AI Supply Chains: Geopolitical implications

Computing power as the driving force of AI progress

All advances in AI today are driven above all by one factor: the application of ever greater amounts of computing power. Looking at the history of AI from the 1950s to the present, the amount of computing power deployed has grown at an extraordinary rate, increasing by orders of magnitude each decade, and has accelerated further over the past 15 years with the onset of the “deep learning era,” the era that enabled AI systems such as ChatGPT. Academic studies consistently find that improvements in computing capabilities have contributed far more to AI progress than algorithmic advances alone. The computer scientist Richard Sutton calls this the “bitter lesson,” his argument being that virtually all gains in computer science have been dwarfed by the raw impact of scaling compute power. Technology leaders across the industry share this view, and their overriding priority today is acquiring the computing power their AI systems need. On a regular basis, leading AI executives have had to publicly apologize to users for being unable to offer the full range of AI services their platforms would like to provide, simply because the computational demands outpace what even the most well-resourced data centers can currently handle.

Today, even the most influential tech companies are constrained by insufficient computing power. Demand for AI video generation, for example, is so computationally intensive that leading companies must limit user access because their data centers cannot keep pace. This demand is equally acute in China, where technology leaders at Alibaba, Tencent, Baidu, and DeepSeek cite access to high-end AI chips as their most pressing challenge. This is one reason why Chinese firms have favored open-source AI models that can be deployed locally on users’ own hardware rather than relying on centralized data centers. The largest companies in the world are now almost all engaged in designing, manufacturing, or deploying AI chips. The fact that the most valuable publicly traded company in Asia is Taiwan Semiconductor Manufacturing Company (TSMC), the world’s most advanced chipmaker, reflects the transformative economic weight of this trend. Meanwhile, in the United States, three major AI companies are expected to go public this year at valuations in the hundreds of billions of dollars, among the largest IPOs ever seen, further illustrating how central computing infrastructure has become to global capital markets.

AI supply chains and military power

However, the AI supply chain is not merely an economic matter; it is deeply intertwined with military power. From the earliest Cold War guidance computers to today’s autonomous weapons, advances in computing have consistently enabled advances in military capability. While guidance, for example of missiles, was the defining military computing challenge of the Cold War era, the frontier today lies in AI and autonomy: systems that can process sensor data, distinguish signal from noise, navigate environments, and make targeting or other decisions with minimal human oversight. In the Russia-Ukraine conflict and in the Middle East, drones deployed in the millions have already demonstrated the transformative battlefield impact of AI and autonomy. There is widespread consensus among all of the world’s leading militaries that AI is central to the future of military power, and investment in autonomous systems is accelerating accordingly. U.S. defense leaders, for example, are already integrating AI into intelligence analysis, using it to rapidly process vast streams of satellite sensor data to identify militarily significant signals.

A historically unusual feature of this competition is that competing military powers, including Russia, China, and the United States, all draw on the same highly international semiconductor supply chain. Every Russian drone captured in Ukraine has been found to contain non-Russian components. Open-source researchers have similarly documented repeated attempts by China’s People’s Liberation Army to access American-designed, Taiwanese-manufactured AI chips for military use. This situation of a single global supply chain simultaneously feeding adversarial military programs is something historically unprecedented. No prior era of great-power competition has seen a single civilian supply chain serve as the key input for competing military powers who have so far been unable to reduce that shared dependence. Despite a decade of intensified geopolitical rivalry between the United States and China, both remain deeply reliant on the same international semiconductor infrastructure, as fragmenting the existing supply chain would be economically costly and technologically difficult.

The structure and fragility of semiconductor supply chains

No country or region is self-sufficient in semiconductor production. Different parts of the supply chain, such as EDA design tools, intellectual property, wafers, fabrication equipment, and chip manufacturing itself, are each highly concentrated in specific geographies. As a result, no region possesses a complete end-to-end supply chain. This concentration is in some ways a stabilizing force, preventing any single country from dominating the full chain, but it also creates choke points that have already been weaponized. China has demonstrated this clearly, imposing a series of opaque export controls on critical minerals, including gallium, germanium, and heavy rare earths, all essential to chip production. While these restrictions have not yet directly curtailed semiconductor output, they signal China’s capacity to cause severe disruption if it chose to cut off supplies it currently monopolizes. Taiwan, meanwhile, is the world’s largest producer of advanced logic chips and manufactures nearly all of the world’s AI chips, while facing sustained and growing military pressure from China across the Strait. This combination of irreplaceable economic centrality and acute geopolitical vulnerability makes Taiwan the single most consequential node in the global AI supply chain, and the one that concentrates risk most acutely for governments and companies worldwide.

The extraordinary complexity of advanced semiconductors makes self-sufficiency remarkably expensive and probably ultimately futile, even for the largest economies. As the COVID-19 pandemic illustrated, even minor shortages of just one or two chip types can have vast economic impacts and cascade into tens of billions of dollars of lost production in downstream industries such as automobiles. The scope of exposure is widening rapidly: a new car today may contain over 1,000 semiconductors, with the fastest-growing share going to safety and autonomous driving systems. As AI costs continue to fall, AI-enabled capabilities are applied across the economy, meaning that the entire legacy manufacturing base is becoming ever more dependent on both leading-edge and commodity chips alike. The stakes of any future supply chain disruption are therefore rising with each passing year. Looking ahead, the expected proliferation of drones across civilian industries and the gradual rollout of semi-autonomous robotics will add further layers of chip dependency across the economy, making the supply chain question even more consequential over the coming decade.

Cybersecurity, connectivity, and the risks of autonomy

Beyond supply disruption, the proliferation of interconnected, semi-autonomous devices raises serious cybersecurity concerns. U.S. intelligence agencies publicly flagged the discovery of unexpected modems on Chinese-made port cranes operating in American ports, raising the possibility that critical logistics infrastructure could be remotely frozen by a foreign government. While a benign technical explanation is conceivable, the incident illustrates a broader vulnerability. Virtually every piece of modern industrial equipment is designed for remote access and over-the-air software updates, opening potential attack surfaces that are difficult to monitor or secure. A Norwegian cybersecurity researcher reinforced these concerns by purchasing a Chinese electric vehicle and finding that 90% of the sensor data it generated, including GPS location, camera feeds, audio, and tire pressure, was transmitted to servers in China. These findings prompted the United States to ban Chinese-made connectivity equipment in automobiles, and the British Ministry of Defense to advise its employees against holding sensitive conversations inside Chinese-made vehicles.

However, the underlying issue extends well beyond cars. As virtually every modern device now requires the ability to receive over-the-air software updates, the cybersecurity-related areas that are exposed to risks due to the deep integration with geopolitical adversaries’ supply chains is vast and growing. The same logic that applies to port cranes and electric vehicles applies equally to agricultural equipment, construction machinery, medical devices, and consumer electronics. Governments have begun to recognize this, with some senior policymakers including President Trump calling for a wholesale review of the national security implications of electronic supply chains. However, awareness has broadly outpaced any coherent or proportionate policy response, and the tools available to address these risks remain limited relative to the scale of the challenge. Compounding the difficulty is the fact that most governments and even most companies still have only a partial understanding of their own electronic supply chains, making it hard to identify which specific interdependencies pose the greatest risk and to prioritize responses accordingly.

De-risking: challenges, costs, and the limits of policy

Weaponization of supply chains is no longer an occasional event but an increasingly regular feature of global competition, practiced by China, the United States, and others alike. The recent case of Nexperia, a Chinese-owned chip firm based in the Netherlands whose access to its own packaging facilities in China was cut off, causing disruptions to auto manufacturing across multiple continents, shows that even legacy chips are already being used as geopolitical leverage. Nevertheless, self-sufficiency is not a viable response. The cost of building new chip fabrication facilities is enormous. Beyond cost, excessive de-risking measures risk slowing the very technological progress that makes advanced AI possible, since it is the highly interconnected global supply chain that enables the pace of innovation. Any effective de-risking strategy must therefore carefully and precisely distinguish between interconnections that are genuinely dangerous and those that are regrettable but unavoidable. Most governments and companies are only beginning to develop the analytical capability needed to make these distinctions with sufficient rigor.

In this regard, China has made substantially more progress toward supply chain independence than Western countries. While the United States, Europe, and Japan have discussed de-risking extensively, China has backed its strategy with sustained government subsidy and investment, accepting real costs in terms of its technology companies’ reduced access to high-end products in exchange for greater strategic autonomy. The power that feels more de-risked will also feel more empowered on the global stage, and by that metric, China has gained ground. On the other hand, regarding U.S. exposure, imports from Taiwan are higher than ever before, showing how deeply AI-driven demand has entrenched U.S. dependence on Taiwanese production.

So long as AI remains the primary engine of both economic and military competition, and so long as computing power is its essential input, the geopolitical implications of AI supply chains will remain central to international politics. Navigating the trade-offs wisely will require governments, companies, and analysts to study and understand these supply chains far more rigorously than most currently do. There are no easy answers, only difficult trade-offs. Carefully studying the dynamics of AI supply chains will hopefully enable stakeholders to make wiser decisions around the trade-offs implied by this deep but problematic interconnection.

Comment

NAMBU Tomoshige:
Computing power now determines economic output and the balance of power among nations, a dynamic that explains why countries worldwide are competing intensively in semiconductor policy. Jensen Huang, the CEO of NVIDIA, has argued that compute equals revenue for companies and gross domestic product (GDP) for countries. If that is correct, the ability to secure and scale computing resources is becoming a measure of national economic strength. Japan has recognized this, actively promoting its semiconductor strategy since 2021 through a combination of attracting foreign investment, such as inviting TSMC to open a facility in Kumamoto prefecture, revitalizing domestic production facilities, and advancing frontier manufacturing projects such as Rapidus, which aims to begin mass production of 2nm logic chips by 2027.

Q:
Is the current wave of AI investment economically justified, or does it represent a speculative bubble?

Chris MILLER:
While those whose business it is to sell AI chips will naturally have an expansive view of AI’s benefits, a sober assessment still points to substantial economic impact. The 1990s application of personal computers and the Internet produced a meaningful and sustained productivity uptick across the U.S. economy, and AI is likely to generate something comparable over the coming decade. From a macroeconomic perspective, even a modest shift from 2% to 2.5% annual productivity growth carries enormous cumulative implications. On the infrastructure side, there is clearly speculative investment underway, but it is important to distinguish between AI training costs and AI inference costs. Profit margins on inference are already positive and broadly comparable to other software products, at around 30% to 40% for leading AI companies. That commercial reality lends credibility to the view that the economic impact of AI will be real and substantial, even after discounting some of the sales-motivated claims coming out of Silicon Valley.

Q:
How should the Trump administration’s AI and semiconductor policy be assessed, and what direction will it take going forward?

Chris MILLER:
As with many aspects of the Trump administration, the key is to focus on what has actually happened rather than what has been said. Three observations stand out. First, support for TSMC’s manufacturing investments in Arizona has continued with strong bipartisan backing, carrying over seamlessly from the Biden administration and, before that, the first Trump term. Second, the absence of significant tariffs on semiconductors to date signals a recognition that self-sufficiency is not achievable in the near term, and that such tariffs could slow the rate of AI progress. Third, the bipartisan consensus in Washington on limiting advanced chip exports to China has remained broadly intact. Despite extensive media coverage suggesting major policy shifts, the gap between what the Biden administration imposed and what the Trump administration has actually implemented is smaller than headlines imply. Across the administration and in Congress, there continues to be a strong degree of bipartisan consensus that selling cutting-edge AI chips to China is strategically unwise.

Q:
What is Japan’s AI strategy, given that the U.S. and China are the dominant players in frontier AI, and how can Japan play an influential role in the AI system?

Chris MILLER:
Japan already plays an influential role in the global AI system. Several of the most valuable Japanese companies by market capitalization are ones without which it would be impossible to produce AI chips at all: Tokyo Electron in chip-making equipment and Shin-Etsu in the specialty materials required for semiconductor fabrication are two prominent examples. Both have benefited significantly from the surge in AI-related demand in recent years. Looking further ahead, however, it is important to recognize that the largest share of AI’s economic benefit will come not from producing chips or building data centers, but from deploying AI effectively across specific industries, including healthcare, logistics, financial services, manufacturing, and others. Each of these domains involves country-specific regulatory environments that create space for distinctive geographically distinct domestic AI applications.

NAMBU Tomoshige:
Prime Minister Takaichi has identified 17 strategic industries, with AI and semiconductors at the top of that list. Physical AI is a central keyword in Japan’s current growth strategy, with efforts underway to promote AI-robotics applications across sectors ranging from medicine and agriculture to construction and logistics. Japanese companies hold rich operational data accumulated over decades of manufacturing excellence, and that data can now be harnessed through newly developed large AI models, creating significant competitive opportunities.

Q:
Could AI displace workers at scale, and how should the ethical risks of AI be managed?

Chris MILLER:
On employment, technology-driven labor market change is the very mechanism of productivity growth and ought to be embraced rather than resisted. Roughly once a decade for the past two centuries, societies have experienced a collective moment of concern about technological displacement, and it has never proven accurate over a sustained period, even if localized disruptions did occur. Looking carefully across the current world economy, it is difficult to find sectors where AI has already produced dramatic reductions in labor demand. That situation could change, and it is right to monitor it closely, but the appropriate response is careful observation rather than pre-emptively slowing AI progress in response to a problem that may not materialize at scale. On ethics, there will be or perhaps already are AI systems with compromised ethics, and there will be cases where people place undue trust in AI to make or influence decisions where they should not. But these challenges are not categorically different from those that arose with earlier technologies such as social media. AI is not the first technology that has challenged society to think carefully about how to manage ethical trade-offs.

Q:
There have been reports suggesting that TSMC’s Kumamoto fab may advance to 3nm production. Should Japan further strengthen its role in leading-edge semiconductor manufacturing?

Chris MILLER:
The move is both unsurprising and entirely rational. Global demand for advanced logic has grown substantially faster than was anticipated several years ago, driven above all by AI, and TSMC’s 3nm capacity is already facing potential shortages as a result. At the same time, TSMC has been very clear that there are real physical limits to how much it can continue expanding production within Taiwan. Droughts and labor shortages have already created operational constraints, and there is simply a ceiling on what can be built on the island before scarcity becomes a binding problem. That means that even setting aside the geopolitical considerations, TSMC would have faced a strategic need to establish secondary production bases elsewhere in the world. Japan has benefited directly from this dynamic, and TSMC leadership has been publicly positive about its experience operating there so far, making further investment a logical next step.