Overview
- Gavin Baker, chief investment officer of Atreides Management, said Tuesday that the profitable ‘‘bottleneck trade’’ of buying owners of scarce AI infrastructure is losing its edge as physical constraints ease.
- He identified the key chokepoints that powered that trade as TSMC wafer capacity, power generation, cooling systems, optics, and networking equipment and estimated unconstrained Nvidia GPU demand could be $2 trillion to $3 trillion a year.
- Baker credited TSMC’s disciplined wafer capacity strategy with helping avoid an oversupply-driven bubble by keeping chip production tight during rapid AI buildout.
- He argued capital should rotate toward hyperscalers and specialized ‘‘neocloud’’ operators that show clear return on AI deployments, while crypto-origin data center firms that shifted from Bitcoin mining must now prove execution and customer relationships to win business.
- The shift could compress margins for firms that relied only on power and cooling and favor operators that combine scale, sales channels, and tight cost control, so watch for more deals by hyperscalers and a test of repurposed sites’ competitiveness.