There can be many bottlenecks forecast for AI boom currently. One such point is the semiconductor fabrication facilities (“Fab”). The industry has traditionally been vulnerable to heavily cyclicality. There is an oligopoly of few chip manufacturers, who have survived the cycles, in a capital intensive domain. TSMC is the biggest and best in the business, ultimately it’s in the control of which AI chips get built and when. Mostly AI chips require the best quality process. They will be very conservative to expand capacity, not risking billions on new factories if demand is less. Nvidia, Broadcom, Google, Arm etc cannot manufacture chips. The chip industry is mostly split very clearly into chip designers and manufacturers. Elon Musk is setting up a TeraFab for this bottleneck. The CHIPS and Science Act tries to get the manufacturing back to the US. TSMC have reportedly dismissed Sam Altman as a “podcast bro”, not taking his unfunded “commitments" seriously.
Thus possibly, chips can be a bottleneck, with fabs like TSMC acting as a safety valve on the AI demand. So, compute may be be constrained for longer-term.
Also setting up fabs and bringing them up to optimum performance (“yield", in terms of chips within specification and without defects), takes 2-3 years minimum, if their supply chain is working fine.
I have worked in the industry but as a lay person, mostly far removed from the actual business. Wrote about it here.
Possible effects of manufacturing bottlenecks on Nvidia Rubin rollout
Disc: invested