Silicon's Slow Fracture

The chip supply chain is not decoupling so much as duplicating — at enormous cost, and with no guarantee the copies will ever run at scale.

A paper silicon wafer is divided by a fine seam into separate planes.
AI-assisted editorial illustration / Northline Journal

Editor's note, September 20, 2026: The U.S. fab cost premium was previously given as 25 to 50 percent above Taiwan and attributed to an unnamed industry consensus; the passage now cites the 2020 SIA/BCG estimate of about 30 percent, Morris Chang's 2023 remarks and TechInsights' 2025 finding that the wafer-cost gap at TSMC Arizona is under 10 percent. We also replaced an unattributed subsidy tally and an unattributed industry timeline with our own sourced estimates, removed an unsourced yield figure, narrowed a claim about staffing-related delays to the documented TSMC Arizona case, and corrected the EUV supply-chain description to ASML's sole optics supplier in Germany. Further review on September 20, 2026: Removed the unreconciled global subsidy total and fixed time-to-parity forecast; distinguished reduced geographic risk from remaining equipment concentration.

Every advanced fab under construction in Arizona, Ohio, or Dresden is, in one sense, a monument to redundancy. The industry spent four decades concentrating the most difficult steps of chipmaking into a handful of campuses in Taiwan, South Korea, and a few coastal cities in Japan, because concentration was efficient: one set of engineers, one supply base for exotic gases and photoresists, one accumulated body of tacit knowledge about why a particular etch tool drifts on Tuesdays. Governments have announced programs that include $39 billion of U.S. CHIPS Act manufacturing incentives, Japan’s pledge of more than ¥10 trillion for chips and AI through fiscal 2030, South Korea’s 33 trillion won support package, India’s ₹76,000 crore semiconductor mission and the state aid approved under the EU Chips Act. These programs have different currencies, time horizons and mixes of grants, loans, credits and AI support; they cannot be added into a defensible total of direct semiconductor incentives without a reconciliation. This is not reshoring in the sense the phrase implies — production coming home to a market it once served. It is duplication: building parallel capacity whose commercial utilization must be assessed separately from its resilience value. The distinction matters, because a duplicated supply chain behaves differently than a relocated one, and the difference shows up first in the numbers nobody wants to put in a press release.

The arithmetic of a second fab

Start with the plain cost premium, which is smaller and more contested than the folklore suggests. The most-cited industry model, a 2020 study by the Semiconductor Industry Association and BCG, put the ten-year cost of owning and operating a new fab in the United States at roughly 30 percent above Taiwan, South Korea or Singapore before incentives; TSMC’s founder, Morris Chang, said in 2023 that Arizona costs ran 50 percent above Taiwan and might be double. But the one public estimate built from a fab that is actually running points the other way: TechInsights’ March 2025 analysis put the cost of processing a wafer at TSMC’s Arizona plant at less than 10 percent above Taiwan, because equipment — priced the same everywhere — accounts for well over two-thirds of wafer cost and labor for under 2 percent. TSMC itself says only that its overseas fabs remain costlier, diluting its gross margin by two to three points in their early years. Some of that gap is one-time: American construction labor is expensive and American permitting is slow, so the capital cost of the building itself runs high but is a sunk cost. The more durable piece is operating cost, particularly the specialized technician and engineering labor that keeps yields climbing after a fab opens — though on the TechInsights numbers that labor is a far smaller share of the bill than the wage differential alone would imply, and the gap it creates does not obviously close with time or subsidy.

The subsidy programs — the CHIPS Act in the U.S., its European counterpart, similar efforts in Japan and India — were designed explicitly to absorb that premium, and on the capital side they mostly do. What they cannot buy is time. A leading-edge fab takes three to five years to build with a further process-specific yield ramp before mature production, the point at which the share of usable chips per wafer is high enough to make a process line profitable. Yield ramp is not a matter of writing bigger checks; it is a matter of thousands of small process corrections made by engineers who have done this specific ramp before, on this specific tool set, under this specific fab’s contamination and vibration profile. That knowledge does not transfer cleanly on paper, and it does not transfer quickly in person.

Where the real bottleneck sits

This is the part of the story that gets underweighted in policy debates, which tend to fixate on the fabs themselves as the unit of sovereignty. The tighter constraint sits one layer up, in tooling and materials, and two layers up, in people. Extreme ultraviolet lithography — the technology that makes the smallest transistor geometries possible — is produced, effectively, by a single company in the Netherlands, ASML, which in turn depends on a single supplier, Carl Zeiss SMT in Germany, for every lens, mirror and optical column and on a handful of other specialized suppliers for the rest. There is no credible near-term plan, in any country’s industrial strategy, to duplicate that capability domestically; every “domestic” advanced fab being built anywhere in the world is domestic only up to the point where it needs an EUV scanner, at which point it rejoins the same single global queue everyone else is standing in. Regionalizing fabrication without regionalizing the equipment and materials stack underneath it can reduce geographic fabrication risk while leaving important upstream dependencies concentrated.

The talent bottleneck is less visible but arguably worse, because it cannot be solved by writing a bigger check to a Dutch equipment maker. Advanced semiconductor manufacturing runs on process engineers and equipment technicians with years of specific, tool-level experience, and the relevant labor markets — outside a few clusters in Asia — simply do not contain enough of these people. Every new fab now competing for the same thin pool of experienced hires bids wages up and stretches onboarding timelines; TSMC pushed the start of production at its first Arizona fab from 2024 to 2025 for reasons its chairman attributed to a shortage of skilled workers rather than to construction. Universities are standing up semiconductor programs, and some of that investment will pay off on a ten-year horizon. It does nothing for the fabs opening in the next eighteen months.

The subsidy programs can buy a building. What they are discovering, slowly and at great expense, is that they cannot buy the decade of institutional memory that makes a building profitable.

What industrial policy can and cannot buy

None of this means the regionalization push is pointless — that would be too easy a conclusion, and probably wrong. There is a real, defensible case for duplication that has nothing to do with cost efficiency: it is insurance against a scenario in which the concentrated supply chain is disrupted by conflict, embargo, or natural disaster, and in that scenario the “wasted” premium on a second fab looks less like waste and more like a put option that finally paid off. Governments buying that insurance should be honest about what they are buying, though, and the honest framing is expensive redundancy purchased for resilience, not a cheaper or faster path to the same chips. Treating it as industrial competitiveness policy — as a way to eventually out-produce the incumbent centers on cost — sets up a comparison that the new fabs are very likely to lose for a long time, possibly permanently, given how far behind the cost curve they start.

The more interesting question, and the one industrial policy has barely engaged with, is what happens to the concentrated centers while the duplication is underway. Taiwan and South Korea are not standing still waiting to be replicated; their leading firms continue to invest at a scale the new entrants cannot easily match, and the process node gap between the frontier and the newly built fabs — which mostly enter production one to two generations behind the state of the art: TSMC’s first Arizona fab began volume production on its N4 node in late 2024, a year before 2-nanometer production began in Taiwan, and 2-nanometer output in Arizona is not targeted until the end of the decade — shows little sign of closing. A subsidized fab that comes online trailing the frontier by two generations is not decoupling its host country from dependence on the frontier; it is adding a slower, costlier node to the country’s own menu while the dependence for cutting-edge chips persists untouched. That is arguably the most important thing to say about the whole project: it changes where some chips are made, not who makes the chips that matter most, and the two things are being sold to legislatures and the public as though they were the same achievement.

The realistic horizon

Our own reading of the published roadmaps and award schedules is that the realistic timelines cluster in a similar place: greater geographic diversification, with uncertain timing, costs and remaining dependencies. We do not have a reproducible model that establishes a universal five-to-seven-year horizon or rules out eventual cost parity. That is not a failure of will or of money. It is what happens when a fifty-year process of optimization toward a small number of locations gets run partway in reverse on a decade’s notice, against physical and human constraints that do not respond to appropriations bills. The fracture is real, and probably permanent in some modest, partial form. It is just slower, costlier, and less complete than the ribbon-cuttings suggest.

Sources