Beijing Launches AI4Chip Action Plan, AI to Penetrate Entire Chip Design and Manufacturing Chain
Nashnova编辑部
Beijing's Yizhuang development zone has released China's first dedicated AI4Chip policy, a three-year plan to systematically embed AI into chip design, wafer fabrication, advanced packaging, equipment, and materials — the core idea is not building chips *for* AI, but using AI to build chips.
What's the difference between "AI for chips" and "AI chips"?
"AI chips" usually means GPUs, ASICs, and other processors that run AI workloads — chips serving AI.
AI4Chip flips the direction: AI participates in designing and manufacturing chips, using machine learning to shorten development cycles, cut physical trial-and-error, and raise yields.
This means → AI here is not the product but the production tool; Beijing is betting on upgrading the method of making chips itself.
How does AI change the chip design process?
The plan calls for deep integration of AI into EDA — electronic design automation, the core software engineers use to lay out circuits and run verification — covering routing, timing optimization, power analysis, and physical layout.
AI agents will plug directly into EDA tools and design knowledge bases, helping engineers debug code, diagnose faults, and optimize design decisions — moving AI from a side tool into the engineering workflow.
Priority product areas include general-purpose GPUs, AI training and inference ASICs, RISC-V AI processors, high-end CPUs, memory chips, and FPGAs; enabling technologies span silicon photonics, co-packaged optics, optical I/O, and high-speed interconnects.
What can AI do on the fab floor and packaging lines?
On the manufacturing side, fabs are required to build digital-twin demonstration lines — virtual models that mirror real production in real time — with AI tuning process parameters, predicting equipment failures, and tracking end-to-end yields.
The plan also covers DTCO (design-technology co-optimization), STCO (system-technology co-optimization), and AI-assisted optical proximity correction (AI OPC), extending AI into lithography-related workflows.
In advanced packaging, AI will be applied to 2.5D and 3D packaging, testing, and manufacturing execution, along with AI-enabled computer-integrated manufacturing systems.
Why bring AI into equipment and materials too?
The plan extends AI to semiconductor equipment, components, and materials, using digital twins and edge AI to shorten equipment qualification and materials development cycles while improving purity and process stability.
It explicitly lists "security, self-reliance, and controllability" as a guiding principle, and proposes using AI to monitor global supplier dynamics and geopolitical risks — adding an intelligence early-warning layer to the supply chain.
This means → AI in the equipment and materials segments is not just an efficiency play; it is a supply-chain security play — using technology to hedge against cut-off risk.
What are the 2028 targets?
The plan sets a goal of establishing an integrated AI-semiconductor development framework by 2028, incubating three to five internationally competitive AI4Chip ecosystem companies, and producing more than ten replicable application scenarios.
In plain terms = Beijing is not just issuing a document; it wants deployable, replicable companies and use cases within three years — this is industrial policy with KPIs attached.
Where is the biggest bottleneck?
AI systems require large, high-quality datasets, but wafer process parameters, equipment fault logs, defect data, and yield records are among a semiconductor company's most sensitive commercial assets.
Cross-company data sharing carries clear IP and confidentiality risks; the plan includes data-cleaning, de-identification, and sharing mechanisms, but execution details remain unspecified.
This means → Whether companies are willing to contribute sufficiently detailed manufacturing data will directly determine if AI4Chip can scale beyond single-company boundaries — data-sharing willingness is the hard constraint on the entire plan.
市场有风险，内容仅供研究参考，不构成投资建议。