What is Liquid Cooling — and Why AI Chips Are Making It Mandatory?
When NVIDIA engineered the GB200 NVL72 rack — 72 Blackwell GPUs and 36 Grace CPUs packed into a single system drawing 132 to 140 kilowatts — it shipped without fans. Not as a premium option. As the hardware specification. That design decision is not unusual engineering ambition. It is physics announcing that air cooling’s thirty-year run as the default thermal medium for data centres is functionally over for AI-class compute.
Why air hits a wall. Water conducts heat roughly 25 times better than air and can absorb approximately 3,500 times more energy per unit volume before warming by one degree. For the enterprise servers of the 2000s, drawing 5 to 20 kilowatts per rack, air was an adequate medium: fans moved sufficient volume, raised floors distributed the flow, and cooling units on the perimeter extracted the heat. That equilibrium held for decades because chip power stayed manageable. It broke when AI training workloads arrived.
An NVIDIA H100 GPU draws roughly 700 watts. A B200 Blackwell draws up to 1,200 watts with liquid cooling enabled. Eight of them in a rack, plus CPUs, networking, and power components, pushes the system to 100–140 kW — five to seven times the density of a fully loaded traditional server rack. At those densities, air requires flow rates that become impractically loud, energy-intensive, and prone to hot-spot formation that throttles chips and degrades model training throughput. The practical ceiling for air is around 15 kW per rack; at 30 kW it needs supplement; above 50 kW it ceases to be viable at scale. NVIDIA’s coming “Vera Rubin” platform is projected above 200 kW per rack — meaning every vendor, ODM, and data centre operator building around it must accommodate liquid cooling as a system-level baseline, not a retrofit.
Three architectures, not one technology. “Liquid cooling” is not a single product. It describes three distinct cooling architectures that differ sharply in deployment complexity, thermal performance, and commercial maturity.
Direct-to-chip cooling — also called cold plate cooling — attaches a metal cold plate directly to the GPU or CPU die. Coolant, typically treated water, pumps through micro-channels in the plate, absorbs heat at source, and routes to a facility-level heat exchanger. Cold plates handle 70–80% of a rack’s thermal load; residual heat from memory, storage, and power components is still managed by air. Critically, servers remain in standard 1U/2U rack form factors, allowing deployment into existing data centre infrastructure with no structural modification. Cold plate represents over 43% of the liquid cooling market today and is the primary solution in operational hyperscaler AI clusters.
Single-phase immersion cooling submerges entire servers in a dielectric fluid — typically an engineered hydrocarbon or mineral oil — that remains liquid throughout. Fans are eliminated entirely, cutting server-level power consumption by 10–15%. Power Usage Effectiveness (PUE) can drop to 1.02–1.03, against 1.2–1.4 for well-run air-cooled facilities. The constraint is infrastructure: immersion tanks are purpose-built and cannot integrate into standard raised-floor environments without major rebuild. It is the fastest-growing segment in liquid cooling.
Two-phase immersion uses fluorinated dielectrics that boil on contact with the chip surface, absorbing heat as latent energy. Vapour rises, recondenses on a cooling surface, and drips back. Thermally the most efficient of the three — but commercially the most constrained. The dominant fluid, 3M’s Novec line, is being discontinued as 3M exits PFAS chemical manufacturing. Alternatives from Chemours and Solvay are more expensive and unproven at hyperscale. Two-phase remains largely in trials and pilot programmes.
AI racks have shattered the air cooling ceiling: rack thermal density by GPU generation
Typical peak rack power draw in kilowatts; air cooling becomes unreliable above 15–20 kW and impractical above 50 kW; liquid cooling is required for all current AI training platforms
Source: NVIDIA GB200 NVL72 reference architecture (NVIDIA, 2025); B200 Blackwell thermal specifications; H100 DGX system power data; enterprise rack averages from Uptime Institute Data Center Survey 2024 (38% average power density increase 2022–2024); Vera Rubin projection from supply-chain industry estimates (October 2026). Scale: 1 kW = 2.3 px.
Asia leads the buildout. China is moving fastest. Envicool, a Shenzhen-listed liquid cooling manufacturer whose customers include NVIDIA, Alibaba, and Tencent, saw its share price more than triple over the past twelve months; Goldman Sachs and UBS both raised price targets on the stock. China’s Ministry of Industry and Information Technology reports 42 operational “10,000-card” intelligent computing clusters with total capacity exceeding 1,590 exaflops — installations that require liquid cooling as a physical necessity at their power densities. The “Eastern Data, Western Computing” initiative channels these AI clusters toward western provinces with renewable energy and greenfield land, where new campuses can be designed around liquid cooling from the first foundation pour rather than retrofitted into existing air-cooled facilities. Alibaba Cloud’s Cube DC 5.0 architecture already supports 200 kW loads through combined direct-to-chip and immersion cooling.
Korea and Japan build the supply chain. South Korea’s LG Electronics, SK Enmove (the speciality fluids arm of SK Innovation), and GRC (Green Revolution Cooling) announced a joint venture to develop next-generation immersion cooling systems for hyperscale deployments across Asia-Pacific. SK Enmove holds the fluids expertise; LG brings manufacturing scale; GRC the immersion tank intellectual property. It is the most consequential cooling supply-chain partnership in Asia outside China. In Japan, KDDI field-tested direct-to-chip cooling and reported a 94% reduction in electricity for temperature control compared to conventional air cooling. The Japanese government has committed USD 11 billion to carbon-neutral data centre infrastructure, subsidising 50% of construction costs for qualifying zero-carbon facilities. Hong Kong’s HKUST operates an eight-tank immersion installation covering 280 GPU and CPU servers; it has reduced cooling energy by over 80% and avoids an estimated 900 tonnes of CO₂ annually.
Liquid cooling architecture comparison: four approaches across five dimensions
Typical performance ranges as of November 2026; market share estimates are indicative; adoption rates vary by deployment type and geography
Source: Global liquid cooling market share: Grand View Research / introl.com analysis (2026); cold plate market share (~43%) from Meticulous Research (2025); PUE ranges from ASHRAE, Uptime Institute, and operator disclosures; KDDI field trial data (W.Media, 2025). PUE = Power Usage Effectiveness; a PUE of 1.0 is theoretical perfect efficiency; most hyperscalers target 1.1–1.2.
“Liquid cooling could represent 40% of all new AI data centre capacity commissioned by end of 2026, up from single digits two years earlier. The pace of adoption has no precedent in data centre infrastructure history.”
Dell’Oro Group — Data Center Liquid Cooling Market Forecast (2026 edition)
Three predictions for 2027–2030. First: cold plate cooling becomes a mandatory hardware specification — not a premium add-on — for all AI server racks above 60 kW by end of 2027. The forcing function is NVIDIA’s Vera Rubin platform, which ships as a fully fanless tray design. When a platform reference design has no fans, every server manufacturer, ODM, and data centre operator building in its ecosystem must accommodate liquid cooling as a system-level requirement. NVIDIA controls enough of the AI accelerator roadmap that its hardware design decisions propagate into the infrastructure supply chain within 12–18 months. AMD’s MI400 series and Google’s next TPU generation will follow the same thermal trajectory.
Second: at least three of China’s five largest hyperscalers will adopt liquid cooling as the default deployment architecture for all new AI training clusters by mid-2027. The enabling conditions already exist: greenfield campuses with no legacy air-cooled infrastructure to protect, government PUE mandates that set thresholds direct-to-chip cooling meets more easily than air, and a maturing domestic supply chain — Envicool, Tier Liquid Cooling, and several others — that can serve volume at competitive cost. This is not a prediction about technology preference; it is a prediction about where rack density mathematics leave no other viable option.
Third, and most contrarian: two-phase immersion cooling materially underperforms its market forecasts through 2028. The reason is fluid supply. 3M’s announced exit from PFAS fluorochemical manufacturing removes the dominant dielectric fluid — Novec — from the market. Replacement fluorinerts from Chemours and Solvay are more expensive, available in smaller volumes, and carry their own regulatory uncertainty as PFAS restrictions expand across the EU and US states. Deployment programmes that modelled two-phase at scale on Novec supply chains are repricing and in some cases switching architectures. Single-phase immersion and direct-to-chip will capture the deployments that two-phase was projected to win in 2026–2028. The two-phase market will not collapse — Microsoft, HKUST, and several hyperscalers have operational two-phase installations — but its projected CAGR of 20%+ will not materialise in this window.
The structural tension no one is pricing. Liquid cooling requires water — and data centres are being built in water-stressed regions. The “Eastern Data, Western Computing” provinces in China’s west receive less rainfall than coastal China. Singapore’s data centre moratorium was driven partly by water and energy constraints. Closed-loop direct-to-chip cooling can be nearly water-free, using dry coolers that reject heat to air. But open-loop and poorly designed systems still consume significant water. The operators that design liquid cooling systems with closed-loop, dry-cooler heat rejection will face fewer regulatory barriers to building new capacity. This is not a thermal engineering question. It is a site selection and permitting question that will differentiate winning operators from those stuck in approvals.
Data centre liquid cooling market trajectory: 2025–2033
Global market size in USD billions; ~20% CAGR driven by AI training cluster build-out; direct-to-chip (cold plate) dominant through 2028, single-phase immersion accelerating beyond
Source: Grand View Research “Data Center Liquid Cooling Market” (2026 edition), market size ~$6.6B in 2025 to ~$29.5B by 2033 at ~20% CAGR; Dell’Oro Group “Data Center Liquid Cooling Market to Approach $7 Billion by 2029” (2026). Figures are for all liquid cooling segments combined including direct-to-chip, single-phase and two-phase immersion. Forecast years interpolated at equal 2-year intervals on a compound growth curve.
Liquid cooling is not an infrastructure trend competing with air cooling for a share of the market. It is the thermal medium that the physics of AI-class compute requires. Every new NVIDIA AI platform from B200 onwards specifies liquid cooling; every hyperscaler building at scale is deploying it; every data centre operator who wants to host AI training workloads is retrofitting or rebuilding for it. The question for Asia is not whether to adopt liquid cooling but which architecture, at what speed, and who controls the supply chain. China is building fastest. South Korea and Japan are building the most defensible supply chain positions. The operators who get the water footprint right — closed-loop, dry-cooler, water-neutral — will face fewer site approvals and build more capacity. The operators who do not will get stranded in permitting queues while their competitors fill contracts.