When playing mobile games for a long time, the phone gets hot, and a computer running a dozen programs simultaneously may also experience roaring fans and screen stuttering. The busier the chip, the more heat it generates; if the heat cannot be dissipated, the device can only proactively throttle down to "take a break".
Scaling this scenario up to AI data centers, the problem is no longer just about warm palms. Training and inference for large models require a massive number of GPUs to run continuously under high loads. If the cooling cannot keep up, it leads to computing power throttling in mild cases, or triggers shutdowns in severe cases, leaving expensive GPUs "idling on the payroll".
Consequently, liquid cooling is transitioning from a bonus feature to a basic configuration for AI servers. NVIDIA's GB300 NVL72 has already adopted a fully liquid-cooled rack-level architecture, integrating 72 Blackwell Ultra GPUs and 36 Grace CPUs on the same platform. In July 2026, Inspur Information released a CPU native liquid-cooled full rack, no longer simply "adding water loops" to air-cooled servers, but redesigning the entire system architecture around computing, communication, and storage.
These moves send out the same signal: liquid cooling is evolving from a cooling accessory for servers into a prerequisite for computing system design.
Liquid Cooling Accelerates in Popularity, Welcoming Two New "Rulers" for the Industry
In July 2026, GB/T 48023—2026 "Technical Specification for Cold Plate Liquid Cooling Systems in Data Centers" was released and will be implemented in February2027; in August, GB/T 48097—2026 "General Specification for Cold Plate Liquid Cooling Cabinets in Data Centers" was released and will be implemented in March 2027.
The former answers "how to build the entire system," while the latter answers "how to manufacture liquid-cooled cabinets." Liquid-cooled servers thus enter the national standard era.
GB/T 48023—2026
The two national standards primarily target cold plate liquid cooling. The system specification does not apply to two-phase cold plate liquid cooling, nor can it be directly applied to immersion solutions.
Why Exactly Now?
Over the past few years, the industry has been solving "whether liquid cooling can be used"; now, it needs to solve "whether it can be applied at scale with low cost and replicability".
Research by MIR shows that in 2025, the liquid cooling penetration rate for AI servers in China was approximately 30%. With the rapid expansion of the liquid-cooled server market size, the liquid cooling penetration rate is expected to exceed 50% by 2027; within liquid cooling solutions, cold plate types account for over 90%.
2025-2028E China Liquid-Cooled Server Market Size and Forecast
Data Source: MIR DATABANK
Demand is shifting from demonstration projects to large-scale deployment, but the industry still exhibits distinct characteristics of project customization.
Even for cold plate liquid cooling, different manufacturers may have different understandings of interfaces, pipelines, coolants, pressure, and testing methods. Cold plates, quick disconnects, manifolds, and Coolant Distribution Units (CDUs) come from different suppliers. Combining them is somewhat like several manufacturers producing plugs, sockets, and appliances separately without unified voltage and interface standards: individual components can work, but the assembled system may not be stable.
The result is that repeated design and verification drive up costs, making component interchangeability difficult; once liquid leakage, condensation, or cooling anomalies occur, there is also a lack of unified acceptance criteria among the participating parties.
What Has Changed with the New National Standards?
First, drawing a clear "system map"
GB/T 48023 divides the system into the primary side and the secondary side: the cold plates, quick disconnects, manifolds, pipelines, and CDUs on the server side constitute the secondary side; the CDU then exchanges heat with the primary side pipelines and cold source equipment.
Simply put, the secondary side "takes away" the heat from the chips, while the primary side "sends the heat out of the computer room".
Composition of Cold Plate Liquid Cooling Systems
In the past, the system boundaries and product naming conventions of various companies were not entirely consistent, which could easily lead to misunderstandings in design, procurement, and acceptance. Now that the map is clearly drawn, there is a common language for equipment selection, system design, and responsibility division.
Second, looking beyond just cooling capability to long-term stability
The standards set requirements for material and coolant compatibility, pressure-resistant sealing, pipeline flow uniformity, leak detection, condensation control, and monitoring alarms.
For example, the liquid supply temperature on the secondary side should not be blindly lowered but should be at least 3°C higher than the ambient dew point temperature; otherwise, the pipelines might "sweat" like the outside of an iced drink cup. The heat exchange capacity, temperature control accuracy, and flow regulation of the CDU also have testable metrics.
This is equivalent to liquid cooling systems evolving from on-site trial and error into engineered products with health check items, pass/fail thresholds, and anomaly alarms.
Third, extending the standards from equipment manufacturing to deployment and O&M
Liquid cooling risks often do not occur at the first boot, but rather in corrosion, blockage, micro-leaks, and water quality changes during long-term operation. The system national standard covers design, manufacturing, deployment, and O&M; the cabinet national standard further supplements structural, performance, quality evaluation, delivery acceptance, and O&M requirements.
Manufacturers' product layouts are also changing in this direction.
Huawei has upgraded the traditional CDU to a liquid cooling thermal management controller (TMU) that undertakes heat exchange, flow regulation, monitoring, and group control functions, emphasizing the extension from a "coolant distribution unit" to a "system control hub".
Envicool's Coolinside full-chain liquid cooling solution extends from cold plates, quick disconnects, manifolds, and CDUs to coolants, pipelines, and cold sources, covering the "factory-to-site" delivery from server manufacturing plants to data center sites.
While laying out liquid-cooled data center products and systems, Sugon Data also took the lead in drafting GB/T 48023. Inspur Information continues to advance from fully liquid-cooled reference designs to CPU native liquid-cooled full racks.
These dynamics indicate that the focus of liquid cooling competition is shifting from selling a single component to delivering and managing a system over the long term.
GB/T standards are recommended national standards and are not immediately mandatory for all projects; however, when they enter the bidding documents of operators, state-owned enterprises, or internet clients, they become a threshold that suppliers must cross.
What Opportunities Does Standardization Bring?
For data center clients, procuring liquid cooling no longer means just listening to each manufacturer's own narrative. Unified metrics and testing methods allow for horizontal comparison of different solutions and also help reduce costs for expansion, replacement, and O&M.
For the liquid cooling industry chain, opportunities belong not only to server OEMs and liquid cooling system providers. Cold plates, quick disconnects, manifolds, CDUs, pumps, coolants, sensors, leak detection, as well as third-party testing, system integration, and O&M services, will all see incremental growth.
However, the national standards do not provide equal support to all enterprises.
After the standards are unified, the space for low-level customization and pure price competition will be compressed, and manufacturers whose products cannot pass reliability verification will face consolidation; enterprises with capabilities in core components, system control, engineering delivery, and long-term services are more likely to expand their advantages.
For ordinary internet users, the change will not manifest as their phones suddenly getting cooler, but will be hidden behind every search, video generation, and smart assistant response. More stable cooling means less throttling and downtime, and also supports more sustainable expansion of computing power.
Returning to the beginning, when a phone overheats, we can put it down for a few minutes; but a 24/7 operating AI data center cannot solve the problem by "taking a break".
The liquid cooling national standards provide not just a set of cooling parameters, but a common language for replicating liquid-cooled servers across more data centers.
MIR believes that the true inflection point is not just about how high the penetration rate rises, but whether the industry can transition from project customization to standardized products, from point cooling to system delivery, and from being usable to being reliably scalable. Obviously, the inflection point has arrived.
When liquid cooling has a unified design language, testing ruler, and acceptance baseline, the heat of AI servers will truly have the opportunity to be stably and scalably dissipated.