When Air Cooling Ran Out of Physics
For four decades, data center operators kept servers cool the same way you keep a kitchen cool in summer: push cold air in, let hot air out. It worked because racks drew 2–3 kW in the 1980s, crept to 5–8 kW through the 2000s, and settled around 8–10 kW for most of the last decade. Fans, raised floors, and precision air conditioning were sufficient. Then generative AI changed the equation entirely.
According to Network World’s analysis of JLL Research data, average rack power density doubled from 8 kW to 17 kW in just two years — and is projected to hit 30 kW by 2027. Meanwhile, AI-specific racks are in a completely different tier: NVIDIA’s GB200 NVL72 — a single rack unit containing 72 Blackwell GPUs — draws 120 to 130 kW in sustained operation. The B200 GPU alone carries a thermal design power (TDP) of 1,000 watts per chip.
The physics of air cooling impose a hard ceiling around 30–50 kW per rack. Above that threshold, the volume of cold air required to carry heat away becomes impractical — fan power budgets balloon, hot-air recirculation causes uneven cooling across the rack, and acoustic noise in densely packed halls approaches occupational hazard levels. At 130 kW, air cooling is not an engineering challenge to be optimized. It is a boundary condition to be replaced.
The industry has responded with liquid cooling — a category that now encompasses three distinct technology families: rear-door heat exchangers, direct-to-chip (DTC) cold plates, and full immersion. Of these, direct-to-chip has emerged as the mainstream solution for the current generation of AI racks, capturing approximately 47% of the AI data center liquid cooling segment by 2025. Understanding why requires examining what each approach actually does — and where each breaks down.
How Direct-to-Chip Cooling Works — and Why It Dominates
Direct-to-chip cooling replaces the thermal interface between a processor and its heatsink with a liquid-filled cold plate. Coolant — typically water mixed with glycol — flows through micro-channels machined into the cold plate surface, absorbs heat directly from the chip package, and carries it to a coolant distribution unit (CDU) that rejects the heat via a building chiller or, increasingly, a dry cooler operating without refrigerant.
The physics advantage over air is not incremental. As Network World reports, water conducts heat roughly 25 times better than air at rest, and liquid-cooled cold plates achieve approximately 15,000 watts per square meter per degree Celsius of heat transfer — compared to air cooling’s roughly 50 W/m²/°C. That is a 300-fold improvement in heat flux density.
In practical terms, this means a direct-to-chip system can extract 50 to 150 kW from a single rack — the exact band occupied by current-generation AI inference and training servers. Single-phase DTC (where the coolant stays liquid throughout) handles the 50–150 kW range comfortably and can be extended to roughly 200 kW in purpose-built rack designs. The technology fits into existing rack form factors, uses standard coolant distribution piping, and can be retrofitted into colocation facilities that already have chilled water infrastructure. This retrofit-friendliness is a key reason DTC leads over full immersion, which requires specialized tanks, dielectric fluids, and purpose-built facilities.
Power Usage Effectiveness (PUE) — the ratio of total facility power to IT equipment power — tells the efficiency story clearly. Traditional air-cooled facilities operate in the 1.4–1.6 PUE range. Direct-to-chip liquid systems achieve 1.10–1.25. Single-phase immersion reaches 1.02–1.10, and two-phase immersion can reach 1.01–1.05. Each step down in PUE is a direct reduction in electricity consumed per unit of compute delivered — a material cost saving at hyperscale, where a 0.1 PUE improvement across a 100 MW campus can save tens of millions of dollars annually in power costs.
Advertisement
The Hyperscaler Blueprint: What Google, Microsoft, and Meta Are Building
The clearest signal that direct-to-chip has crossed from niche to default comes from deployment decisions at the companies building the largest AI infrastructure in the world.
Microsoft has moved all new data center designs to closed-loop liquid cooling systems. The company reports saving more than 125 million liters of water annually per facility compared to evaporative cooling systems — a significant advantage as water scarcity becomes an operating constraint in many data center markets. According to Network World’s coverage of hyperscaler cooling strategies, Google has deployed liquid cooling across more than 2,000 TPU pod deployments, achieving twice the chip density compared to air-cooled equivalents. The density gain is not merely a cost efficiency — it is a competitive variable that determines how much AI compute a given facility footprint can deliver.
Meta committed $800 million to a liquid-cooled data center in Indiana featuring 140 kW liquid-cooled racks — one of the first public benchmarks for a purpose-built, large-scale AI training facility built around DTC as the baseline. The choice of Indiana is itself infrastructure-revealing: the site has access to abundant water for cooling systems, renewable energy contracts, and proximity to the fiber backbone connecting Meta’s training clusters to its production inference fleet.
These decisions are not isolated experiments. The GlobeNewswire market report published May 2026 pegs the global data center liquid cooling market at $4.07 billion in 2026, with a trajectory toward $27.65 billion by 2033 — a 31.5% compound annual growth rate that reflects the entire hyperscaler build-out pipeline converting to liquid-first design.
The vendor ecosystem is consolidating in response. In March 2026, Ecolab — the global water and hygiene technology company — announced a definitive agreement to acquire CoolIT Systems, a specialist in high-density liquid cooling for AI servers. The acquisition signals that liquid cooling is no longer a boutique infrastructure play but an industrial-scale supply chain category being absorbed into the broader data center services ecosystem.
What Data Center Operators Should Do Now
The operational question for any organization running or procuring data center capacity in 2026 is not whether to move to liquid cooling, but how to sequence the transition without stranding existing air-cooled assets or locking into infrastructure that cannot scale to the next generation of AI hardware.
1. Audit Rack Power Density and Map Against the Three Cooling Thresholds
The starting point is an honest inventory of current and planned rack power draws. As JLL Research’s framework sets out, air cooling is viable up to roughly 20 kW; rear-door heat exchangers extend viability to around 100 kW; above 175 kW, immersion cooling becomes the recommended approach. Direct-to-chip DTC occupies the critical middle band — 50 to 175 kW — where most AI inference and smaller training clusters currently operate.
Map each rack or cluster against this framework. Racks running AI inference workloads on current-generation GPU hardware (H100, A100, Gaudi 3) are almost certainly in the 40–90 kW band today and will cross 100 kW with the next model generation. Knowing your density distribution tells you which aisles need DTC retrofit now, which can run rear-door exchangers as a bridge, and which will require full immersion infrastructure in the 18-to-36-month planning horizon.
2. Evaluate Colocation and Cloud Contracts for Liquid Cooling Commitments
Any colocation agreement signed in 2026 that does not include a credible liquid cooling roadmap — or at minimum, a committed power density ceiling above 50 kW per rack with a defined upgrade path — is a constraint that will bind within two to three years. Only 45% of data centers now run purely air cooling, down from 48% in 2024, and 59% of operators plan to implement liquid cooling. This means the majority of new capacity coming online is being built with liquid infrastructure — and facilities that cannot offer it will lose AI workload customers to those that can.
When evaluating colocation bids, require vendors to specify: whether their existing chilled water plant can support CDU connections; what their maximum supported rack density is under liquid cooling contracts; and what their retrofit timeline is for aisles currently provisioned for air. These are now standard procurement questions, not advanced requests.
3. Plan the Water Supply and Sustainability Equation in Parallel
Direct-to-chip cooling dramatically reduces energy consumption — a system achieving PUE 1.15 versus a legacy facility at PUE 1.5 reduces total power draw by 23% for the same IT load — but it shifts consumption from electricity to water, at least in designs using cooling towers or evaporative rejection. Microsoft’s 125 million liter annual water saving figure comes from moving to closed-loop systems that recirculate coolant without evaporation; achieving similar results requires pairing DTC cold plates with dry coolers or adiabatic cooling rather than wet cooling towers.
Sustainability reporting requirements in Europe, and emerging disclosure frameworks globally, are beginning to require data-center-level water usage effectiveness (WUE) reporting alongside PUE. Building the measurement infrastructure now — metering coolant flow, tracking evaporative losses at the facility level — avoids a compliance retrofit when reporting becomes mandatory. The most efficient implementations today achieve WUE below 0.5 liters per kWh, compared to the industry average of roughly 1.5 liters per kWh for evaporative cooling.
The Bigger Picture: Infrastructure Decisions Made Now Will Compound for a Decade
The AI hardware roadmap does not plateau at current densities. NVIDIA’s Vera Rubin platform, the successor to Blackwell, is designed for liquid cooling at 45°C supply temperature — a spec that implies not just DTC but the use of warm-water cooling, where data center coolant runs warmer than the traditional 18–24°C chilled water range. Warm-water cooling enables economizer modes in mild climates, substantially reducing or eliminating mechanical refrigeration for large portions of the year — PUE values approaching 1.05 without full immersion.
The market trajectory cited in the MarketsandMarkets direct-to-chip report projects the direct-to-chip segment alone growing from $3.33 billion in 2026 to $17.31 billion by 2032, at a 26.5% CAGR. This capital flow is building a supply chain — CDU manufacturers, cold plate suppliers, coolant chemistry specialists, leak detection systems — that will make liquid cooling progressively cheaper and more standardized over the next five years.
The data center decisions made in 2026 — which facilities to build, which colocation contracts to sign, which hardware generations to commit to — carry depreciation timelines of 10 to 15 years. Organizations that treat liquid cooling as a future concern rather than an immediate design constraint are essentially designing facilities around hardware that already exists rather than hardware that will exist when those facilities are three, five, or eight years old. The question is not whether to make this transition. It is whether to make it proactively or reactively — and reactivity in data center infrastructure is expensive.
Frequently Asked Questions
What is direct-to-chip cooling and how does it differ from immersion cooling?
Direct-to-chip (DTC) cooling uses liquid-filled metal cold plates that attach directly to processor packages — CPUs, GPUs, and AI accelerators — and extract heat via circulating coolant. Immersion cooling submerges entire servers in a bath of dielectric fluid. DTC handles the 50–150 kW per rack range efficiently and can be retrofitted into standard rack form factors; immersion is more effective above 175 kW but requires purpose-built tanks, specialized dielectric fluids, and significant facility modifications. For most 2026 AI infrastructure deployments, DTC is the practical choice — immersion is reserved for the most extreme-density use cases.
Why can’t modern data centers simply use more powerful air conditioning?
Air cooling faces a physical limit: the heat transfer coefficient of air is roughly 300 times lower than that of water. At 130 kW per rack, sustaining safe chip temperatures with air would require moving impractically large volumes of cold air at high velocity — creating noise, vibration, and air resistance that damage equipment. The fan power alone for air-cooling a 130 kW rack would consume a significant fraction of the IT power budget, defeating the purpose. Liquid’s superior thermal conductivity makes it the only viable medium at AI rack densities.
How does liquid cooling affect data center water consumption?
The answer depends on which liquid cooling architecture is deployed. Evaporative cooling systems — including cooling towers used to reject heat from liquid-cooled racks — consume significant water. However, closed-loop direct-to-chip systems paired with dry coolers or adiabatic heat rejection can dramatically reduce water use compared to conventional raised-floor air cooling. Microsoft reports saving more than 125 million liters of water annually per facility by moving to closed-loop liquid cooling, compared to equivalent evaporative cooling infrastructure.
Sources & Further Reading
- Further Reading
- Why AI rack densities make liquid cooling nonnegotiable — Network World
- Data Center Liquid Cooling Market: $4.07B to $27.65B by 2033 — GlobeNewswire
- Data Center Direct-to-Chip Cooling Market worth $17.31 billion by 2032 — Yahoo Finance / MarketsandMarkets
- Data Center Trends & Cooling Strategies to Watch in 2026 — AIRSYS North America
- Data Center Liquid Cooling Market to Witness 28.7% CAGR — GlobeNewswire
🔗 Related Intelligence
Liquid Cooling Goes Mainstream: The Race to Cool AI’s Insatiable Heat
200 kW Racks: Why Liquid Cooling Is No Longer Optional for AI Infrastructure in 2026
The Cooling Tipping Point: Liquid Cooling Now Powers 57% of AI Servers in 2026
Two-Phase Direct-to-Chip Cooling: How Data Centers Tame 250kW AI Racks in 2026
