AVGO
MRVL
The AI hardware script is getting rewritten
For the last couple years, GPUs have been the shiny object in AI infrastructure. But this piece argues the next act may belong to custom AI chips — the bespoke, built-for-one-job silicon that lets big tech tailor performance and cost instead of renting the whole GPU buffet.
The headline claim is simple: custom AI chips are poised to outship GPUs by 2027. If that happens, the winners won’t just be the usual hyperscaler suspects. The supply chain could tilt toward select semiconductor names that help design and build these chips.
Who stands to win?
Broadcom and Marvell are called out as key beneficiaries. That makes sense: when mega-cap tech wants custom silicon, it doesn’t sketch it on a napkin and hope for the best. It leans on specialized chip designers, IP, and manufacturing expertise.
But there’s a catch, because of course there is. The article flags elevated valuations, which is Wall Street’s way of saying the easy money may already be partially baked in.
Why investors should care
This is less about one company’s quarterly print and more about where AI dollars flow next. If custom chips take more of the budget pie, it could reshape who gets paid across the semiconductor stack.
- Mega-cap tech may gain more control over AI costs and performance
- Broadcom and Marvell could see demand tailwinds from custom silicon work
- ETFs may be the calmer way to play the theme if you don’t want to marry a stretched multiple
Big picture: the AI chip trade may be moving from “buy the biggest GPU sellers” to “follow the custom silicon plumbing.”