TSMC CEO C.C. Wei just delivered bad news for anyone hoping the chip shortage would end soon: it won't. Even with new US facilities coming online, the company can't meet AI chip demand for years.

Meanwhile, Broadcom's earnings miss triggered a trillion-dollar selloff across chipmakers. Nvidia fell 6.19%. Micron dropped 13.25%. The market is realizing the AI hardware crunch is getting worse before it gets better.

Everyone talks about the AI race, but the bottleneck isn't algorithms - it's silicon. And you can't just "build more fabs."

TSMC's Arizona facilities represent $40 billion in investment and years of construction. They'll help, but Wei says even with that capacity, they won't fulfill US customer demand. The gap between what AI companies need and what semiconductor companies can produce is widening.

Here's why fabs can't scale quickly: they require extreme precision manufacturing in clean rooms. You're building transistors measured in nanometers. A single particle of dust can ruin wafers worth millions. The equipment is specialized, expensive, and takes years to install and calibrate.

Then there's the talent problem. You need engineers who understand both semiconductor physics and manufacturing processes. They're in short supply globally, and competition for them is fierce.

Broadcom's revenue of $22.19 billion missed estimates, and their AI chip revenue forecast of $16 billion disappointed investors. The selloff wasn't about the numbers being bad - it was about them not being good enough given the AI hype.

This creates interesting dynamics. If supply is constrained for years, then companies with chip allocation have massive competitive advantages. Google, Microsoft, and Amazon secured long-term capacity agreements early. Smaller players are scrambling.

That's part of why Google is paying SpaceX $920 million monthly. When you can't get chips from TSMC, you lease them from whoever has inventory. The constraint isn't money - it's atoms.

Some companies are exploring alternatives. Groq built custom inference chips. Cerebras designed wafer-scale processors. Google developed TPUs. But these alternatives take years to develop and deploy at scale.

The geopolitics matter too. Most advanced chips come from Taiwan. That concentration creates supply chain vulnerability, which is why the US is investing heavily in domestic production. But building fab capacity takes 5-7 years minimum.

Wei's timeline of "years" means we're looking at this shortage lasting through 2028 or beyond. That's a long time in AI development cycles.

The technology is impressive. The question is how you build AI infrastructure when you can't get the hardware. Right now, the answer seems to be: slowly, expensively, and with massive competitive advantages for those who planned ahead.