TSMC Equipment Demand Nearly Doubles In Seven Months As AI Buildout Strains Supply Chain
TSMC says its equipment demand has risen from 1x at the end of last year to 1.9x by July, the fastest pace in three decades, and it now has close to 20 fabs under construction. Substrate capacity and advanced packaging have emerged as the binding constraints on AI hardware shipments.
TSMC senior vice president and co-COO Cliff Hou said he has not seen demand grow at this pace or frequency in thirty years, illustrating the point with the company's own equipment needs: 1x at the end of last year versus 1.9x by July, a 90 percent increase in just over half a year. The company is building 13 fabs in Taiwan and another five to six overseas, close to 20 at once, against a historical norm of four or five, and Hou said even that is not enough to meet demand.
The bottleneck has moved downstream of wafer fabrication. Unimicron's chairman said the substrate supply chain is at a critical juncture as AI requirements push larger body sizes, higher layer counts and tighter coefficients of thermal expansion and dielectric constants, while key production tools and materials including ABF and glass-through-via technologies remain heavily dependent on small and medium-sized Japanese suppliers. Advanced packaging capacity, rather than front-end wafer output, is what gates accelerator shipments.
AI demand is also reshaping the mix at the leading edge. TSMC's 3nm ramp is being driven by Nvidia's Rubin platform and follow-on orders from AMD and Broadcom, with monthly 3nm starts heading to about 180,000 wafers in early fourth quarter. The next-generation A16 angstrom-class process is expected to be production-ready in the second half of this year, with Apple and Nvidia seen among first customers and OpenAI's custom silicon as another candidate.
Mature nodes are not being left behind. Suppliers point to high-density bonding, specialty memory, logic-based dies, silicon interposers with deep-trench capacitors and silicon photonics as areas where older nodes gain from AI systems, framing advanced and mature technologies as complementary rather than competing in the AI buildout.