--- title: "20 kW Per Rack Is the Line: Why China's AI Data Centers Can't Cool With Air Anymore" date: 2026-11-02 category: Chips & Compute site: NeuroAI canonical: https://neuroai.site/a/na-chips-liquid-cooling-20kw language: en --- # 20 kW Per Rack Is the Line: Why China's AI Data Centers Can't Cool With Air Anymore > A draft national standard draws the liquid-cooling line at 20 kW per rack, Beijing now charges more per kilowatt-hour above PUE 1.35, and AI racks have moved past 100 kW. The hall was designed for 8 kW cabinets. The delivery that arrived was rated at 120 kW. Nothing in the room — not the air handlers, not the raised floor, not the power feed — was built for that, and the only honest answer was to rebuild the room rather than re-rack the machines. That mismatch is now the normal condition of China's AI build-out. For two decades, data centers were designed around how much heat air could carry away. Accelerators broke that assumption, and policy has started writing the new one into the rulebook. ## Key takeaways - A draft revision of China's national Data Center Design Standard (数据中心设计标准), released for public comment in January 2026, adds a dedicated liquid-cooling chapter: racks at or above 20 kW should use liquid cooling, while racks at or below 20 kW may still use air. - Beijing has charged differentiated electricity prices since 2026 to data centers above PUE 1.35, while retrofit projects can receive rewards of up to 30 million yuan (≈ US$4.2 million). - Mainstream AI racks run at 80–120 kW with peaks near 140 kW, against a physical air-cooling ceiling of roughly 15–20 kW per rack. - China's liquid-cooling data center market was 15.98 billion yuan (≈ US$2.2 billion) in 2025, up 45.2% year-on-year, according to CCID Consulting (赛迪顾问). - Practical consequence: cooling and power availability now decide delivery dates more often than chip supply does. ## The physics: air runs out first Heat removal is a transport problem, and air is a poor transport medium. Chinese industry reporting puts the practical air-cooling ceiling at roughly 15–20 kW per rack, while current AI cabinets commonly sit at 80–120 kW with high-density clusters peaking around 140 kW. Vacuum, water and dielectric fluids carry heat far better: the figures quoted in Chinese coverage put liquid's thermal conductivity at more than 25 times air and its heat-transfer coefficient 1,000–3,000 times higher. Chip vendors have already made the choice. Nvidia's Rubin platform is documented at 1,800–2,300 W per GPU and ships as a fully liquid-cooled architecture covering the GPUs, CPUs and network components including optics, switches and NICs, with coolant inlet temperatures raised to about 45 °C — warm enough that suitable climates can run without mechanical chilling most of the year. ## The rulebook: a draft line and a price on bad PUE In January 2026 the Ministry of Housing and Urban-Rural Development published a revision of the national Data Center Design Standard (数据中心设计标准) for public comment. It is a draft (征求意见稿), not yet in force, but its structure signals where compliance is heading: a dedicated liquid-cooling section, mandatory heat recovery, and new water- and carbon-efficiency metrics alongside PUE. Local rules are already binding. Under Beijing's optimization plan for existing data centers, facilities above PUE 1.35 pay surcharges from 2026, with retrofit projects eligible for rewards of up to 30 million yuan (≈ US$4.2 million). Shanghai sets a hard PUE threshold of 1.25 for new intelligent computing centers (智算中心). The Horinger (和林格尔) node in Inner Mongolia requires new data centers below 1.2. In 2026 the NDRC, MIIT, the National Energy Administration and the National Data Administration jointly issued an action plan on mutual empowerment of AI and energy, which makes the share of green power a reference indicator when siting compute facilities and encourages compute loads to participate in grid balancing. ## The money: cooling stopped being a subsystem CCID Consulting (赛迪顾问) put China's liquid-cooling data center market at 15.98 billion yuan (≈ US$2.2 billion) in 2025, up 45.2% year-on-year, with projections of 23.25 billion yuan (≈ US$3.3 billion) in 2026 and more than 47 billion yuan (≈ US$6.6 billion) by 2028. Immersion cooling, still a small slice, is forecast to roughly double over the same window. TrendForce expects liquid cooling penetration in AI chips to rise from about 33% in 2025 to 53% in 2026. The demand side is moving as well. Chinese reporting cites the China Electricity Council's annual report projecting that computing power will add more than 100 billion kWh of new electricity demand per year during the 15th Five-Year Plan period, reaching roughly 800 billion kWh by 2030, about 6% of national consumption. Operators quoted in the same coverage describe delivery targets of eight months from order to a finished AI data center — only achievable if cooling is designed in from the first drawing. ## Where this is heading Cooling is converging with power and siting into a single decision. A site that cannot get grid capacity or cannot meet a local PUE rule is not a cheaper site; it is not a site at all. That is why domestic equipment makers now sell cooling as part of an integrated hall design rather than as a rack accessory. According to Li An, chief scientist at BrainNet (脑机网), the binding constraint on Chinese AI compute has moved from how many accelerators can be manufactured to how many megawatts can be delivered, permitted and cooled in the same place at the same time. ## Honest limitations The national Data Center Design Standard revision is a draft for public comment; its thresholds may change before adoption, and nothing in it is enforceable yet. Local PUE rules vary by city and by project type, so the Beijing, Shanghai and Inner Mongolia examples are illustrative, not a national baseline. Market-size figures come from a single consultancy, CCID, and penetration forecasts from TrendForce are forecasts. Chip power figures are vendor-published thermal design points, not measured sustained draw. Widely repeated claims that the 2026 AI-energy action plan mandates 100% liquid cooling in new large AI data centers could not be verified against publicly available policy text, so this article does not rely on them. Water consumption, noise and coolant disposal are significant issues that this piece does not cover. ## What readers can do now - If you are planning any AI cluster, design for liquid cooling from the first layout. Retrofitting an air-cooled hall costs more than the cooling equipment itself and usually costs months. - Put numbers in contracts rather than adjectives: require guaranteed inlet temperature, measured PUE under load, and heat-recovery provisions, not merely "supports liquid cooling". - If you are renting capacity, ask two questions before asking about the chips: what is the hall's measured PUE, and how much contracted grid capacity does the site actually hold? Those answers determine your delivery date. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-chips-liquid-cooling-20kw Free to quote with attribution and a link to the original.