When Rack Power Approaches 500kW, Air Cooling Hits Its Ceiling

As AI chips push past 2,000W in 2026 and single-rack power approaches 500kW, air cooling has run out of breath. TrendForce data shows liquid cooling penetration in global AI data centers climbing from 14% in 2024 to 33% in 2025, with 2026 forecasts raised from 40% to as high as 53%; the global liquid cooling market is projected at US$12.7 billion in 2026 and US$42.2 billion by 2030. Yet few liquid-cooling concept stocks have real competitiveness. This article zooms in on the system's most valuable device, the AI GPU liquid cooling CDU, and uses it as a yardstick to measure Envicool, Shenling Environment and Goaland.

In one sentence: if the cold plate carries heat away from the chip, the CDU is the central hub of the entire liquid CDU cooling system, circulating coolant, isolating the server-side loop from the facility-side loop, and managing temperature, pressure and flow. An internal heat exchanger thermally separates the two circuits, so facility water never contaminates servers and pressure fluctuations never propagate across systems. The chain: server cold plate → hot coolant return → CDU heat exchanger → facility water absorbs the heat → cold coolant returns to the server.

CDUs deploy in-rack, in-row or as standalone units, which is why the modular air liquid cooling CDU has become the standard building block of AI data center cooling. A CDU accounts for 30%–40% of a liquid cooling system's total value: 500kW-class units cost roughly RMB 75,000–150,000; 1.4MW–2MW units run RMB 500,000–1,000,000; custom solutions go higher.

Inside a data center air liquid cooling CDU, the pump is the costliest component at 35%–50%, the plate heat exchanger 10%–20%, and piping, valves, sensors, controls and drives the rest. System value concentrates in cold plates, CDUs, quick connectors and manifolds, so a company that outsources all core parts is an assembly plant with thin margins and no pricing power. Only full-chain self-development builds a real moat.

Envicool, the first mainland China vendor to pass NVIDIA's NPN Tier1 system-level certification, is a rack air liquid cooling CDU supplier in the truest sense: cold plates, CDUs, dry connectors and piping all self-developed, with a leak rate below 0.01% and UQD connectors listed in NVIDIA's MGX reference design. Customers include NVIDIA, Google, ByteDance, Alibaba and Tencent. 2025 overseas revenue hit RMB 849 million at a 52.64% gross margin, versus 23.83% at home. 2026 H1 revenue was RMB 3.017 billion (+17.24%), with Q2 marking a clear inflection (revenue +56.67% QoQ, net profit +1,934% QoQ). Its weakness: immersion capabilities are thin, and margins slipped on product mix.

Shenling Environment, Huawei Ascend's core liquid cooling supplier for fifteen years, is a server rack air liquid CDU manufacturer covering both cold-plate and immersion routes, serving Huawei, ByteDance, Tencent, Alibaba, Baidu, Meituan, Kuaishou, JD.com and the telecom operators. 2025 data services revenue reached RMB 2.343 billion (+51.42%), 55.67% of total, with orders up about 72% YoY; in-hand orders exceed RMB 2 billion. Its edge is the order book, but it leans toward system integration, carries lower margins (20%–24%), and its overseas breakthrough is still early.

Goaland is one of the few companies covering cold-plate, immersion and containerized routes, with a strategic partnership with Midea Building Technologies. Note that its headline Saudi contract (RMB 361 million) is UHV flexible-DC cooling, not AI liquid cooling. 2025 revenue was RMB 989 million (+43.09%) and turned profitable; its Huizhou AI liquid cooling base (30,000 systems a year) comes online in Q3 2026. As high density data center cooling becomes the norm: by near-term scale and the NVIDIA ecosystem, Envicool leads; by domestic substitution and the Huawei ecosystem, Shenling has the most upside; Goaland is distinctive but smaller in absolute terms.

Risks remain. Cold plate is mainstream today, but immersion is irreplaceable above 500kW per rack. At that density, air cooling hits its ceiling even with solutions like the rack rear door heat exchanger. Customer concentration is another risk (Huawei accounted for 18.38% of Shenling's data services sales), as is capacity ramp-up. The investment logic: near term, cold-plate delivery; mid term, overseas breakthrough; long term, immersion penetration. Liquid cooling is a long-term necessity of AI infrastructure, and the companies that survive and scale will keep compounding value.

To navigate these industry risks and capture long-term liquid-cooling opportunities, operators need a versatile hardware partner capable of covering both near-term cold-plate deployments and future high-density requirements. The more AI evolves, the more heat it produces—and that heat has to go somewhere. EXTRCOOL's liquid-cooling range covers everything from CDUs and cold plates to manifolds and the full piping network. You can pick from our ready-made stock or commission a design custom-built for your installation. Our approach delivers outstanding heat dissipation, lower energy draw, and a reduced environmental footprint—demonstrated through reliable service at gigawatt-scale sites around the world. When every minute of uptime matters, EXTRCOOL stands as your trusted partner. Take a look at our full portfolio or submit a request for a tailored quote at: https://www.extrcool.com