Cooling spent decades as the supporting act: air plus fans, and air conditioning when that was not enough. Then AI chip power went nearly vertical. In 2026 liquid cooling is not a design choice, it is the price of admission.
Physics does not negotiate
| Chip or platform | Power | Cooling |
|---|---|---|
| Nvidia H100/H200 | 700W | Air is viable |
| Nvidia GB200 | 1200W | Hybrid: cold plate on chips, air on storage and IO |
| Nvidia GB300 (Blackwell Ultra) | 1400W, 1600W at the top end | Full liquid, every component |
| Nvidia Rubin (VR200/VR300) | 1800W / 2700W | 100% liquid, fanless |
| Google TPU v7 (Ironwood) | 980W | Mandatory 100% liquid |
| Air-cooling limit, per Vertiv | Chip TDP under 1000W, rack under 80kW | Beyond that, no amount of air works |
The rack jump is easier to feel: 2-10kW traditional, 42kW for an air-cooled H100, 120kW for GB200 NVL72, 120-150kW for GB300 NVL72. Vera Rubin ships in the second half of 2026 near 300kW per rack, and Rubin Ultra passes 1 MW in 2027. Put a GB300 on air and thermal protection strips at least 40% of its compute.
Penetration: 2026 is the year of scale
| Measure | 2024 | 2025 | 2026 | Later |
|---|---|---|---|---|
| Global AI data center liquid cooling penetration (TrendForce) | 14% | 33% | 40% | |
| China market penetration | 20% | 37% | Over half in 2027; 82% in 2030 | |
| Liquid cooling share in new AIDC builds (IDC and Gartner) | Under 15% | Over 45% | ||
| New Chinese AIDC penetration (broker forecasts) | Above 80% |
The measurement bases differ; the direction does not. The global liquid-cooled server market crosses US$8.5 billion in 2026, about RMB 61.5 billion. JPMorgan puts AI server liquid cooling above US$17 billion, from US$8.9 billion. China Galaxy Securities sees the China market above RMB 30 billion, Orient puts the global space at RMB 68.8 billion, and Guoxin expects US$53.5 billion by 2030. The Wind liquid-cooled server index jumped 11.41% in a week, its biggest move in 76 weeks.
Cold plate leads, immersion takes the extremes
| Dimension | Cold plate | Immersion |
|---|---|---|
| Principle | Plate bonded to the CPU or GPU, coolant through microchannels; storage and IO stay air-cooled | Whole machine submerged in non-conductive coolant, every component cooled by contact |
| Share | Above 90% in 2026, over 80% of new builds | 25-30% on another caliber, aimed at ultra-high density |
| Strengths | Most mature, easy retrofit into air-cooled rooms, good supply chain fit | Best efficiency, PUE below 1.1, no fan noise, no hot spots |
| Best fit | 1000-1400W chips such as GB300, 100kW class racks | Racks above 30kW, and the 200kW to 1MW class ahead |
| Cost | Cold plate 32%, connectors 28%, CDU 25%, so 85% in core parts; piping and valves 10%; coolant 5% | Coolant is a large share, and tank and sealing demands are severe |
| Barrier | Micron-scale microchannels, vacuum brazing, connectors rated for tens of thousands of mate cycles, CDUs built for 7x24 | Chemical stability and dielectric strength of the coolant, and the sealing system |
Sealing failures account for 60% of all liquid cooling faults, the largest risk. Coolant purity and temperature swings are brutal on seals, now fluororubber and perfluoroelastomer, stable from -45 to above 300 degrees C.
Tighter PUE rules: the second driver
PUE is total data center energy divided by IT equipment energy, and closer to 1 is better. Cooling takes 20-30% of the total, the main lever.
| Policy or benchmark | Target |
|---|---|
| Green and Low-Carbon Data Center Action Plan | National average below 1.5 by the end of 2025; new large projects within 1.25; hub nodes at or below 1.2; green power above 80% at new hub sites |
| Domestic reality | The eight hub clusters average about 1.3, and the best data centers reach 1.04 |
| US and EU | New data centers at 1.2 or lower |
| Industry specification | The Global Computing Consortium calls for new AIDC racks at 50kW or more with liquid cooling as standard |
| Four-ministry document, March 2026 | MIIT, the NDRC, SASAC and the National Energy Administration jointly promote liquid-cooled servers and phase-change liquid cooling |
Air cooling is exiting on compliance grounds as much as technical ones: a new AIDC needs liquid cooling to pass its efficiency review.
Supply chain and procurement view
- System builders. Inspur Information tripled cold plate shipments year on year and sits in the first tier, with Supermicro's rack-level delivery and Dell.
- Infrastructure. Vertiv is now a full-stack provider of CDUs and long-life coolant. Envicool posted 2025 revenue of RMB 6.068 billion, up 32%, has joined Nvidia's MGX ecosystem, builds CDUs for Google, and detects leaks at 0.1 ml per minute against an industry average of 1 ml.
- Certification. Full-system Nvidia certification became the 2026 benchmark, and Nvidia opened its supplier list at the GB300 stage.
- Coolant. After 3M exited fluorinated liquids, substitution runs on fluorinated synthesis or on silicone, mineral and synthetic oils, gated by base chemicals and UL or ASHRAE validation.
- Price war. Envicool's first quarter brought RMB 1.175 billion in revenue, up 26%, but net profit of just RMB 8.6576 million, down nearly 82%. Guoxin judges the industry early and unconsolidated, so quotes have more room than in 2025.
Three directions the technology is moving
- Greener coolant. PFAS rules are tightening, and the market is moving to synthetic fluids.
- Waste heat. The 40-60 degrees C hot water feeds district heating in Europe and northern China, little value in Guangzhou.
- Standard interfaces. OCP and ODCC work advanced in 2026, so CDUs and cold plates from different brands can interoperate.
What this means for SMEs
Few will buy a GB300, but the shift reaches small business three ways.
1. IDC hosting price and availability
A liquid-cooled rack costs more to host, but an air-cooled rack will not take high-power AI cards at all. Many IDCs surcharge above 8-10kW and refuse anything above 20kW on air. State your per-rack power need before you shop.
2. The ceiling on a self-built room
A typical enterprise room with 2-4 racks, precision cooling and a UPS supports 5-8kW per rack. That sets hard limits:
| Equipment | Power per unit | Feasible in-house |
|---|---|---|
| 1U/2U general-purpose server | 300-600W | Yes. A 42U rack takes 15-20 units, about 6-10kW |
| 4-card GPU inference server, RTX 4090 class | 2.5-3.5kW | Yes, but 2-3 per rack. Confirm cooling headroom |
| 8-card GPU training server, A100 or H20 class | 5-8kW | Marginal. Needs 10kW or more per rack, which most older rooms lack |
| GB200/GB300 NVL72 rack | 120-150kW | No. Liquid-cooled room or colocation |
Inference servers up to 4 cards can be self-hosted after a cooling and power assessment; 8 or more belong with a licensed IDC. Self-building's hidden costs usually overtake colocation within three years.
3. Ask about cooling when you buy
For GPU servers bought after 2026, check rack power density, inlet temperature, liquid interface needs and rack load, since a liquid-cooled rack weighs 30-50% more. Customers have bought 8-card servers only to find their air conditioning cannot hold them, throttled at 60% for life.
Our call: who should go liquid now
- Move now: GB200 or GB300 class compute, racks above 20kW, new intelligent computing centers, and hard PUE targets.
- Optimize first: rooms at 10-20kW per rack. Airflow work and aisle containment save 20-30% of cooling energy for far less than a retrofit.
- Stay on air: ordinary rooms at 5-8kW per rack. Air plus containment is enough, and UPS redundancy pays back more.
ACCPC builds and services server rooms: power and cooling design, cabling, UPS and distribution, precision cooling, airflow optimization, racking, monitoring, and IDC colocation with dual feeds, UPS, diesel and 24-hour staffing. If cooling is the worry, we will run a free power and cooling calculation: how many units fit today, and what would let you add more. Call 020-39029800.