Huawei scientist warns that NVIDIA's chip scaling is nearing physical limits.

ME AI message: Liao Heng, chief scientist of Huawei Semiconductor and one of the core founders of the Ascend AI chip, issued a warning in a recent public interview, stating that Western chip giants like NVIDIA are approaching physical limits as they pursue more powerful processors—“Expanding scale by continuously adding more compute chips and HBM inevitably has a ceiling. The industry is still pushing forward, but once this physical limit is crossed, a cascade failure will occur.” Liao Heng compared the entire AI value chain to an “18-story pagoda,” aligning with NVIDIA CEO Jensen Huang’s “layered cake” framework, emphasizing that China must build collaborative capabilities across every layer—particularly fostering close cooperation between chip manufacturers and AI model developers. Liao also revealed that Huawei will soon launch its first smartphone chip designed under the Tau Scaling Law framework, enabled by LogicFolding technology. Once analyzed by institutions such as SemiAnalysis and TechInsights, “the world will clearly understand by the end of this year how this alternative path can help close the gap.” Tau Scaling Law is an innovative design philosophy introduced by Huawei, centered on shifting focus toward improving transmission speeds between components in a computing system as the inherent advantages of chip scaling gradually diminish. Liao also highly praised Liang Wenheng, founder of DeepSeek, attributing the key to training top-tier models with extremely low compute power in 2025 to innovative model architecture design. He likened Chinese AI innovation to maximizing space in a small apartment, while Western counterparts live in much larger villas. Liao noted that Huawei itself is developing AI chips better suited to support efficient architectures through co-design mechanisms between chips and models: “We must invest more effort into design, trading higher complexity for reduced computational resource consumption.” (Source: BlockBeats)