AI Geopolitics Capital: The Silicon Fault Line

The story of AI geopolitics capital begins at a chokepoint, and it ends at a fabrication plant. Everything between is commentary.

Markets keep debating which model reasons best. That debate misreads the board.

The precedent: 1987, and the logic of the chokepoint

In 1987, Washington penalized Toshiba. The company had sold precision machine tools to Moscow, tools that quietened Soviet submarine propellers.

The mechanism was blunt: deny the rival a manufacturing capability, and that capability decays across a decade. Congress moved with speed, the sanctions passed, and the doctrine entered permanent memory.

Control the machine that makes the weapon, and you control the weapon. The context in 2025 differs entirely, yet the structure repeats with uncanny fidelity. The chokepoint has shifted from machine tools to lithography, and the stakes have risen from submarines to the compute that trains frontier systems.

Three precedents are sufficient to call it a pattern. CoCom in 1949, Toshiba in 1987, and the October 2022 export rules form one continuous line. Each denied a manufacturing input to a strategic rival. Each aimed at the factory rather than the finished product.

The pattern today: fabs, lithography, and the GPU

The current configuration is legible for anyone reading primary documents. The US Bureau of Industry and Security rule of October 2022 restricted advanced GPUs and the tools that make them.

ASML holds an effective monopoly on extreme ultraviolet lithography, the single machine required for the most advanced nodes. TSMC and Samsung convert those machines into wafers. The rest of the supply chain orbits these facts.

The SEMI World Fab Forecast, a source most readers would fail to open, tracks capacity additions node by node. It shows a concentration of leading-edge capacity in a narrow geographic band around Taiwan and South Korea. That concentration is the real map of power in this cycle.

Compute is the raw material of frontier AI. The raw material has a physical origin, and that origin sits inside a contested strait.

The mechanism: why access beats capability

Models can be copied. Weights leak, architectures diffuse, talent migrates across borders within months.

Fabs behave differently. A leading-edge facility costs upward of twenty billion dollars and takes years to reach yield. The knowledge embedded in a mature fab resists replication for reasons closer to craft than code.

This is causation, and it deserves the distinction. Whoever governs access to advanced fabrication governs the ceiling on adversary compute. That ceiling determines the pace of frontier training runs, which determines relative capability across a decade. The GPU restrictions on Huawei are the most geopolitically consequential move of the past five years, heavier than the tariff war of 2018.

The counter-argument deserves a hearing. Skeptics argue that domestic Chinese fabs will close the gap through sheer state capital. That argument underestimates the depth of the lithography bottleneck, which lacks a domestic substitute at the leading edge.

The dollar angle the market fails to connect

Capital flows follow strategic certainty. When Washington weaponizes the semiconductor supply chain, it signals the same posture it holds toward the payment rails.

The dollar share of global reserves fell from 71.4% in 2001 to roughly 58% in 2024, absent any single crisis event. That is a structural drift, and export controls accelerate it by teaching emerging economies the cost of dependence.

Read the two together. A state that fears exclusion from compute will hedge against exclusion from clearing. The BRICS+ payment architecture and the fab race are branches of one tree: the search for autonomy from American chokepoints. Investors who treat them as separate stories will misprice both. Our earlier work on dollar reserve fragility traces this drift in detail.

My position, and what would change it

My standing thesis is explicit: the US-China AI contest is fundamentally a semiconductor access problem, rather than a model capability problem.

Whoever controls the fabs controls the outcome. The evidence sits in capex allocation, in the geography of EUV deployment, and in the sequencing of export rules.

I hold this position at high conviction, and intellectual honesty demands the falsifier. Two developments would force revision. First, a credible Chinese EUV alternative reaching commercial yield at the leading edge. Second, an algorithmic breakthrough that collapses compute requirements by an order of magnitude, dissolving the chokepoint through efficiency. Absent those two signals, the fab remains the fulcrum. I track both quarterly, and neither has crossed the threshold that would move me. The structural read holds.

Three implications for capital

The pattern carries direct portfolio consequences across a defined horizon.

• Family Office and Sovereign Wealth (36 months): overweight the lithography and equipment layer, underweight pure-play model developers exposed to commoditization.
• Chief Risk Officer (24 months): add a Taiwan Strait supply shock to the VAR model, a scenario absent from most current calibrations.
• CFO and Investor Relations (18 months): stress-test any AI growth narrative against a compute-access constraint, because that constraint governs the ceiling.

Each implication carries its own clock. The equipment layer rewards patience across three years, while the risk-model gap demands attention inside the fiscal year.

Board members should ask one question. Which strategic plan assumes uninterrupted access to leading-edge compute? Our semiconductor supply risk primer frames the exercise.

The prediction

Here is the verifiable claim. Washington will extend advanced-chip export controls to additional Southeast Asian transshipment hubs before the horizon closes, tightening the diversion routes that presently soften the 2022 rules.

Confidence: Medium-High. Orizzonte: 300 days. Verifica: a new BIS entity-list action or interim rule naming intermediary jurisdictions. This is a structural extension of an existing doctrine, rather than a discretionary policy whim.

What to watch

Three leading indicators will confirm or weaken the thesis across the coming quarters.

• ASML order backlog and China revenue mix in quarterly filings.
• TSMC leading-edge capex allocation between Taiwan, Arizona, and Japan.
• New BIS entity-list additions covering transshipment intermediaries.

The divergence between model hype and fab reality resolves in one direction. The question is timing. Watch the machines.