Since the release of ChatGPT in November 2022, the breakneck pace of progress in artificial intelligence has made it nearly impossible for policymakers to keep up. But the AI revolution has only just begun. Today’s most powerful AI models, often referred to as “frontier AI,” can handle and generate images, audio, video, and computer code, in addition to natural language. Their remarkable performance has prompted ambitions among leading AI labs of achieving what is called “artificial general intelligence” (AGI). According to a growing number of experts, AGI systems equaling or surpassing humans across a wide range of cognitive tasks—the equivalent of millions of brilliant minds working tirelessly at the top of their fields at machine speed—may soon be capable of unlocking scientific discoveries, enhancing economic productivity, and tackling tough national security challenges. With advances once in the realm of science fiction now in the realm of possibility, the United States has no time to spare in crafting a coherent and truly global strategy.
Given its game-changing potential, the countries that best innovate, integrate, and capitalize on frontier AI—especially as it approaches AGI—will accrue significant economic, military, and strategic benefits. But for the United States, failure to influence how the technology diffuses around the world carries two profound risks. One is the prospect that uncontrolled diffusion of advanced AI could empower the world’s most dangerous state and nonstate actors, and potentially rogue autonomous AI systems themselves, to develop catastrophic cyber- and bioweapons or unleash other existential national security threats. The other is that authoritarian powers, most notably China, could come to dominate the global AI technology stack in ways that embed censorship, surveillance, and other antidemocratic principles throughout the digital ecosystems that have come to define our lives.
The Trump administration is well positioned to take advantage of the AI policies put in place by the Biden administration to ensure that the United States and its democratic allies win the global AI competition. But doing so will require more than just doubling down on the United States’ technological edge. It will also necessitate partnering with the private sector to up the country’s AI offering, both at the frontier and in “good enough” AI, to outcompete China around the world. The Trump administration can either choose to lead in shaping the rapidly emerging AI future—or watch as this brave new world is built by Beijing.
Leading AI labs such as Anthropic, Google DeepMind, and OpenAI partner with U.S. hyperscalers Amazon Web Services (AWS), Google Cloud, and Microsoft Azure to provide the computational resources (or “compute”) needed to train and run frontier AI models, while Meta and xAI combine proprietary data centers with external cloud services. These data centers rely heavily on advanced semiconductors, particularly graphics processing units, known as GPUs. U.S. companies Nvidia and AMD originally designed and developed GPUs to render video game graphics, but AI labs found that they excel in performing the massive number of simultaneous calculations needed to train deep learning models. Amazon and Google have designed their own specialized chips in an effort to make AI workloads even more efficient.
Progress in frontier AI has relied heavily on scaling compute and data. U.S. companies are banking on this trend continuing. Last year, Elon Musk’s xAI constructed its Colossus data center, with 100,000 Nvidia H100 GPUs to train the company’s Grok models, in Memphis, Tennessee, and has raised $5 billion to increase the center’s cluster of GPUs tenfold. Other leading U.S. AI labs and hyperscalers are planning similarly massive data centers.
Emerging frontier models have challenged the maxim, common among AI technologists, that inference—using trained models to respond to queries, make predictions, and generate outputs based on new, unseen data—is less compute intensive than training. Frontier AI models have come to rely on “test-time” compute, in which a model dedicates more resources during inference to engage in chain-of-thought “reasoning” and improve performance on complex tasks. The proliferation of models with larger context windows (the amount of text a model holds in its memory) and a rapidly growing user base are further driving escalating demands for compute.
Because compute is central to frontier AI, Washington has focused on restricting China’s access to advanced AI chips and chipmaking equipment. The Trump administration devised this “denial” strategy in 2018 and 2019, when the United States successfully pressured the Netherlands to block China’s acquisition of extreme ultraviolet lithography equipment, exceedingly complex machines critical in the creation of advanced semiconductors, made by the Dutch company ASML. Starting in October 2022, the Department of Commerce’s Bureau of Industry and Security (BIS) intensified these controls, initially restricting the sale of top GPUs, such as Nvidia’s A100 and H100 chips, along with other AI accelerators, to China. To extend the territorial reach of U.S. controls, the Biden administration also imposed a Foreign Direct Product Rule covering foreign-made items derived from U.S. semiconductor technology. A year later, the administration expanded the measures to cover advanced GPUs that had been only slightly modified to satisfy previous restrictions and, in December 2024, it added high-bandwidth memory chips, older immersion deep ultraviolet (DUV) lithography machines, and other critical chipmaking software and tools. Implementing these controls has required significant and sometimes contentious negotiations with U.S. allies, especially the Netherlands; Japan, home to equipment makers Tokyo Electron and Nikon; South Korea, home to semiconductor producers Samsung and SK Hynix; and Taiwan, home to the world-leading chipmaker Taiwan Semiconductor Manufacturing Company (TSMC).
Progress in frontier AI has relied heavily on scaling compute and data.
These restrictions have undeniably slowed China’s access to advanced chips and hindered its ability to produce substitutes. SMIC, China’s most prominent chipmaker, has used existing DUV machines to manufacture some advanced chip nodes for smartphones. It also reportedly produced Huawei’s Ascend 910 AI chips, which Huawei asserts match the performance of Nvidia’s widely used A100s. But domestically manufacturing such chips with older DUV machines is expensive, reduces yield, and undermines reliability. Moreover, Huawei’s supposedly SMIC-produced Ascend 910B chip sets actually contained chips produced by TSMC, which TSMC had unknowingly sold to a Huawei front, casting doubt on SMIC’s true capabilities. In November, BIS directed TSMC to end all sales of its most advanced AI chips to China and has since blacklisted Sophgo, the Huawei cutout.
Meanwhile, U.S. chip designers are pulling further ahead. Nvidia’s leading, TSMC-manufactured H100s and H200s and new Blackwell chips are substantially faster than China’s best. Experts generally assess China to be at least five years behind leading-edge chip producers, with export controls slowing Beijing’s catch-up effort.
Nevertheless, the computing power gap has not stopped Chinese tech giants such as Alibaba and Tencent, and startups such as 01.AI, DeepSeek, Moonshot AI, and Zhipu AI, from releasing high-performing generative AI models. Chinese firms have capitalized on data centers equipped with Nvidia chips prior to the United States’ imposition of export controls, used downgraded chips not covered by U.S. controls, and optimized software to maximize less capable hardware. Crucially, many successful Chinese AI models rely on open-source models already released by U.S. labs or use outputs from U.S. models for training.
Despite these achievements, U.S. AI labs likely remain one or two years ahead at the frontier, especially since many not-yet-released models are closed-source and therefore harder for Chinese companies to emulate. And as long as scaling state-of-the-art computing power remains vital for frontier AI progress, U.S. companies will expand their lead. As DeepSeek’s CEO Liang Wenfeng has acknowledged, China’s difficulties competing with U.S. AI firms boil down to Washington’s “bans on shipments of advanced chips.”
Not content with second place, Chinese entities have attempted to circumvent U.S. controls, including by smuggling thousands of advanced semiconductors and purchasing chips through intermediaries not subject to export controls and, most consequentially, taking advantage of loopholes in the U.S. export control regime to remotely access overseas cloud providers. According to Reuters, some state-run Chinese entities and research institutes have accessed computing resources and AI models via the cloud from AWS and Microsoft Azure. And in some instances, Chinese customers seem to have masked their true identity to use foreign data centers for AI training.
U.S. officials worry that global AI data center expansion could provide indirect pathways to restricted technologies. The Chinese firm ByteDance, which owns TikTok, for example, plans to spend $7 billion to access controlled Nvidia chips, including the next-generation Blackwell chips, by using cloud services in Southeast Asia and elsewhere that are not covered by U.S. controls. Proposed construction of large-scale AI training clusters in Saudi Arabia and the United Arab Emirates (UAE) have also drawn considerable scrutiny. Despite close security ties to the United States, both countries maintain deep economic and technological relationships with China, leaving many in Washington to worry that their grand AI ambitions could end up benefiting Beijing.
To mitigate these concerns, BIS expanded its export controls in October 2023 and again in April 2024 to reduce the risk of diversion to China, widening the same blanket controls applied on China to include a select group of third countries, predominately in the Middle East, requiring local entities to obtain licenses and prove they had no connections to China. These measures were successful: under pressure from the Biden administration and members of Congress, Abu Dhabi‒based AI firm G42 removed Huawei equipment from its infrastructure, divested from Chinese companies, committed to complying with U.S. export controls, and agreed to stringent security measures to prevent unauthorized Chinese access to its advanced technologies. These changes paved the way for Microsoft to invest $1.5 billion in G42, and the two companies announced a $1 billion initiative to expand Internet connectivity in Kenya, as well as conduct AI research, develop local-language AI models, and build a state-of-the-art Microsoft Azure data center in Kenya for a new East Africa Cloud Region.
Following a year of negotiations, in September 2024, BIS unveiled a Validated End User (VEU) program for data centers to ease shipments of AI chips to the Middle East. The rule allowed approved facilities to receive designated items under a general authorization, rather than requiring individual licenses for each shipment. To secure VEU status, data center operators are required to provide detailed information about their customers, business activities, facility access controls, and cybersecurity measures; permit on-site reviews by U.S. officials; and pledge to use the chips securely and responsibly.
Chinese entities have attempted to circumvent U.S. chip controls.
The rule set the stage for Biden’s closing AI act. In its final days, the administration released its long-anticipated Export Control Framework for the Diffusion of Artificial Intelligence outlining how Washington intends to manage worldwide access to crucial AI technologies. The future of U.S. AI leadership could depend on whether the Trump administration decides to follow it.
At its core, the framework introduces a three-tiered global licensing regime for advanced AI chips. Tier 1 comprises the United States and 18 close U.S. security allies and partners—many of which play critical roles in the global AI supply chain. These countries face no AI chip restrictions and are expected to help enforce the framework. Tier 2, which includes most of the world, receives managed and mediated access to advanced AI chips. Tier 3 countries, such as China, Russia, Iran, North Korea, and Venezuela, are under U.S. arms embargoes and remain subject to stringent export controls.
The framework’s most debated aspect is a quota system for Tier 2 countries, requiring validated end users to act as gatekeepers for large-scale AI chip access. It identifies two types of VEUs: “universal VEUs,” which are globally authorized companies headquartered in Tier 1 nations, such as U.S. cloud companies; and “national VEUs,” authorized firms headquartered in Tier 2 countries. Both types of companies must meet the same rigorous safety and security standards, but they face different limitations. Universal VEUs are expected to maintain at least 75 percent of their total AI computing power in Tier 1 countries, with U.S. firms guaranteeing 50 percent in the United States, but can allocate up to seven percent of their global capacity to any Tier 2 country. They may also store advanced closed-source model “weights”—the numerical parameters an AI model uses to make predictions and decisions—in Tier 2 countries, provided strict cybersecurity measures are in place.
In contrast, national VEUs face a cap limiting their chip imports in each Tier 2 country—up to 320,000 chips through 2027—and cannot store advanced closed-source AI model weights. Companies in Tier 2 countries without VEU status must compete for licenses to import a smaller pool of AI chips, capped at a total of 50,000 H100-equivalents per country from 2025 through 2027. Governments that sign commitments with Washington to align their export control, energy, and technology security arrangements with the United States can double their cap, to 100,000 chips. Finally, a licensing exemption also allows for more limited chip imports (roughly 1,700 H100 equivalents), covering most conceivable academic research and government needs in Tier 2 countries.
Any comprehensive AI strategy must ensure that the most advanced models are trained in secure data centers housed in the United States and allied countries, while denying China and other adversaries the computing power to catch up. At the same time, it must assertively promote the global adoption of U.S. and allied AI technologies to shape AI deployment and norms. Nowhere is this balance between denial and diffusion more critical than in the countries across Africa, Latin America, the Middle East, and South and Southeast Asia turning to AI to catalyze economic growth to meet development goals such as food security, public health, education, energy security, climate resilience, and efficient service delivery. The competition for AI supremacy is not just technological. It will also hinge on whether Washington or Beijing can convince countries across the global South of its vision for the AI future.
Digital sovereignty concerns drive many global South nations to seek local AI data center construction to protect privacy, comply with regulations, and reduce dependence on foreign infrastructure. Locally hosted AI models also enable more reliable performance on important applications, including high-frequency financial transactions and critical infrastructure monitoring, than data centers located continents away. And many nations see the construction and operation of local data centers as drivers for employment and homegrown AI ecosystems.
But building large-scale, cutting-edge AI data centers in every country is financially and logistically infeasible given the enormous upfront costs, uneven availability of skilled labor, and high demands for land, energy, and water. Moreover, from a strictly technological perspective, it is unnecessary; most countries can access AI models and applications remotely via the cloud. Many low- and middle-income countries would benefit most from focusing scarce resources on improving electricity reliability and the availability of broadband and small connected “edge” devices that access AI systems.
The competition for AI supremacy is not just technological.
A hybrid approach that consolidates foundational AI research in advanced cloud hubs while building regional and local data centers to handle specialized tasks would better balance cost, performance, respect for sovereignty concerns, and the need for more equitable AI access across the world. The Biden administration’s strategy aimed to encourage the emergence of such a structure.
Critics of Biden’s approach argue that limiting AI chip exports will rankle U.S. partners with AI ambitions, especially those in the Middle East and South and Southeast Asia, pushing them into China’s waiting arms. As Reva Goujon recently observed in Foreign Affairs, capping the sale of U.S.-made AI chips abroad risks weakening the United States’ long-term leverage in fast-growing markets. Trump’s former National Security Adviser Robert O’Brien, along with leading companies Nvidia and Oracle, likewise contends that restricting U.S. companies from freely selling advanced AI chips may allow China to corner the AI market.
This criticism overlooks two key points. First, the Biden administration framework still allows for building data centers in Tier 2 countries that could house hundreds of thousands of cutting-edge chips—leaving minimal hypothetical gaps for China to fill. Second, since U.S. export controls have hampered China’s ability to produce cutting-edge GPUs, Chinese companies, which lack the technology and capacity to produce high-quality advanced chips at volume, cannot easily supply the global market at the high end of AI.
The real challenge for Washington lies not in backfilling unmet global demand for cutting-edge chips but in dominating “good enough” AI systems deployed around the world. China’s downgraded GPUs and domestically produced chips, though weaker than Nvidia’s best, remain sufficient for most commercial AI tasks. Even if the world ends up preferring U.S. and allied cloud companies for large-scale AI model training and ultra-high-capacity inference (where access to the latest chips remains critical), China’s chips may remain attractive to many countries in the global South.
China is already well positioned in this midrange market, particularly in the global South, where Chinese firms have helped build “smart” cities, composed of interconnected networks of cameras, sensors, and communication devices, and digital port and logistics networks that integrate Chinese AI models and surveillance capabilities. China has also trained thousands of AI engineers in two dozen countries through its Luban Workshop program, demonstrating a focus on developing AI talent that the United States lacks. Chinese firms also openly release their AI models, encouraging countries and companies around the world to build on top of Chinese AI software.
Beijing has long pushed AI solutions through its Belt and Road Initiative, and Xi Jinping recently made China’s Global AI Governance Initiative one of the cornerstones of the country’s grand strategy. As China continues to lag behind the United States at the high end, Beijing will likely double down on offering midrange AI data centers around the world, encouraging global South customers to adopt its AI models and cloud services. To cut into the roughly 70 percent global cloud market share of AWS, Google Cloud, and Microsoft Azure, Chinese companies Alibaba, Huawei, and Tencent have plans for significant expansions of AI-focused data centers in Africa, Latin America, and Southeast Asia. Chinese smartphone makers Xiaomi, Oppo, and Vivi also have strong market positions in many global South countries, potentially allowing them to integrate AI apps directly into local consumer devices.
China has made AI hubs in the global South a foreign policy priority.
Although China’s strength in “good enough” AI could grant it a foothold, the generalizability and highly adaptable nature of the U.S. frontier AI models, if bundled with required infrastructure, could provide a compelling value proposition for many global South countries. Nothing in the Biden administration’s diffusion framework prevents U.S. companies from seizing this competitive advantage. But market forces alone will not guarantee that the United States will adequately counter China’s AI offering.
Large U.S. tech companies are increasingly investing in such countries as Brazil, India, Indonesia, Kenya, Malaysia, and South Africa. But U.S. firms remain wary of high-cost, long-term projects in countries with emerging markets that lack the scale or purchasing power to justify large investments without external support. By contrast, China has made AI hubs in the global South a foreign policy priority and uses subsidized financing and state-backed loans to fund affordable AI applications and the required underlying infrastructure, including stable Internet and electricity, at cut-rate prices.
The Biden administration recognized the need to rival China in this lane. In September 2024, it launched the Partnership for Global Inclusivity on AI alongside major U.S. firms to expand access to AI models, compute credits, and open-source tools, while mobilizing billions in digital infrastructure investments. But without continued U.S. public-private collaboration and targeted policies, China could end up dominating these emerging markets even as the United States maintains its technological lead.
Biden’s AI strategy has helped the United States protect its edge over China at the AI frontier—an achievement carrying enormous economic, national security, and geopolitical implications. Nevertheless, opponents, including some in the Trump administration, may dispute parts of this strategy, particularly the global diffusion framework, arguing that fewer export restrictions would unlock partnerships with wealthy and energy-rich nations like Saudi Arabia and the UAE, which could prove crucial in outcompeting Beijing.
But dismantling current restrictions would be counterproductive. Simply flooding the zone with the world’s best AI chips without establishing guardrails against diversion and misuse by U.S. adversaries risks unleashing a torrent of those chips—and the remote high-performance compute capacity enabled by them—flowing back to China.
Instead, the Trump administration should continue to refine export controls to stay ahead of China’s adaptations, coordinating closely with allies and offering countries willing join the U.S. AI ecosystem a rigorous, but clear pathway into Tier 1 status. Negotiations with Middle Eastern partners and critical swing states such as Brazil, India, and Malaysia could offer Trump useful leverage to shape partner behavior and protect U.S. national security.
Meanwhile, the Trump administration should take more ambitious actions to distribute the benefits of U.S. frontier AI models and compete with China’s delivery of “good enough” AI and digital infrastructure across the global South. To start, the new administration should form deeper partnerships with the private sector, allied governments, international organizations, and philanthropic institutions to offer significantly more generous cloud computing credits to entrepreneurs and researchers in the global South. Access to these credits would enable the training and fine-tuning of AI models through trusted U.S. and allied cloud providers, enhancing the global appeal of U.S. AI while allowing for the detection of malicious AI uses that could threaten U.S. national security.
Biden’s AI strategy has helped the United States protect its edge over China.
The Trump team can also proactively counter China’s AI outreach in the global South by promoting additional investment by U.S. companies in AI and related digital infrastructure. The U.S. International Development Finance Corporation, the Export-Import Bank, and other government-backed finance institutions can increase low-interest loans, loan guarantees, export credits, and equity investments in digital projects to free up billions with which to compete with Beijing’s digital initiatives. Political risk insurance and partial credit guarantees should also be deployed to mitigate concerns about expropriation, regulatory changes, and currency convertibility— thereby unlocking private capital that might otherwise be deterred by these uncertainties.
U.S. government agencies can also support or fund feasibility studies, environmental assessments, and market research to provide information and lower project costs for firms considering investments in global South infrastructure. And the Trump administration should consider offering tax benefits to encourage U.S. firms and offset the risk of expansion into new markets.
Skeptics may doubt whether Trump, with his penchant for transactionalism, will prioritize investments in countries that may not yield immediate profit for the United States. But there is precedent suggesting Trump understands the importance of nonaligned countries in the competition with China. In 2019, the first Trump administration launched the Blue Dot Network to promote high-quality, sustainable infrastructure development projects in the global South to counter the Belt and Road Initiative. Biden expanded the effort through the G-7’s Partnership for Global Infrastructure and Investment, with the goal of marshaling hundreds of billions of dollars in public and private capital for infrastructure investments, including for digital infrastructure projects, across the global South.
It is now time for Trump to build on Biden’s approach. Winning the global AI competition will require protecting and increasing the United States’ technological lead. But it also necessitates more active engagement with the nations that will ultimately determine how—and whose—AI is deployed worldwide. Failing to balance both risks a paradoxically grim future in which the United States beats China to the AI frontier, only to cede global leadership of the emerging AI world order to Beijing.
Loading...
Please enable JavaScript for this site to function properly.