China’s Liquid Cooling Orders Are Filling Up as AI Racks Overwhelm Air Cooling

Chinese liquid cooling suppliers reportedly have orders stretching through year-end, while a new immersion cooling facility is testing four fluids under operating conditions. The rush signals a conflict that data center operators can no longer postpone. AI racks are becoming too dense for conventional air cooling, but the most aggressive alternative still lacks mature standards and extensive field evidence.

A 36Kr news item, citing CCTV Finance, reported that China’s first immersion-cooled computing pod began operating in Guiyang, Guizhou. The facility submerges servers in electrically insulating coolant and tests silicone oil, synthetic oil, recycled transformer oil, and fluorinated fluid. Researchers want to identify a coolant that balances heat transfer, equipment compatibility, safety, and operating cost.

The development does not establish that immersion cooling has already displaced cold plates. It shows that coolant validation has become an infrastructure priority. NVIDIA’s GB200 NVL72 can draw approximately 120 kilowatts per rack, according to its system documentation. At that density, cooling is part of the computing architecture, not a background facility service.

Cold plates remain the main liquid cooling route for many current AI deployments. They transfer heat from chips into a liquid loop without submerging the complete server. Immersion removes more air-moving equipment and can cool additional components, but it also exposes boards, cables, connectors, seals, and optical equipment to fluid for years.

That creates the industry’s central contest: proven cold-plate deployment versus broader immersion cooling. The winning route will not be determined by laboratory thermal performance alone. Reliability, service procedures, coolant supply, hardware warranties, and lifecycle economics will decide what operators deploy at scale.

China’s Immersion Cooling Trial Moves Coolant Risk Into the Data Center

The Guiyang project matters because it tests the weakest link in immersion cooling, the long-term relationship between fluid and server materials.

The facility resembles a modular computing pod rather than a conventional server hall. Its servers sit inside tanks filled with dielectric coolant, which transfers heat without conducting electricity. The reported experiment compares four fluid families instead of optimizing a single commercial formulation.

Silicone oil offers electrical insulation and relatively stable thermal behavior. Synthetic oils can be engineered around viscosity, flash point, and oxidation resistance. Recycled transformer oil offers a potential cost advantage, but its consistency and contamination profile require careful control. Fluorinated fluids can support demanding thermal designs, although environmental characteristics and supply conditions vary by chemistry.

The trial’s value lies in comparison under similar operating conditions. A coolant that looks effective in a short thermal test can still create problems after prolonged contact with server components. It can swell elastomers, weaken adhesives, affect cable jackets, dissolve labels, alter connector materials, or carry contaminants around the tank.

The Open Compute Project has published immersion requirements, fluid specifications, equipment design guidance, and material compatibility work. That body of documentation illustrates how many interfaces require validation. An immersion system combines IT equipment, fluid chemistry, pumps, heat exchangers, tanks, controls, maintenance processes, and building infrastructure.

The reported Guiyang facility is therefore more than a cold bath for hot servers. It is a materials and systems test bed. Researchers will need to monitor fluid properties, component degradation, thermal stability, electrical behavior, contamination, and maintenance requirements over time.

The claim that participating suppliers have orders booked through the end of 2026 indicates immediate commercial pressure. However, the public report does not identify every supplier, disclose order values, or separate firm contracts from preliminary reservations. It should be read as a demand signal, not audited market data.

The same caution applies to descriptions of a trillion-yuan cooling opportunity. A large addressable market can include equipment, fluids, installation, retrofits, controls, maintenance, and supporting infrastructure over many years. That framing is not equivalent to current annual revenue.

What the project clearly establishes is narrower and more useful. Chinese operators are investing in practical evidence about coolant choices because AI hardware has made thermal management a deployment constraint. Testing four fluid families also suggests that no single chemistry has settled the market.

The experiment creates the article’s central tension. Immersion cooling promises to remove thermal limits that constrain dense computing, yet it moves reliability risk from fans and airflow into chemistry and materials. Operators gain thermal headroom only by accepting a new set of engineering dependencies.

AI Rack Density Is Turning Cooling Into a Compute Constraint

Liquid cooling demand is rising because modern AI systems concentrate far more heat inside each rack than conventional server rooms were designed to remove.

A traditional air-cooled data center moves chilled air across server components and carries heated air back toward cooling equipment. This arrangement remains practical at moderate rack densities. It becomes harder to manage when accelerators, high-bandwidth networking, and power equipment concentrate substantial electrical load inside a confined footprint.

Every watt consumed by a server eventually becomes heat. More computing inside one rack therefore means more heat must leave the same area. Operators can increase airflow, enlarge fans, separate hot and cold aisles, lower supply temperatures, or spread equipment across additional racks. Each response introduces limits involving energy, floor space, noise, pressure, and infrastructure.

NVIDIA’s approximate 120-kilowatt figure for the GB200 NVL72 provides a useful scale. It is not representative of every data center rack, but it shows where rack-scale AI systems are heading. The rack also includes liquid cooling and leak detection as integrated design concerns.

This changes procurement. A customer cannot simply order accelerators and decide later how to cool them. Power delivery, liquid distribution, heat rejection, rack layout, monitoring, and service access must be planned together. A delay in one supporting system can leave expensive computing hardware unable to enter production.

The wider energy trend adds pressure. The International Energy Agency expects global data center electricity demand to more than double to roughly 945 terawatt-hours by 2030. AI is the largest driver in that outlook, according to its energy analysis.

The IEA also reported that data center electricity use increased 17 percent during 2025, while demand from AI-focused facilities grew even faster. Its April 2026 data center update described physical bottlenecks involving grids, transformers, chips, and project approvals.

Cooling cannot eliminate the electricity consumed by computation. It can reduce supporting energy, preserve performance, and make dense hardware deployable. Those gains matter because operators measure infrastructure efficiency through power usage effectiveness, or PUE. PUE compares total facility energy with the energy delivered to IT equipment.

China’s policy environment reinforces that pressure. National green data center criteria published for 2025 called for qualifying facilities to maintain PUE at or below 1.30. The criteria also encouraged continued optimization of cooling architecture, power systems, water use, and operating controls.

Dense racks make that target harder to reach with brute-force airflow. Larger fans and colder air consume additional facility energy. Liquid transports heat more effectively than air and can move it closer to an external heat exchanger without forcing enormous volumes of air through server chassis.

The immediate pressure falls on data center developers, colocation providers, server manufacturers, and thermal equipment suppliers. Developers must reserve space and power for cooling infrastructure. Colocation providers must support customer hardware with different liquid interfaces. Server manufacturers must qualify components for new thermal environments.

Thermal suppliers face a different challenge. Demand can expand faster than their installation capacity, but customers expect multiyear reliability. A crowded order book is useful only if vendors can deliver pumps, coolant distribution units, connectors, tanks, fluids, monitoring systems, and field support without weakening quality control.

That is why cooling orders extending through year-end are plausible as an industry signal. AI infrastructure schedules are pulling thermal equipment into the critical path. The open question is which liquid architecture captures most of that spending.

Cold Plates Have the Lead, but Immersion Targets More Heat

Cold-plate cooling offers a lower-risk transition from air, while immersion cooling seeks to remove more heat with fewer airflow dependencies.

A cold plate is a metal heat exchanger attached directly to a processor or accelerator. Coolant flows through internal channels and carries heat into a facility loop. Other server components can remain air-cooled, allowing operators to preserve familiar rack layouts and service procedures.

That hybrid quality gives cold plates an adoption advantage. Operators can address the hottest chips without placing every component in fluid. Technicians can still remove standard server trays, replace parts, and use established diagnostic workflows. Hardware vendors can adapt existing platforms rather than redesigning the entire server around a tank.

Cold plates also have their own engineering risks. A pressurized loop introduces connections, manifolds, seals, corrosion concerns, water-quality requirements, and leak exposure near electronics. Open Compute Project guidance calls for pressure testing and compatibility checks across metals, polymers, elastomers, and cooling fluid.

Immersion takes a different approach. In a single-phase system, servers remain submerged in a fluid that stays liquid while absorbing heat. Pumps move the warmed fluid toward a heat exchanger. In a two-phase system, fluid boils at component surfaces and condenses elsewhere in a controlled enclosure.

Submerging the server can cool memory, voltage regulators, networking components, and other heat sources that cold plates may not cover directly. It can reduce dependence on server fans and provide more uniform temperatures. It can also support dense layouts that would create complex airflow problems.

Those benefits explain why the Guiyang experiment focuses on fluid rather than tank geometry alone. Coolant becomes part of the server’s operating environment. Its viscosity affects pumping and heat transfer. Its chemical stability influences useful life. Its interaction with materials affects reliability and warranties.

The four reported fluid categories represent different compromises. A lower-cost liquid has limited value if it requires frequent replacement or damages equipment. A high-performing fluid can still fail commercially if supply is constrained or environmental handling becomes difficult. No thermal metric settles those tradeoffs by itself.

The route comparison also depends on the facility. A new AI data center can be designed around liquid loops, higher floor loading, heat exchangers, and maintenance areas. A retrofit must work around existing pipes, raised floors, power distribution, aisle geometry, and operating workloads.

For an existing air-cooled facility, cold plates often provide a more incremental path. The operator can convert selected high-density racks while retaining air cooling elsewhere. Full immersion can demand new tanks, lifting equipment, fluid-handling procedures, modified cabling, and different spare-parts practices.

For a modular computing pod, immersion becomes more attractive. Designers control the enclosure, rack substitute, cooling loop, and service zone as one system. The Guiyang facility’s pod format therefore fits the technology being tested. It avoids forcing tank-based cooling into a hall built around conventional racks.

Immersion vendors argue that falling coolant costs and better system integration will narrow the operational gap. That proposition remains under evaluation. The new trial can help determine whether less expensive formulations maintain thermal and material performance over extended periods.

Cold plates, meanwhile, continue to improve. Suppliers are developing higher-flow designs, better quick disconnects, improved manifolds, and facility loops for hotter water. Server platforms increasingly arrive with liquid cooling designed into the product, reducing the amount of custom integration required on site.

This is not a simple sequence in which immersion automatically replaces cold plates. Different approaches can serve different densities, facilities, and maintenance models. Cold plates can remain dominant for mainstream AI racks while immersion expands in specialized clusters, modular systems, edge installations, or environments where airflow is especially constrained.

The reported Chinese trial still increases competitive pressure. If it validates a lower-cost fluid across common server materials, immersion vendors gain a stronger deployment case. If degradation, handling, or service issues persist, cold plates preserve their advantage despite immersion’s thermal potential.

Coolant Compatibility Is the Risk Behind the Order Boom

The biggest uncertainty is not whether immersion removes heat, but whether complete systems remain safe, serviceable, and reliable for years.

Material compatibility is an immediate concern because servers contain many substances that were never selected for continuous immersion. Printed circuit boards, solder masks, thermal interface materials, cable insulation, gaskets, labels, plastics, and optical components can respond differently to the same fluid.

The Open Compute Project’s immersion design guidance warns that chemical and physical interactions can degrade material properties or component function. It recommends testing cables, circuit boards, packages, optical fibers, and passive components under conditions that represent actual use.

Testing must also account for time. A server can operate normally during an early demonstration and still develop failures after fluid absorbs contaminants or components undergo repeated thermal cycles. Operators need evidence covering aging, oxidation, acidity, viscosity changes, dielectric strength, and particulate transport.

Fluid maintenance creates another uncertainty. Operators must establish sampling intervals, filtration practices, storage rules, spill procedures, end-of-life handling, and replacement criteria. Staff need equipment and training that differ from work in an air-cooled server hall.

Serviceability can affect downtime. A technician removing an immersed server must drain or contain residual fluid before diagnosis and repair. Components may require cleaning before shipment to another facility. Spare parts and replacement units must be compatible with the selected coolant.

Warranty coverage matters as much as technical compatibility. An operator can validate a fluid internally and still face restrictions from server, drive, cable, or networking vendors. Commercial adoption becomes easier when hardware makers qualify complete combinations of equipment, fluids, tanks, and operating conditions.

Standards are developing, but they remain incomplete. ASTM opened a 2026 work item for specifications covering two-phase immersion liquids. Its stated rationale says the market lacks a product specification for those fluids, highlighting the need to standardize safety and quality.

ISO is also developing guidance for applying liquid cooling in data centers. The proposed work addresses architectural, mechanical, electrical, and communication considerations. Its early status shows that deployments are advancing while international guidance is still being assembled.

Independent testing organizations are responding. UL Solutions operates an immersion cooling program covering equipment integration, electrical safety, fire safety, and fluid evaluation. Such programs can reduce uncertainty, but certification does not replace site-specific engineering.

Environmental performance requires similar care. “Liquid cooling” is not automatically synonymous with low environmental impact. Outcomes depend on coolant chemistry, pumping energy, heat-rejection equipment, water use, leakage, fluid lifespan, disposal, and the electricity supplying the data center.

Fluorinated fluids deserve particular scrutiny because chemicals within that broad category can have very different environmental profiles. Operators need formulation-specific information rather than conclusions based only on a family name. Regulatory changes can also alter supply and disposal requirements during a facility’s operating life.

Other fluids introduce different risks. Hydrocarbon-based formulations can raise questions involving flammability, oxidation, and material interaction. Silicone fluids have distinct viscosity and contamination considerations. Recycled transformer oil requires confidence in purification, consistency, and trace contaminants.

The Guiyang project’s inclusion of recycled transformer oil is commercially interesting because coolant expense can influence immersion economics. Yet a lower acquisition cost cannot compensate for shortened component life, higher maintenance, or unpredictable batches. The trial needs to measure lifecycle performance, not only initial thermal results.

The report does not disclose test duration, exact fluid formulations, server configurations, evaluation thresholds, or independent review arrangements. It also does not provide failure data. Those omissions do not invalidate the project, but they limit what readers can conclude from its launch.

Order-book claims carry their own uncertainty. Customers can reserve production capacity before final data center schedules are fixed. Grid connections, chip deliveries, financing, construction, and permits can change deployment timing. Suppliers may therefore report strong demand while recognizing revenue over a different schedule.

The most useful reading is cautious. The order boom confirms that cooling has become urgent. It does not prove that every supplier can scale, every project will arrive on time, or immersion will replace cold plates across the market.

The Liquid Cooling Supply Chain Is Becoming Part of AI Infrastructure

The shift from air to liquid distributes value and risk across a much wider supply chain than the cooling unit itself.

A liquid-cooled AI deployment needs heat exchangers, coolant distribution units, pumps, valves, manifolds, hoses, quick disconnects, sensors, controls, and facility piping. Immersion adds tanks, dielectric fluids, filtration, fluid monitoring, handling equipment, and compatible server designs.

This creates opportunities for thermal specialists, chemical producers, component manufacturers, engineering contractors, and data center operators. It also creates coordination risk. A failure at one interface can delay an entire cluster even when accelerators and networking equipment are available.

Connectors provide a simple example. They must support repeated maintenance, control pressure loss, and resist leaks across many service cycles. A connector that performs well in one coolant or temperature range may not suit another. Standardized interfaces can reduce custom engineering, but only after vendors agree on requirements.

Coolant distribution units create another dependency. These systems isolate the facility water loop from the technology cooling loop and regulate flow, temperature, and pressure. Their controls must coordinate with servers and building systems, especially when workload changes cause rapid shifts in heat output.

For immersion systems, chemical suppliers become infrastructure partners. They must provide consistent formulations, technical data, compatibility evidence, quality control, replenishment capacity, and end-of-life guidance. Data center operators cannot treat coolant as a generic consumable if changing it affects warranties and material performance.

Server manufacturers are also being pushed toward closer thermal integration. Systems designed for immersion can remove unnecessary fans and modify component placement, connectors, coatings, and service access. Those changes can improve performance, but they reduce the convenience of placing standard air-cooled hardware into a tank.

Colocation providers face a portfolio problem. Customers can arrive with air-cooled servers, direct-to-chip cold plates, rear-door heat exchangers, or immersion systems. Supporting every approach increases operational complexity. Supporting too few approaches can make a facility unsuitable for high-density customers.

This tension favors modular infrastructure. Operators can dedicate zones or pods to particular cooling architectures without redesigning an entire campus. Modular systems can also be assembled and tested before arrival, although site power and heat rejection still require careful planning.

China’s domestic market adds scale and policy pressure. New intelligent computing centers are being built while older facilities evaluate retrofits. National efficiency criteria encourage better cooling performance, and dense domestic AI systems increase the need for thermal equipment.

The market will not expand evenly. New facilities designed around AI loads can adopt liquid cooling more quickly than older sites. Large operators can demand custom engineering and long-term support. Smaller operators may wait for packaged systems, clearer standards, and broader hardware warranties.

Competition can also compress margins even as orders rise. Pumps, tanks, heat exchangers, and distribution systems can become standardized over time. Suppliers will need differentiation through reliability, controls, integration, field service, fluid expertise, or verified efficiency.

The phrase “trillion-yuan market” can obscure these divisions. There is no single liquid cooling product or uniform revenue pool. The category includes equipment sales, construction, fluids, replacement parts, monitoring, service, and retrofit work. Each segment has different economics and competitive barriers.

Investors and enterprise buyers should therefore look beyond announced orders. Backlog quality, delivery schedules, customer concentration, acceptance tests, warranty provisions, and repeat deployments reveal more than a broad market forecast. Suppliers that survive the transition must convert thermal demand into dependable operating systems.

For engineers and technology leaders, the lesson is equally practical. Infrastructure decisions around coolant, connectors, monitoring, and maintenance can constrain future hardware choices. Teams need a durable record of vendor specifications and test results. A searchable knowledge base can help preserve those decisions across facilities, vendors, and equipment generations.

Liquid cooling is becoming part of AI infrastructure because compute cannot operate without it at emerging densities. Yet the supply chain will mature only when vendors provide interoperable equipment, repeatable installation, verified materials, and service practices that ordinary data center teams can manage.

What to Watch After the Guiyang Trial

Three signals will show whether China’s immersion cooling surge becomes a durable infrastructure transition or remains concentrated in demonstrations and specialized deployments.

The first signal is published compatibility evidence from the four-fluid trial. Useful results should identify test duration, server materials, thermal conditions, fluid-property changes, failure criteria, and component outcomes. Simple statements about successful operation will not resolve lifecycle risk.

Long-duration evidence would strengthen the case for immersion if multiple fluid families maintain dielectric, chemical, and thermal stability. Results showing swelling, contamination, corrosion, or difficult maintenance would reinforce the cold-plate advantage.

The second signal is repeat commercial deployment. A pilot proves that a system can operate. A second or third deployment by an independent customer shows that design, installation, training, and support can transfer beyond the original engineering team.

Buyers should watch for deployments that publish rack density, cooling architecture, uptime, maintenance procedures, and energy measurements. Repeat orders from operating customers are more meaningful than reservations attached to facilities still awaiting chips or grid capacity.

The third signal is broader hardware and standards support. Immersion adoption becomes easier when server manufacturers qualify products for named fluids and tank systems. Published warranty terms, connector specifications, coolant testing methods, and safety requirements can reduce project-specific negotiation.

The ASTM fluid work item, emerging ISO guidance, and existing Open Compute Project documents show that this process is underway. However, developing a document is not the same as achieving interoperability. The market needs equipment and fluids tested against shared requirements.

Cold plates will continue advancing during this period. Higher-capacity plates, improved manifolds, warmer-water loops, and factory-integrated racks can raise their practical ceiling. Immersion must outperform that moving target, not the air-cooled systems of the previous decade.

Operators should also distinguish facility efficiency from computing efficiency. A low PUE can indicate reduced supporting energy, but it does not show how much useful computing work the facility completes. Workload utilization, accelerator performance, water use, and total energy remain essential.

The Guiyang pod represents a meaningful step because it treats coolant compatibility as an operating problem. It does not settle which fluid will win, how quickly costs will fall, or whether immersion will replace direct-to-chip cooling.

The reported year-end order backlog shows that customers are not waiting for every standard to mature. Their AI hardware schedules are forcing decisions now. That urgency benefits liquid cooling suppliers, but it also raises the cost of choosing a system without adequate lifecycle evidence.

The next few months should bring clearer answers. Watch for detailed fluid-test results, repeat customer deployments, and explicit hardware qualification. Together, those signals will reveal whether China’s liquid cooling boom is building a repeatable operating model or merely racing ahead of its reliability data.

For enterprise buyers, the immediate action is not to declare a winner. It is to demand evidence that connects thermal performance with maintenance, warranties, materials, and measured facility results. As dense AI racks move into production, the best cooling architecture will be the one that keeps computing available long after the demonstration tank is filled.