The loudest debates in AI infrastructure center on which company ships the most advanced accelerator, or who wins the memory bandwidth race. The Globe and Mail argues investors are looking in the wrong place. The real constraint — and the under-followed opportunity — is data center cooling.

The logic is physical rather than speculative. Modern AI racks draw far more power than the air-cooled designs of the previous decade could handle, and every incremental kilowatt of compute has to leave the building as heat. Once power densities cross the threshold where fans and chilled air no longer suffice, operators must move to liquid: direct-to-chip cold plates, immersion tanks, and the pumps, manifolds, coolant distribution units and heat rejection equipment behind them.

That shift turns cooling from a facilities afterthought into a gating factor on AI capacity. A GPU order only converts into revenue-generating inference if the building can dissipate the heat the silicon produces. In that sense, the thermal chain is becoming as commercially decisive as the supply of accelerators or high-bandwidth memory — the two categories most investors already track closely.

The category's appeal, as The Globe and Mail frames it, is that it sits outside the crowded chip trade. Investors have bid up GPU designers, foundries and memory makers on AI demand. Far fewer have priced the companies that keep those components from thermal throttling. Yet cooling spend scales directly with AI compute deployment, and the transition from air to liquid multiplies the content per rack — more plumbing, more liquid handling, more engineering — rather than simply tracking server unit volumes.

The argument also carries a durability angle. Accelerator architectures turn over on one- to two-year cadences, and memory vendors face brutal cyclical pricing. Liquid cooling infrastructure, by contrast, is installed for the life of the facility. Once an operator commits to a liquid-cooled design, retrofitting or swapping vendors is costly, which can make design wins stickier and margins less commoditized than in the components everyone watches.

None of this makes cooling a guaranteed winner, and the piece does not pretend otherwise. The sector spans established HVAC and industrial equipment makers adding liquid-cooling lines, specialists in cold plates and CDUs, and the tooling suppliers around them. Competition is intensifying as hyperscalers qualify multiple vendors, and exact revenue splits within the segment remain difficult for outsiders to verify. That opacity is precisely the point of the thesis: the market is unsung because the numbers are hard to see, not because the demand is small.

For semiconductor and systems investors, the takeaway is a portfolio question rather than a stock pick. If AI buildouts are constrained not by chip supply alone but by the ability to power and cool the resulting racks, then thermal management belongs in the same analytical bucket as advanced packaging and HBM — pick-and-shovel exposure to compute growth, one layer down the stack from the headlines.

How the category evolves will depend on rack power trajectories and how quickly hyperscalers standardize liquid-cooling architectures — dynamics that will determine whether the niche specialists capture the value or the large industrial players fold it into existing product lines.