Cooling outlay climbs from $26.3B in 2025 to $31.4B in 2026 and a projected $128.3B by 2033 at a 22.3% CAGR, driven by AI training racks crossing 70 kW per rack and exceeding 100 kW in the largest 2026 hyperscale builds [S3].
Cooling systems already consume 30-40% of a facility's total electricity, meaning a 100 MW campus burns 30-40 MW just to reject heat, which makes the air-to-liquid transition a pure OpEx lever rather than a sustainability talking point [S3]. Asia Pacific holds 36.9% of 2025 cooling revenue, the largest regional pool of cooling capex, while North America is the second-largest and the fastest-growth segment in 2026 sits in Latin America and parts of Southern Europe where new AI zones are being zoned and powered [S4].
Anchor Specs: PUE, Density, WUE, and MTTR
The Uptime Institute's 2025 Global Data Center Survey reports a weighted average PUE of 1.54, unchanged for the sixth consecutive year, while the best hyperscale campuses operate between 1.03 and 1.10 and Google reports a fleet average of 1.10 [S3]. WUE, measured in litres of water per kWh of IT load, has moved from a sustainability footnote to a first-class procurement metric, with operators now publishing litres-per-kWh alongside PUE in 2026 RFPs [S4].
Buyers writing 2026 cooling RFPs should lock four numbers before contacting suppliers: target PUE (mechanical-only versus total-facility), sustained rack density in kW per rack, on-site water consumption expressed as WUE, and mean-time-to-repair inside the buyer's region [S2]. Operators who delay cooling procurement by even one quarter are now reporting 12-18 month delivery slip on CDU skids from major OEMs, which is why the spec gate must be set before, not after, supplier shortlisting [S4].
Density Decision Grid: Air, Rear-Door, Cold-Plate, Immersion
Air cooling (CRAC, CRAH, in-row, and in-rack doors) remains the baseline for legacy raised-floor white space and typically handles 10-25 kW per rack, while rear-door heat exchangers extend air-side infrastructure to 40-60 kW per rack by pairing DX or chilled water with a rack-mounted coil [S2]. Direct-to-chip cold plates sit in the middle band and are the default for new GPU pods above 40 kW per rack, while single-phase and two-phase immersion baths unlock 200-400 kW per rack densities and dominate new AI training pods at the cost of dielectric fluid aging, pump seal leaks, and condenser fouling that few legacy service contractors are trained on [S3].
The workload dictates the architecture: general-purpose enterprise stays on air or RDHx, high-density AI training moves to cold-plate or immersion, and edge or telecom sites with leak intolerance can run sealed passive heat-rejection modules that are pump-free and water-free [S2]. Direct liquid cooling is growing at a 24.7% CAGR through 2035, while air cooling's 5.2% CAGR significantly lags the overall market rate of 12.3% [S3]. Cold-plate and rear-door heat-exchanger rollouts are the default for new GPU pods above 40 kW per rack; immersive single-phase tanks remain a minority path in 2026 but show the steepest growth curve in vendor pipeline data [S4].
Vendor Map: US Reps, US OEMs, and Chinese System Houses
The supplier base is splitting into three models. US rep and integrator houses offer design, installation, and lifetime field service for everything from single racks to server farms, a service-heavy model common among North American independents that resell multiple brands rather than manufacturing their own coils and compressors [S2]. US OEMs ship proprietary heat-rejection units with explicit differentiators such as pump-free and water-free loops, targeting edge and HPC deployments where leaks and pumped loops are unacceptable [S2].
Chinese system builders such as Soeteck Power in Shenzhen now ship precision cooling, UPS, telecom power, AIDC racks, and energy storage as a single package, reflecting the Asian bid to export full-stack white-space solutions and mapping directly onto the AIDC reference designs being tendered across Asia in 2026 [S2]. Mainland-China sourcing directories list more than 100 vendors tagged under data center environment monitoring and adjacent cooling categories, with manufacturers that also produce temperature/humidity, LoRa, and wireless sensors for DCIM integration [S3]. M&A signal: Ecolab paid $4.75B to acquire CoolIT Systems, a direct-to-chip liquid cooling specialist, which consolidates the D2C supply chain under a single industrial-water major [S3].
Standards and Instrumentation Buyers Should Anchor To
Specifiers writing 2026 cooling RFPs are consolidating around four reference frameworks, with ASHRAE TC 9.9 remaining the operating-envelope bible for inlet classes and allowable temperature/humidity ranges [S4]. Operators in water-stressed US states and parts of EMEA are specifying closed-loop hybrid systems with adiabatic-mode bypass to drive site WUE down without raising inlet temperatures beyond ASHRAE TC 9.9 allowable classes [S4].
The practical engineering knock-on is a surge in instrumentation on the secondary loop: flow meters, pressure transmitters, and temperature sensor strings sized for 25-40% glycol mixtures, plus leak-detection cable under every cold-plate manifold [S4]. Specifying instrumentation in line with ISA 5.1 for tag nomenclature and IEC 60079-x for hazardous-area chillers is becoming standard rather than optional on 2026 builds, and each retrofit tier adds roughly 30-60 measurement points per MW, which is the hidden cost line most owners underestimate before a refresh [S4]. For the underlying sensing hardware on the secondary loop, the pressure transmitter and flow meter categories now ship with glycol-compatible wetted parts and HART diagnostics as stock items rather than specials. For on-rack and white-space airflow monitoring, the data logger category covers rack-inlet and return-air strings that have to log against ASHRAE TC 9.9 envelope points.
Demand Side: AI Server Pull and CDU Lead Times
Carrier reported Q1 2026 data center orders up more than 500% year-over-year with backlog fully covering its $1.5B full-year target, and Trane reported Americas commercial applied equipment bookings up more than 160% in Q1 2026, both of which confirm that chilled-water plant and CDU demand is outrunning nameplate capacity [S3]. Major cloud providers have standardized direct liquid cooling across next-generation AI data center designs, with Google TPU v5 clusters, Microsoft Azure AI infrastructure, and Meta AI racks all specifying cold-plate or single-phase loops at the rack level [S3].
For the silicon-side pull behind that cooling demand, the NAND Flash Procurement: Cell Type, Interface Tier, and Supplier Terms in 2026 piece tracks how substrate bottlenecks are forcing server OEMs to lock in cooling capacity ahead of accelerator shipments. On the thermal mapping side of the white space, Temperature recorder sizing: channel count, span, isolation, accuracy covers how to size rack-inlet and return-air logger channels so the ASHRAE TC 9.9 envelope check survives an audit. Buyers should verify CDU skid delivery slots in writing before signing the server PO, since 12-18 month slip is now the floor rather than the ceiling on major-OEM quotes [S4].
Who This Procurement Path Is For, and Who It Is Not For
This density-driven, PUE-and-WUE-anchored path fits operators planning new AI training pods above 40 kW per rack, hyperscalers standardising on cold-plate loops, and colocation providers writing 2026 RFPs in water-stressed regions where WUE is a contract clause [S4]. It also fits enterprise buyers doing a brownfield refresh where rear-door heat exchangers can extend existing air plant to 40-60 kW per rack without re-piping the chiller yard [S2].
It is the wrong path for single-cabinet edge sites with sub-5 kW loads where a sealed passive heat-rejection module is more cost-effective than any pumped loop [S2]. It is also wrong for operators whose facility water system cannot accept a 25-40% glycol mixture, since cold-plate and RDHX loops require freeze-tolerant coolant and that constraint rules out several legacy plant tie-ins [S4]. Buyers who treat nameplate power as a guaranteed heat-load measurement will oversize the plant; rack heat-load documentation should distinguish normal load from anticipated peak load, with assumptions flagged for supplier validation [S1].
Trackable Signals for the Next Two Quarters
Three signals will tell whether the 2026 cooling market stays supply-constrained or loosens: Q3 2026 Carrier and Trane backlog conversion rates against the $1.5B and 160%-growth bookings already reported, CDU skid lead-time quotes from at least two independent OEMs for identical 1-2 MW cold-plate packages, and any 2026 Q4 hyperscale WUE disclosure in litres per kWh that breaks below 0.5 L/kWh at a campus above 50 MW IT load [S3][S4].