Huawei takes fresh path to advance chips

As global chip companies approach physical limits in their pursuit of more powerful processors, Huawei Technologies Co's emphasis on improving transmission speeds between system components could offer an alternative path, experts said.

The comments come as industry watchers recognize a systematic, replicable innovation ecosystem taking shape across China's tech landscape, spanning semiconductor design, memory chip fabrication, artificial intelligence frameworks and system integration.

Liao Heng, chief semiconductor scientist at Huawei, said in a recent four-hour interview that the decades-long practice of continuously shrinking transistors and packing more into integrated circuits is running out of runway, as transistors are now almost as small as atoms, nearing the physical extreme of contemporary manufacturing capabilities.

While Western companies have multiyear road maps enabled by ASML's advanced lithography machines, Huawei — barred from accessing such equipment due to sanctions imposed by the United States government — has had to pursue a more creative approach, Liao said.

That approach is Huawei's "Tau Scaling Law", proposed in May, which shifts focus from shrinking chip dimensions to improving transmission speeds between system components.

Huawei is set to unveil its first smartphone chip designed under this framework in September, using a technology called LogicFolding.

Liao said the world will have a clear understanding later this year of how Huawei's alternative approach helps narrow the gap with rivals, when the new smartphone chip is dissected by industry research companies.

He likened the AI value chain to an "18-story pagoda", positioning it as a more comprehensive framework than Nvidia CEO Jensen Huang's five-layer cake model.

Huang's model, which he has articulated during multiple public appearances, divides the AI industry into five layers: energy, chips, computing infrastructure, cloud/AI models and applications. It is a top-down business framework designed to illustrate how value flows through the AI stack, culminating in economic returns at the application layer.

Liao's "18-story pagoda", by contrast, is a bottom-up engineering and production framework that spans the entire semiconductor value chain. It begins with foundational elements such as basic theory, energy, mining and raw materials, silicon, transistor design, wafer fabrication, lithography and equipment, and advanced packaging. The middle part covers chip architecture, runtime environments and quantization. The top layers encompass AI training methods, foundation large language models, AI agents, products and commercial applications.

Crucially, Liao said the overall performance ceiling of the industry is determined by the shortest layer — the weakest link in the chain. The weakest link varies based on geographies.

Zhou Jianjun, a professor at Shanghai Jiao Tong University's School of Integrated Circuits, said the strategic implications are profound. Huawei is looking for a different path to improve performance — not by shrinking transistors any further, but by making groups of chips and the overall AI system work better together.

Xiang Ligang, director-general of the Information Consumption Alliance, a telecom industry association, said that while Huang's five-layer cake frames the AI industry as a market opportunity to be captured layer by layer, Liao's 18-story pagoda frames it as an integrated engineering challenge where no single layer can be optimized in isolation.

For Huawei, which cannot access the advanced lithography equipment, this holistic view is very practical. Huawei's breakthrough lies in cross-layer collaboration — using the advantages of system clusters to compensate for the hardware shortcomings of single chips, Xiang added.

According to the latest report by the US market research company International Data Corp, a dramatic shift occurred in China's AI accelerator market in 2025, driven by the US export controls. Domestic Chinese suppliers, led by Huawei, saw their combined market share surge to 41 percent. Nvidia's share fell to 55 percent, a significant drop from its near-monopoly levels.

masi@chinadaily.com.cn