The AI Diffusion Rules - Broader Implications

Dear Reader,

We cannot not talk about the AI Diffusion Rules. On 13th January, just days before Donald Trump’s inauguration at the White House, the Biden Administration released the Framework for AI Diffusion, its final policy salvo. This Interim Final Rule (IFR) aims to control the movement of two commodities, namely, large quantities of advanced chips (equivalent to Nvidia’s H100s) and the model weights of closed-weight AI models trained using more than 1026 computational operations. The stated objective underlying this framework is to safeguard the United States’ national security interests by maintaining leadership in foundational technology essential to innovation and to prevent the movement of commodities that can make a significant contribution to the military prowess of other countries to its detriment. At first blush, this appeared to be but the newest addition to the United States’ ever-expanding armoury aimed at haemorrhaging the strategic capabilities of its biggest geopolitical adversary, China. But upon closer examination, it is becoming clearer that the implementation of this new framework could very well mark the onset of a truly global compute scarcity that is far more manufactured than before, and characterised by the cementing of a stronger ideological position on US digital trade relations with the rest of the world.

The IFR specifically relies on a handful of key components for its operationalisation. First, it categorises countries into 3 distinct tiers. This categorisation is of utmost importance for it determines if a country can import American chips and under what specific set of conditions. The US and 18 of its allied partners have been placed in Tier 1, which signifies the total absence of import restrictions under the framework. Much of the globe, including India, has been placed in Tier 2. Finally, arms-embargoed countries like China, Russia, and geopolitically precarious regions like Macau, have been placed in Tier 3, which signifies a complete import ban.

Second, in the case of advanced chips, the IFR creates two separate mechanisms for regulating the export activities of companies headquartered in Tier 1 and Tier 2 countries respectively. Companies headquartered in Tier 1 countries are permitted to deploy AI chips in data centres located in both Tier 1 and Tier 2 countries through a ‘Universal’ authorisation process, subject to a few thresholds calculated as percentages of the companies’ total computing capacity. Specifically, such companies are required to ensure that at least 75% of their total compute power is located in Tier 1 countries, out of which a minimum of 50% must be located in the United States itself. Further, no such company is permitted to deploy more than 7% of its total compute in any individual Tier 2 country. On the other hand, companies headquartered in Tier 2 countries must undergo a ‘National’ authorisation process such that they must obtain separate authorisations for each specific country (barring Tier 3) they seek to deploy to. In this case, such companies are subject to a pre-determined maximum number of chips that can be exported through 2025 (~100,000 chips) per country. Similar caps have been determined for 2026 and 2027. In the absence of such national authorisation, however, one-time shipments of up to 50,000 chips are permitted to such countries, in addition to a set of 1,700 chips that are not counted towards this cap.

Third, the IFR introduces export controls on model weights, which are numerical parameters capable of being tweaked to reflect stronger or weaker neural connections to optimise the outputs of an AI model. The export restrictions however specifically target the weights of AI models that have been trained on more than 1026 computation operations, a training threshold that no contemporary model can claim to meet. Further, they only target those models whose model weights have not been openly published in the public domain. As per the IFR, companies in Tier 1 countries can freely deploy such weights to both Tier 1 and Tier 2 countries, provided they fulfill specified storage security compliances. Companies in Tier 2 countries in this regard are subject to far more stringent safety safeguards to prevent unauthorised access or model weight theft and maintain independence from Tier 3 countries. At the same time, the IFR requires authorised cloud service providers to take measures to prevent unauthorised training of controlled models by entities in Tier 2 and 3 countries.

Evidently, the AI Diffusion Rules, which are open for comments through the next three months, are by far the most complex and expansive US export controls on AI technology. It therefore comes as no surprise that they have already become the subject of nuanced debate both in the US and elsewhere. However, one of the more immediate questions being raised about the rules is regarding their very life expectancy under a new Trump administration. This is especially the case since, unlike a whole host of Biden-era Executive Orders (EOs) that have been revoked in the past week – including the EO on federal diversity hiring (2021), the EO on AI Safety (2024), and close to 17 EOs and Presidential Memorandums from January 2025 itself – the Diffusion rules have not been put through the meat grinder as yet. While this alone does not definitively answer the question, the ideological essence of the IFR is shared by many of the export control policies that were introduced under the first Trump administration. It would therefore be reasonable to expect the rules to not undergo any significant dilution in the near term.

If these rules do in fact survive the initial months of the Trump administration and manage to stick around, there are three key implications for both India and the world at large.

First, the rules very clearly divide the world into AI haves and have-nots, and who falls where will be decided by the United States. This is fundamentally problematic, as unlike say nuclear power in an earlier age, AI is a general-purpose technology that can be used for a wide variety of purposes, military but also civilian. The AI development chain, from the minerals required to fashion chips, to the chip design and manufacture and energy sources required to power data centres, is itself inherently global in nature. By seeking to impose export controls on the final aspects of AI, the rules disincentivize several country nodes in this chain from participating in the entire process, thereby potentially slowing down the development of the technology itself. Trying to impose global restrictions on AI’s trade and development is akin to trying to control the spread of the internet or electricity. It is neither practical nor desirable.

Second, precisely because these rules attempt to divide the world’s access to foundational AI technology as mentioned above, they are unlikely to achieve their aims of safeguarding America’s leadership in AI. It is more likely that the opposite happens. It is not difficult to imagine that even countries currently in Tier 1 will be uncomfortable with an arrangement where their ability to develop AI capabilities is dependent on American beneficence. Most countries will quietly go along with this situation for the time being because American companies are truly the leaders in this space. However, every country that can do so, including the US’s closest allies, will be looking to invest in developing indigenous capabilities. Those who cannot invest the necessary resources will then look to the next best option, China, which is very unlikely to impose a similar global mandate. While America remains the predominant AI power, China is catching up fast. Previous sanctions against China that sought to impede its AI progress have not achieved much apart from forcing China and Chinese companies to invest in indigenous capabilities, which is now beginning to pay off. The best example of this is a little-known open-source model called DeepSeek V3, which with a paltry training cost of $6 million, is as good as, if not better than the likes of Meta’s Llama and OpenAI’s O1 models. It seems almost like a no-brainer that given the complexity of dealing with the American export control regime, a lot of countries might choose to go with Chinese chips and products, even if they are marginally less effective than American ones, reducing the global market share of American companies in the medium term. The opposition from within the US tech industry itself has come most dominantly from chipmaking giants like Nvidia who argue that the IFR is misguided in that such restrictions will isolate the US tech industry from servicing the global AI market, giving Chinese counterparts a default right of way to do so. Prominent voices from within the US public policy space also suggest as much.

Third, for India specifically, these rules are unlikely to have any negative effects in the short term, as India’s current appetite for compute is far below the given cap for Tier 2 countries. For instance, the Indian government, through the National AI Mission, is looking at purchasing 10,000 GPUs. When combined with the private sector’s plans, the potential compute capacity in the country is unlikely to go beyond 30,000-40,000 GPUs in the near future. It will likely be at least 2-3 years before India starts coming close to the cap. The question before India however, much like the rest of the world, is not what are the immediate practical implications, but what are the philosophical implications that will have a knock-on effect on policymaking. There is no doubt that the Indian government will assume the US government is going to get increasingly uncomfortable as India’s AI capabilities keep getting better, thereby providing it greater impetus to focus on indigenisation of as much of the AI compute ecosystem as possible. The use of industrial policies in the digital world will also convince Indian policymakers, as well as other countries, that they are well within their right to impose their own ring-fencing policies in the digital sphere. Finally, the assumption that the world could share in American technology development and leadership, a sort of tech trickle-down, is now rested forever. While no country, not even China, can be completely isolated from global techno-economic relations, every nation will try to do one of two things: either carve out an unassailable niche for themselves the way Taiwan or the Netherlands have, or try and reduce their dependence on non-local firms for as much of the technology in question as possible.

The AI Diffusion Framework not only marks a pivotal moment in the US’s increasingly solidifying control over the global movement of emerging tech, but also underscores the urgency for countries like India to rethink their external dependencies on even strategic partners like the United States.

Thank you for reading. Have a great weekend.

Best,

Shashank & Shruti