This group makes the plumbing that keeps AI computers from cooking themselves. Modern AI chips (the GPU racks NVIDIA sells, e.g. its "GB200 NVL72" rack — GPU means graphics processing unit, the chip that does AI math) throw off roughly 60–120 kilowatts of heat per rack — far more than a fan and chilled air can carry away. So data centers now pipe liquid (water or a special non-conductive fluid) right up to the chips, either onto metal "cold plates" bolted to each chip (called direct-to-chip, or DLC) or by dunking whole servers in fluid (immersion cooling). The core hardware is the CDU (coolant distribution unit — the pump-and-heat-exchanger box that moves the liquid loop), plus manifolds (the branching pipes that split flow to each rack), rear-door heat exchangers (a radiator on the back of a rack), hoses and piping. The named makers here are Vertiv (VRT), nVent Electric (NVT), and Modine (MOD), with Carrier (CARR) and Generac (GNRC) as diversified entrants. The product is needed because air cooling stops being practical above roughly 30 kilowatts (kW) per rack, and new AI data centers are being built above that line.
The facts, stated neutrally: liquid cooling went from optional to effectively required for the newest AI hardware in under three years, and the scan reports orders growing faster than the supply chain can build CDUs, piping and engineering teams. What limits supply, per the scan, is not raw material but manufacturing capacity, qualification of new vendors, and skilled thermal-engineering labor. On price: the pure-play names trade at a high multiple of today's sales and earnings — i.e., you hand over several dollars of market value per $1 of current annual revenue. est. That is arithmetic describing what buyers pay today for future growth; whether it is worth it is left to the reader.
The product is a heat-removal system sold into a data center build. Think of it as the cooling loop for AI computers: a CDU (the central pump-and-heat-exchanger unit) circulates liquid through manifolds and hoses to cold plates sitting on each chip, carries the heat away, and dumps it into the building's water or chiller system. Immersion cooling is the alternative form, where servers sit submerged in a non-conductive fluid. The "unit" of demand is best thought of as cooling content per megawatt (MW) of IT load — i.e., how many dollars of cooling hardware a data center has to buy for each megawatt of computing it installs. As racks get hotter, that dollars-per-MW figure rises, because more of the cooling job shifts from cheap air handling to higher-value liquid hardware.
The companies earn cash three ways. First, selling the equipment up front (CDUs, manifolds, heat exchangers, racks pre-plumbed for liquid) — a one-time hardware sale per build. Second, engineering, integration and installation services, because liquid loops have to be designed and qualified (tested and approved) for each deployment. Third, an aftermarket tail: maintenance, replacement parts, fluids and service contracts over the life of the data center. The up-front equipment sale is the bulk of revenue today; the service tail grows as the installed base grows. Margins (the cents of profit kept per dollar of sale) are typically higher on the engineered systems and services than on commodity sheet-metal pieces like plain piping or racks.
Source: 500-stocks scan, Data Center Liquid Cooling section — /Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/04-data-centers-infrastructure.html
Demand today is large and accelerating, per the scan. The scan describes liquid cooling as "the single most transformative shift in DC physical infrastructure driven by AI," notes that "every new AI data center being built today includes liquid cooling from day one," and that "Vertiv reported liquid cooling orders up 50%+ YoY in 2025." The scan states the market is "growing from ~$2B in 2024 to a projected $15-20B+ by 2028." That market-size path is an approximate, not-live-verified figure. est.
The driver, through the AGI lens: demand for liquid cooling is close to a mechanical function of how much AI compute the world installs and how hot each chip runs — and, given that AI capability is scaling and compute / physical-AI demand is rising, both of those are increasing. More high-density GPU racks get deployed, and each new chip generation runs hotter, pushing rack densities up. That does two things at once: it raises the number of racks needing liquid, and it raises the dollars of cooling per MW on each one (the unit of demand). There is also a one-time conversion wave on top of organic growth: a facility moving from air to liquid is a step-change in spend, not a gradual ramp. This is why the scan frames liquid cooling as a "prerequisite" rather than an optional upgrade.
Who the buyers are: hyperscalers (the largest cloud and AI operators building their own campuses — Amazon, Microsoft, Google and similar), colocation/data-center developers (firms that build the buildings and rent space), and the server makers (often called OEMs, original equipment manufacturers) who assemble racks. Because NVIDIA's reference designs (e.g. GB200 NVL72) specify liquid cooling, the buying decision is largely forced by the chip roadmap rather than chosen.
Forecast vs known fact: the order-book growth (Vertiv +50%+ YoY in 2025) is a backward-looking data point taken from the scan. The $2B → $15–20B+ market path is a forecast and should be read as a direction-of-travel estimate, not a contracted number.
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Supply is the constrained side of this market, per the scan. The scan answers "Supply constrained?" with a flat "Yes," and explains that the industry is "scaling from a niche specialty to a mass-market requirement in under 3 years," with "CDU manufacturing capacity, specialized piping, and engineering talent" all called out as bottlenecks. Its closing line is direct: "The shift from air to liquid is happening faster than the supply chain can adapt."
What actually limits supply, per the scan, is not a scarce raw material — it is factory throughput, vendor qualification (getting a new supplier's parts tested and approved), and people. CDUs are engineered assemblies that must be built, tested and certified; standing up new manufacturing lines and qualifying new suppliers takes time. Specialized piping and manifolds need precision fabrication. And the thermal/mechanical engineering talent to design and validate each deployment is in short supply. None of these can be scaled overnight by simply buying more steel or copper; they require capital build-out and trained labor, which is why lead times (the wait between order and delivery) stretch.
Capacity expansion: the incumbents are described as expanding — Vertiv's reported order growth implies it is investing to add liquid-cooling output, and nVent and Modine have been redirecting plant capacity toward data-center thermal products. Exact capacity figures and expansion timelines are not in the provided files, so any specific expansion numbers here would be approximate and not live-verified. est.
Market-share structure: on a US-listed basis this is a concentrated field. Vertiv is the broadest and most-referenced supplier in the scan (it also appears under UPS — uninterruptible power supply, the battery backup system — racks, and DCIM — data center infrastructure management software). nVent and Modine are the next most-exposed relatively focused players. Carrier and Generac are large diversified firms entering through acquisitions or new product lines (Carrier's "Abound" liquid cooling; Generac "via acquisitions" per the scan), so they hold smaller shares of this specific product today. Private and Asia-based suppliers also serve the market but are outside this US-listed group. Any share percentages would be approximate and not live-verified. est.
Source: 500-stocks scan, Data Center Liquid Cooling section (same file as above).
Putting the two sides together: by the scan's account the product is currently short — demand is outrunning the industry's ability to build and qualify capacity. The evidence is the combination of forced (not optional) adoption, order growth above 50% year-over-year at the largest supplier, and the scan's explicit statement that supply cannot adapt fast enough. In a short market you typically see firm or rising prices, high utilization of plants, lengthening lead times, and growing backlogs (unfilled orders); the scan's qualitative signals point that way, though it does not quantify price or backlog.
| Factor | Demand side | Supply side |
|---|---|---|
| Direction | Rising fast (effectively required for newest AI racks) | Expanding, but lagging (per scan) |
| Pace | Niche → mass-market in <3 years | New lines/qualification take time |
| Hard signal | +50%+ YoY orders (VRT, 2025, per scan) | CDU capacity, piping, talent cited as bottlenecks |
| Binding limit | Chip roadmap forces liquid | Factory throughput & engineers, not raw material |
| Net today (per scan) | Short — demand > supply | |
When could it flip to oversupply? A shortage like this usually unwinds when (a) capacity catches up after the incumbents' expansions come online, (b) the air-to-liquid conversion wave finishes and demand reverts to slower organic growth, or (c) new entrants and low-cost (often Asian) suppliers add commodity capacity that compresses prices on the less-differentiated pieces (piping, basic CDUs) even while engineered systems stay tight. None of those flip points has a contracted date in the provided files; the timing is a forecast question, and the scan's framing is that supply remains behind through at least the near term. The timing here is an estimate, not live-verified. est.
| Ticker | Company | What it makes here | Exposure to liquid cooling | Rough size | Position (per scan) |
|---|---|---|---|---|---|
| VRT | Vertiv Holdings | CDUs, full thermal systems, plus UPS (battery backup), PDUs (power distribution units — the strips that feed power to racks), racks, modular builds, DCIM software | Diversified data-center supplier; liquid cooling a fast-growing slice, not the whole company est. | ~$40–50B mkt cap est. | Broadest portfolio; +50%+ YoY orders (per scan); most-referenced supplier |
| NVT | nVent Electric | Liquid cooling, racks/enclosures, cable management | Meaningful and growing data-center thermal exposure; rest is electrical connection/protection est. | ~$12–20B mkt cap est. | Present in enclosures + rack-level liquid integration |
| MOD | Modine Manufacturing | Data-center cooling / CDUs & heat-transfer (plus legacy auto/industrial thermal) | Data-center is a rising share of a historically vehicle-thermal company est. | ~$5–10B mkt cap est. | Heat-transfer engineering heritage redirected to AI cooling |
| CARR | Carrier Global | "Abound" liquid cooling, plus HVAC (heating/ventilation/air conditioning) and fire suppression | Small slice of a large HVAC conglomerate est. | ~$50–60B mkt cap est. | Scale + HVAC sales channel; new to this product (per scan) |
| GNRC | Generac Holdings | Liquid cooling via acquisitions, plus backup generators | Minor slice; core business is power generation est. | ~$8–12B mkt cap est. | Entering through M&A (mergers/acquisitions); least focused on this product here |
Source: company list and exposure tags from the 500-stocks scan (rows for VRT, CARR, NVT, GNRC). MOD added because it is a named key ticker for this group but does not appear in the scan's liquid-cooling list. Market-cap ranges and percent-of-revenue exposure are general-knowledge estimates, not live-verified.
In plain money terms, the question is: how many dollars of market value are you handing over today for $1 of these companies' current sales and profit? For the broad-platform leader (VRT) and the more exposed industrials (NVT, MOD), the market has been paying a high multiple — on the order of several dollars of market value per $1 of this year's revenue, and a high multiple of current annual earnings — because the price embeds the forecast that revenue keeps compounding. est. The diversified firms (CARR, GNRC) trade at lower multiples, but there your money buys a whole HVAC or generator business, and liquid cooling is only a small fraction of what it is exposed to. est. Specific current multiples were not available offline; the statements here are general-knowledge shape, not live-verified, and should be confirmed against live quotes and filings.
Money-in / money-out shape: these are hardware-and-engineering businesses, not capital-light software. Money goes in as factory capacity, working capital (the cash tied up in inventory and in receivables — bills sent but not yet paid — to fulfill big orders), and engineering headcount (the cost of hiring engineers). Money comes out as equipment-sale revenue with a growing service tail. The economics sit in the middle of the spectrum: more capital-intensive (needing more upfront spending on plant and equipment, often called capex) than software, but more asset-light and cash-generative than building the data centers or power plants themselves. The incumbents generate free cash flow (operating cash left after the capex needed to run and grow the business), and per the scan's growth picture a chunk of that cash is currently being reinvested into expanding capacity to chase the order book rather than returned to owners. The arithmetic for the reader to watch is simply whether revenue and cash grow fast enough to justify the high multiples already being paid — stated as the math, with the judgment left to the reader.
Source: general-knowledge valuation shape (not live-verified) plus the scan's qualitative business descriptions. Specific current multiples were not available offline and should be confirmed against live quotes and filings before relying on them.
For a company-level deep-dive, one informative ordering is by how focused each name is on this product and by its supply position:
For the supply-bottleneck angle specifically, the most informative question across all names is CDU manufacturing capacity and qualified engineering headcount — that is what gates how much of the demand each player can actually convert to cash.
Used:
Hard vs approximate: