What China Can Learn from Domestic Large‑Model Compute: Challenges and Strategic Recommendations

With the rapid rise of ChatGPT, large AI models have become a hot topic worldwide, and numerous domestic models have emerged in China. While these models have achieved impressive progress, the article highlights that Chinese large‑model development still faces significant technical and operational challenges.

The discussion covers three main aspects:

Computing characteristics: Overview of domestic platforms, system bottlenecks, operator implementations, fault tolerance, and the parallelism support provided by existing frameworks.

System challenges: Limited compute scale, insufficient network bandwidth and latency, fragmented data resources, and a lack of coordinated R&D efforts.

Future algorithmic directions: Need for more efficient training algorithms and better hardware‑software co‑design.

To address these issues, the article proposes six strategic recommendations:

Increase compute scale by accelerating the “East‑Data West‑Compute” initiative, expanding compute‑network infrastructure, and improving network speed to lower latency.

Strengthen data management through national standards, clear data‑usage rules, and breaking data silos between vendors to ensure high‑quality, domain‑specific training data.

Establish a national large‑model R&D team that consolidates top talent and resources, reducing fragmented, low‑efficiency projects and conserving compute and energy.

Boost financial investment by creating a dedicated national large‑model fund for research and training.

Enhance policy support with favorable tax measures, such as allowing state‑owned enterprises to count double their AI R&D spending toward net profit.

Accelerate core technology development, especially AI‑chip design and manufacturing, to resolve the “bottleneck” problem.

These recommendations aim to leverage China’s institutional strengths, improve top‑level design, and increase coordinated planning and resource allocation, ultimately moving Chinese AI from a follower to a leader.