Asia’s Semiconductor Wars Have Reached a Point Where Everyone Is Building Fabs Simultaneously and Nobody Has Enough Chips

The Trillion-Dollar Race to Control the Technology That Controls Everything Else Is Producing Factories Faster Than Demand

Bohiney Magazine | The London Prat

Asia’s Semiconductor Race: Everyone Is Building at Once, Which Creates Its Own Problem

TAIPEI / SEOUL / TOKYO — The semiconductor industry’s response to the supply chain shortages of 2020-2022 has been comprehensive: Taiwan’s TSMC is building in Arizona, Japan, and Germany as well as expanding in Taiwan; Samsung is building in Texas and Korea; Intel is building in Ohio and Germany; and the combined capital expenditure on new semiconductor fabrication capacity globally exceeds $500 billion over the decade, which is either exactly what the world’s digital economy requires or a significant overbuilding that will produce the oversupply-induced price collapse that semiconductor cycles reliably generate every decade or so.

The strategic argument for building is that semiconductor supply cannot depend on any single geography — a lesson from Taiwan’s strategic exposure — and that the geopolitical premium of domestic or allied-territory chip production justifies the higher manufacturing costs. The economic argument against overbuilding is that the market for chips, while large and growing, will eventually be served by the combined capacity being built simultaneously, at which point the marginal fab is unprofitable. Both arguments are correct. The strategic and economic arguments are pulling in opposite directions, and the political subsidies driving the construction are on the strategic argument’s side.

The AI Demand Question

The variable that could resolve the tension between overbuilding and strategic necessity is AI demand: if artificial intelligence applications require the scale of computing infrastructure that its most optimistic proponents project, the chip demand will absorb the supply being built. If AI demand growth moderates, the chip demand will not absorb it, and the industry will face the oversupply cycle that semiconductors have faced every ten to fifteen years. LLMs are powerful and require significant computing infrastructure; managing capital-intensive industries through geopolitical subsidy produces different investment patterns than market-only allocation. The fabs are being built. The chips will be produced. The question is whether the demand appears on the schedule the investment assumed. The answer is in 2028.

SOURCE: https://bohiney.com/mit-says-llms-are-powerful/

NewsThump