AI’s Next Bottleneck Isn’t the GPU. It’s the Cooling Loop.

A single GB200 rack draws about 120 kilowatts. Air cooling quits around 40. That gap determines which clusters come online, and it is being priced into the cooling supply chain.

A single Nvidia GB200 NVL72 rack draws about 120 kilowatts. The GB300 that followed runs in the 120 to 140 kilowatt range, per SemiAnalysis. A conventional enterprise rack ran closer to 5–15 kilowatts only a few years ago.

That jump breaks the old cooling model. Traditional air cooling generally tops out at 20–35 kilowatts per rack and starts to lose efficiency beyond roughly 40, even with hot-aisle containment, according to Introl’s 2025 thermal guide. Rear-door heat exchangers can stretch that to about 40–60 kilowatts. Past that, the physics is unforgiving. You can't push enough air through a cabinet to pull the heat off 72 GPUs running flat out.

So the frontier of AI compute crossed the air-cooling line all at once by a factor of 3 to 4, not just a few kilowatts. There’s no practical air-cooled path to a GB200-class rack. At that density, the liquid has to come into contact with the chip.

This is bottleneck migration, the same pattern I wrote about in last week’s megawatt story. The scramble to secure power for a site is real and unsolved. But once those megawatts reach the building, the constraint moves downstream. Securing power was the first bottleneck. Rejecting the heat it throws off becomes the next.

Liquid cooling isn't something operators bolt onto an air-cooled hall the way they swap a power supply. Once rack density jumps, the change cascades through every system around it.

A 130-kilowatt rack needs direct-to-chip cold plates. Those cold plates need a coolant distribution unit (CDU) to manage flow, pressure, and heat exchange. The CDU then needs facility water piping, an external heat-rejection loop, floor loading rated for the added weight of liquid manifolds, and a powertrain sized for the new draw. Change the density, and you’ve changed the building.

That cascade is why the supply chain seized at the narrowest link first. The CDU sits between the facility water system and the sealed loop that runs to the chips, the control point for the whole cooling architecture. Direct liquid cooling revenue jumped 156% year over year in Q2 2025, according to Dell'Oro, and the market still couldn't deliver units fast enough. Dell’Oro reports CDU demand is several times the available supply, with some vendors quoting lead times of nearly a year.

When demand for one component runs that far ahead of supply, it stops being a line item and becomes a constraint. A cluster can have chips on the dock and power in the yard, but if the CDU is missing, the rack still can't come online.

Cooling has always been a quiet slice of the budget. It accounts for a low single-digit share of total data center spending, within the roughly 12% that goes to facility infrastructure, per IoT Analytics. That share is about to matter more than its size suggests, because data center capex is on track to reach about $1 trillion a year by 2030, with a large slice going to accelerated servers, according to IoT Analytics and Dell'Oro. A small piece of a trillion-dollar flow, growing far faster than the base, is a serious market.

The order books show it. Vertiv, one of the largest thermal and power vendors, closed 2025 with a $15 billion backlog against $10.2 billion in revenue, per its quarterly filings. Third-quarter organic orders rose about 60% year over year, and the book-to-bill ratio sat near 1.4, meaning new orders kept arriving faster than the company could ship.

The friction in the cooling loop has become a fee. Whoever owns the scarce link in that loop, the CDU, the cold plate, the quick disconnect that cannot leak onto a live board, collects on every rack that has to get cold before it can earn.

The clean-sweep story oversells how fast this transition can happen. Liquid cooling may be the answer for frontier AI racks, but most of the installed base is still air-cooled. Only about 22% of data centers had adopted liquid cooling as of 2024, with another 61% considering it, per the Uptime Institute. Most of that base can't be converted cheaply, which is why Schneider Electric and others are pushing hybrid designs that run air and liquid in the same hall. A full-liquid retrofit is closer to a gut renovation than a component swap. For much of the existing base, the realistic path is incremental, with rear-door heat exchangers and selective liquid deployments first, then full CDU loops only where the economics justify it.

A supply response is coming too. Dell'Oro has already asked whether the CDU market is heading toward saturation as LiquidStack, CoolIT, Vertiv, and others double manufacturing capacity. A shortage that attracts this much capital rarely stays a shortage forever. Today's supply gap may be a 2025–2026 condition, not a permanent moat, and companies that capitalize it into a 10-year margin story may be overearning.

The architecture isn't settled either. Two-phase cooling, which boils a dielectric fluid to achieve an even higher density, hit its own wall. 3M completed its exit from PFAS manufacturing at the end of 2025, ending production of the Novec fluids the approach relied on, and Microsoft and Meta have effectively paused two-phase immersion research, per ServeTheHome and Data Center Dynamics. As of mid-2026, no hydrocarbon two-phase fluid is qualified for data-center-scale production. The transition is real, but the winning loop isn't guaranteed. Betting on the wrong one is a direct path to stranded capex.

The risk isn't only technical. At CES 2026, Jensen Huang said Nvidia's Rubin chips can run on water warm enough to skip traditional chillers. Shares of Trane Technologies and Johnson Controls fell on the comments, as investors weighed how warmer water affects demand for their gear, per Capacity. The cooling trade and the chip roadmap are tied together, and the chip roadmap can move the cooling spec out from under a supplier.

None of this is a reason to chase the cooling trade at any price. The useful frame is separating temporary shortage from structural scarcity. Which links in the loop stay tight after the capacity expansions land, and which collapse into commodity economics?

The most defensible parts of the stack sit closest to the chip. Cold plates and full CDU systems carry the most engineering risk because the tolerance for a leak near a live board is unforgiving, and that makes them hard to commoditize. Generic piping, basic manifolds, and heat-rejection hardware are where excess capacity tends to show up first. The question is which parts of the loop keep pricing power and which get competed away.

The next density step determines whether this is one cycle or a steady state. Nvidia's Vera Rubin is quoted at 180–220 kilowatts per rack, with the Rubin Ultra roadmap pointing past 600 kilowatts and the industry already planning for megawatt racks, per SemiAnalysis and The Register's CES coverage. If density keeps climbing that curve, air never comes back, and heat rejection recurs in every generation. The industry doesn’t clear this bottleneck once; it pays for the cooling loop in every cycle.

Three questions decide the trade:

For now, the answer is simple. The chips are ready. The power is ready. The cooling loop decides which clusters come online this year. That is an odd place for a billion-dollar bottleneck to sit, and exactly where the pricing power has moved.