Nvidia Gets the Headlines. Broadcom Gets the Checks.

If you have been reading about AI chips, you have been reading about Nvidia. Blackwell. The trillion-dollar club. The whole world watches Jensen Huang do his leather jacket thing on stage.

But there is a parallel AI chip story unfolding — one that is even more interesting, and far less covered. It is happening at Broadcom, and it involves the most lucrative design wins in the history of semiconductors.

Broadcom custom AI ASIC revenue has surged 106% year-over-year. That is not a rounding error. That is a business doubling in size because the world's biggest tech companies are tired of buying off-the-shelf GPUs and want chips built exactly to their specifications. Broadcom's total TTM revenue stands at $68.3 billion — a 29.5% jump — and the market has rewarded it with a $1.96 trillion market cap.

The Hyperscaler Gold Rush

Google uses Broadcom-designed TPUs for AI training and inference — now on their seventh generation. Meta MTIA chips, which the company just committed 1 gigawatt of infrastructure to, are designed by Broadcom. Amazon Trainium and Inferentia rely on Broadcom networking silicon. And OpenAI reportedly partnered with Broadcom for a $10 billion custom AI processor project.

When hyperscalers want custom silicon, they call Hock Tan.

This is the stealth AI trade. While everyone debates Nvidia's P/E ratio, Broadcom collects design wins across Google, Meta, Amazon, and OpenAI. Each win locks in years of recurring revenue. The switching cost for a hyperscaler to redesign its entire AI infrastructure around a different chip partner? Astronomical.

Why Custom ASICs Matter Now

The off-the-shelf GPU model works — until it does not. Nvidia H100 and Blackwell are incredible general-purpose AI accelerators. But hyperscalers running at planetary scale need chips optimized for their specific workloads: Google TensorFlow models, Meta recommendation engines, Amazon Alexa inference pipelines.

Custom ASICs offer 3-5x better performance per watt for specific workloads. When you are spending $10 billion+ on AI infrastructure, that efficiency delta is the difference between leading and lagging.

The Numbers

Broadcom AI revenue is on track to hit $12-$15 billion in fiscal 2026 — up from roughly $6 billion last year. At that run rate, Broadcom's AI business alone would be a Fortune 200 company. Tom's Hardware recently called Broadcom the undisputed leader in the custom AI ASIC space. With $14.2 billion in cash on hand and an enterprise value of $2.01 trillion, Broadcom has the balance sheet and the backlog to keep winning.

The stock is up 19.4% year-to-date, with a consensus analyst target of $480.49 — another 16% upside from current levels. Of the 44 analysts covering AVGO, the consensus rating is a Strong Buy.

The contrarian take: Nvidia's moat in training GPUs is real and durable. But the custom ASIC market — where hyperscalers design their own chips for inference and specialized AI workloads — belongs to Broadcom. And that market is not just growing. It is exploding.

AVGO is the stealth AI stock. Everyone watches Nvidia. The real money is being made at Broadcom.