The US House Select Committee on China has found that the “DeepSeek moment” was enabled with extensive technical assistance from Nvidia. This was when a Chinese company, DeepSeek, built a reasoning model that achieved near-parity with leading models from US labs at a fraction of the training cost and time. This sent shock waves across the world, raising concerns about possible Chinese dominance in AI and even triggering a massive sell-off in Nvidia stock.
The export controls on semiconductors over the past two years have seen a lot of flip-flops. Reversing previous export controls, the Trump administration recently allowed the sale of advanced chips to China with 25% tariffs. Meanwhile, China has softened its own stance. After banning domestic companies from using Nvidia’s AI chips to promote domestic alternatives, Beijing has now granted exemptions to three of its largest tech companies - ByteDance, Alibaba and Tencent.
Multiple factors are driving these shifts in policy decisions. Here’s an effort to make sense of some of these developments.
- We still don’t understand how large language models, which are essentially predicting words based on training on huge datasets, generate outputs that make sense. But the initial understanding that scaling compute and data would lead to increasingly powerful models far superior to humans at most tasks is no longer true. As the gains from current AI approaches plateau, the real value lies in AI’s diffusion across sectors across the economy. Such applications include coding, research, education, healthcare, agriculture and even military use cases.
- AI is a dual-use technology. A Reuters review of patents, tenders, and research papers shows Deepseek being used for military applications. AI use for military applications can broadly be classified into three functions: automation, prediction and analytics. Applications include target identification systems, lethal autonomous weapons systems, decision support systems, battlefield healthcare, logistics and maintenance, and data analysis. Underlying many of these use cases is a technology with similar civilian applications.
- Unlike other dual-use technologies, such as space or nuclear, whose development was state-led, frontier AI development is being led by the industry. Firms are competing and collaborating across borders for a global market. Their incentives are not always aligned with those of the state, and this tension also influences the direction of policy.
- US restrictions on chip exports to other economies, including the now-repealed diffusion framework that divided the world into three tiers of AI chip access, have many unintended consequences. It frays US diplomatic relations with other countries and has also led to global efforts to build resilience in the semiconductor supply chain. Export controls on chips have been extremely difficult to enforce. There have been several instances of chips being diverted to China or data centres built in Southeast Asia that address Chinese AI computing needs.
- In some tasks across various sectors, we can expect a sudden increase in productivity. For instance, coding has become something AI excels at. There are bound to be applications in the military as well to realise such gains. Given the dual-use nature and industry-led development, it is difficult to restrict such military applications. The speedbreakers are likely societal or organisational structures that determine how these technologies are applied.
- China’s approach of initially banning its companies from using US chips and then creating an exemption for some of its largest companies is an effort to balance two strategic priorities: developing strategic autonomy in the semiconductor supply chain while staying globally competitive in AI development. However, developing a domestic semiconductor supply chain will be a long and gradual process. While US export controls on advanced AI chips might have been relaxed, there are several other bottlenecks elsewhere in the value chain. Most notably, this includes the ban on the sale of EUV lithography machines to China. These are crucial for manufacturing advanced chips (below 7nm), and a Dutch company holds a monopoly.
The US relaxation of chip restrictions on China seems to be a response to multiple of these factors. This delicate dance of balancing strategic autonomy and innovation is likely to continue, and the limiting factor for military applications of AI is likely to be organisational and societal constraints.