Speaking at the Taiwan Printed Circuit Association show in Taipei, Unimicron chairman S.C. Chien put a number on the wait for the equipment that makes AI chip substrates: up to thirty months.
More than 80 percent of the machines his company buys have longer lead times than before, he said, and for over 20 percent of them the delay exceeds double the original schedule. Low-expansion glass cloth, copper-clad laminate and ABF film are all tight, with the materials expected to fall about 20 percent short of demand in 2027.
The gap in the ramp
Chien sketched how unevenly the AI supply chain is expanding. CoWoS packaging capacity should grow 6.8 times by 2028 from its 2024 level, and hyperscaler capital spending 6.3 times. Substrate capacity is set to merely double, and DRAM only 1.8 times.
He cautioned that those multiples are not directly comparable because each starts from a different base. The point is the arithmetic of assembly. Every accelerator needs a substrate, however fast the packaging lines run.
Bigger boards, fewer of them
Physical size compounds the problem. Substrates have grown from about one reticle in early designs to three in 2025, with five targeted for 2026 and five to nine by 2027 and 2028. Chien said twelve to fourteen reticles is conceivable.
A larger board means fewer fit on a panel and more of them fail inspection. The machines that pattern and test them are the same ones with the thirty-month queue.
