The Federal Reserve Bank of Richmond has published an analysis arguing that surging AI chip demand has not translated into higher utilization at U.S. semiconductor fabs, with the most advanced accelerators still produced in Taiwan.

The piece, titled "Made in Taiwan, Measured in America: Why the AI Boom Isn't Showing Up in U.S. Fab Utilization," challenges a common assumption about the AI build-out's geographic reach. Rather than spreading across the U.S. manufacturing base, the AI cycle remains concentrated in a small set of Taiwanese fabs operated by TSMC, the global foundry leader.

What does the Fed's framing imply?

The Richmond Fed's title suggests a decoupling: AI model training and inference consume compute, but the silicon behind that compute has not moved geographically alongside the demand. Frontier AI accelerators — including Nvidia's H100 and H200 GPUs built on TSMC's 4nm and 3nm processes — ship from Taiwanese lines. The economic spillover shows up in foundry revenue and capex announcements, not in U.S. fab throughput.

Why hasn't the U.S. response lifted utilization?

The U.S. has moved aggressively to rebuild domestic capacity. The CHIPS and Science Act of 2022 authorized $52.7 billion for semiconductor manufacturing, research, and workforce, with a separate $24 billion investment tax credit. Construction is underway in Arizona, Ohio, Texas, and New York. But new fabs take years to come online, and headline capacity does not become running wafers until tools, process maturity, and customer qualifications align.

That is the timing gap the Richmond Fed appears to highlight. While the U.S. capacity pipeline grows, actual production of AI accelerators — the segment defining the boom — stays centered in Taiwan. Domestic fabs currently running produce logic, analog, MEMS, and mature-node chips that don't see the same demand surge as 3nm AI silicon.

What does this say about U.S. policy?

If the AI demand spike isn't reaching U.S. fabs, the policy question becomes one of waiting versus accelerating. New fabs under TSMC's Arizona umbrella are slated to start at 4nm, with later phases targeting more advanced nodes — but the timeline extends into 2025, 2026, and beyond. Intel's 18A node in Ohio and Arizona is also a multi-year ramp.

The Fed's premise implicitly raises a question for the policy community: does subsidizing capacity equate to capturing the AI cycle? The answer depends on whether U.S. fabs ramp in time to absorb the next wave of AI silicon demand, or whether that demand continues to flow to Taiwan and to other Asian foundries expanding capacity.

How exposed is the supply chain?

Taiwan's concentration in advanced AI chip manufacturing remains the dominant share. TSMC's CoWoS advanced packaging — the bottleneck for AI accelerators that combine multiple dies — is itself constrained, which is why Nvidia, AMD, and others have publicly disclosed order backlogs stretching into 2025. The Fed's framing suggests that until U.S.-based production of leading-edge logic and advanced packaging comes online at scale, AI demand will keep bypassing American fab utilization metrics.

What changes if Taiwan's role narrows?

Geopolitical pressure on Taiwan — from export controls to potential disruption scenarios — has driven the U.S. and its allies to subsidize domestic capacity. The Richmond Fed's analysis argues this pressure has not yet produced a measurable shift in fab activity tied to the AI cycle. If Arizona and Ohio fabs reach high-yield production of leading-edge AI silicon within the next 24 to 36 months, the next iteration of this study may read differently.

Until then, the AI boom and U.S. fab utilization appear, by the Richmond Fed's reading, to be two separate stories — one measured in revenue at a handful of Taiwanese fabs, the other still under construction on American soil.