From: redpointai
Dylan Patel, a prominent thinker on hardware and AI, shared his insights on the “AI diffusion rule” and its geopolitical ramifications during an episode of Unsupervised Learning [00:00:15]. This rule, primarily targeting the semiconductor industry, aims to shape the future of global power dynamics [00:01:38].
The AI Diffusion Rule: Context and Goals
The initial October 2022 regulations were “primarily on the semiconductor industry,” but their wording clearly indicated a desire to regulate AI due to its anticipated rapid advancement [00:01:43]. According to White House special adviser Ben Buchanan, the government knew the rapid scaling of AI was coming and intentionally acted [00:02:04]. The overarching goal of these regulations is to ensure the US maintains its lead over China in AI, believing that “the next five years of progress… are going to shape the next… century of like hegemony for the world” [00:02:22].
While well-intentioned from a perspective that AI will transform the world in the next five years, these regulations could be detrimental to long-term US competitiveness if the transformation takes longer [00:02:41]. Subsequent rounds of regulations in 2023 and December continued to patch loopholes in the initial rules [00:03:07].
Key Loopholes and Their Implications
Despite successive rounds of regulation, several loopholes remain [00:03:12]:
- Chinese companies can still acquire GPUs from foreign firms [00:03:16]. This has led to companies like Oracle expressing concerns, as one of their largest cloud customers is ByteDance [00:03:20].
- The ability to build data centers in non-US ally countries, like Malaysia, where significant data center capacity is being constructed by Chinese companies (some re-registering as Singaporean firms) [00:03:31]. This capacity is substantial, with Malaysia’s planned builds from 2024 to 2027 roughly equivalent to Meta’s entire global footprint at the beginning of 2024 [00:03:39].
The regulations are “very far-reaching” as they attempt to regulate overseas clouds and limit what foreign companies can buy [00:04:34]. A key constraint is the “7% rule,” which dictates that only 7% of a US company’s data center capacity can be located in a country not considered a US ally [00:05:53]. This rule significantly impacts companies like Oracle, which had planned to have 20% of their capacity in Malaysia [00:06:00].
Impact on Competition and Innovation
The regulations have “reduced competition massively in the market” and “created sort of a monopoly” for major US hyperscalers like Microsoft, Meta, Amazon, and Google [00:06:47]. These companies have the vast majority of their AI data center capacity in the US, allowing them to absorb new international capacity without violating the 7% rule [00:06:33].
This approach parallels historical industrial policies, where the US would make companies like Henry Ford “absolutely freaking rich” by having them produce essential wartime materials [00:07:07]. In the context of AI, this means “here you go Sacha Nella like you you get to do this in AI” [00:07:20].
While potentially beneficial for American startups in terms of model development, the regulations are “pretty bad” for hardware infrastructure and smaller cloud providers, which are “very much hit hard” [00:07:37]. Companies like CoreWeave, which have grown close to hyperscaler levels, are relatively insulated, but smaller “mini-cloud” ecosystems and Sovereign AI firms in countries like Malaysia, Singapore, India, and the Middle East are heavily impacted [00:07:56]. These regulations make it clear that countries developing their own AI expertise face the risk of being cut off [00:20:00].
Before these regulations, the landscape was largely “the wild west” [00:09:40]. A US executive order only required notification for models over 26 teraFLOPs (twice the size of Llama 405b), and this only applied within the US [00:09:42]. Other countries could freely build and rent GPUs from anywhere [00:10:00].
The new rules significantly restrict:
- Exporting model weights outside US-trusted clouds (i.e., hyperscalers) for large foundation models like Llama 4 and GPT-5 [00:10:46].
- Protecting against synthetic data generation, which Chinese companies extensively use to improve their models (e.g., generating data from GPT-4 to post-train) [00:11:04].
- Accessing clouds, impacting many smaller cloud companies whose business model relied on renting GPUs to Chinese firms like ByteDance [00:12:20]. ByteDance, while using GPUs for TikTok, is also developing language and generative video models, leading to fears of manipulation [00:12:35].
China’s Response and Future
China faces significant challenges but has one “obvious loophole”: buying fewer than 1,700 GPUs per transaction, which doesn’t count towards the 50,000 GPU country cap [00:13:26]. This could lead to spinning up “a bajillion… shell companies” to route GPUs to China [00:13:48].
Despite this, Chinese companies will need to “innovate” significantly [00:14:07]. While companies like DeepSeek are excellent engineers, they face a massive compute deficiency compared to US labs like OpenAI or Anthropic [00:14:10]. They must be “way, way, way, way better at engineering” to compensate for limited compute [00:14:30].
The focus on “test-time compute” for models requires immense training, including generating vast amounts of data, verifying it, and using complex reward models [00:14:51]. While training runs by frontier labs in the US cost billions of dollars and are rapidly increasing, test-time compute is a newer area where Chinese companies might “out engineer” with limited resources [00:15:53]. However, scaling these models for inference, which is crucial for widespread AI impact, will be extremely difficult due to GPU limitations [00:17:39].
The cost of developing software in China is “way cheaper than in the US,” allowing for internal development even if it’s not “best in class” [00:18:47]. This cheap labor challenges the “platform SAS” model, where a single solution is developed and sold widely [00:19:00].
Geopolitical Strategy and Future of Regulations
The current regulations are a “hard-handed approach” designed to significantly limit China’s compute capabilities, aiming for “order of magnitude less compute” than US labs [00:23:00].
The situation presents a “Goldilocks zone” challenge [00:23:23]:
- Too lenient, and China benefits too much from Western technology.
- Too strict, and China might be forced to build its own competitive domestic supply chain, which could increase geopolitical risks like action on Taiwan [00:23:41].
- The ideal scenario is to keep China “behind but not so far behind there’s hope of catching up” [00:24:29].
The US uses its hyperscalers to enforce its “sphere of influence” over AI [00:21:20]. For instance, G42, the UAE’s Sovereign AI champion, faced threats of being cut off before forming a partnership with Microsoft [00:20:41]. This partnership now sees G42 deploying GPUs in the US and Europe, aligning with American oversight objectives [00:21:13].
The current Biden administration’s regulations include a focus on human rights [00:21:39]. However, a future Trump administration would likely retain the rules but might remove the human rights aspect, shifting priorities towards “American economic influence or like industry or like… purchasing energy and weapons from the US government” [00:21:47]. This highlights how AI policy and regulation implications can serve as a “weapon for sort of the arsenal of democracy” [00:22:12].
Conclusion
The evolving landscape of AI regulation and safety is complex and far-reaching, fundamentally altering global technology and geopolitical dynamics [00:20:00]. The US aims to maintain its AI hegemony through strict export controls and strategic partnerships, pushing the development of massive compute clusters within its sphere of influence [00:02:22]. While this creates a near-monopoly for US hyperscalers and poses significant challenges and future of AI interpretability and regulation for countries like China, it also forces innovation and the exploration of new approaches to AI development under constrained conditions [00:14:07]. The long-term implications for global AI safety and regulation, economic competition, and geopolitical stability remain an ongoing, complex experiment [00:24:29].