China’s breakthrough cooling technology can drop temperatures 30 degrees in 20 seconds: could an AI data center cooling revolution reshape the global competitive landscape?

Chinese scientists have recently published a “dissolution-pressure cooling” technology that claims to deliver an ultra-fast 30°C temperature drop in 20 seconds, a breakthrough aimed squarely at the severe cooling challenges faced by AI data centers. Global annual electricity consumption by AI data centers is expected to reach 90 terawatt-hours in 2026, roughly one-seventh of Japan’s total electricity use, and traditional cooling systems can no longer handle the heat generated by next-generation GPUs that now exceed 30 kilowatts per rack. If commercialized, the technology could cut AI workload costs by 30% to 40% and potentially redraw China’s strategic position in the global AI infrastructure race.

Pressure-driven chemistry: ammonium thiocyanate overturns conventional cooling logic

The technology, developed by a team led by Professor Li Bing at the Institute of Physics, Chinese Academy of Sciences, relies on the unusual dissolution behavior of ammonium thiocyanate (NH₄SCN) in water to create an innovative system that combines the efficiency of solid-state cooling with the flow properties of a liquid. Data published in Nature Communications show that when pressure is applied to a saturated solution, a large amount of ammonium thiocyanate dissolves and releases heat; when the pressure is released, the salt recrystallizes and rapidly absorbs surrounding heat. At room temperature, it can produce a temperature drop of nearly 30 Kelvin, and at higher temperatures the cooling range can reach 54 Kelvin. William Averdieck, co-founder of Cambridge spinout Barocal, said this barocaloric technology merges the refrigerant and the heat-transfer medium into one, solving the long-standing heat-transfer bottleneck that has plagued solid-state cooling. Experimental measurements show that the system can deliver 67 joules of cooling per gram of refrigerant, with a cooling efficiency as high as 77%, far exceeding existing solid-state barocaloric materials.

The breakthrough lies in mimicking the physical process of squeezing a wet sponge: the pressurization stage is like wringing water out of a sponge, as the salt dissolves and releases heat; the depressurization stage is like the sponge soaking water back up, as the salt recrystallizes and absorbs heat. When pressure is applied to the material, long molecular chains shift from a disordered state to an ordered one, releasing heat as energy levels change. The process is reversible at ambient temperature and does not depend on chemical refrigerants with high global warming potential, unlike conventional vapor-compression cooling. Dr. Kwan Pui Man, director of the Information Technology Services Center at Hong Kong University of Science and Technology, said such zero-carbon cooling technologies complement the school’s largest immersion cooling system in Hong Kong, deployed in October 2024, which has already cut cooling energy use by more than 80% and brought power usage effectiveness, or PUE, below 1.1. The research team said the liquid system overcomes the fatal weakness of poor thermal conductivity in solid-state barocaloric materials, because the liquid acts as both refrigerant and heat-transfer carrier and can make direct contact with the surface being cooled.

Solving AI computing power overheating: a technical race for a market worth hundreds of billions of dollars

The global data center cooling market is growing at explosive speed, and is expected to rise from $10.141 billion in 2025 to $31.344 billion in 2034, a compound annual growth rate of 13.43%. Liquid cooling solutions, including immersion cooling and direct-to-chip cooling, already account for more than 38% of new high-density deployments. The reason is simple: AI chip power consumption is climbing exponentially. NVIDIA’s H100 has a thermal design power of 700 watts, while the newly launched Blackwell B200 series pushes past 1,000 watts, rendering traditional air cooling systems ineffective. Microsoft announced in September 2025 that it had successfully tested an on-chip microfluidic cooling system, and in server tests simulating Microsoft Teams meetings, cooling performance was three times better than advanced cold plate technology, while temperature rise inside the GPU chip fell by 65%. Christian Belady, vice president of cloud operations and innovation at the company, said microfluidics make it possible to build for higher power density in a smaller footprint while improving PUE and lowering operating costs.

As Hong Kong positions itself as a data center hub in Asia-Pacific, the cooling market there is projected to grow from $189 million in 2026 to $466 million in 2031, a compound annual growth rate of 19.72%. Market researchers said Schneider Electric has combined its acquisition of Motivair with the integration of cooling distribution units into its Galaxy power system lineup to offer turnkey liquid cooling solutions tied into EcoStruxure monitoring. LiquidStack, which focuses on immersion cooling, has set benchmark PUE records of 1.01 in multiple markets. TrendForce analysts forecast that as North American cloud providers ramp up spending and sovereign cloud projects expand globally, AI server shipments in 2026 will rise more than 20% from a year earlier, pushing penetration of third-generation semiconductors such as silicon carbide, or SiC, and gallium nitride, or GaN, in data center power systems from 17% in 2026 to more than 30% by 2030.

Deloitte’s Technology, Media and Telecommunications Predictions 2025 says generative AI is driving a surge in data center energy consumption, with annual electricity use by global AI data centers expected to reach 90 terawatt-hours in 2026, equal to the yearly power consumption of medium-sized countries such as the Netherlands or Argentina. The International Energy Agency says that once AI is fully integrated into search engines such as Google, the electricity needed for a single query could rise from 0.3 watt-hours for traditional search to 2.9 watt-hours, based on ChatGPT. At 9 billion searches a day, that would add up to 10 terawatt-hours of extra annual power demand. That energy pressure is forcing the industry to look for a breakthrough cooling solution, and the 20-second temperature-drop feature of China’s supercooling technology fits the sudden high-load bursts that come with AI training.

The self-reliance puzzle: building a full chain from chips to cooling

China’s supercooling breakthrough fits closely with its semiconductor self-sufficiency strategy, especially after the United States continued tightening export controls on advanced chips and high-bandwidth memory, or HBM, to China. TrendForce reported that YMTC, China’s largest NAND maker, is using through-silicon via, or TSV, advanced packaging to move into the DRAM market, with the goal of producing HBM chips needed for AI processors. It plans to put part of the capacity at a new third wafer fab in Wuhan into DRAM production. At the same time, CXMT, China’s leading DRAM maker, has resumed large-scale capital spending and is expanding DDR5 DRAM and fourth-generation HBM3 lines at its Hefei site, with mass production of HBM3 expected by the end of 2026. Tom’s Hardware reported that China plans to begin domestic HBM3 production by the end of 2026, a key piece in its AI chip self-reliance puzzle.

This vertical strategy, from chips to cooling, reflects the deeper logic of China’s technology policy. Think tank analyses say the Chinese government is continuing to invest through the third phase of the National Integrated Circuit Industry Investment Fund in areas such as AI chip design platforms and semiconductor manufacturing equipment, with the aim of reaching a 70% self-sufficiency rate for AI chips by 2027. With imports of key components such as HBM constrained, Chinese companies including Baidu, Alibaba, Huawei and Cambricon are accelerating development of their own AI chips, and the ability to master efficient cooling will directly affect how reliably those chips run in hyperscale compute clusters. The Institute of Physics team at the Chinese Academy of Sciences specifically noted in its paper that the liquid barocaloric cooling system can achieve significant cooling effects with just 1 bar of pressure change, a low-pressure operating feature that improves safety and commercial viability.

It is also worth noting that China’s East Data West Computing project is pushing data centers toward energy-rich regions, and AI data centers equipped with energy storage systems are becoming standard in large campuses. Market research shows that global installed capacity of energy storage systems for AI data centers is expected to jump from 1.57 GWh in 2025 to 8.8 GWh in 2030, a compound annual growth rate of 46.1%, with the share of medium- and long-duration storage systems of one to four hours rising sharply to support power arbitrage and grid services. This infrastructure-level upgrade, combined with the cost cuts that supercooling could bring, would give China’s AI industry a meaningful TCO advantage.

Commercialization outlook: a long road from lab breakthrough to industry use

Although the lab data are striking, the commercial case for this supercooling technology still faces multiple hurdles. Public information does not disclose specific prototype details, cycle-life test data or a commercialization timeline, in contrast with Microsoft’s microfluidic cooling, announced in September 2025, and HKUST’s immersion cooling system, launched in October 2024, both of which have already moved into large-scale deployment. Averdieck of Barocal has acknowledged, when commenting on similar technologies, that turning theoretical potential into a commercial solution depends on safety, reliability, maintainability and energy efficiency. Liquid barocaloric cooling systems will need precise pressure-control mechanisms, corrosion-resistant materials and long-term stability validation, all of which are part of the route from lab to data center.

Still, rapid consolidation in the global data center cooling industry is creating room for new technologies. Market leaders Vertiv and Schneider Electric have reinforced their positions through continued R&D investment and strategic acquisitions, and together account for nearly 35% of the market. But innovators focused on AI and HPC workloads, such as Envicool, iTeaq and Deep Green, are gaining ground in niche segments with specialized offerings, including Deep Green’s model of using waste heat from data centers to heat swimming pools. Chinese players such as Shenling have maintained a strong position in the domestic market through government partnerships and climate-adapted custom solutions. Industry analysts say AI remains the main force driving data center growth, but power, water and land constraints are pushing the sector to look for breakthrough solutions.

For corporate decision-makers, the strategic value of the technology lies in the possibility that it could redefine the economics of AI infrastructure. HKUST’s immersion cooling case shows that advanced cooling can cut PUE from 1.5 to 2.0 in a traditional data center to below 1.1, while reducing energy use by as much as 80%. If China’s supercooling system can deliver similar performance while keeping its rapid response characteristics, it would be highly attractive for cloud providers planning to deploy hundreds of thousands of GPUs over the next five years. JLL forecasts that global data center capacity will add nearly 100 gigawatts from 2026 to 2030, roughly doubling existing scale and implying about $200 billion in infrastructure investment, with cooling systems accounting for about 15% to 20%, or a potential market of $30 billion to $40 billion.

A new variable in the geotech race: how cooling could shape AI dominance

The timing of China’s supercooling development is notable, coming as the global race for AI computing power intensifies and energy bottlenecks become more visible. If commercialized successfully, the technology would reinforce China’s broader push in robotics, semiconductors and AI chips. Scholars at Tsinghua University in Beijing say the Chinese government sees semiconductors as the foundation of modern technology policy, using state investment funds to push self-reliance across the full chain, from chip design and manufacturing equipment to materials, with the goal of building a semiconductor industry entirely dependent on domestic intellectual property by 2030. In that context, cooling technology becomes a key infrastructure element for scaling domestic AI chips.

The link between technical breakthroughs and geopolitical advantage is not linear. Microsoft and NVIDIA deepened cooperation in 2025, deploying Blackwell GPUs with GB200 NVL72 rack-scale systems, and using AI to identify new coolant prototypes with immersion-cooling potential in 200 hours, a process that would normally take months or years. That shows the systemic edge Western tech giants still hold in AI-driven innovation loops. Microsoft’s microfluidic cooling technology has already been tested on real workloads and can send coolant directly into microchannels inside the chip, achieving unprecedented heat removal efficiency. By contrast, China’s supercooling technology has not yet published any data center environment test results or any comparison against existing liquid cooling systems.

Over the next three years, data center cooling is likely to evolve along a path of short-term liquid cooling expansion, medium-term hybrid systems, and long-term chip-level thermal management. MIT research has found that training a large language model such as GPT-3 can produce more than 300 tons of carbon dioxide equivalent emissions, and the energy efficiency of cooling systems directly shapes the carbon footprint of the AI industry. In that context, zero-carbon refrigerant supercooling does have environmental advantages, but its commercial competitiveness will ultimately depend on system integration costs, maintenance complexity and compatibility with existing data center infrastructure. The real question in this race is not who first masters a single breakthrough technology, but who can move innovation fastest into a complete AI infrastructure stack, from chips and cooling to power and software optimization.

For data center operators in Hong Kong and the broader Asia-Pacific region, a prudent strategy would be to stay technology-neutral and watch how multiple cooling approaches are validated. As HKUST plans to extend immersion cooling to its eight-story high-performance data center, the local market is becoming a live testbed for different innovations. Whether this supercooling technology can move from lab breakthrough to industry standard will ultimately depend on whether it can show a scalable prototype within the next 12 to 24 months and reach commercial deployment thresholds across reliability, cost and energy efficiency, just as AI itself must prove its promise in the real world.