Walk into any modern AI data center, and you’ll immediately be hit in the face by heat. Even before you’ve set your eyes on the racks, you feel the output of rows upon rows of servers pushing air through increasingly dense racks that were never designed for processors drawing the sheer power needed to run huge AI models. Multiply that across dense training clusters, and the thermal envelope becomes one of today’s greatest engineering constraints.
Now, according to reporting by the South China Morning Post, China is accelerating investments in liquid cooling technology. Dozens of Chinese companies — including Envicool, whose shares have more than tripled over the past year — have announced plans to expand into liquid-cooling systems to prevent AI racks from overheating. With thermal envelopes riding, this isn't a problem unique to Chinese data centers, either.
Since 2022, Beijing has pushed its “Eastern Data, Western Computing” strategy, which shifts data processing toward western provinces with access to renewable energy. According to China’s Ministry of Industry and Information Technology, the country has already built 42 “10,000-card” intelligent computing clusters, with total capacity exceeding 1,590 exaflops. Those clusters aren’t populated with low-power CPUs but built around high-end AI accelerators that concentrate enormous heat loads into increasingly compact footprints.
Air cooling won’t cut it
As an example, Nvidia’s B200 operates at more than 1,000 watts per processor, and its GB200 NVL72 rack-scale platform, which integrates 72 Blackwell GPUs and 36 Grace CPUs, ships in a liquid-cooled configuration as standard and out of necessity.
As chip thermal design power continues to climb, air cooling faces two limits. First, the volumetric heat capacity of air constrains how much thermal energy can be removed per unit time without driving airflow and energy use to impractical levels. Secondly, high airflow requirements complicate rack density and overall facility design. At multi-kilowatt per rack densities, these air-cooled systems become increasingly inefficient.
Liquid, by contrast, offers orders of magnitude higher heat capacity and thermal conductivity. Direct-to-chip cold plates circulate coolant through sealed loops attached to processors, removing heat at the source. More aggressive approaches, including immersion and spray cooling, place components in direct contact with dielectric fluids. This has created an all-new bill of materials consisting of cold plates, manifolds, leak-detection systems, and cooling distribution units, among other components.
Global suppliers like Vertiv are prominent in this stack, but Chinese firms are racing to localize capabilities. Envicool, whose customers include Nvidia, Alibaba, and Tencent, has become one of the most obvious domestic players, with Goldman Sachs and UBS both raising price targets on its stock, citing demand tied to AI and energy efficiency.
Growing model sizes
(Image credit: Tom's Hardware)
... continue reading