Trade and Economic Dependencies

Lena Müller spends her days inside a cleanroom in Veldhoven, a small Dutch town most of the world has never heard of, assembling machines that most of the world will never see. She works for ASML, the only company on Earth that builds extreme ultraviolet lithography systems—the machines that etch the finest features onto the most advanced computer chips. Each one costs roughly $200 million, weighs as much as two buses, ships in dozens of freight containers, and takes months to assemble and calibrate. Inside, droplets of molten tin are struck by a laser fifty thousand times a second to produce light so short in wavelength it can draw circuit lines a few atoms wide. No competitor has ever built a working equivalent. When people talk about the "AI supply chain," they usually picture Nvidia's chips or the giant data centers filling with them. But every one of those chips begins its life passing under the light of a machine that only Lena's employer knows how to make.

That is a strange fact to sit with. The entire trillion-dollar edifice of modern artificial intelligence—the models, the data centers, the valuations, the geopolitical anxiety—rests on a chain of suppliers narrow enough to name on a single page. A handful of companies, in a handful of countries, control components so specialized that if any one of them stopped shipping, the whole system would seize up within months. For most of the past thirty years, almost no one outside the industry worried about this. Concentration looked like efficiency. Now it looks like exposure. And the choke points that efficiency created have become the targets that governments reach for when they want leverage over one another.

A chain you can name on one page

Start with the shape of the thing. The global market for AI chips ran into the hundreds of billions of dollars by 2026, part of a broader semiconductor market climbing toward a trillion dollars before the decade is out. But money is not the binding constraint. The binding constraints sit at a small number of nodes where a single firm or a single country holds a share so large that no near-term substitute exists.

Read down that "where" column and a pattern jumps out: the AI economy has a geography, and it is astonishingly small. Taiwan for the logic, the Netherlands for the light, South Korea for the memory, China for the raw materials that make the whole thing possible. Each is a single point of failure in a system that most of the world's economic growth is now betting on.

The bottleneck that bites first is not even the most famous one. It is memory and packaging. You can have all the leading-edge fabrication capacity in the world, but an AI accelerator also needs stacks of high-bandwidth memory bonded to the logic die through a delicate packaging process, and both are in chronic short supply. Nvidia and a few other giants have locked up blocks of that capacity years in advance, which means a startup or a national lab cannot simply buy its way in—the supply is spoken for before it exists. This is the texture of the "giga cycle" analysts describe: AI demand reshaping the economics of compute, memory, and networking all at once, on a foundation that was never built to bend.

Why the chain got so thin

None of this concentration was planned as vulnerability. It was the natural endpoint of comparative advantage taken to its logical extreme. Making advanced chips is one of the hardest things humans do. The learning curves are brutal, the capital costs astronomical, the tacit knowledge accumulated over decades. In a world where goods and capital moved freely and no one expected the shipping lanes to close, the rational move was to let whoever was best at each step do all of it, for everyone. Taiwan's TSMC got better at fabrication than anyone because it did nothing else and did it at planetary scale. ASML consolidated EUV because the physics was so unforgiving that spreading the effort across several firms would have meant no firm ever cracked it. Specialization was not a mistake; it was the source of the cheap, abundant compute that made the AI boom possible in the first place.

The rare earth story is the same logic wearing different clothes. China did not corner gallium and germanium refining because the deposits are uniquely Chinese—they aren't; these are byproducts of aluminum and zinc processing found in many places. China cornered the refining, the technically demanding and environmentally dirty middle of the value chain, because for decades it was willing to bear the pollution and undercut the price while Western producers, facing stricter environmental rules and higher costs, quietly exited. The West offshored not just the mining but the messy, valuable processing—and told itself this was progress. Resilience had no line item. In an era that measured supply chains by cost and speed, redundancy looked like waste, and a second supplier looked like money left on the table. The bill for that omission is now coming due.

The mineral squeeze

In October 2025, Beijing added several more rare earth elements to its export control list and signaled that licenses would be withheld from foreign arms manufacturers and selected chipmakers—a move read across Western capitals as a deliberate reminder of who holds which cards. It was not the first such signal, and the mineral most watched is gallium, where China controls roughly 98% of primary production, alongside germanium, where it refines around 60% of global supply. These are not exotic add-ons. Gallium goes into the high-performance compound semiconductors used in power electronics, radar, and increasingly in AI hardware; germanium into fiber optics and infrared systems.

What would a sustained embargo actually do? Not an instant blackout—buyers hold stockpiles, and shortages call forth substitution and frantic new investment. But stockpiles are a buffer, not a solution, and rebuilding refining capacity outside China is a matter of years, not months: it takes permits, specialized facilities, waste-handling infrastructure that no longer exists at Western scale, and workers who know how to run it. A prolonged cutoff would not stop chip production overnight, but it would force a costly, disorderly reorganization of the whole materials base of the industry, and the pain would arrive well before the fix did.

The deeper point is symmetry. The United States can throttle China's access to advanced chips; China can throttle the West's access to the minerals those chips require. Interdependence, it turns out, is not a one-way leash. Each side has a hand around something the other needs, and that mutual grip is the real structure of the current standoff.

The chip blockade, and its critics

On the other side of the ledger sits the American export control regime—the most ambitious peacetime attempt to deny a rival a whole category of technology. For years the rule for exporting the most capable AI chips to China ran on a "presumption of denial": applications were assumed rejected. In January 2026, the U.S. Commerce Department's Bureau of Industry and Security shifted to a case-by-case review, allowing some exports of high-end chips provided they met strict conditions on supply, security, and testing. Around the same time, the administration wielded a second instrument, moving toward a 25% tariff on semiconductors above certain performance thresholds under a Section 232 national-security investigation.

The strategic logic is straightforward on its face: capable chips accelerate military AI, mass surveillance, and autonomous weapons, so denying them slows an adversary. But the policy is caught between two masters. American chip companies want access to their largest potential market; national security officials want that market closed. The case-by-case compromise satisfies neither and adds a cost of its own—nobody knows which licenses will clear, how long reviews will take, or what conditions will attach. That is the tell that trade policy has become national-security policy: for the first time, industry leaders ranked tariffs and trade rules above talent as their top concern.

And the critics are not fringe. At a January 14, 2026 House Foreign Affairs Committee hearing on "winning the AI race" against China, members and expert witnesses across the spectrum voiced deep skepticism, and one widely circulated analysis called the new policy "strategically incoherent and unenforceable." Their objection has a specific mechanism, not just a mood. Export controls, they argue, burden American allies who depend on U.S. chips while failing to stop a determined China from obtaining technology through smuggling, reverse engineering, or—most consequentially—building its own. Every restriction is also an industrial-policy gift to Beijing's domestic champions: it guarantees them a captive market, removes the foreign competition that would otherwise undercut them, and hands them a national mission. Cut off from Nvidia, Chinese firms have every incentive and every subsidy to make the homegrown alternative work. The controls may buy time; they may also be the best thing that ever happened to China's indigenous chip industry. Which of those dominates is genuinely unknown—and that uncertainty sits at the center of the whole debate.

What we can't actually measure

Here honesty demands a pause. We do not know how well the export controls are working, and the reasons we don't know are structural, not temporary. To measure the effect you would need a counterfactual—how fast Chinese AI would have advanced without the controls—and no such world is available for inspection. China does not publish reliable figures on its chip output, smuggling volumes, or the true performance of its domestic accelerators. Public evidence points both ways: Chinese firms have demonstrated more capability than the controls were meant to permit, which critics read as proof of failure, while defenders read the same facts as proof that China is being forced onto slower, more expensive, less efficient hardware—a real tax on progress even if it is not a wall. Both stories fit the data. Anyone who tells you with confidence that the controls are clearly working, or clearly backfiring, is reading their prior into a genuinely ambiguous record.

The same caution applies to the dramatic "months to collapse" timelines that hang over Taiwan and ASML. It is true that leading-edge chip production would be crippled within months if TSMC went dark or ASML stopped shipping—there is no second supplier at scale. But "months" is an estimate resting on assumptions that deserve daylight: how large the stockpiles are, how far demand would compress under rationing, and how much substitution older or lesser nodes could absorb. Change the stockpile assumption and the number moves. A world with two years of buffered inventory and aggressive triage of non-critical uses degrades far more slowly than one running just-in-time. The direction of the risk is not in doubt. The precise clock is softer than the confident phrasing suggests.

The island the world depends on

Which brings us to Taiwan, the most strategically consequential island on the planet. TSMC makes the overwhelming majority of the world's most advanced chips—north of 90% at the leading edge. If its fabs went offline, from earthquake, cyberattack, blockade, or invasion, the global AI industry would face collapse within months, and the damage would not stop at AI: consumer electronics, automotive, aerospace, and defense all draw from the same well. Samsung and Intel are spending enormous sums to close the gap, and they remain years behind. There is no credible Plan B measured in weeks.

Western governments understand this, which is why the 2022 CHIPS Act put $52 billion behind domestic fabrication, TSMC is building in Arizona, Samsung is expanding in Texas, Intel is investing in Ohio, and Europe has launched its own initiatives. These are worth doing. They are also insufficient as near-term hedges, and it is important to be clear-eyed about why. The Arizona fabs are not expected to reach full stride until late in the decade, and even at capacity they will not reproduce the depth of Taiwan's ecosystem—the dense web of suppliers, packagers, and accumulated know-how that took forty years to grow. Domestic fabs are insurance against catastrophe, not a replacement for the thing being insured. Genuine diversification—not a token fab but a real alternative ecosystem, including domestic materials refining and a second credible source of high-bandwidth memory—is a project measured in a decade or more, not a few years. Until then, the AI economy needs Taiwan to stay accessible, stable, and aligned. That is not a strategy. It is a hope with a semiconductor attached.

The sixfold surge—diversifying, or just digging in?

Governments are not standing still. Industrial-policy interventions motivated by national and economic security—tariffs, export controls, direct equity stakes, subsidies—rose more than sixfold between 2021 and 2026. That is a Cold War scale of state involvement in what was, a decade ago, treated as ordinary commerce. The question worth asking is what all that intervention has actually bought.

The uncomfortable answer, so far, is that much of it reinforces existing dependencies rather than replacing them. Securing a relationship is easier and faster than building a redundant one. Locking in preferential access to TSMC, signing offtake agreements with Korean memory makers, striking minerals deals with friendly governments—these entrench the map that already exists. Building a genuinely new node—a fully domestic refining base, a second-source EUV supplier, an HBM industry outside the incumbents—is slower, dearer, and politically thankless, because the payoff arrives in someone else's term of office. So the six-fold surge has, to date, mostly hardened the chain it was meant to loosen. Dependencies are being fortified faster than they are being diversified. Redundancy remains the thing everyone endorses and few will fund.

Why uncertainty is itself the tax

There is a subtler cost running underneath all of this, and it falls on everyone regardless of which bloc they sit in. Uncertainty is not a side effect of the new trade regime—it is one of its most damaging outputs. When a company cannot know whether its next shipment will clear an export license, whether a tariff will jump, or whether a review will conclude in six weeks or sixteen months, the rational response is to wait. Paused investment defers innovation, delays deployment, and opens gaps that better-positioned rivals move into. The cruel irony is that the interventions meant to secure the supply chain manufacture the very instability they were meant to prevent, and the advantage flows to whoever is least constrained by the rules—often the actor the rules were aimed at.

graph LR
  A[Trade-rule uncertainty] --> B[Firms delay investment]
  B --> C[Deferred innovation & deployment]
  C --> D[Gaps exploited by less-constrained rivals]
  D --> E[Security policy erodes the security it sought]

And the tax is not levied evenly. A hyperscaler with warehouse stockpiles, deep supplier relationships, and a floor of lobbyists can ride out a disruption that would kill a startup, a university lab, or a national AI program in a lower-income country. When memory tightens, the giant has already booked its allocation; the small buyer simply cannot get parts at any price. When a license stalls, the giant has lawyers and alternatives; the small actor has a stalled project. Every episode of supply-chain stress therefore widens the gap between those who can absorb shocks and those for whom the same shock is fatal—compounding the inequalities in AI access explored elsewhere in this book. Scarcity is not neutral. It is a sorting mechanism, and it sorts in favor of the already-large.

Three futures, and the fine print

Where does this go? Three broad scenarios are worth holding in mind, though—as we'll see—the neat boundaries between them are more a convenience of analysis than a feature of reality.

The first is managed interdependence. Governments negotiate frameworks, bilateral or multilateral, that permit continued trade in critical AI inputs under agreed conditions—limits on military end-use, reciprocal verification, predictable licensing. Supply chains stay globally integrated but sit under layered oversight. This preserves efficiency and lowers the odds of catastrophic rupture, but it demands sustained diplomatic trust that today's climate makes hard to conjure.

The second is bifurcated fragmentation. The U.S.-aligned and China-aligned worlds each build parallel, largely self-sufficient chains with minimal overlap—incompatible standards, divergent architectures, a global AI ecosystem split down a geopolitical seam. Mutual vulnerability falls; the cost is staggering duplication of capital, slower innovation from the loss of shared knowledge, and a brutal choice forced on every country in between.

The third is abrupt decoupling: a shock—most plausibly a conflict over Taiwan—collapses the existing architecture fast. Fabrication halts, shipments stop, years of investment strand, and recovery proceeds under crisis conditions at costs no one can forecast with confidence.

Now the fine print the three-way framing tends to hide. These scenarios are neither exhaustive nor mutually exclusive, and the most likely real future is a blend. We could easily land in managed interdependence for minerals while fragmentation hardens in chips—a partial decoupling that runs along some links of the chain and not others. Bifurcation is not a switch that flips; it is a dial that has already been turning for years and could keep turning at any speed. And the whole picture is acutely sensitive to near-term politics—one election, one strait crossing, one retaliatory license denial can shove the system from one regime toward another. Treat the three scenarios as landmarks for orienting, not as a menu from which exactly one dish will be served.

What we owe each other

Underneath the forecasting sit questions that data alone cannot settle. The first is the efficiency-resilience trade-off. Redundancy is not free—a second refining base, a duplicate fab, a strategic stockpile all mean paying more for the same output, deliberately accepting some economic inefficiency as the price of insurance. How much is defensible? A reasonable answer is that the acceptable premium scales with how catastrophic and how irreversible the failure would be, and few failures are as catastrophic as losing leading-edge chips. But someone pays that premium, and here the politics turn sharp: left to the market, the cost lands on consumers through higher prices and on taxpayers through subsidies, while the concentrated benefit of security is diffuse and hard to see. A resilience policy that quietly taxes the many to insure a system dominated by a few incumbents needs to say so out loud, and design accordingly.

Then there is bifurcation itself—not a prediction now but a value judgment. Is a world of two incompatible AI blocs something policy should fight to prevent, or an acceptable, even healthy, result of legitimate security competition? The case against is real: fragmentation slows the science, strands capital, and forces poorer nations into a loyalty test they never asked to take. The case for is also real: some separation may be the honest price of not depending on a rival for the tools of your own security. The book's stance is that fragmentation carries large, underpriced costs that its advocates tend to wave away, and that the burden of proof sits with those who would accelerate it—but it is a judgment, not a theorem, and it should be argued as one.

Finally, the chokepoint holders. What does the world owe Taiwan and the Netherlands, and what can they owe in return? The honest answer is that their leverage is also their vulnerability—ASML's dominance makes it a target for pressure from every direction, and Taiwan's "silicon shield" is both its protection and the reason it sits in the crosshairs. They can offer transparency, reliability, and good-faith participation in shared frameworks. They cannot offer a guarantee against events beyond their control—an earthquake, an invasion, a coercion they lack the power to refuse. Any assurance they extend is real but bounded, and a global economy that has staked its future on their continued good fortune should understand exactly how thin that assurance is.

Key Takeaways

1. The AI supply chain has a tiny geography. Taiwan for advanced logic, the Netherlands for EUV lithography, South Korea for memory, China for refined gallium and germanium—disruption at any single node would cascade across the entire AI economy, and no substitute can be stood up quickly.

2. This concentration was efficiency, not accident. Decades of comparative advantage and environmental arbitrage produced the cheap compute that made the AI boom possible. Resilience was never valued or funded; that omission is now the central vulnerability.

3. Export controls are a genuinely contested bet. They may slow China in the short run while guaranteeing its domestic industry a captive market and a national mission—possibly accelerating the very independence they aim to prevent. We cannot yet measure which effect dominates, and the methodological obstacles to knowing are structural.

4. Interdependence cuts both ways. The U.S. can restrict chips; China can restrict the minerals those chips need (≈98% of gallium, ≈60% of germanium refining). Mutual vulnerability produces mutual leverage.

5. Uncertainty is itself a tax. Unpredictable licenses, tariffs, and rules delay investment and hand advantage to the least-constrained actors—and the costs fall hardest on small firms, labs, and developing-world programs, while hyperscalers absorb shocks that would be fatal to smaller players.

6. The sixfold rise in industrial policy has mostly reinforced dependencies, not replaced them. Securing existing relationships is faster than building redundancy, so the chain is being fortified rather than diversified. Real resilience—Arizona fabs at depth, domestic refining, alternative HBM—is a decade-plus project, not a near-term hedge.

7. The three scenarios are landmarks, not a menu. Managed interdependence, bifurcated fragmentation, and abrupt decoupling are neither exhaustive nor mutually exclusive; the likely future is a blend, acutely sensitive to near-term political shocks. The central challenge is to manage strategic dependencies without triggering the disruptions that supply-chain security policy is meant to prevent.

Sources

• The New AI Chip Export Policy to China: Strategically Incoherent and Unenforceable | Council on Foreign Relations
• BIS Revises Export Review Policy for Advanced AI Chips | Morgan Lewis
• Administration Policies on Advanced AI Chips Codified | Mayer Brown
• Adjusting Imports of Semiconductors | The White House, January 2026
• Artificial Intelligence and the Critical Minerals Crunch | FP Analytics
• 2026 Semiconductor Industry Outlook | Deloitte
• Semiconductor industry enters unprecedented 'giga cycle' | Tom's Hardware
• AI-Boom Drives Semiconductor Industry Confidence | KPMG
• The Geopolitical Forces Shaping Business in 2026 | BCG
• AI geopolitics and data in the era of technological rivalry | World Economic Forum

Last updated: 2026-08-14