It is incredibly expensive to expand microchip production capacity. You need extremely large, extremely clean clean rooms filled with expensive machines from ASML and other vendors for the well over 1.000 different steps that need to be done for a modern chip. And then you need staff to operate these machines.
I remember seeing those figures on some video from AMSL about EUV machine. They also mention about how many planes are needed to transport those machines. But I don't remember the values.
And while everyone wants more investment in capacity, presumably no-one wants to be an investor who's left holding the bag if capacity overshoots and supply exceeds demand in 1-3 years' time.
You'd be surprised. You can almost _always_ sell excess supply given enough time. These parts... take up very little space. Even if it has to sit in a warehouse for 5 years eventually someone will buy them and while they may not break even, they'll get close.
I don't think you understand the scale at which these companies act.
A 300mm wafer holds 150ish chips, and TSMC makes 13 million wafers a year. Or roughly 2 trillion chips per year.
Now let's say TSMC expands to 2.2 trillion chips/year with a new fab.
If there is no customer for the 200 billion chips / year, you need to keep building newer and newer warehouses constantly. That's just not a number that can be solved with one or two warehouses.
You could fit 200 billion chips in an Amazon warehouse with room to spare, easily. If your customer disappears, you can turn off the faucet. You can sell equipment, land, IP, break contracts, etc.
The problem with supply exceeding demand isn't unsold inventory/capacity, it's that your profit margin gets slim enough that you can no longer fund the R&D and fab construction necessary to remain on the leading edge. This is especially apparent for memory chips that are more commoditized than processors. Boom and bust cycles frequently result in at least one major bankruptcy, merger or pivot away from leading-edge processes. Qimonda, Elpida, Winbond all had their heyday as major DRAM manufacturers. For logic fabs, casualties include Chartered, GloFo, UMC, Motorola/Freescale—all of whom were once able to manufacture competitive processors.
