A UC Berkeley research team argues that controlling the supply chain for advanced AI chips could force a moratorium on training new AI models. By targeting critical manufacturers like ASML and Zeiss, regulators might restrict the hardware needed for massive model development. This proposal underscores the extreme concentration of semiconductor manufacturing and raises significant questions about global AI regulation and geopolitical cooperation.
A new study from the University of California, Berkeley, presents a strategy to slow the development of frontier artificial intelligence by focusing on the hardware that makes it possible. The researchers suggest that rather than regulating software algorithms directly, authorities could manage the hardware bottleneck by targeting the semiconductor supply chain. This approach leverages the reality that high-end AI development depends on a small number of specialized suppliers.
The semiconductor industry, particularly at the high end, is highly concentrated. Producing chips capable of training massive AI models requires specialized lithography machines exclusively supplied by companies like ASML and optics from Zeiss. The research suggests that because this supply route is limited and monitorable, policymakers could enforce a long-term pause on model training by restricting the production of hardware capable of such intense computational work.
The proposed regulatory framework would create a clear distinction between two types of hardware: training-capable chips and inference-only chips. Under this plan, authorities would limit the manufacturing of hardware capable of large-scale training while continuing to allow the production of inference-only chips. These inference-only processors would be designed to run existing, approved AI applications efficiently but would be physically incapable of the heavy lifting required for creating new, frontier-level models. By phasing out current training-capable hardware, the researchers argue that nations could manage the risks associated with rapid, unchecked model scaling.
For investors and the broader market, this research highlights the strategic and economic importance of the semiconductor equipment sector. The extreme concentration of the supply chain means that any significant regulatory or export control changes in this area could have a meaningful impact on the revenue and growth outlook for semiconductor equipment manufacturers. The study acknowledges that such a system would require rigorous international oversight to monitor chip fabrication and the flow of raw materials.
The feasibility of such an agreement faces significant hurdles, including geopolitical tensions between major powers like the United States and China. However, the study argues that training large-scale AI models is energy-intensive and requires thousands of chips operating simultaneously, which makes clandestine development difficult to hide. If such regulations were to be adopted, investors should monitor for shifts in semiconductor export policies, changes in international chip manufacturing regulations, and the long-term impact on global demand for high-end chip fabrication equipment.