Liquid Cooling Goes Mandatory:
The Data Center Rebuild and What It Means for Chinese IDC Counterparties

1. Why Air Cooling Just Hit the Wall

If you have been procuring data-center capacity in China for more than five years, you probably remember when a 5–8 kW rack was normal and “precision air conditioning” was state of the art. That world is gone. The arrival of 100 kW+ AI super-nodes — and the policy apparatus that is now enforcing greenness by the kilowatt-hour — has turned liquid cooling from a CTO’s pet project into a regulatory and engineering precondition for every new hyperscale facility.

Three things changed in 2026, all at once:

• A real national standard. On 30 July 2026, the State Administration for Market Regulation published GB/T 48023-2026, Technical Specifications for Cold-Plate Liquid Cooling Systems of Data Centers — China’s first national standard for data-center liquid cooling. Led by Sugon Shuchuang under the China Electronics Standardization Institute, drafted with 80 organizations including Huawei Digital Power, Alibaba Cloud, Baidu, ByteDance’s Volcengine, China Mobile Design Institute and Tsinghua University, the standard covers system design, CDU (coolant distribution unit), quick-disconnect fittings, cold plates, coolant, hydraulic calculation and deployment. It takes effect on 1 February 2027. Drafting is underway on a twin standard for single-phase immersion cooling.
• A harder PUE ceiling. The joint MIIT / NDRC / NEA Data Center Energy Efficiency Improvement Action Plan (2026–2028) now requires all newly-built large and ultra-large data centers to hit PUE ≤ 1.1, with 2028 as the retrofit deadline for the legacy stock. National-hub nodes in the “East Data West Computing” program already sit under a ≤ 1.2 cap. Beijing has started levying a punitive 0.2–0.5 RMB/kWh surcharge on data centers whose PUE exceeds 1.35. Multiple provincial governments have simply stopped approving new air-cooled permits in AIDC projects.
• Chips that cannot be air-cooled. NVIDIA confirmed in June 2026 that the Vera Rubin platform uses a 100% fanless, 100% liquid-cooled architecture, with coolant entering at 45 °C and exiting around 55 °C — high enough that sites in suitable climates can eliminate chiller plants altogether. Single-GPU TDP is heading past 1,500 W (Rubin) toward ~2,300 W and potentially 3,500 W for Rubin Ultra; a Rubin NVL72 rack sits around 130–220 kW. Domestic equivalents such as Huawei’s Ascend 950 SuperPoD, Sugon’s 100,000-card Dengfeng cluster, and every major super-node shown at WAIC 2026 (Huawei, ZTE, H3C, Inspur) are now liquid-cooled by default.

2. The Numbers Behind the Rebuild

The industry has moved from “talking about it” to “shipping it” faster than almost any analyst predicted. The composite picture from CICC, CAICT, IDC, TrendForce and the China Center for Information Industry Development is unambiguous: 2026 is the inflection year.

For context, the China Academy of Information and Communications Technology (CAICT) updated its forecast at the 2026 China Computing Power Conference on 11 September, putting China’s total liquid-cooling market at RMB 71.6 billion in 2026 (+70% YoY). All 40+ new intelligent-computing centers that broke ground in China in Q1 2026 specified liquid cooling in the tender documents; none procured pure air-cooled racks. CICC estimates that on Rubin-era super-nodes, the liquid-cooling BOM alone is worth US$140,000–160,000 per rack, implying RMB 4.9–7.7 billion of liquid-cooling equipment for each GW of IT load.

Two technical routes, two risk profiles

3. Three Structural Impacts on Chinese IDC Operators

Standard, regulation and silicon together are rewriting the asset value, cost structure and compliance profile of every Chinese data center. For an overseas customer evaluating a Chinese counterparty, the consequences fall into three buckets — each of which is a diligence question in its own right.

4. An Eight-Step Verification Checklist for Overseas Tenants

The following list is what we run for international clients before they sign a three- to five-year colocation, managed-AI or leased-capacity agreement with a Chinese IDC. None of these steps are technically complex — most rely on public registries, third-party certifications, and a careful read of the operator’s corporate filings — but collectively they separate serious operators from marketing decks.

1. Pull the operator’s official business registration and annual reports. Verify registered capital, paid-in capital, operating scope (must explicitly cover IDC / internet data center services with a valid value-added telecom license — the IDC Zeng Zhi Dian Xin Ye Wu Jing Ying Xu Ke Zheng), shareholder structure, foreign-investment status, and any pledged equity or frozen shares. This is the same baseline we deliver with an official enterprise credit report.
2. Check administrative penalties for energy / environmental / fire-code violations. Run the operator and each of its project subsidiaries through the National Enterprise Credit Information Publicity System (gsxt.gov.cn), the local Ecology and Environment Bureau, and the fire department’s public penalty announcements. Look specifically for excess-energy consumption fines, failure-to-pass fire-protection acceptance, and construction-without-permit records — these are the issues that cause a site to be powered down mid-lease.
3. Request — and independently verify — measured PUE records. Ask for the site’s PUE as measured and filed with the local big-data / industry-and-information department (which every operating DC above a certain size is now required to report). Cross-check against the province’s published list of compliant data centers and green-data-center certifications. PUE claimed in marketing collateral is routinely 0.1–0.2 lower than the audited number.
4. Map energy quota and land tenure to the specific building, not the parent company. Each data center hall needs a separately allocated energy quota, land-use certificate, construction planning permit and fire-protection acceptance. Confirm these for the exact module you are being sold, not just for the campus as a whole. Projects are still being marketed in 2026 where the operator has approval for Phase 1 but is pre-selling Phase 2 capacity without a finalized energy allocation.
5. Audit liquid-cooling IP and standards participation. Pull the operator’s patent portfolio via CNIPA for keywords around CDU, cold plate, manifold, leak detection, liquid-cooled cabinet, and coolant. Operators that genuinely deliver liquid cooling at scale almost always appear as co-drafters of national / industry / association standards (e.g., GB/T 48023-2026, the YD/T series, ODCC specifications). Presence on a standards committee is a stronger signal than a product datasheet. Our intellectual-property inquiry service can map this in a single workstream.
6. Verify the liquid-cooling supply chain and O&M capability. Who supplies the CDU? Who supplies the quick-disconnect fittings? Is there a redundant coolant / spare-parts stock on site? What is the documented leak-response protocol? How many technicians on the overnight shift are factory-trained by the liquid-cooling vendor? Ask for three reference customers running liquid-cooled workloads on the same campus for at least twelve months — and actually call them.
7. Trace the management team’s build-out history. Liquid-cooled data centers are hard projects, and past delivery is the single best predictor of future delivery. Run background checks on the general manager, the CTO / engineering lead and the project director — prior tenures at established operators (GDS, Chindata, VNET, 数据港 etc.), and references from commissioned liquid-cooled sites, matter more than CV rhetoric.
8. Check tendering and government-project records. For operators that serve state-owned enterprises, cloud SOEs and local governments, public bid-and-award records (zhaobiao) reveal the true contracted capacity, delivered-PUE commitments and any liquidated-damages clauses. An operator that has consistently won government AIDC tenders and delivered on them has passed a far more rigorous audit than any customer reference tour.

None of these items show up on the English-language sales brochure. Most of them are public information inside China, but they are fragmented across a dozen different government portals and are rarely presented in English. That is exactly why a consolidated diligence package — corporate filings, risk records, IP, bidding history and executive background — is worth running before the LOI stage, not after a rack has been paid for and a deployment delayed.

5. Reading the Chinese IDC Landscape in 2026

Chinese third-party IDC is already a concentrated market. CAICT and Kezhi Consulting rank the major carriers along these lines as of 2026 — but rankings alone are not enough; what matters for an overseas tenant is how each operator’s liquid-cooling strategy and asset base map to your workload.

A note on smaller players: 2026 has seen a rush of steel, real-estate, chemical and energy companies entering the AIDC space — they bring land, power, and balance sheet, but many have no prior data-center operating record. If your counterparty is a first-time IDC operator spun out of an industrial group, the eight-step checklist above is not optional; it is the minimum.