Oluwadi
Investigating the Overlooked
Region
What
Who For
Analysis
Oluwadi
Investigating the Overlooked
The person who built OpenAI's data centers just left — and he did not go to another AI lab. He went to Nvidia, to run the part of the company that certifies power and cooling, not chips. Chris Malone, OpenAI's head of data centers, joined Nvidia this month as vice president of its DSX Platform.[1] DSX is Nvidia's program for "AI factory" reference designs — the electrical, thermal, and facility standards a building has to meet to run dense GPU systems at all. Where a top operator chooses to go is a truer map of the real bottleneck than any earnings call, and Malone's map does not point at the silicon. It points at the room the silicon lives in.
The scarcity everyone talks about in AI is chips: who gets Nvidia's allocation, how many, when. But watch the revealed scarcity instead of the stated one. The man who spent the last few years finding, powering, and cooling the buildings OpenAI's chips run in did not move to a chip designer in order to design chips. He moved to certify power and cooling. When your best operator leaves to work on the envelope around the GPU rather than the GPU itself, the envelope is the constraint.
The physics explains the career move. A modern GPU rack draws power and rejects heat at densities the existing data-center fleet was simply never built to handle. That is why Nvidia had to stand up a qualification program in the first place: the old data-center certifications do not apply to its newest "factory" designs, so it created new ones and put a senior operator in charge of them. The chip is only worth anything inside a building that can feed it enough power and carry away enough heat — and most buildings, including some very good ones, cannot.
I have been inside that gap. As a director at a managed-services provider operating inside RagingWire — later NTT — in Ashburn, the heart of Data Center Alley, I watched some of the most advanced colocation space in the world fail to deliver the water cooling or hold the temperatures a full GPU rack demands. We ended up leaving four rack units unpopulated between GPU machines just to give them enough air. When the densest, best-resourced halls in the country hit that wall, Nvidia certifying the room is not a formality — it is Nvidia admitting, in public, that the room is the gate.
It also rhymes with everything else happening on the ground. Oracle just declared force majeure on an unfinished data center over a late pipeline. Google is launching TPUs into orbit to escape the grid and the cooling problem entirely. And the best data-center operator at the most important AI lab just moved to the chipmaker to work on power and cooling. Different actors, same conclusion, reached independently: the binding constraint has moved off the chip and onto the power-and-thermal envelope wrapped around it.
So here is the practical read. If you want to know where the value and the scarcity actually sit in AI infrastructure, stop counting GPU shipments and start watching where the operators go. Chips are designed by a handful of firms and, eventually, made in volume. A building that can land tens of megawatts and cool a hall of GPUs is slow, physical, permitted, and rare. Malone went to the room. So should your attention.