China's first engineering white paper covering the full life cycle of domestically built 10,000-GPU intelligent computing pools came out in Guangzhou on November 13. Nanfang Net reports that the "Engineering White Paper on Domestic 10,000-Card Intelligent Computing Resource Pools," put together by the Guangdong Telecom Planning and Designing Institute, is the first attempt at a modular, standardized and process-driven engineering system for these pools. It answers three questions — what to build, how to build, and how to evaluate. The backdrop is hard to ignore. By September, China's total intelligent computing capacity had reached 2,592 EFLOPS (FP16), with over 20 domestic 10,000-card clusters already running and 100,000-card pools entering planning.

Interior of a large AI computing data hall, illustration
Power, networking and cooling of 10,000-card pools are becoming standardized engineering subjects

Scale, it turns out, widens the engineering gaps. A single cabinet in a domestic 10,000-card cluster can draw close to 350 kW, according to the white paper. Fiber deployment runs more than 10 times that of traditional data centers, and one model archive alone can flood storage with terabytes of data. Projects of the same scale have seen construction timelines differ by more than six months. That spread is exactly why standardized delivery capability has become the core bottleneck for mass deployment of domestic computing power.

"The bottleneck of 10,000-card AI infrastructure is no longer single-point technology, but a standardized engineering methodology that runs through planning, design, delivery and operations," said Pang Zhipeng, a senior engineer and computing-industry expert. The six-chapter white paper draws on the institute's own delivery experience and centers on three systems. One is a modular architecture selection system, which proposes "supernode training zones, rack-based inference zones" and sets forward-looking requirements for 100,000-card evolution. Another is a standardized design baseline that unifies four-plane networking, three-tier storage, high-voltage power and liquid cooling, with cross-generation reserve criteria for upgrading. The third is a process-based delivery framework, featuring a five-stage "power-first" path and seven-party coordination, plus evaluation metrics that include full cost per token, grid-readiness and utilization.

The institute behind the document was founded in 1953 as one of the first seven class-A survey and design institutes of the former Ministry of Posts and Telecommunications. Its track record includes national projects such as the Xiong'an supercomputing center, the Helingeer intelligent computing center and the National Supercomputing Guangzhou Center, and it has been investing early in 800V HVDC, global memory pooling and hollow-core fiber. As 10,000-card clusters shift from demonstration to mass delivery, expect the competition to move past chips. Engineering methodology is quietly becoming the second frontier.