From: redpointai
The United States has implemented significant regulations aimed at limiting China’s progress in artificial intelligence (AI), viewing AI as a critical factor for global hegemony in the coming century [00:31:00]. These regulations, particularly the “AI diffusion rule,” are the most far-reaching seen by some experts, regulating clouds overseas and limiting what foreign companies can buy [00:44:00].
Regulatory Landscape
The initial October 2022 regulations primarily targeted the semiconductor industry with the explicit goal of regulating AI due to its anticipated rapid advancement [01:43:00]. These regulations were intended to maintain the US lead over China in AI, believing the next few years of progress will shape the next century of global dominance [02:22:00]. While well-intentioned from a perspective of immediate AI transformation, these regulations could negatively impact US competitiveness long-term if AI development takes 20 years instead of five [02:51:00]. Subsequent rounds of regulations in 2023 and December continued to patch loopholes [03:07:00].
Key Restrictions and Loopholes
Current regulations impose strict caps on GPU purchases, with each country limited to buying 50,000 GPUs over four years, a negligible amount compared to Nvidia’s total production [13:30:00]. A notable loophole exists where purchases of 1,700 or fewer GPUs do not count towards the 50,000 unit cap [13:43:00]. This could allow for the use of numerous shell companies to acquire GPUs and route them to China [13:50:00].
Furthermore, there are explicit prohibitions on exporting model weights outside of the US or trusted cloud environments, which are primarily US hyperscalers [10:46:00]. Regulations also target synthetic data generation, a technique widely used by Chinese companies to improve their models using data from advanced models like GPT-4 [11:05:00]. Accessing cloud services is also heavily regulated [11:27:00].
Impact on Chinese Companies and Cloud Providers
The “AI diffusion rule” has significantly hampered Chinese AI players and various cloud companies [13:07:00]. Before these regulations, the AI development landscape was largely unregulated globally [09:40:00]. Now, Chinese companies are heavily restricted in the size of clusters they can obtain and face notification requirements and demands for observability into their workloads [12:54:00].
Many smaller cloud companies, particularly those in foreign countries, are severely impacted as their business model relied on selling GPU access to Chinese firms like ByteDance [08:32:00]. ByteDance, which previously rented GPUs from various global clouds, is now heavily limited in cluster size and access to compute [12:17:00].
Some Chinese companies like Alibaba have their own hyperscale partners, but many new players such as Moonshot and DeepSeek lack such partnerships [11:54:00]. Foreign firms, including American companies, are also impacted; for example, Oracle faced challenges due to a rule limiting data center capacity in non-US ally countries to 7% [05:51:00].
Chinese Response and Future Outlook
Facing these restrictions, China’s numerous skilled engineers will be forced to innovate [14:08:00]. Companies like DeepSeek are noted for their strong engineering capabilities, allowing them to outperform competitors with similar compute levels [14:11:13]. However, American labs’ compute continues to scale rapidly, while China’s cannot keep pace [14:26:00]. This necessitates a “way, way, way, way better” engineering approach from China to overcome this compute deficiency [14:32:00].
Focus on Test-Time Compute and Inference
The push towards test-time compute (also known as inference) could offer a path for China [14:34:00]. While test-time compute requires significant training, particularly post-training with synthetic data generation and reward models, it represents a less mature “rung of the ladder” compared to traditional training [15:00:00]. This means Chinese companies might be able to achieve significant gains through engineering improvements with their limited compute resources [16:08:00].
However, scaling inference is crucial for AI to “change the world,” and China’s GPU limitations will make this very difficult [17:48:00]. The high costs associated with inference, even with low margins, translate to massive hardware investments that exceed the caps imposed by regulations [18:14:00].
Domestic Supply Chain and Geopolitical Considerations
The stricter regulations force China to either build its own domestic supply chain for semiconductors and AI hardware or access limited Western sphere resources [23:41:00]. Some believe that if these regulations are highly effective in preventing China from developing its own semiconductor industry, or if China builds a competitive one, it could increase the likelihood of action on Taiwan [23:48:00]. The “Goldilocks zone” for policy would be to keep China behind but with enough hope of catching up to avoid inducing greater risk [24:26:00].
The US government’s approach reflects a desire to maintain hegemony, even if it means creating a semi-monopoly for large US tech companies in the global AI infrastructure market, akin to Henry Ford producing tanks during World War II [07:07:00]. This approach centralizes power and could potentially stifle innovation in hardware and infrastructure outside of the major players [07:32:00].
Sovereign AI Initiatives
Many countries are pursuing “Sovereign AI” initiatives, seeking to develop their own AI expertise and models internally [08:36:00]. American startups serving these firms in regions like Malaysia, Singapore, India, and the Middle East are heavily impacted by the regulations [08:40:00]. The case of G42 in the UAE illustrates this, where initial concerns over Chinese links led to a partnership with Microsoft, demonstrating how the US can enforce its sphere of influence through hyperscalers [20:11:00]. Future US administrations may tweak these rules, potentially shifting focus from human rights considerations to economic influence [21:50:00].