Almost every portfolio in the US owns Nvidia.
Directly. Through an index. Through a 401(k).
The whole market is long the same trade. And the trade assumes one thing.
Nvidia keeps winning as AI scales.
It won’t.
Not because Nvidia gets worse.
Because its customers got rich enough to stop being customers.
The math is simple. When one supplier takes 70% margins on something you buy for $50 billion a year, you build it yourself.
Google ran the numbers first.
In 2013, it couldn’t afford to scale neural networks on Nvidia chips. So it built the TPU. Today it runs 1.6 million of them.
Amazon followed with Trainium. Meta with MTIA. Microsoft with Maia.
None of them built these chips alone.
They hired Broadcom.
The market is watching the wrong fight.
AMD vs Nvidia is theater. Two GPU companies fighting for a market that is shrinking.
The real fight is ASICs vs GPUs. And Broadcom isn’t fighting Nvidia. It’s designing the chips that replace Nvidia inside the 4 largest data centers on earth.
Here is the number nobody is watching:
Broadcom will own ~60% of the custom AI chip market by 2027.
The signals are already here:
-> Broadcom AI revenue: $8.4B last quarter. Up 106% YoY. $10.7B guided for next.
-> Custom chips ship 3x faster than GPUs in 2026.
By 2028, they overtake GPUs entirely.
Inference is two-thirds of all AI compute. Nvidia’s share of it falls from 90% to 25% by 2028.
Nvidia’s margins didn’t just attract competition. They funded it.
Every chip sold to a hyperscaler paid the engineers who built its replacement.
The trade isn’t short Nvidia. Nvidia stays huge. It owns training. It owns CUDA. It owns the enterprise.
The trade is long Broadcom.
The market sees a networking company. It’s the toll collector on every serious AI chip in production.
This is public. It’s in the 10-Ks. It’s in the earnings calls.
In 2027, this will be obvious.
