The AI Diffusion Framework — How US Export Controls Are Reshaping Global AI Leadership

The Most Ambitious AI Export Control in History

In one of its final acts, the Biden administration introduced what the Center for Strategic and International Studies (CSIS) calls “the most ambitious exercise of technological statecraft in modern history.” The Framework for Artificial Intelligence Diffusion — commonly known as the AI Diffusion Rule — transforms what was once a simple export restriction targeting China into a sweeping global system that dictates who can import advanced GPUs, where they can be deployed, under what security conditions, and how AI capabilities can be shared.

Written by Barath Harithas, CSIS Senior Fellow, the analysis reveals a framework that simultaneously pursues two objectives in fundamental tension: preventing advanced AI chips from reaching China while maintaining the overseas infrastructure expansion that US hyperscale companies need to sustain their dominance. The result is a complex three-tiered classification system that has the potential to reshape global AI development — but also risks alienating allies and inadvertently strengthening the very competitors it aims to contain.

For organizations navigating the new landscape of AI investment and infrastructure decisions, understanding the AI Diffusion Rule’s mechanics, implications, and potential evolution is now essential strategic intelligence. The framework does not merely regulate chip exports — it restructures the entire global architecture of AI compute access.

Understanding the Three-Tier AI Diffusion Framework

The AI Diffusion Rule organizes the world’s nations into three tiers, each with dramatically different levels of access to the advanced AI chips that power frontier models. The metaphor CSIS employs is apt: first class, coach, and no ticket.

Tier 1 comprises 18 nations that receive near-frictionless access to US AI technology. These are close military and intelligence allies: Australia, Belgium, Canada, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Netherlands, New Zealand, Norway, South Korea, Spain, Sweden, Taiwan, and the United Kingdom. For these countries, the framework preserves essentially the pre-existing level of access, with some new reporting and compute distribution requirements.

Tier 2 encompasses the vast middle ground — nations that are neither close allies nor adversaries. This includes major economies like India, Brazil, Singapore, Malaysia, Indonesia, Saudi Arabia, and the UAE. These countries face strict quantitative caps on GPU imports and are deliberately kept a generation behind the AI frontier through allocation limits that shrink in real terms as hardware improves.

Tier 3 includes the “usual suspects” — China, Russia, Iran, North Korea, and other arms-embargoed nations — facing a presumption of denial for any AI chip exports. For these countries, the framework represents a hardening of existing restrictions rather than a new policy direction.

GPU Allocation Caps and the Shrinking Reality

The headline numbers in the AI Diffusion Rule appear more generous than they are. Tier 2 countries receive a total processing performance (TPP) cap equivalent to 49,901 NVIDIA H100 GPUs through 2027. This is not an annual allocation — it is a cumulative cap. Once exhausted, no additional chips can be imported until the policy is revised after 2027. Small purchases below 1,699 H100-equivalents do not require a license and do not count toward the cap.

However, the critical detail that the CSIS analysis highlights is how these caps shrink in real terms as hardware improves. Machine learning hardware performance doubles approximately every 1.9 years, according to Epoch AI research. This means that as NVIDIA releases more powerful chips, the same TPP cap buys progressively fewer physical units:

• 50,000 H100s → 21,987 B200s → 13,192 GB300s → approximately 6,373 chips in 2027-equivalent terms
• The small purchase threshold of 1,700 H100s shrinks to roughly 217 chips in 2027 terms
• Even Tier 1 NVEU allocations of 320,000 H100s compress to about 40,773 in 2027 terms

The practical effect is that Tier 2 countries find themselves on a technological treadmill: the cap appears stable on paper while delivering steadily less computational capability in practice. For nations planning multi-billion-dollar data center investments, this shrinking reality creates profound uncertainty about whether their infrastructure will have access to the chips needed to remain competitive.

At approximately $1.25 billion for 50,000 H100s, the financial stakes are enormous. For context, training GPT-4 required roughly 25,000 A100 chips — a significant portion of a Tier 2 nation’s entire allocation even before accounting for the inference compute needed to serve a trained model to users.

Tier 1 Nations — The Inner Circle of AI Access

Tier 1 nations enjoy privileged access but face their own constraints. US companies operating in Tier 1 countries must maintain at least 50% of their total AI compute within the United States and at least 75% across all Tier 1 nations combined. No more than 7% can reside in any single Tier 2 country. These compute distribution requirements ensure that AI infrastructure remains concentrated in allied nations even as it expands globally.

To qualify for maximum access, data center operators can pursue Universal Validated End User (UVEU) status — but this requires meeting 19 separate certification and policy requirements covering physical security, personnel vetting, cybersecurity standards, and government access provisions. The compliance burden is substantial, creating a significant barrier even for companies in friendly nations.

The US currently hosts more than 50% of global AI data center capacity, with Tier 1 nations collectively controlling over 75%. North America’s planned data center capacity through 2027 is approximately four times larger than either Europe/Middle East or Asia Pacific. Ten or more gigawatt-scale campuses are projected in the US alone over the next two years, ensuring that the concentration of AI infrastructure in allied nations deepens rather than disperses.

Among Tier 1 beneficiaries, Australia emerges as the standout winner. Positioned to capture displaced demand from restricted Southeast Asian markets, Australia’s data center sector is rapidly scaling to absorb investment that might otherwise have flowed to Malaysia, Indonesia, or Singapore. With approximately 3 GW of planned capacity and colocation comprising 75% of the market, Australia’s AI infrastructure is growing precisely when competitors face new constraints.

Tier 2 Countries — Strict Caps and Strategic Frustration

The Tier 2 experience is defined by a fundamental mismatch between ambition and allocation. Nations like Malaysia, India, Singapore, and the UAE have invested billions in AI infrastructure on the assumption of continued access to frontier hardware. The AI Diffusion Rule retroactively restricts that access, potentially transforming purpose-built data centers into stranded assets.

Malaysia illustrates the problem most dramatically. The country’s data center capacity exploded from 100 MW in 2023 to a projected 3.5 GW by 2027, putting it on track to become the world’s third-largest data center market by 2026. NVIDIA signed a $4.3 billion partnership with Malaysian conglomerate YTL. ByteDance committed to leasing 628 MW. Oracle announced $6.5 billion in planned investments. All of this infrastructure now faces uncertainty about whether it will have access to the advanced chips it was designed to house.

India faces similar constraints despite its strategic importance to the United States. With approximately 3 GW of planned capacity — including Mukesh Ambani’s 3 GW mega data center in Jamnagar representing a $20-30 billion investment — India’s AI ambitions collide with Tier 2 allocation limits that may prove insufficient for a country of 1.4 billion people pursuing AI-driven economic transformation.

The NVEU pathway offers Tier 2 data centers a route to higher allocations, but the gap between UVEU approval (available to Tier 1 operators) and NVEU approval creates economic pressure. Tier 2 operators face months or years of uncertainty about their authorization status, during which they cannot plan hardware procurement, customer commitments, or infrastructure expansion with confidence.

Winners and Losers Under the AI Diffusion Rule

CSIS’s analysis maps the uneven global impact with striking clarity. Among the biggest losers, Malaysia stands out, followed by India and the broader Southeast Asian region. Singapore faces particular irony: having imposed a 2019 moratorium on new data centers to manage energy consumption (data centers are projected to consume 12% of its electricity by 2030), it now finds its limited capacity further constrained by export controls it had no hand in designing.

Brazil’s Scala Data Centers, operating a São Paulo campus exceeding 350 MW with an audacious “AI City” proposal of 4.75 GW (approximately $90 billion), faces similar questions about hardware access. Gulf states including the UAE (G42/Khazna with 406 MW across 13 campuses and a 5 GW pipeline) and Saudi Arabia (Google eyeing a 3 GW pipeline near King Salman Energy Park) find their long-term AI infrastructure plans subordinated to US foreign policy objectives.

The Tier 1 winners list includes Australia, the United Kingdom, Japan, Ireland, Germany, Canada, South Korea, Netherlands, and Spain. Nine of 18 Tier 1 countries have more than 1 GW of planned AI data center capacity, with five exceeding 2 GW — compared to only four Tier 2 countries with more than 1 GW planned. The framework accelerates a concentration of AI infrastructure in allied nations that was already underway but is now backed by regulatory force.

China’s AI Hardware Gap and Strategic Response

The AI Diffusion Rule’s ultimate target is China, and CSIS’s assessment of China’s current hardware position provides important context. Huawei’s best available AI chip, the Ascend 910B, delivers 280-400 TeraFLOPS compared to NVIDIA Blackwell’s 2,250 TeraFLOPS — a gap of 5.6 to 8 times. Of 263 documented AI models with known hardware configurations, only 2 used Huawei Ascend chips, while 31 Chinese organizations built their models on NVIDIA hardware. DeepSeek, China’s most prominent recent AI achievement, was trained on NVIDIA H800s.

China is projected to produce approximately 1.8 million Huawei Ascend 910B/910C GPUs by the end of 2025. The United States will deploy 14.3 million AI accelerators that are significantly more performant. Huawei is estimated to be one to two generations behind in both GPU performance and fabrication capacity — a meaningful but not insurmountable gap.

The critical question is whether China’s hardware disadvantage translates into a lasting AI capability gap. The hardware performance differential is clear today, but China’s track record of rapid technological catch-up in other domains — telecommunications, high-speed rail, space technology — suggests that a static assessment of the gap may be misleading. If China’s chip fabrication capabilities advance faster than US export controls assume, the framework’s core assumption of sustained hardware superiority could erode.

The Open-Source Paradox and Third Way Risk

Perhaps the most strategically dangerous unintended consequence of the AI Diffusion Rule is what CSIS identifies as the open-source paradox. By restricting access to advanced chips for closed, proprietary AI systems, the framework inadvertently incentivizes migration to open-source AI models that can run on less restricted hardware. Open-source models, once released, operate beyond the reach of export controls — they can be downloaded, modified, and deployed without hardware-specific restrictions.

China is strategically positioned to exploit this dynamic. Through projects like DeepSeek and other open-source initiatives, Chinese AI labs are positioning themselves as patrons of the global open-source AI community. If Tier 2 nations, frustrated by US restrictions on proprietary systems, turn to Chinese-led open-source alternatives, the AI Diffusion Rule could achieve the precise opposite of its intended outcome: expanding rather than constraining China’s AI influence.

The “Third Way” risk compounds this concern. Tier 2 nations may not choose between the US and China — they may chart an independent path. Not open rebellion but slow, deliberate drift from US-controlled AI infrastructure toward sovereign alternatives. Mid-tier compute stacks built on RISC-V architectures, specialized ASICs, and locally developed AI frameworks could gradually reduce dependence on both US hyperscalers and Chinese alternatives.

CSIS draws an empire analogy: the benefits of the system must seem within reach even to those who will not immediately grasp them. If Tier 2 nations perceive the framework as permanently consigning them to second-class AI status, resentment rather than cooperation will drive their strategic choices. The difference between a durable system and a brittle one lies in whether subordinate members see a credible path to advancement — or only a ceiling on their aspirations.

Three Conditions for Framework Stability

CSIS identifies three conditions that must hold simultaneously for the AI Diffusion Framework to remain stable and achieve its objectives. If any one breaks, the entire structure could unravel:

First, compliance costs must remain lower than switching costs. As long as it is cheaper and easier for Tier 2 nations and their data center operators to comply with US restrictions than to build or migrate to alternative compute stacks, the framework holds. But if compliance becomes sufficiently burdensome — through allocation reductions, additional certification requirements, or unpredictable enforcement — the calculation tips toward alternatives.

Second, alternate compute stacks must remain inefficient. The framework depends on the assumption that non-US AI hardware (Huawei, RISC-V-based alternatives, specialized ASICs) cannot deliver competitive performance at scale. If alternative architectures close the performance gap faster than expected, Tier 2 nations gain viable options outside the US-controlled ecosystem. The history of technology suggests that architectural diversity often produces surprising competitive outcomes.

Third, China’s AI development must proceed slower than expected. If China demonstrates frontier-level AI capabilities despite hardware restrictions — as the DeepSeek development suggests is possible with algorithmic efficiency gains — the entire rationale for restricting Tier 2 access weakens. Why accept second-class status in a US-led system if China can deliver comparable AI capabilities through alternative channels?

The fragile equilibrium these conditions describe means that US policymakers face a continuous balancing act. Too much restriction drives nations toward alternatives; too little restriction enables the very diversion the framework was designed to prevent. Finding the sustainable middle ground requires diplomatic finesse that technological statecraft has historically struggled to achieve.

CSIS Recommendations for Strengthening US AI Leadership

The CSIS analysis concludes with three substantive recommendations designed to preserve the framework’s objectives while reducing its counterproductive effects.

First, establish clear graduation requirements for Tier 2 countries. Creating a “Tier 2A” classification would give nations with significant at-risk AI investments a path to higher allocations in exchange for increased export control enforcement cooperation. This addresses the resentment problem directly — countries that demonstrate commitment to preventing chip diversion receive tangible rewards, creating positive incentives alongside the restrictions. Annual revision of TPP thresholds, particularly as US mega-projects like Stargate scale up and demonstrate that domestic capacity is secure, would signal good faith.

Second, reassert US leadership in open-source AI. Rather than ceding the open-source landscape to China by default, the US should proactively shape it. Incentivizing international AI research collaborations, providing US university public compute access for vetted Tier 2 researchers, and developing certification standards for “trusted” open-source models that meet security requirements would maintain US influence in the open-source ecosystem that the framework inadvertently strengthens.

Third, modernize the Bureau of Industry and Security (BIS) and its enforcement capabilities. The US export control regime has grown by approximately 1,000 pages over 30 months, creating complexity that overwhelms existing enforcement infrastructure. Deploying automated TPP tracking systems for real-time GPU deployment visibility, creating standardized APIs for reporting and monitoring, and building automated early warning systems for potential diversion would transform enforcement from a paper-based compliance exercise into a dynamic, technology-enabled capability.

Together, these recommendations sketch a path toward an AI diffusion framework that maintains US leadership while offering credible partnership rather than mere subordination to the nations whose cooperation is essential for the framework’s long-term viability. Whether US policymakers adopt this more nuanced approach — or maintain the current blunter instrument — will shape the global AI landscape for decades to come.