Our thoughts on why TSMC is the gating factor on an AI bubble:

Recently there’s been ink spilled on the topic of "capex overhang" and whether there’s even enough money to build more fabs.

This likely misdiagnoses the challenge. Money isn't the challenge. There's plenty of money now, and plenty of ways to get more money in future. The challenge is managing that supply and demand.

If the hyperscalers decide that AI is a flop, they cut their capex and FCF jumps right back up. Private investors take a bath and OAI goes back to being a research lab.

But if TSMC (and the memory vendors) builds capacity in line with Sam or Jensen's hyperbolic forecasts, it will face years of overcapacity. See the analog players who still haven't recovered from the 2021 boom cycle.

The time from breaking ground to first wafer out is like 3 years for the leading edge node. A lot can happen in 3 years. TSMC is not going to build 50k or 100kwpm (~$100b capex) extra capacity for 2028 on a promise from Sam Altman.

If TSMC's customers start agreeing to 5Y+ NCNR offtakes, then we know they are getting serious about AGI. But this is unlikely to happen because 1) NCNR is how analog got into trouble 3Y ago in the first place, and 2) NCNR is only worth the paper it is printed on if the customer has the money to pay for the product regardless of whether or not they want it.

Paradoxically, it is hard for a full blown AI bubble to form if TSMC continues to act like the rational adult in the room, as it is the single bottleneck to flooding the market with compute. There are many solutions to the power challenge, and 3 memory vendors is 2 too many to ensure rational HBM supply. There is only one leading edge foundry that can reliably produce at scale.