TSMC Collects Twenty-Two Billion Dollars From a Chokepoint Nobody Broke

Every advanced chip the AI industry needs runs through one company’s fabs, and that company just collected twenty-two billion dollars in a single quarter.

It is true that TSMC’s second-quarter results — net profit up 77 percent year-over-year to NT$706.56 billion, or US$21.98 billion, with gross margin reaching 67.7 percent — read, from a portfolio-management standpoint, like a well-run company executing on strong demand. TSMC described itself as “capacity-constrained” and said it is accelerating expansion in both advanced process nodes and advanced chip packaging. The trouble is that “capacity-constrained” is a polite description of a structural dependency the AI industry has not solved and policy has not broken. TSMC manufactures the overwhelming majority of the world’s leading-edge logic chips — those fabricated at process nodes below 7 nanometers — and there is no second source at equivalent scale. Samsung’s foundry division has struggled with yield problems at its own advanced nodes, with 3nm yields reportedly stuck at 50 percent even three years into mass production. Intel Foundry Services remains years from competitive volume production at leading-edge processes. TSMC’s fabs in Arizona, Japan, and Germany, once at volume, will scatter production across three countries — but the corporate chokepoint remains because the same company controls each site. The result is a chokepoint: one company, one set of fabs, one bottleneck through which every dollar of AI infrastructure spending must pass.

This is not a new observation. Morningstar flagged TSMC as undervalued back in June, noting the gap between the company’s structural dominance and its market pricing. What the earnings report clarifies is the price of the dependency itself. TSMC’s 67.7 percent gross margin — the highest in more than two decades — is not simply operational efficiency meeting strong orders. It is what monopoly pricing power produces when one firm controls the production bottleneck for a critical global input. Nvidia’s market capitalization, Apple’s silicon roadmap, Meta’s training-cluster buildout, the entire agentic-AI transition TSMC’s own executives cited as a growth driver — all of it passes through fabs in Hsinchu and Tainan, and TSMC’s quarterly earnings report tells you exactly what the dependency costs.

This is what Tim Wu describes as the curse of bigness: when market concentration reaches a certain depth, the dominant firm’s margins become a structural tax on every downstream activity, and the downstream firms — even the ones with trillion-dollar market capitalizations — have no leverage to renegotiate.

The policy response has been, by any honest assessment, insufficient. The CHIPS and Science Act, signed in August 2022, authorized $52.7 billion in semiconductor manufacturing subsidies and investment tax credits. TSMC received $6.6 billion to build three advanced fabs in Arizona. The first began producing 4-nanometer chips in early 2025, behind schedule and over budget. The second and third — intended to produce 3-nanometer and 2-nanometer process nodes, the chips that actually matter for AI training and inference — remain years from high-volume production. Samsung’s planned facility in Taylor, Texas has been delayed repeatedly. Intel’s Ohio campus, announced with considerable fanfare, is a decade from meaningful leading-edge output.

The deeper problem is that the subsidies bought incremental capacity but did not purchase structural independence. The Arizona fabs will produce advanced chips for the American market, but TSMC’s bleeding-edge research and highest-volume production remain in Taiwan — on an island ninety miles off the Chinese mainland that represents the semiconductor supply chain’s most acute geopolitical concentration risk. The next reauthorization must tie funding to measurable second-source milestones before approving new awards.

Cory Doctorow’s framework of the four forces that historically constrain monopoly power — competition, regulation, self-help interoperability, and labor — explains why the chokepoint persists. Competition: Samsung cannot match TSMC’s yields at leading-edge nodes; Intel Foundry cannot match its volume; below 5 nanometers at commercial scale, no third option exists. Regulation: the CHIPS Act subsidizes but does not restructure; export controls on advanced chip-making equipment to China constrain a potential competitor without creating an alternative supplier. Self-help interoperability: semiconductor fabrication is physical manufacturing at atomic scale — the physics of extreme ultraviolet lithography, the light-based printing process that etches circuit features smaller than a virus, do not admit the kind of adversarial interoperability that can disenshittify a software platform. Advanced chip packaging and design-rule translation are more exposed to competition, which is precisely why TSMC is racing to expand its CoWoS capacity and lock down those bottlenecks too. Labor: TSMC’s process-specific manufacturing expertise is institutionally embedded, built over decades of iteration, and transfers slowly if at all to new geographies.

This is the structural fact the AI industry’s capital expenditures rest on, and the latest round of chip-sector earnings has made it impossible to avoid. Memory-chip makers like Micron posted strong quarters, quieting AI skeptics, but memory is a different market with different concentration dynamics. In advanced logic — the chips that perform the computation artificial intelligence requires — the dependency is total. TSMC’s Taipei-listed shares rose 37 percent during the second quarter, bringing year-to-date gains to nearly 60 percent. Citi Research raised its target price to NT$3,800 from NT$2,875. The investor community has correctly priced the chokepoint. The policy community has not broken it.

The CHIPS Act reauthorization cycle opens in the next two years. The reauthorization should tie subsidies to measurable leading-edge capacity targets in multiple geographies, with structured procurement commitments from the Department of Defense and major cloud providers that guarantee the new fabs a customer base independent of TSMC’s existing client relationships. That is the minimum viable industrial policy for a supply chain this concentrated. Whether it happens depends on whether the next authorization is structured to build a bypass or to continue subsidizing the checkpoint.

TSMC’s 67.7 percent gross margin is not a sign that the market is working. It is a sign that one company in one country has the world’s AI supply chain by the throat, and nobody — not Samsung, not Intel, not the CHIPS Act — has managed to pry the grip loose. The margins are the invoice. The question is whether anyone intends to pay for a different arrangement.