As Moore’s Law approaches its physical limits, the global AI chip race has shifted from pure process-node scaling to a contest over system-level architectural innovation and control of advanced manufacturing ecosystems. Huawei’s so-called “Tau Law” is not merely an alternative to Moore’s Law—it is a forced reimagining of semiconductor design under external constraints. By integrating LogicFolding, optical interconnects, and heterogeneous chiplet integration, Huawei aims to approach the system-level performance of 5nm or even 3nm chips using mature nodes like 7nm or older. The goal is not to catch up with TSMC’s transistor density but to render transistor density less relevant.
Meanwhile, NVIDIA is leveraging TSMC’s System-on-Integrated-Chips (SoIC) advanced packaging to push the compute density of its Blackwell and upcoming B100/B200 GPUs to unprecedented levels. According to TrendForce, the Blackwell Ultra—built on TSMC’s 4NP process with SoIC-H—delivers a 40% increase in HBM3e bandwidth and exceeds 20 petaflops of FP8 performance per card. This “advanced node + 3D packaging” combination creates a formidable moat: it requires exclusive access to TSMC’s CoWoS capacity and deep co-design between NVIDIA’s NVLink interconnect and memory subsystems. By 2025, TSMC’s monthly CoWoS capacity is expected to reach 200,000 wafers, over 70% of which are already reserved by NVIDIA through 2027. This is no longer just a commercial arrangement—it is a geotechnical alliance made tangible.
Huawei’s predicament underscores the exclusivity of this alliance. Although its Ascend 910B uses SMIC’s 7nm process, it lacks access to both advanced packaging and high-bandwidth HBM memory. As a result, its single-card FP16 performance remains around 2.5 petaflops—less than one-tenth of Blackwell’s. More critically, Huawei cannot license TSMC’s SoIC or InFO-3D technologies. Even if China achieves domestic 5nm production in the near future, without 3D integration capabilities, system performance will hit an “interconnect wall,” not a “process wall.”
The true ambition of Tau Law lies in breaking this dependency chain. At its 2024 Developer Conference, Huawei demonstrated its “Nebula” AI cluster, which uses board-level optical interconnects to link thousands of Ascend chips, reducing latency by 60% and improving energy efficiency by 2.3x. This “replace copper with light, trade architecture for nodes” philosophy seeks to establish a high-performance computing paradigm independent of TSMC. Yet two major hurdles remain: optical interconnects still cost 5–8 times more than electrical ones, limiting adoption in mainstream AI training; and Huawei’s MindSpore software stack lags far behind CUDA, with developer numbers differing by over two orders of magnitude.
I judge that the next three years will see a dual-track AI chip landscape: NVIDIA and TSMC will dominate large-scale model training through their iron triangle of process, packaging, and software, while Huawei and its Chinese ecosystem partners carve out niches in edge inference and vertical-specific AI via Tau-inspired architectures. However, if TSMC expands SoIC access to AMD, Intel, or other non-Chinese clients, Huawei’s window for breakthrough could close rapidly.
More alarmingly, the U.S. Department of Commerce is evaluating whether to add advanced packaging tools—such as hybrid bonding equipment—to its export control list for China. If implemented, not only would Huawei’s chiplet strategy be crippled, but China’s entire “post-Moore” semiconductor roadmap would suffer structural damage. Technological autonomy has never been just about transistors—it is about mastering an integrated stack of manufacturing, packaging, materials, and tools. As the world’s most advanced AI compute increasingly relies on 3D stacking and heterogeneous integration, whoever controls the vertically integrated manufacturing ecosystem will define the boundaries of next-generation AI. Can Huawei’s Tau Law crack the wall jointly built by NVIDIA and TSMC? The answer may lie not in Shenzhen, but along the deepening geotechnical rift between Hsinchu and mainland China.