Liquid Cooling Deep Dive: How Vertiv, Boyd, Rittal, CoolIT Systems, and Asetek Built the Most Concentrated AI Data Center Thermal Stack of the Cycle

Liquid cooling is the binding AI data center thermal stack of the cycle - air cooling is no longer sufficient for AI accelerators >700W TDP, and the 2026-2028 liquid cooling demand is the cleanest single read on the AI capex cycle's thermal leg. The 5 leading-edge liquid cooling suppliers are Vertiv (~30% global share), Boyd (~10%), Rittal (the German precision enclosure leader, ~10%), CoolIT Systems (the cleanest pure-play, recently acquired by Eaton in 2025, ~10% share post-acquisition), and Asetek (~5%, the cleanest pure-play publicly listed in Norway). The 2025 liquid cooling market is ~$7B, projected to grow to ~$30B by 2028E at ~60% CAGR, driven by the AI accelerator capex cycle.

The customer base is structurally concentrated in the 4 US hyperscalers + a few second-tier. Microsoft (the #1 AI data center capex spender in 2025-2026, ~$80B/year), Alphabet (#2 at ~$75B/year), Amazon (#3 at ~$100B/year with AWS), Meta (#4 at ~$60-100B/year), plus the second-tier (Oracle, CoreWeave, Lambda, xAI). The top-4 hyperscalers are ~70% of liquid cooling demand, and the customer concentration is the structural reason the 2026-2028 liquid cooling capacity is ~100% allocated.

The 2025-2026 liquid cooling demand is the cleanest single read on the AI capex cycle's thermal leg. The demand is driven by: (1) the AI accelerator power density ramp - Nvidia H100 at 700W TDP, H200 at 700W, B100 at 700W, B200 at 1000W, GB300 at 1200W, Rubin at 1800W+; (2) the data center rack power density - 15-20 kW/rack for legacy CPU racks, 50-80 kW/rack for AI training racks, 100+ kW/rack for the leading-edge AI training clusters, 500+ kW/rack in 2028E for the largest AI training clusters; (3) the data center PUE (power usage effectiveness) mandate - the hyperscalers are targeting PUE 1.05-1.10 (vs 1.4-1.5 for legacy air-cooled data centers), which is the binding reason for liquid cooling adoption. The 2026-2028 liquid cooling demand at ~60% CAGR is the cleanest single read on the AI capex cycle's thermal leg.

Liquid cooling technology is built on 5 binding layers. (1) Heat capture: the binding technology is how the heat is captured from the AI accelerator. The 3 leading-edge architectures are: (a) Direct Liquid Cooling (DLC, cold plate) - a copper or aluminum cold plate is attached to the AI accelerator (GPU + HBM), and the coolant flows through micro-channels in the cold plate. DLC is the cleanest mid-cycle technology (used at 70% of 2026 liquid cooling capacity), and the cold plate is a precision-machined copper part with ~50-200 micron micro-channels. (b) Single-phase immersion - the entire server is immersed in a dielectric coolant (typically a fluorinated hydrocarbon like 3M Novec or Chemours Vertrel), and the heat is captured by natural convection. Single-phase immersion is used at ~20% of 2026 liquid cooling capacity. (c) Two-phase immersion - the entire server is immersed in a low-boiling-point dielectric coolant (typically a fluorinated ketone like 3M Novec 7000 or Chemours Opteon), and the heat is captured by phase change (boiling). Two-phase immersion is used at ~5% of 2026 liquid cooling capacity and is the cleanest ultra-high-density architecture. The heat capture technology is the binding moat for Vertiv (cold plate), Boyd (cold plate + immersion), CoolIT Systems (cold plate), and Asetek (cold plate + D5LC pump).

(2) Coolant: the binding material is the coolant chemistry. The 3 leading-edge coolants are: (a) PG-25 (propylene glycol 25%) - the cleanest water-based coolant for DLC, ASP $5-15 per liter, supplied by Chevron (the largest supplier), Dow, BASF, and Clariant. (b) Dielectric fluorinated hydrocarbon (3M Novec, Chemours Vertrel) - the cleanest single-phase immersion coolant, ASP $200-500 per liter, supplied by 3M (the original supplier, but exiting PFAS production by 2025-2026), Chemours, and a few specialty chemical suppliers. (c) Dielectric fluorinated ketone (3M Novec 7000, Chemours Opteon) - the cleanest two-phase immersion coolant, ASP $300-600 per liter. The dielectric coolant is the binding material for immersion cooling, and the 3M PFAS exit is the binding 2025-2026 supply constraint. (3) Manifold + distribution: the coolant distribution unit (CDU) is the binding infrastructure for DLC. The CDU is a rack-level or row-level system that pumps the coolant through the cold plates, and the CDU is dominated by Vertiv, CoolIT Systems, Rittal, and Boyd. The CDU ASP is ~$10-50K per CDU, and the 2026-2028 CDU demand is the cleanest single read on the liquid cooling infra leg.

(4) Heat rejection: the heat captured by the coolant is rejected to the outside environment via a chiller, cooling tower, or dry cooler. The heat rejection technology is dominated by the same chiller + cooling tower suppliers (Vertiv, Trane, Carrier, Johnson Controls) and the dry cooler suppliers (Baltimore Aircoil, EVAPCO, SPX). The heat rejection capex is ~$1-3M per MW of heat rejected, and the 2026-2028 heat rejection demand is the binding leg of the data center capex. (5) Controls: the liquid cooling controls (CDU controls, leak detection, predictive maintenance) are dominated by the same data center DCIM (data center infrastructure management) suppliers (Vertiv, Schneider Electric, ABB). The 5 layers together are the binding liquid cooling technology stack, and the 5+ year R&D moat in each layer is the structural reason the 5 leading-edge suppliers control ~65% of the market.

Vertiv is the largest liquid cooling supplier with ~30% global share, dominating the DLC + CDU + heat rejection integrated stack. Vertiv's liquid cooling advantages are: (1) the broadest liquid cooling product portfolio (DLC CDU + immersion + RDHx + heat rejection + controls); (2) the largest installed base in the hyperscaler market; (3) the deepest R&D budget (~$200M/year in liquid cooling R&D); (4) the most global service network. Vertiv's liquid cooling revenue is ~$2-2.5B/year in 2025-2026, with the liquid cooling business growing at ~50% CAGR. The liquid cooling business is concentrated at the Westerville OH + Bangalore R&D centers, with ~1,000 liquid cooling engineers globally.

CoolIT Systems) is the cleanest pure-play at ~10% global share. CoolIT was the original liquid cooling pure-play (founded 2001, headquartered in Calgary, Canada), and was acquired by Eaton in 2025 for ~$1.5B. CoolIT's liquid cooling advantages are: (1) the cleanest DLC CDU product line (the CHx + Liquid to Liquid CDU family); (2) the deepest relationship with the AI startup ecosystem (CoreWeave, Lambda, xAI, Anthropic, OpenAI); (3) the cleanest North American installed base. The CoolIT/Eaton liquid cooling revenue is ~$700M-1B/year in 2025-2026, with the business growing at ~60% CAGR. The CoolIT business is now part of Eaton data center segment, with the CleanIT brand maintained.

Boyd is at ~10% global share, dominated by DLC + immersion. Boyd is a $2B+ private industrial conglomerate (the cleanest US-listed parent is Boyd Corp), with the liquid cooling business the cleanest single exposure to the AI accelerator + data center market. Rittal is at ~10% global share, dominated by DLC CDU + precision enclosures. Rittal is the German precision enclosure leader (the parent is Friedhelm Loh Group), with the cleanest single European data center installed base. Asetek is at ~5% global share, the cleanest pure-play publicly listed in Norway. Asetek is the original D5LC (direct liquid cooling) pump supplier, with the cleanest single supply position in the pump + cold plate + manifold sub-segment.

The 2026-2028 liquid cooling demand is the cleanest single read on the AI capex cycle's thermal leg. The demand is driven by: (1) the AI accelerator TDP ramp - Nvidia Rubin at 1800W+ in 2026-2027, Rubin Ultra at 2300W in 2028, AMD MI350 at 1000W in 2026, MI400 at 1500W+ in 2027-2028; (2) the data center rack power density ramp - 50-80 kW/rack for AI training racks in 2025, 100+ kW/rack for the leading-edge AI training clusters in 2026E, 200-500 kW/rack for the largest AI training clusters in 2028E; (3) the data center PUE (power usage effectiveness) mandate - the hyperscalers are targeting PUE 1.05-1.10 (vs 1.4-1.5 for legacy air-cooled data centers), which is the binding reason for liquid cooling adoption. The 2026-2028 liquid cooling demand at ~60% CAGR is the cleanest single read on the AI capex cycle's thermal leg.

The customer base is structurally concentrated in the 4 US hyperscalers + the AI-native neoclouds. Microsoft (the #1 AI data center capex spender in 2025-2026, ~$80B/year capex, ~40-50% liquid cooling mix by 2026E), Alphabet (#2 at ~$75B/year, ~30-40% liquid cooling mix), Amazon (#3 at ~$100B/year with AWS, ~30-40% liquid cooling mix), Meta (#4 at ~$60-100B/year, ~30-40% liquid cooling mix), plus the AI neoclouds (CoreWeave, Lambda, xAI, Anthropic, OpenAI). The top-4 hyperscalers are ~70% of liquid cooling demand, and the customer concentration is the structural reason the 2026-2028 liquid cooling capacity is 100% allocated.

The 2026-2028 liquid cooling capex is the cleanest single read on the AI capex cycle's thermal leg. The combined liquid cooling capex 2025-2027 is ~$15B+ (Vertiv $3B, CoolIT Systems $2B, Boyd $1B, Rittal $1B, others $8B for the hyperscaler in-house liquid cooling programs). The capex is concentrated at the AI accelerator + data center buildout, and the 2027-2028 liquid cooling capex is the binding reason. The liquid cooling capex is roughly 5-8% of the total AI data center capex, which is the cleanest single read on the thermal share of the AI data center capex pie.

The liquid cooling expert pool is structurally concentrated in 3 hubs. At Vertiv, the top liquid cooling experts are: CTO Stephen Liang, VP of Liquid Cooling Giordano Albertazzi, and VP of Data Center Solutions Peter Lambrecht. The Vertiv liquid cooling team is ~1,000 engineers at Westerville OH + Bangalore. At CoolIT Systems, the top liquid cooling experts are: CEO Patrick McGinn (former), CTO Anthony Dubreuil, and Head of Product Management Geoff Lyon. The CoolIT team is ~400 engineers at Calgary + Toronto. At Asetek, the top liquid cooling experts are: CEO André Sloth Eriksen, CTO Ole Madsen, and Head of D5LC Engineering Lars Bøgild. The Asetek team is ~150 engineers at Aalborg, Denmark. The 3-hub expert pool is ~1,550 liquid cooling engineers globally, and the geographic concentration is in the US (Vertiv) + Canada (CoolIT) + Denmark (Asetek) + Germany (Rittal).

The 2026-2028 liquid cooling capex is the cleanest single read on the AI capex cycle's thermal leg. The combined liquid cooling capex 2025-2027 is ~$15B+ (Vertiv $3B, CoolIT Systems $2B, Boyd $1B, Rittal $1B, others $8B for the hyperscaler in-house liquid cooling programs). The capex is concentrated at the AI accelerator + data center buildout, and the 2027-2028 liquid cooling capex is the binding reason. The liquid cooling capex is roughly 5-8% of the total AI data center capex, which is the cleanest single read on the thermal share of the AI data center capex pie.

The 2026-2028 liquid cooling read-through is concentrated in 4 trades. (1) Vertiv is the cleanest pure-play with ~30% share; the AI data center + liquid cooling cycle is the cleanest single re-rating catalyst. (2) Eaton (via the CoolIT Systems acquisition) is the cleanest diversified liquid cooling exposure with the cleanest single AI neocloud installed base. (3) Asetek is the cleanest publicly listed pure-play at ~5% share; the D5LC pump + cold plate + manifold is the binding sub-segment. (4) Schneider Electric is the second largest data center infrastructure supplier (after Vertiv) with ~5% liquid cooling share, and the cleanest European data center exposure. The 2026-2028 liquid cooling demand at ~60% CAGR is the cleanest single read on the AI capex cycle's thermal leg.

The first tell is the AI accelerator TDP ramp. Nvidia Rubin at 1800W+ in 2026-2027 is the cleanest single demand driver for liquid cooling; a clean Rubin launch is a re-rating catalyst for the liquid cooling cohort, a delay is a multiple-compression event. Watch the quarterly Nvidia TDP disclosures as the cleanest read on the liquid cooling cycle.

The second tell is the data center rack density. The 100+ kW/rack in 2026E and 500+ kW/rack in 2028E are the cleanest single demand drivers; a clean density ramp is a re-rating catalyst for the liquid cooling cohort, a density stall is a multiple-compression event. The third tell is the cold plate vs immersion mix. The cold plate (DLC) is ~70% of 2026 mix and the cleanest mid-cycle architecture; the immersion is ~25% and the cleanest ultra-high-density architecture. A clean immersion adoption in 2027-2028 is a re-rating catalyst for the immersion specialists (Boyd, Green Revolution Cooling). The fourth tell is the dielectric coolant supply. The 3M PFAS exit in 2025-2026 is the binding 2025-2026 supply constraint for the dielectric coolant; a clean Chemours + other supplier ramp is a re-rating catalyst, a supply constraint is a multiple-compression event. The fifth tell is the hyperscaler in-house liquid cooling programs. Microsoft, Alphabet, Amazon, and Meta are all building in-house liquid cooling capabilities (the cleanest single is Microsoft's Project Natick-style data center); a clean in-house program is a re-rating catalyst for the in-house team (and a multiple-compression event for the external suppliers), a slow program is a re-rating catalyst for the external suppliers. The sixth tell is the CoolIT Systems integration. The Eaton acquisition of CoolIT closed in 2025; a clean integration is a re-rating catalyst for Eaton, an integration failure is a multiple-compression event.