AI investment is no longer limited to GPUs; it's expanding into the entire data center infrastructure.
Nvidia management emphasizes that AI opportunities have extended from GPUs to CPUs, networks, and system-on-a-chip (SoCs); Bernstein believes its data center opportunities remain enormous. Broadcom better exemplifies a "second AI theme": its custom AI chip business is expanding from Google to more customers like Anthropic, and management expects related demand to continue growing in 2027. In other words, future beneficiaries of AI will not only be companies selling GPUs, but also custom chip, network interconnect, and data center system manufacturers.
However, the real limiting factor is shifting from "whether there are AI chips available" to "whether the entire supply chain can scale up production together."
Nvidia mentions that current constraints already involve wafers, CoWoS advanced packaging, HBM and other memory components, optical modules/transceivers, and data center power and server room space. (Bernstein report, September 21)
