For years, the AI industry's attention has been almost entirely on stacking GPU compute, but a once-overlooked piece of the puzzle is stepping into the spotlight: thermal management. With NVIDIA's Blackwell platform pushing per-chip power into the kilowatt range and the next-generation Rubin architecture climbing even higher, traditional air cooling is approaching its physical limits, and liquid cooling is going from “optional” to “mandatory.” For US stock investors, the structure and elasticity of this supply chain deserve a close look.
The More Compute, the More Heat: Why Cooling Became the New Bottleneck
Let's do the math first. The H100 chip of the previous-generation Hopper architecture draws about 700 watts, which beefed-up air cooling could barely handle. In the Blackwell generation, the B200 and B300 cross the kilowatt threshold per chip. A single NVL72 rack packs in 72 GPUs and can draw over 120 kilowatts at full load, and next-generation platforms will only go higher.
Traditional data center racks, however, are typically designed for just a few to a dozen kilowatts. With hot and cold aisle containment plus computer room air conditioners “blowing cold air,” the realistic ceiling for reliably cooling a single rack is around 40-50 kilowatts. As rack power doubles and doubles again, air cooling has only two ways out: pile on fans and AC units while power bills and noise spiral out of control, or throttle the chips down—and the pricey compute you bought shrinks on the spot.
That is the core logic behind cooling's promotion from “supporting role” to “make-or-break”: if the cooling solution can't keep up, even the most powerful chips can't run at full tilt. For NVIDIA, liquid cooling as standard on high-end GPUs is a done deal; for cloud providers, the maturity of cooling solutions directly affects how fast new racks can be delivered and what they cost to run.
Liquid Cooling's Three Technology Routes, Explained in Plain Language
Cold Plate: Today's Absolute Mainstream
Cold plate (Direct-to-Chip) attaches metal plates with internal flow channels directly onto the chip surface, with coolant circulating through the piping to absorb heat. It doesn't change the server form factor and retrofit costs are relatively manageable, making it the default solution for today's AI racks. The supply chain built around it includes manifold piping inside the rack, coolant distribution units (CDUs) outside the rack, quick disconnects, hoses, and pump modules—every link is added value.
Immersion: The Most Thorough but Also the Most Radical
Immersion cooling submerges entire server motherboards in non-conductive dielectric fluid. It offers the highest cooling-efficiency ceiling and can push energy-efficiency metrics very low, but it demands heavy upfront investment and a relearning of operating routines, so today it appears mostly in supercomputing and specialty computing. The industry generally believes immersion will regain attention as power density keeps rising, but in the short-to-medium term cold plate remains the main battlefield.
Rear-Door Heat Exchanger: A Transitional Fix for Legacy Data Halls
A rear-door heat exchanger is essentially a “water-cooled AC unit” mounted on the rack's rear door, significantly boosting per-rack cooling capacity without overhauling the whole data hall—a compromise choice for retrofitting existing facilities. The three routes aren't mutually exclusive; real-world projects often mix and match them.
Map of the US-Listed Thermal Supply Chain: Who Gets a Slice
- Vertiv (VRT): A veteran player integrating thermal management and power, with product lines covering CDUs, racks, and heat exchangers. It holds a leading position among hyperscale data center customers and is one of the most-discussed names in the liquid cooling space.
- nVent (NVT): Started out in industrial electrical protection; in recent years it has entered liquid cooling piping and CDUs through acquisitions and organic expansion, with ties to multiple cloud providers and server makers.
- Modine (MOD): Began with automotive thermal management; its Airedale brand is deeply rooted in data center cooling. Walking on both air-cooling and liquid-cooling legs, its upside has drawn plenty of investor attention.
- Super Micro Computer (SMCI): The most aggressive liquid-cooling pioneer among server OEMs, selling complete cold-plate rack solutions directly to cloud customers, with liquid-cooling penetration leading its peers.
- Traditional HVAC giants: Trane Technologies, Carrier, Johnson Controls and the like feed on the “broader perimeter” of data center chiller plants and cooling towers.
- Materials and components: Specialized suppliers are also emerging in niches such as dielectric coolants, quick disconnects, pumps, and sensors; two-phase coolant maker Chemours (CC) often makes watchlists.
One caveat: a batch of core suppliers such as Boyd and CoolIT are not listed in the US, so the listed names are more like “supply chain proxies” with varying degrees of exposure. Before investing, tell apart the companies that genuinely book liquid cooling revenue from those merely riding the story.
The Investment Thesis: A Triple Drive of Penetration, Value Content, and Orders
The appeal of the liquid cooling space can be summed up as a “triple play.” First, penetration: AI servers are moving from “partial liquid cooling” to “full liquid cooling,” and industry analysts broadly expect the market to grow about 30% a year over the next several years. Second, value content: the cooling value in a fully liquid-cooled rack can reach several times that of an air-cooled solution at the same power, so cooling's share of total system cost is rising structurally. Third, order visibility: cooling equipment is typically tendered in step with data center construction, so leading vendors' orders in hand and backlog figures are high-frequency signals for tracking industry momentum.
From a broader view, cloud giants' capex keeps flowing into AI infrastructure, and data center construction is moving from “putting up buildings” to “installing equipment”—cooling and power are the two areas with the greatest elasticity in equipment spending. That's why whenever a cloud provider raises capex guidance, liquid cooling and power stocks tend to react at once.
A Sober Look: This Business Isn't a Sure Thing
- Technology route risk: Cold plate, immersion, and phase-change routes are all still evolving; betting on the wrong link could mean getting hit by product iteration.
- Intensifying competition: The big prize is drawing HVAC giants, server makers, and startups into the arena all at once, and prices and gross margins may come under pressure.
- Customer concentration: Orders rely heavily on a handful of hyperscale cloud providers; any slowdown in expansion by one of them will bring earnings swings.
- Operations and standards: Leak risks, water quality management, and maintenance complexity remain concerns for cloud customers, and industry standards are not yet fully unified.
- Valuations aren't cheap: Some names are already fully priced in; chasing highs calls for caution about earnings delivery falling short of expectations.
A Beginner's Watchlist
To track this space, watch four things: first, the evolution of cooling solutions for NVIDIA's new platforms—rack power and liquid cooling attach rates are the hardest anchor; second, capex guidance in cloud providers' earnings reports; third, order and backlog data disclosed by leaders like Vertiv; fourth, liquid cooling penetration and shipment figures published by industry research firms. String these clues together and you can roughly gauge the temperature of the thermal supply chain.
In one sentence: the compute arms race has reached the point where “after chips come power, and after power comes cooling.” Liquid cooling is stepping from behind the scenes to center stage. For US stock investors, this supply chain carries both the certainty of the AI narrative and a test of timing and valuation discipline. Reading the landscape clearly, controlling position sizes, and tracking orders may matter more than chasing rallies and dumping dips.
