The Chip Designers — Custom Silicon

The default AI brain — GPU king

NVIDIA designs the GPUs that train and run nearly every major AI model . CUDA software lock-in keeps developers captive. The source's note is honest: this is the most-owned, most-talked-about AI stock on Earth — own it as an anchor, but the asymmetric upside is in the smaller chokepoint names around it.

Custom-ASIC duopoly — Tomahawk + Jericho networking

Broadcom designs custom chips for hyperscalers (Google's TPUs, Meta's MTIA, etc.) and dominates AI networking switches. When Google says 'we don't need NVIDIA,' the chip they're using was probably designed with Broadcom. Half the AI router market lives here.

The other custom-ASIC house

Marvell is Broadcom's main competitor in custom AI silicon . Has 18 active hyperscaler programs in the pipeline — these are multi-year, multi-hundred-million-dollar engagements. Every program that ramps to production is a step-function for revenue.

Only credible merchant GPU alternative to NVIDIA

AMD's MI series GPUs are the only realistic 'plan B' if you don't want to buy NVIDIA . Hyperscalers are increasingly using them as a pricing lever and a hedge. Software stack still trails CUDA, but it's improving fast and the price-performance is real.

U.S. government's de facto semi arm

Intel is now strategically backed by the U.S. government as the only domestic leading-edge fab. The bullish twist: AI inference (running models, not training them) increasingly happens on CPUs, and Intel just flipped the AI server CPU-to-GPU ratio from 1:8 to 1:1 — meaning more CPUs per server, more revenue.

Royalty model — riding the custom-silicon wave

Arm licenses chip designs and collects a royalty on every chip shipped . Every custom AI chip from Broadcom and Marvell, every smartphone, and increasingly every AI server CPU pays Arm. As edge inference grows, so does the toll.