Mad Money

Hock Tan on why co-designed silicon beats the general-purpose GPU

Hock Tan· President and CEO of Broadcom at Broadcom
·~10 min·English·CNBC
AI InfrastructureGPUInferenceBusiness Strategy
TL;DR

Broadcom's CEO sees no AI slowdown: a business built on six frontier-lab customers, a $230 billion custom-accelerator forecast for 2028, and a bet that silicon co-designed for one lab's workloads beats a general-purpose GPU.

01The Demand Call

Not in the Least

Asked whether the AI-slowdown debate gives him any pause, Broadcom's CEO answers <strong>'not in the least'</strong> - he sees demand for AI compute as strong and durable.

We see the demand for compute infrastructure for AI, development of AI frontier models, and inference for the products that they feed to the world as continuing to be very strong and, I believe, very durable.

— Hock Tan, Mad Money
Key Insight
He is making a demand call on the entire AI complex on the very day it sold off - leaning into the fear rather than away from it, because Broadcom only wins if the buildout is real.
02The Customer Base

Six Customers, Only Six

Broadcom's entire custom-silicon business rides on <strong>six frontier-lab customers, and only six</strong> - a concentrated bet, not a broad market.

We have six customers, only six.

— Hock Tan, Mad Money
Key Insight
Custom silicon has no long tail: unlike a general GPU sold to thousands, a co-designed chip exists for one buyer, so Broadcom's growth is hostage to whether these six labs keep spending.
03The Financing

Cash-Flow Poor, Valuation Rich

Two of those customers are <strong>cash-flow poor but valuation rich</strong>, so Broadcom lines up private credit (Apollo, Blackstone) to fund the buildout its chips depend on.

And these companies are Apollo, Blackstone, who loves this sort of opportunity, steps in because the rates are better than going treasury for sure.

— Hock Tan, Mad Money
Key Insight
The real product here isn't a chip - it's the financing. By arranging credit for customers who can't pay cash, Broadcom turns itself from a vendor into the platform that makes its own multi-year forecast fundable.
04The Forecast

$230 Billion, on AI Alone

On AI alone - custom accelerators plus the networking sold beside them - Broadcom expects to <strong>hit $230 billion in 2028</strong>.

We believe we will hit $230 billion alone for those products in 28.

— Hock Tan, Mad Money
Key Insight
Calling out the networking alongside the accelerators is the tell: Broadcom monetizes the whole rack, not just the compute die, which is how a 'chip company' reaches a data-center-scale number.
05The Technical Bet

Co-Design Beats the General-Purpose GPU

A chip co-designed around one lab's workloads, Tan argues, <strong>performs much better than a general-purpose GPU</strong> - the structural case against Nvidia's moat.

Your chip will perform anytime much better than a general purpose GPU. That's the underlying premise behind it all.

— Hock Tan, Mad Money
Key Insight
The bet is narrow specialization beating general capability - and it lives or dies on inference. Tan hedges on training demand but is certain on inference, because an inference workload is stable enough to design a fixed chip against.
06On Safety

A Tool, Not a Wild Animal

AI is <strong>a tool, not a wild animal</strong>: it needs governance and safeguards, but not the slowdown some frontier labs are calling for.

It's important to put governance's safeguards on how we use the tool, but it is still, at the end of the day, a tool that will make our society, humanity reach a better level of living, but still a tool.

— Hock Tan, Mad Money
Key Insight
Reframing AI as an ordinary tool quietly moves the safety debate away from existential risk toward routine governance - the one framing under which his customers can agree to safeguards and keep building at full speed.