Taiwan and the Semiconductor Chokepoint, Explained: Why Reshoring Does Not Close the Gap

Situation Briefing. Updated September 13, 2026.

Advanced semiconductors are made in very few places, and the most advanced are made in essentially one. Taiwan Semiconductor Manufacturing Company controls approximately 70 percent of global foundry revenue and over 90 percent of the world's leading-edge production at three and five nanometres. Its most advanced facilities sit in Hsinchu and Tainan, about 100 miles from mainland China.

The dependency is not confined to consumer electronics. Every high-end AI accelerator, server CPU and flagship mobile processor in production traces back to the same handful of fabs. So does a large share of the compute underpinning the current AI buildout, which means the capital expenditure cycle driving global equity markets rests on an island in a contested strait.

What makes this different from other chokepoints is that the constraint is not extraction or transit. It is accumulated process knowledge, measured in decades, held by a workforce and supplier ecosystem that clusters geographically for reasons nobody has successfully legislated away.

It was not strategic design. It was thirty years of specialisation working exactly as economics predicts.

The foundry model, which TSMC effectively invented, separated chip design from manufacturing. Designers could compete without building fabs; the foundry could capture scale economies across every customer. Each generation of process technology cost more than the last, so the leading edge consolidated until only three firms could fund it, and then effectively one could execute it reliably.

Yield is the reason. Getting a new node to commercial yield is a problem of tacit knowledge, accumulated failure data and engineering culture, not blueprints. That is why buying the same ASML lithography machine does not buy the same output.

Governments noticed late. The US CHIPS and Science Act committed $52.7 billion in 2022. The EU passed its own. Japan backed Rapidus. China poured money into indigenous capability and, through Huawei's design arm and the foundry SMIC, achieved a seven-nanometre breakthrough that surprised most observers.

By 2026 the results are real and insufficient. TSMC's Arizona investment now stands at roughly $165 billion, with a second fab producing three-nanometre chips from late 2026 and Apple committed to buying Arizona-made silicon from 2027. Intel is building in Ohio, Samsung in Taylor, Texas, TSMC also in Kumamoto and Dresden.

But the timeline is unforgiving. Arizona targets two-nanometre volume production by 2030, and the completed complex is expected to supply about 30 percent of TSMC's most advanced chips. Full diversification of leading-edge production, if it happens at all, is a 2030 to 2035 story.

This is the part that gets lost in the reshoring coverage, and it is where the analysis actually lives.

Packaging is the hidden bottleneck. A finished wafer is not a finished chip. Advanced packaging, which stacks and interconnects dies, is its own specialised capacity and it did not relocate alongside the fabs. Arizona wafers can still require Taiwanese packaging, which means the dependency survives the geography change.

The tool and materials chain is equally concentrated. ASML's extreme ultraviolet lithography systems, specialty chemicals, photoresists. A fab in Phoenix running Dutch tools and Japanese photoresists has moved one link of the chain.

Mature nodes matter more than anyone admits. Cars, appliances, medical devices and industrial equipment run on 28 to 90 nanometre chips. China is strong in exactly that space. A crisis would strike both ends of the chain at once, and the mature-node end is the one that stops automotive assembly lines.

And the fabs need inputs Taiwan does not have. This is the connection almost nobody draws. Semiconductor manufacturing is extraordinarily energy and chemical intensive, and Taiwan imports nearly all of its energy. During the 2026 Strait of Hormuz disruption, the island was running on roughly eleven days of gas reserves.

Eleven days. That is the entire margin between normal operations and a power crisis at the facilities producing 90 percent of the world's advanced chips, and it was reached by a war two continents away.

Scenario work increasingly converges on the view that if Beijing moved, it would most likely use a quarantine rather than an amphibious assault.

The logic is straightforward. An invasion is the hardest operation in modern warfare, would unify opposition instantly and risks catastrophic failure. A quarantine, meaning customs inspections, declared exclusion zones, selective interdiction, achieves much of the coercive effect while staying below the threshold that triggers automatic response. It puts the burden of escalation on the other side.

Against a fab complex dependent on imported energy, chemicals and gases, and on exports moving by air and sea, a quarantine does not need to be total. It needs to be credible enough that insurers, shippers and airlines withdraw. The Bab el-Mandeb has just shown how quickly commercial traffic prices a threat: transits there halved before anyone took the island.

This reframes the question decision-makers should be asking. Not the probability of war, which is genuinely unknowable, but the gap between the damage a quarantine would inflict and the capacity of diversification to absorb it. That gap remains material through at least 2028 and no announced programme closes it sooner.

China wants reunification on its terms, indigenous chip capability, and the elimination of a dependency it regards as a leash. Its self-sufficiency push is also a hedge: the better SMIC gets at mature and near-advanced nodes, the less a crisis costs Beijing. Military capabilities continue advancing toward a widely discussed 2027 readiness threshold, which is a milestone rather than a deadline.

Taiwan depends on what is often called the silicon shield, the argument that its centrality makes it too valuable to attack. The shield has an obvious flaw. Every fab built in Arizona, Kumamoto or Dresden thins it, and Taipei is under sustained pressure from customers and allies to keep thinning it.

The United States wants production on its soil and technological denial to China, pursued through subsidies and escalating export controls that now extend to third-country firms. Those aims conflict: export controls accelerate Chinese indigenisation, which over time reduces the leverage the controls were meant to provide.

The EU has chosen selective autonomy, concentrating on automotive and industrial nodes where Infineon, NXP and STMicroelectronics already hold position, rather than contesting the leading edge. That is the most realistic strategy any bloc has adopted.

Japan and South Korea hold genuine capability and sit inside the threat radius. ASML and the tool makers benefit from every capacity buildout regardless of location, which makes them the least exposed participants in the entire chain.

Tension has not resolved in any direction since 2022. Beijing continues to restate reunification as a goal, including references to the possible use of force. Arizona is producing at four nanometres and moving to three. The reshoring effort is real and will take a decade to meaningfully reduce dependence.

Market pricing is the interesting anomaly. Prediction markets have implied something in the region of a 22 percent probability of conflict by 2027, while geopolitical risk premia across equities, rates and FX do not appear to reflect a shock of the modelled magnitude. Bloomberg Economics has put the first-year cost of a full Taiwan conflict at roughly $10 trillion.

A 22 percent chance of a ten-trillion-dollar event is not a tail risk. Markets are treating it as one because there is no straightforward way to hedge it.

Advanced packaging capacity outside Taiwan. The single most informative indicator and the least reported. Wafer capacity without packaging capacity is a partial hedge that looks like a complete one.

Taiwan's energy reserve margin. Days of LNG storage is a harder number than any military assessment and it moves with events in the Gulf. Watch it during the next chokepoint disruption anywhere in the world.

Whether Arizona's 2030 two-nanometre target slips. Semiconductor construction schedules slip as a matter of routine. Each quarter of slippage extends the exposure window.

SMIC's progress at and below seven nanometres. Rapid Chinese advance reduces Beijing's cost of a crisis, which raises risk. Slow progress preserves the mutual dependency that has kept things stable.

Quarantine rehearsal behaviour. Customs inspection regimes, declared exercise zones and interdiction drills are the observable precursors. They also function as tests of commercial nerve, and shipping response is measurable in transit counts.

Insurance and shipping pricing into Taiwanese ports. As with Hormuz and Bab el-Mandeb, underwriters move before governments do. War-risk premiums are the earliest honest indicator available.

The structural conclusion is that the world spent four years and hundreds of billions solving the visible half of this problem. Fabs are being built, and that is genuine progress. But the chain also runs through packaging plants, photoresist suppliers, EUV tooling, a specific engineering culture, and the LNG tankers that keep the island powered. Concentration risk was never only about where wafers are etched.

They built a shield out of silicon and forgot it runs on imported gas.

This Situation Briefing is updated monthly. Last revised September 13, 2026.