Hey {{first_name|Investor}} -
Cooling AI chips has stopped being a cost line and become an industry. That's the lesson today, and the numbers behind it are simple.
One task handed to an AI coding agent uses an estimated 150 watt-hours of electricity. That number is Zeke Hausfather's, published August 5: he logged 1,138 prompts of his own coding-agent use over 8 weeks and put each one between 60 and 290 watt-hours. It's estimated from tokens rather than measured at the wall, and it's one person's usage, so take that as what you will. A chat reply is about a quarter of one watt-hour. Your phone battery holds about 15, so one agent task is roughly 10 phone batteries, and OpenAI's new Astra model, released September 3 for exactly this kind of multi-step work, is built to run those tasks all day.
Only part of that energy goes into the chip doing the thinking. The rest keeps the building from cooking.
The one number that measures it
Data centers report a figure called PUE, power usage effectiveness. It's the total electricity the building draws divided by the electricity that reaches the IT equipment, meaning the servers, the storage and the networking.
A PUE of 1.00 would mean every watt reaches that equipment. The Uptime Institute's 2025 survey puts the weighted average at 1.54 across 681 operators reporting their own numbers, so for every dollar of power that reaches the racks, the building spends another 54 cents on overhead. Cooling is usually the biggest piece of that, with power conversion and distribution taking most of the rest.
Google reported 1.09 across its fleet for 2025. Meta's most recent published figure is 1.08 for 2024, and Microsoft's is 1.17 for the year to June 2025. Different years and different scopes, so read them as a range rather than a ranking. That gap is one reason the same AI task costs less to run in a hyperscaler's building than in an older one; the chips, the utilization and the power contract do the rest.
Think of a bakery. The ovens are the chips. The air conditioning that stops the bakery from becoming an oven itself is the overhead. An old bakery with a wall unit spends half again as much on cooling as it spends on baking. A modern one, with the ovens vented straight outside, spends almost nothing.
Why air stopped working
For 20 years, data centers cooled computers with cold air blown across them. A typical rack of servers draws 10 to 30 kilowatts, and air carries that much heat away comfortably.
An AI rack is a different object. NVIDIA's GB200 NVL72 is a 120-kilowatt rack by design and the GB300 version runs to about 142. NVIDIA hasn't published a rack figure for Vera Rubin, the generation arriving now, though it has said the platform is all liquid-cooled with no fans at all. Its Kyber design, 576 Rubin Ultra dies in one rack, targets around 600 kilowatts in the second half of 2027, and NVIDIA's 800-volt work is meant to support 1-megawatt racks from 2027. Hold the dates loosely: SemiAnalysis reported in July that Kyber has slipped to 2028, which is the date I used two weeks ago, and NVIDIA says the roadmap is intact.
A 120-kilowatt rack is about 80 hair dryers running flat out inside a phone booth. Air doesn't stop working at one particular number, but somewhere in the high tens of kilowatts the fans, the airflow and the floor space stop being worth what they cost, and rack-scale AI systems are well past that.
Liquid isn't. It carries far more heat per unit of volume than air, so the industry pipes it straight to the chip: a cold plate sits on the processor, warm liquid leaves the rack, and a cooling distribution unit sends it back cool. NVIDIA says its Rubin-generation design takes coolant as warm as 45°C at the rack. That sounds hot until you remember the chips run far hotter, and in the right climate liquid that warm can be cooled outside for much of the year without running a chiller.
The tax becomes a product
When liquid replaces air, the cooling line stops being a cost the operator minimizes and becomes equipment somebody sells.
NVIDIA's own comparison: a Blackwell site with a fifth of its capacity air-cooled runs at a PUE of 1.25, and a fully liquid-cooled Vera Rubin site at 1.1. Same building, same power budget, and the liquid-cooled version fits about 14% more IT load behind the same connection. NVIDIA published that in March, before Rubin shipped, so the 1.1 is a projection written in the present tense. It claims a further gain, up to 30% higher GPU throughput inside the same power envelope, from a mode it calls Max-Q, which is a separate design choice. Both are NVIDIA's numbers, and both describe its own equipment. Huang's line at GTC Taipei in June was "throughput per watt is revenues."
So cooling is the fastest-growing line in data center infrastructure. Dell'Oro has thermal management up nearly 50% year over year in the first quarter of 2026, ahead of every other part of the build, and puts liquid cooling at close to $3 billion of manufacturer revenue in 2025, roughly double the year before, reaching about $7 billion by 2029. TrendForce has liquid cooling on about a third of AI chips in 2025, 53% this year, and close to 60% next year.
The results:
Vertiv ($VRT), the largest listed name in the layer, grew sales 24% last quarter and raised its full-year net sales forecast to $13.8 to $14.2 billion. It closed 2 thermal acquisitions this year, ThermoKey in June and Strategic Thermal Labs in April, neither priced publicly, and it is working with NVIDIA on the 800-volt power designs the Vera Rubin racks need.
Modine ($MOD), a 110-year-old radiator maker, grew data center sales 90% year over year to $349 million in the quarter ended June 30. It split data centers out as its own reporting segment in April, which tells you where the company thinks it lives now.
nVent ($NVT) grew total sales 53% last quarter, expects data center sales above $2 billion this year, more than double last year's, and has leased its third liquid-cooling plant in 3 years, 160,000 square feet in Blaine, Minnesota, producing from the first half of 2027.
Eaton ($ETN) completed a $9.5 billion purchase of Boyd Thermal in March, at 22.5 times Boyd's estimated 2026 adjusted EBITDA. Boyd forecast $1.7 billion of sales this year, $1.5 billion of it in liquid cooling. Ecolab, a water-treatment company, closed its $4.75 billion purchase of CoolIT, another cold-plate maker, on July 2. Checks that size from industrials that big mean the line item has become a market.
The water problem behind the heat problem
Cooling with water creates a second bill. Google consumed 10.9 billion gallons across its data centers and offices in 2025, up 34% on the year, and the data centers are almost all of it.
The pushback is local and it is growing. Calaveras County paused new data center applications in August and Mendocino County followed in September, both citing power and water. Neither is a one-off, and a pause like that adds months to a site nobody has broken ground on yet.
Microsoft's answer is a closed-loop design that consumes no water for cooling once it's filled, launched in August 2024 and used in its new designs since, with the first sites coming online this year and the wider rollout late in 2027. Microsoft says the trade is a nominal increase in PUE, because the heat leaves with electricity instead of evaporation. A site can soften that with recycled water, dry coolers or a cooler climate, but the choice between water and watts is now a local political decision as much as an engineering one.
What to take from this
When a cost line grows faster than the thing it supports, it becomes an industry. Liquid cooling was a specialist product a few years ago. It roughly doubled in 2025 to about $3 billion, is forecast to reach $7 billion by 2029, and 2 of the largest industrials paid billions to buy their way in, Eaton at 22.5 times forward EBITDA.
Two numbers to learn to read: a data center's PUE, which tells you how much of its power bill is overhead, and a rack's kilowatts, which tells you how hard it is to cool. Air handles a 10 to 30 kilowatt rack. Past the high tens of kilowatts it gets expensive and awkward, and the rack-scale AI systems shipping now start at 120.
The cooling layer gets paid every time a new rack goes in, whichever chip is inside it and whichever model runs on it. Few layers in this cycle can say that, and it's why a radiator company is growing 90%.
Stay disciplined, Koh
Make Your Own Alpha is live. Today's issue handed you one lens: follow a cost line that's growing faster than the product it serves. The book teaches the full set of frameworks I use to find those layers and decide which ones earn a place in my portfolio. Get your copy here.
Disclaimer: Nothing in this email, the Reset and Invest newsletter, the Make Your Own Alpha book, or any course or digital product from Starshine Media LLC constitutes investment advice or a recommendation to buy or sell any security. Numbers and observations are as of publication. I may hold positions in companies discussed. Always do your own research and consult a licensed financial advisor before making investment decisions.