TSMC capacity crunch
- Reports say TSMC cannot keep up with AI-chip demand, with shortages projected to extend beyond 2027. - The foundry is investing heavily in new fabs, but supply constraints and memory shortages persist for the near term. - Continued scarcity means the most advanced SoC and memory capabilities will land first in premium devices, preserving device-class stratification (tweaktown.com).
Taiwan Semiconductor Manufacturing Co. says artificial-intelligence chip demand is still outrunning supply, and new capacity will take years to catch up. (investor.tsmc.com) TSMC reported first-quarter 2026 revenue of NT$1.134 trillion and net income of NT$572.48 billion on April 16, with Chief Executive C.C. Wei telling investors that AI-related demand remains “extremely robust.” (cnbc.com) The company now plans to spend nearly $56 billion in 2026, and Wei said new fabs still need two to three years to build and another one to two years to ramp to volume production. (eetimes.com) A foundry is the factory that makes chips designed by other companies, and TSMC is the one Nvidia, Advanced Micro Devices, Apple, and many cloud-chip startups rely on for the most advanced production. (cnbc.com) The bottleneck is not just the silicon wafer itself. Advanced packaging called Chip on Wafer on Substrate, or CoWoS, is the step that bolts a logic chip to stacks of high-bandwidth memory so AI processors can move data fast enough for training and inference. (eetimes.com) Memory is tight too. TrendForce said in December 2025 that stronger-than-expected AI server deployments had created server DDR5 shortages and pushed suppliers to raise 2026 high-bandwidth-memory pricing forecasts. (trendforce.com) TSMC said advanced chips accounted for about 75% of wafer revenue in the March quarter, a sign that the company’s mix is shifting deeper into the leading-edge nodes used in data-center accelerators and premium consumer devices. (cnbc.com) Wei said TSMC is expanding 3-nanometer capacity for the first time after a node had already reached its planned scale, with new 3-nanometer fabs in Taiwan, the United States, and Japan scheduled for 2027 and 2028 starts. (eetimes.com) That leaves the near-term queue largely unchanged: cloud companies buying AI accelerators stay at the front, while the newest system-on-chip designs and fastest memory packages reach premium phones, PCs, and servers before cheaper devices. (investor.tsmc.com; trendforce.com) TSMC’s message in April was simple: demand is already here, the factories are still being built, and the shortage clock runs longer than one product cycle. (investor.tsmc.com; eetimes.com)
