The Geography Of Advanced Chipmaking

AAlmost nothing else in modern manufacturing is as concentrated as the fabrication of advanced logic chips. The devices themselves weigh less than a coin and can be flown across the world overnight, which by the ordinary logic of trade ought to let their production spread to wherever land and labour are cheap. It has done the opposite. A plant able to print circuits at the finest dimensions on sale now costs between eighteen and twenty-four billion dollars, takes four years to build, and is almost always put up within an hour's drive of another plant of the same kind. Since the shortages of 2021 emptied car showrooms on three continents, governments have treated that pattern as a strategic problem rather than an academic curiosity, and have begun paying very large sums to break it. What nobody has settled is what holds the pattern in place, and the answer matters, because the two explanations now on offer imply quite different amounts of patience.

BPavel Osterberg, an economic geographer at the Nordhavn School of Economics, attributes the stubbornness of the pattern to the way yield is learned. Yield, the share of finished wafers that work, begins low and climbs as engineers discover which of several thousand process steps is quietly spoiling the batch, and Osterberg's contention is that this discovery travels in conversation rather than in documents. Her group assembled operating records from 74 plants commissioned between 2004 and 2018 and measured how long each took to reach a yield of eighty per cent. Plants built within thirty kilometres of an existing fabrication site reached that mark in a median of nine months; for the rest of the sample the figure was fourteen. Written procedures were transferred in both cases, often by the same vendors, which is the observation Osterberg leans on hardest. She is careful to add that her records cover leading-edge logic only.

CThe account yields a prediction, and Osterberg has tested it. If the decisive knowledge moves with people, then an isolated plant should stay slow even when its equipment is identical to that of a clustered one, and its difficulties should ease only when people who have done the work elsewhere arrive in person. Among the twelve most remote plants in her sample the median climb to eighty per cent took twenty-one months, and rapid improvement was confined to the four sites that had recruited a team of engineers from an established cluster. The pattern held once the age of the equipment had been taken into account, and none of the twelve stood more than three hundred kilometres from a port, so access to shipping cannot account for the difference. Osterberg nonetheless presents the twenty-one months as a description of how long the disadvantage lasts, not as a demonstration that a new cluster cannot be built.

DMateo Ferreira, who studies industrial policy at the Alvorada Institute, rejects the inference, on the ground that what Osterberg records as knowledge in the heads of engineers is mostly service in the hands of toolmakers. The machines that print the circuits come from a handful of firms, and when one drifts out of specification the plant stands idle until a vendor engineer arrives. In Ferreira's survey of interruptions at eleven plants, sixty-two per cent of unplanned stoppages were cleared within twelve hours where the supplier kept an office within fifty kilometres, against a median of four days elsewhere. Proximity, on this reading, buys repair rather than insight. He concedes two things. His plants all made memory rather than logic, and the length of each stoppage was reported by the operators themselves, since no independent register of downtime exists.

EThe disagreement is no longer confined to journals. Legislation passed in 2022 in several jurisdictions committed public money to plants deliberately sited away from existing clusters, and ministries now cite academic work when defending where they build; public contributions of two or three billion dollars towards a single site have become routine. Osterberg points out, awkwardly for her own case, that nothing in her figures shows the disadvantage to be permanent, since a gap that closes in twenty-one months is a delay and not a barrier. Ferreira answers that the distinction is lost in practice, because a minister shown a slow ramp reads it as a failure of the site and withdraws support before the ramp ends. Both dislike the habit of writing up single celebrated plants instead of the ordinary run of them, which they say has left the literature with a collection of successes and almost no record of the plants that were quietly abandoned.

FSomething like a settlement is now forming, in which the two explanations are attached to different parts of the industry rather than ranked against each other. At the leading edge, where a process is still being tuned, the human channel Osterberg describes appears to dominate; in mature production and in the packaging of finished chips, where the process is stable and the main risk is a stopped machine, the vendor network Ferreira describes matters more. That division would explain why several countries have managed to attract packaging plants while failing to attract the fabrication that precedes them. The open question is whether the advantage of being close by is eroding as diagnostic equipment is watched from a distance, and no dataset assembled so far runs late enough to show it.