Remember when investing in AI was basically like playing a video game with cheat codes? Just buy Nvidia, sprinkle in some Microsoft and Amazon, maybe throw in a wild card like Super Micro, and boom – you’re basically printing money. Nvidia alone shot up over 1,100% since early 2023. It was almost too easy.
Well, plot twist: the game developers just patched that exploit.
While retail investors are still mashing the same buttons expecting infinite lives, the big tech companies are quietly rewriting the entire playbook. They’re done paying Nvidia’s eye-watering 75% profit margins when they can build their own chips for way less cash.
Why Everyone’s Breaking Up With Nvidia
Here’s the thing – running massive AI models on Nvidia’s general-purpose GPUs is like using a Ferrari to go grocery shopping. Sure, it works, but you’re paying for a twin-turbo V8 when all you really need is trunk space and decent gas mileage.
The hyperscalers (Google, Amazon, Microsoft, Meta) have figured out they can design custom chips that slash costs by 30-50% per operation. And they’re not just talking about it – they’re actually doing it:
- Google’s been using their TPU chips for most of their AI training
- Amazon just launched Trainium2 chips that supposedly beat Nvidia on price-performance
- Microsoft is rolling out custom Maia accelerators in their data centers
- Meta is literally in talks to buy billions worth of Google’s TPUs instead of Nvidia’s gear
Translation: The “infinite budget” phase of AI is dead. Welcome to the “efficiency” era, where the generalist always loses to the specialist.
So Where’s All That Money Going?
If $100+ billion is shifting away from Nvidia, it’s got to land somewhere, right? Enter the “enablers” – companies that sell the blueprints, cables, and lasers that make custom chips possible.
Broadcom (AVGO) is basically the pick-and-shovel play here. Google can’t build TPUs without Broadcom’s intellectual property. They provide the critical tech that moves data on and off chips at lightning speed. No Broadcom, no custom AI chip. Period.
Credo Technology (CRDO) makes the special cables that connect all these chips. When Elon Musk needed to wire up 100,000 GPUs for his Colossus supercomputer, guess who got the call? Credo. Their revenue grew 272% year-over-year last quarter. Not too shabby.
Lumentum (LITE) makes the lasers that power fiber optic connections. As AI clusters scale up, you need light-speed connections, and light needs lasers. Simple physics.
Arm Holdings (ARM) is the ultimate royalty play. Every custom chip needs their underlying architecture, and they collect 1-2% on every chip shipped. It’s like owning the patent on wheels.
The Bottom Line
The AI revolution isn’t over – it’s just growing up. The “dumb money” is still chasing the GPU shortage. The “smart money” is building the factory that makes GPUs obsolete.
Don’t get trapped in yesterday’s trade. The capex budgets for 2026 are already being written, and they heavily favor custom silicon. Time to follow the money.