AI Diffusion Strategy Under Strain?

Implications for the Developing World.

Published Feb 14, 2025

⋅

⋅

Investigation by

Rafael

Jonas

Joanna

Introduction: A January Trifecta of AI Events

In January 2025, a number of high-profile events in AI development and governance took place globally. These events set the tone for a turbulent year and carry great significance for developing countries and their strategic considerations, as they are compelled to navigate an ever-changing, high-stakes environment. While the ability of AI diffusion to bring about extreme economic growth remains under debate ¹ (OECD, 2023) it is clear that, as a transformative technology, AI will rearrange socio-economic realities as well as geopolitics.

A key way for states to reap benefits and stay ahead of the curve is by controlling AI diffusion, that is, the domestic and international proliferation of AI technologies. This can involve accelerating diffusion through large-scale public and private investments domestically or controlling diffusion to other countries, depending on geopolitical realities.

An example of the former is Project Stargate, a large-scale AI investment launched by SoftBank, OpenAI, and Oracle founder Larry Ellison, with the regulatory blessing and infrastructure support of the U.S. government. A major event illustrating the latter is the activation of the U.S.’s new Framework for Artificial Intelligence, which aims to control access to U.S. AI technology through a complex, tiered system.

Both of these efforts were somewhat upended by the release of DeepSeek-R1, an open-source LLM from the Chinese startup DeepSeek, which is claimed to be significantly more cost-effective than competing U.S. models.

In this brief opinion piece, we will analyze these three events from an AI diffusion perspective and then provide our take on what they might mean for the economic and policy landscape of developing countries.

What Happened? And What It Means for Diffusion

U.S. Framework for AI Diffusion

In the final week of the Biden administration, the U.S. Department of Commerce released a new interim final export control rule that aims to “regulate the global diffusion” of advanced artificial intelligence chips and models. It is a tiered-licencing mechanism that leverages the access to key enablers to strengthen U.S. AI leadership.

It is evident that the framework is by far the most comprehensive in its category. However, from a diffusion theory perspective, it is not entirely exhaustive. As noted in the research paper by our Head of Economic Research, Rafael ²(Andersson Lipcsey, 2024), AI diffusion occurs through several prominent pathways, including international trade, research collaboration, and inter-firm knowledge transfers.

With regard to international trade, the U.S. framework addresses the trade of AI chips and AI licenses through restrictions on non-public model weights. Although this approach initially seems both effective and realistic, the policy, whether by oversight or design, appears to overlook certain complexities of international trade.

In today’s globalized world, indirect effects should not be underestimated. Supply chains are complex, and multi stage, passing through several countries from the initial stage of raw material extraction to eventually becoming a final product or service. Export control regimes fail to capture such complexity, as they mostly target the flow of goods or services to or from a singular country. Similar to how sanctions on Russia have shown limited impact, or how price caps aimed at curbing inflation often fail, AI advancements will spread even to Tier 3 countries, albeit more slowly. Export control regimes usually target

The second major pathway, research collaborations, is not explicitly addressed in the document. Although the framework’s implementation may reduce the number of collaborations in AI or related fields, strong global commitments to academic freedom and unhindered research could actually elevate the significance of this pathway, particularly for countries facing restrictions on accessing novel developments.

Finally, inter-firm knowledge transfers are also not addressed directly, but they will undoubtedly be affected by limits on access to model weights and any subsequent spin-off effects. Nevertheless, we expect that active knowledge exchange among the U.S., Tier 1, and Tier 2 countries will largely continue on platforms like Substack and through direct communication between companies.

The Three Pathways of AI Diffusion from Conference Presentation of “AI Diffusion to Low-Middle Income Countries; A Blessing or a Curse?” at the 2024 Cambridge Conference of Catastrophic Risk

Project Stargate

Project Stargate is a massive AI infrastructure initiative, backed by OpenAI, SoftBank, Oracle, and MGX, and was announced by President Trump on January 21. Its primary goal is to hyperscale advanced data centers for AI, thereby enhancing AI capabilities and supporting the development of next-generation AI models.

Scheduled for completion by 2029, this USD 500 billion investment will receive substantial support from the U.S. government. President Trump has pledged the necessary energy access, invoking an emergency declaration to expedite infrastructure development ³ (Barron’s, 2025). The government has also promised regulatory backing and the creation of a comprehensive national AI strategy, which is intended to reinforce the United States’ global leadership in AI (White House, 2025) and offer greater regulatory certainty for developers working on frontier AI in the U.S.

In terms of AI diffusion, Project Stargate is likely to act as a significant accelerator both in the United States and abroad. While it may be argued that the initiative is well aligned with the U.S. diffusion strategy to maintain global leadership in AI, it is also expected to spark similar large-scale AI investments elsewhere, particularly in China, thereby intensifying the emerging AI race between the two countries.

Nonetheless, many uncertainties remain. These include highly publicized concerns, such as whether the project will secure the necessary funding (citing Elon Musk’s comments) and whether the U.S. government can overcome resistance to create the promised regulatory framework for the companies involved.

A less obvious but equally important consideration is that Project Stargate relies on the assumption that “more data equals more advanced models” will hold true for the next several years. While current signs point in that direction, there is no absolute certainty that this assumption will remain valid.

DeepSeek R1

The arrival of Chinese startup DeepSeek’s R1 large language model on January 21 has not only disrupted markets but also raised questions about the United States’ dominance in AI development and the effectiveness of using trade restrictions to control AI diffusion.

The R1 model performs similarly to OpenAI’s O1 and is both free and open source. Furthermore, if DeepSeek’s claim of a mere USD 5.6 million training cost is accurate, then its development cost was significantly lower than the approximately USD 1 billion needed to train leading U.S. frontier models ⁴ (Sellman, 2025).

All these claims, however, should be taken with a grain of salt. Development costs have not yet in a verifiable way been confirmed to be as low as claimed, nor do we know whether the Deep Seek breakthrough is one that will fully set the tone for Chinese AI advancements in the future. After all, it must not be forgotten that Deep Seek was trained using Nvidia chips that had entered China before the US’s chip export control policy was up and running.

That being said, DeepSeek R1 carries significance, and it raises the projected likelihood of China becoming an important source of AI diffusion in the next few years, a scenario that perhaps was almost unimaginable a month ago. Furthermore, the emerging AI race between the United States and China is poised to further accelerate this process.

Equiano Viewpoint - What does this all mean for Global South Nations?

Given the above context, a few salient questions come up for the developing world. Firstly, on a broad level, what does this mean for AI diffusion to the Global South? Secondly, what do the recent events have in store for access to critical AI development enablers for these countries? Thirdly, how will they impact geopolitical relations between Global South countries and the two main sources of AI diffusion: USA, and China?

Impact 1: AI Diffusion Rates to the Global South

Although the launch of a comprehensive diffusion control framework may initially suggest a significant slowdown in the proliferation of AI technologies, the situation is far from straightforward.

While the new framework does indeed pose some hurdles for accessing novel AI technologies, it nonetheless provides a structured set of guidelines that apply to almost every country worldwide. By contrast, and acknowledging the substantial differences between the two, past global trade agreements considered key facilitators of AI diffusion, such as the World Trade Organization’s Information Technology Agreement, excluded a large number of developing countries, including almost all of Africa.

Through its clearly defined access tiers and despite a high regulatory burden, the U.S. framework reduces complexity and could make it easier for developing countries to access the latest AI technologies than if they had to rely solely on bilateral agreements with the United States. This is likely the goal. As Lennart Heim ⁶ explained in his expert assessment of the framework, it aims to balance the twin objectives of maintaining U.S. leadership in AI while facilitating AI diffusion ⁶ (Heim, 2025).

However, only a few days after the release of this carefully balanced approach, the scales have already begun to tilt with the introduction of DeepSeek R1. While many questions remain, one point is clear: U.S. leadership in AI could be lost if the United States fails to remain the primary source of AI diffusion to developing countries.

DeepSeek’s emergence may underscore this concern among U.S. policymakers and could over time prompt either a reshuffling of U.S. AI trade policy or, at the very least, a more “friendly” interpretation of the criteria for Tier 2 countries to access AI knowledge and resources. We are aware that the news of DeepSeek R1 has recently led to cries for an increase in stringency of the current export control policies amongst US lawmakers. However, our prediction is that even if steps will be made in this direction initially, due to the above-detailed reasons, they will likely be rolled back over time.

Therefore, our view is that despite the new diffusion framework, AI diffusion to the developing world will likely accelerate over the next few years. This could be driven by a more relaxed U.S. stance on AI diffusion, an increased diffusion rate from China, or both.

Impact 2: Access to critical AI enablers

The trio of events that occurred in January appears to confirm earlier forecasts of computing power becoming centralized in the hands of a few key players developing frontier AI models. This development deepens the dependence of less AI-resourced economies on these powers. An overreliance on chips undermines the sovereignty of Global South nations and reinforces a bipolar, or possibly tripolar, world order dominated by the U.S., China, and potentially Europe.

That said, it is important to note that chips are only one critical enabler of AI innovation. Local talent and a robust data ecosystem are other key factors that Global South countries can and should, continue to develop and leverage. The focus should be on fostering stronger regional cooperation around critical AI enablers through harmonized regulations and established knowledge- and data-sharing networks. More broadly, prioritizing safe and efficient digital public infrastructure to facilitate responsible AI diffusion is a crucial policy direction, as is increasing collaboration with AI labs to co-create models that empower Global South communities to innovate.

Furthermore, a future where AI chips and massive data centers are the most important keys to progress is not guaranteed. Algorithmic developments may, over time, diminish the pivotal role of chips, and the distributed training of AI models, which has recently seen significant efficiency gains ⁷(Douillard et al., 2025), could enable cross-national AI development among resource-constrained countries.

Impact 3: Funding

The U.S. has recently scaled back funding and deprioritized development initiatives, including those led by USAID. While USAID previously committed $3.8 million⁸ to strengthen capacity building, research, and the deployment of safe and responsible AI in partner countries, including efforts to mitigate risks from synthetic content, there is now a critical need for stronger leadership in AI governance. To sustain meaningful global cooperation, it is essential to reconfigure international development funding mechanisms, such as AI4D (UK) and IDRC (Canada), leveraging them through the G7 AI program to ensure continued support for responsible AI innovation.

Impact 4: US-China Balance In Geoeconomic Relations

As mentioned, DeepSeek’s R1 challenges several assumptions about the current geopolitical landscape of AI, particularly the United States’ hegemonic role as a leader and primary source of AI diffusion. If the U.S. adheres strictly to its newly released diffusion framework, a more affordable, equally capable, and open-source model becomes a highly attractive alternative for resource-constrained firms in the Global South.

According to our research, and as observable in the figure below, many of these countries are already deeply integrated with China, especially via global value chains. Consequently, trade barriers are relatively low for firms in developing nations to source AI chips and models from China, assuming China achieves equal standing with the United States in AI, and any diplomatic costs of challenging U.S. influence are outweighed by the benefits of inexpensive AI technology.

Figure 1: Backward Global Value Chain Integration with China and the USA (2020) from AI Diffusion to Low-Middle Income Countries; A Blessing or a Curse?

Much of how this situation unfolds will depend on the extent to which Chinese AI labs will be able to keep up momentum, and whether Chinese foreign policy capitalizes on these advancements. Will China adopt a strategy similar to that of the Belt and Road Initiative? If so, the United States could lose geopolitical influence, and developing countries may be drawn closer to China, echoing trends seen in other economic sectors over the past decade.

Conclusion

The events of January 2025 mark a pivotal point in the global AI landscape, even against the already rapid pace of developments seen in recent years. In many respects, they reinforced the widely held view of the United States as the leading force in AI innovation and diffusion, supported by renewed investments and policies designed to secure its leadership. At the same time, while “Sputnik” comparisons may be hyperbolic, Deep Seek R1 still caught analysts and policymakers off guard, prompting fresh questions about the effectiveness of the United States’ AI strategy so far. Our analysis indicates that these events will likely accelerate the global rate of AI diffusion, despite tighter U.S. regulations on AI chips. This, combined with China’s emergence as a serious AI contender, reshapes the geopolitical outlook for Global South countries, bringing both new opportunities and potential risks. On one hand, AI adoption by firms in the Global South could lead to substantial productivity gains and, over time, positive GDP impacts. These gains may become even more accessible given the availability of robust, open-source models like Deep Seek R1. On the other hand, the potential benefits could be curtailed by market externalities tied to the increasing concentration of compute resources among a few major powers, growing U.S.-China geopolitical competition, and a renewed “America First” policy stance within the United States.

A collaborative global effort to ensure responsible AI innovation and diffusion benefits not only the Global South but all stakeholders. A “winner takes all” approach is unsustainable, since, over time, the only true “winner” of an unchecked AI race could be a superintelligent, uncontainable AI model itself, coupled with any particular state(s) or corporation.

Instead, the current shift in the AI power balance should serve as a wake-up call for intensified global cooperation. In November 2024, the International Network of AI Safety Institutes demonstrated how mutually beneficial AI evaluations can foster solidarity around AI safety. Initiatives like this, as well as the Hiroshima AI Process⁹, must be prioritized to promote regional and multinational cooperation that actively includes all developing countries.