than relationship language, which makes them the more solid half of today's news: a certified component ships, a signed partnership has to be converted. The strategic read is that the AI buildout's supply chain is consolidating around a short list of vendors a hyperscaler can qualify once and deploy everywhere — and LG just got onto two of the lists in one morning.
What to watch: whether the Microsoft pact turns into named orders with volumes attached, and whether Nvidia's DSX Ready roster stays short enough to be meaningful.
If the binding constraint on AI is cooling and power rather than GPUs, does the market still price the chipmakers as the whole story? Tell us in the comments.
