AI data center cooling: chillers, CDUs and liquid cooling supply

AI servers produce so much heat that many racks now need liquid piped to the chip, which adds new equipment on top of chillers: coolant distribution units, manifolds and cold plates. Filings show record cooling orders and backlogs, but suppliers rarely publish lead times for this gear, so treat any single figure with care.

Key numbers at a glance

Figures are labeled by who said them and when. Most come from company filings, earnings calls and company blogs, which describe each company's own business and products.

• Trane Technologies total backlog: 12.1 billion dollars, up 70% from a year earlier, with about 6 billion dollars scheduled for 2027 and beyond (Trane Q2 2026 earnings call, 30 July 2026, company statement).
• Trane Applied HVAC bookings: up 130%, the fourth quarter in a row above 100% (same call).
• Trane Applied capacity: expanded four times over the last three years (same call).
• Trane's Stellar modular chiller plant business: about 500 million dollars of 2026 revenue expected, basically all data center focused (same call).
• Eaton purchase of Boyd Thermal: completed 12 March 2026 for 9.55 billion dollars, net of cash acquired (Eaton Form 8-K exhibit, Q2 2026).
• Eaton Electrical Global sales: 2,517 million dollars in Q2 2026, up 44%, including 25 percentage points from Boyd Thermal; segment backlog up 103% from June 2025 (same filing).
• Vertiv Q2 2026 net sales: 3,274.3 million dollars, up 24.1%; full year 2026 sales guidance raised to 13.8 to 14.2 billion dollars (Vertiv earnings release, 29 July 2026).
• Rack density then and now: about 20 kW per rack in older data centers, over 135 kW in hyperscale facilities, and 120 kW for one GB200 NVL72 rack (Nvidia blog, 22 April 2025, company statement).
• Cooling share of data center electricity: up to 40% (same Nvidia blog).
• One CDU model's capacity: 2 MW of cooling at a 5°C approach, CoolIT CHx2000 for GB300 NVL72 (same Nvidia blog).
• Rack power where 54 volt power designs strain: above 200 kW; next target 1 MW racks starting in 2027 (Nvidia developer blog, May 2025, updated July 2025).

Why AI chips changed the cooling job

Every watt of electricity that a server uses turns into heat. A data center drawing 100 MW must remove close to 100 MW of heat, around the clock.

For many years that was done with air. Fans in the servers pushed warm air out, and large air handlers cooled the room. This works until racks get dense. Nvidia's blog of 22 April 2025 says data centers once ran at about 20 kW per rack, that hyperscale facilities now support over 135 kW per rack, and that a full GB200 NVL72 rack, which joins 36 Grace CPUs and 72 Blackwell GPUs, runs at 120 kW. Nvidia's product page describes that rack as liquid cooled.

The same blog explains why air runs out. Liquid is nearly 1,000 times denser than air, so it carries heat far more effectively, and cooling a 120 kW rack with air would take below freezing air or near gale force airflow. Nvidia's developer blog on 800 volt power adds that racks above 200 kilowatts start to hit physical limits in power delivery, and that its designs target racks of 1 MW and above starting in 2027, alongside its Kyber rack systems. A rack is a metal cabinet about the size of a refrigerator. Putting a megawatt of heat into one cabinet is far beyond what air can carry away.

A short timeline of rack density

The shift happened quickly, and it explains why so much cooling equipment is suddenly on order.

• Older enterprise and cloud halls: around 20 kW per rack, cooled with air (Nvidia, April 2025).
• Blackwell generation, 2025: about 120 kW for a GB200 NVL72 rack, liquid cooled; Schneider Electric's liquid cooling infrastructure is described as supporting up to 132 kW per rack (Nvidia, April 2025).
• Hyperscale today: over 135 kW per rack supported (Nvidia, April 2025).
• From 2027: 1 MW racks and beyond, paired with 800 volt direct current power (Nvidia developer blog).

Each step multiplies the heat per square metre. Buildings designed in 2022 for air cooled racks often cannot simply be refitted, which is one reason operators order new, purpose built cooling plants rather than adding units to old halls.

The parts of the cooling system, from the chip outward

It helps to follow the heat from the chip to the outdoors.

Cold plate. A metal plate with tiny channels sits directly on a processor. Liquid flows through it and picks up heat. This is called direct to chip or direct liquid cooling.

Manifold and quick connects. Pipes inside the rack gather liquid from many servers. Special fittings let a technician pull a server without spilling coolant.

Coolant distribution unit, or CDU. This is the key new piece. A CDU is a pump and heat exchanger that sits between two separate loops. On the inside is the clean, tightly controlled coolant that touches the servers. On the outside is the facility water loop. The CDU passes heat from one to the other without letting the two liquids mix. It also controls flow and temperature and watches for leaks. CDUs come as in rack, in row and large floor units.

The term approach shows up in CDU specifications. It is the temperature gap between the facility water coming in and the coolant going back to the servers. A smaller approach means the CDU is more effective, so a site can use warmer facility water. Nvidia's April 2025 blog cites CoolIT's CHx2000 CDU at 2 MW of cooling capacity with a 5°C approach for GB300 NVL72 systems. At that size, one floor standing CDU serves many racks.

Chillers and dry coolers. A chiller is a refrigeration machine that cools water using a vapor compression cycle, the same principle as a household fridge. Nvidia's blog calls chillers highly energy intensive and notes that water cooled chillers are often paired with cooling towers to reject heat. A dry cooler is a radiator with fans that cools water using outdoor air, with no compressor. Whether a site needs chillers depends on the local climate and on how warm the water can be. Nvidia says direct to chip cooling allows warmer water, which can reduce or remove the need for mechanical chillers in many climates.

Air handlers. Not every part of an AI hall can be liquid cooled. Memory, networking and power parts still release heat to the air, so most designs are hybrid and keep some air cooling.

Immersion cooling. Some designs submerge servers in a special liquid. On its July 2026 call, Trane's management said there are some challenges in immersion, while noting the company holds patents in that space.

This list is a general description of how these systems are commonly built. The specific figures in this article come from the sources named.

What the filings show about demand

Trane. The company, a maker of chillers and other building climate equipment, reported record numbers on its 30 July 2026 call. Total backlog was 12.1 billion dollars, up 70 percent, with about 6 billion dollars slated for 2027 and beyond. Americas commercial HVAC orders rose 50 percent and Applied HVAC bookings (large, engineered systems like chillers) rose 130 percent. Management said Applied capacity has expanded four times over three years, that it is investing further in its Grand Rapids site and that "we're not turning away orders." It said four orders above 100 million dollars were booked in the quarter, one of them in Stellar, and that it held design days with hyperscale customers covering the full thermal management system. Management also said "we're in the CDU business" and that the business is performing.

It said many hyperscalers now have master purchase agreements or long term capacity commitments with Trane, which gives it more visibility. Hyperscalers are the very large cloud companies that build data centers at scale. Stellar, a maker of modular chiller plants, came with about 1 billion dollars of backlog when Trane bought it, about half of which is expected to ship this year.

Management noted that growth was broad across its markets, so not all of the increase is data centers. It said designs are shifting toward air cooled and closed loop systems, and that press claims about data centers wasting water are "not necessarily true in the majority of the cases." That last statement is the company's view, and communities and regulators sometimes see it differently.

Eaton. Eaton completed its purchase of Boyd Thermal on 12 March 2026 for 9.55 billion dollars, net of cash acquired. Its filing describes Boyd as a U.S. based supplier of thermal components, systems and ruggedized solutions serving data center, aerospace and other markets, with more than 6,000 employees. Nvidia's April 2025 blog lists Boyd as a supplier of CDUs, cooling loops and cold plates. In the second quarter of 2026, Eaton reported Electrical Global sales of 2,517 million dollars, up 44 percent: 18 points organic, 25 points from Boyd and 1 point from currency. Twelve month rolling orders in that segment rose 33 percent organically and backlog rose 103 percent from June 2025. The filing does not split out liquid cooling.

Vertiv. The company sells power and cooling equipment for data centers. Its second quarter 2026 sales were 3,274.3 million dollars, up 24.1 percent, with organic growth of 17.8 percent and a 5 percent contribution from acquisitions. By region, the Americas rose 29.2 percent, Asia Pacific 28.5 percent and EMEA 1.7 percent. It raised the full year sales range to 13.8 to 14.2 billion dollars and said capital spending will be about 4 percent of revenue, at the high end of its range, as it accelerates capacity expansion. It also noted minor revenue timing shifts, primarily from temporary supply chain congestion and multi phase projects. The release ties contingent payments in the quarter to its PurgeRite acquisition but does not break out cooling sales.

Taken together, three suppliers report demand and backlog far above last year, and all three are adding capacity or buying it. None of the three publishes a lead time in weeks for cooling units in the materials cited here.

What is known about lead times

No supplier cited here gives a primary, dated lead time for chillers or CDUs. Trane's management said "Are we seeing constraints? Yes, but we always do," and that it plans with suppliers on a three year horizon. Vertiv's mention of supply chain congestion points the same way. Both suggest suppliers are working under pressure, though neither is a delivery time.

The long term capacity agreements Trane describes are the clearest signal. When the largest buyers reserve factory output years ahead, smaller buyers face longer queues for the same equipment, even if no one publishes the number of weeks.

If you are a customer or investor trying to understand schedule risk, the sensible approach is to ask suppliers for dated written quotes for each of the four items: chillers or dry coolers, CDUs, rack plumbing (manifolds and connectors) and cold plates. Each has a different supply base.

Who supplies what

The sources cited here name these companies in cooling and related roles. This is a list, not a ranking.

• Chillers, air handlers and CDUs: Trane Technologies, which also owns Stellar.
• Thermal components, CDUs and cold plates: Eaton, through Boyd Thermal.
• CDUs and cold plates: CoolIT Systems, per Nvidia's April 2025 blog.
• Power and cooling for data centers: Vertiv, which has published a GB200 NVL72 reference architecture, and Schneider Electric, both also named by Nvidia among data center power system partners.
• Chip and rack designer: Nvidia sets many of the cooling requirements through its reference designs. Its public material does not say which companies supply cold plates for any given customer, so this article makes no claim about that.

Water, power and the chiller question

Cooling is a large share of a data center's power bill. Nvidia's April 2025 blog says cooling can account for up to 40 percent of a data center's electricity use. That is why designers try to avoid running chillers. Using warmer water in the liquid loop lets more heat be thrown away with outdoor air. Trane's remark that chillers are "getting smarter," with controls that support free cooling modes and the right mix of water temperatures, points the same way. Fewer compressor hours means lower power use but does not remove the equipment from the project, so chillers can still hold up a schedule.

Water use is a separate matter. Closed loop systems recirculate coolant and use little water in operation. Evaporative cooling towers do consume water; Nvidia's blog says evaporative cooling can use millions of gallons of water per megawatt per year. Nvidia also claims its liquid cooled GB200 NVL72 gives over 300 times more water efficiency than traditional air cooled designs, and Vertiv's reference architecture for that rack claims a 25 percent cut in annual energy use. These are vendor statements about their own products, not independent measurements. Which approach a site uses depends on the climate, local rules and the operator.

What is firm and what is uncertain

Firm enough. Rack densities have risen from around 20 kW to over 100 kW in a few years, and racks of 1 MW are planned from 2027, so air alone is not enough in many AI halls. Major suppliers report record backlogs and bookings. Capacity is being added and bought: Trane says Applied capacity has expanded four times in three years, and Eaton paid 9.55 billion dollars for Boyd Thermal.

Uncertain. The exact lead time for chillers, CDUs, manifolds and cold plates. How much of the reported growth is data centers versus other buyers. Whether designs end up mostly liquid, mostly air or a mix at the facility level. The cost premium of liquid cooling, for which no primary figure is cited here. The long run water effect, since the main efficiency figures come from vendors.

What to watch next

• Vertiv's and Trane's third quarter results for backlog, capacity additions and any numbers on lead times.
• Whether Eaton reports Boyd Thermal results separately.
• The first 800 volt direct current rack designs in 2027 and the cooling that comes with 1 MW racks.
• Any regulator or community decisions on data center water use.

For power side context, see transformers and switchgear lead times and backup power explained.

Safeguard does not sell cooling equipment, but it can scan the dependencies of software that controls it and produce an SBOM, and there is a free scan at https://gold.safeguard.sh/scan.

Quick check

Sources

• Trane Technologies Q2 2026 earnings call transcript, 30 July 2026 (via Webull)
• Eaton Form 8-K exhibit, second quarter 2026 results
• Vertiv Q2 2026 results release, 29 July 2026
• Nvidia blog, Chill Factor: Blackwell platform boosts water efficiency, 22 April 2025
• Nvidia GB200 NVL72 product page, accessed 19 October 2026
• Nvidia developer blog, 800 VDC architecture for AI factories, 20 May 2025, updated 31 July 2025

Last reviewed: 19 October 2026.