AMD, Nvidia target substrate capacity amid 4x expansion constraints

The AI chip bottleneck has quietly migrated from architectural supremacy to physical substrate control. ABF organic substrates—the carrier interconnecting chiplets in CoWoS and EMIB packages—are now the industry's new "wafer." Ibiden, Unimicron (Taiwan, China), and AT&S collectively command over 80% of high-end substrate capacity, and the 4x expansion ceiling means AMD's and Nvidia's multi-billion-dollar commitments won't yield meaningful supply relief before 2026. Technically, line-width requirements below 10μm push substrate yield to the edge; any process drift directly erodes OSAT margins. Geographically, concentration in Japan and Taiwan, China creates a single-point-of-failure that export-control escalation or a natural disaster could weaponize across the entire compute chain—a more隐蔽 chokepoint than lithography. Strategically, TSMC controls CoWoS throughput yet remains a substrate buyer, handing structural pricing leverage to a three-player oligopoly. Expect substrate ASPs to replicate the 2021 memory supercycle within 12–24 months. AMD and Nvidia's "capacity investments" are, in substance, long-term offtake agreements designed to prevent rivals from intercepting supply. The deeper structural shift: packaging is no longer a back-end process—it is now a strategic variable on par with silicon architecture itself.