AI Bottlenecks and Why They Matter

AI’s growth is ultimately constrained by the supply of advanced chips, specialized materials, electricity and data-centre equipment, making physical bottlenecks the clearest measure of both scalability and investment risk.

• TSMC is the only company able to manufacture the most advanced AI chips; leading-edge fabs require up to $20 billion and years to build, while TSMC is now producing below customer orders after the post-pandemic overexpansion.
• Advanced CoWoS packaging capacity has increased roughly tenfold but remains fully booked into 2027, while Samsung, SK Hynix and Micron have sold their high-speed memory production through the end of 2026.
• Lam Research supplies almost all of the equipment used to drill the microscopic holes needed for stacked memory, with supply expected to remain tight until 2028; ASML produces only 50 to 60 advanced lithography machines annually, each costing up to $400 million and requiring two to three years to assemble and calibrate.
• Semiconductor inputs are concentrated in unexpected suppliers: Shin-Etsu provides silicon wafers, Ajinomoto insulating film and TOTO ceramic plates, while helium supply was reduced by roughly 40% after disruption at Qatar’s largest hub, although recycling rates have reached 80-90%.
• U.S. grid-connection queues exceed 2 terawatts and can take up to five years, transformers require roughly two to five years versus less than one year before the pandemic, and gas-turbine backlogs exceed 100 gigawatts; data-centre electrical equipment also carries one- to three-year waits.

Takeaway: Investors should follow chips and power rather than headline AI demand, while recognizing that shortages attract capital and can eventually become surplus, leaving today’s bottleneck beneficiaries exposed to overbuilding and forecasting errors.