🤖 The $2 Billion Frenemy

Good morning, AI arms dealers.

Last month, Nvidia wrote a $2 billion check to a company whose biggest customers are racing to build their way off Nvidia.

Today, we’re breaking down why that check is about Nvidia turning its rivals into customers, and why the real winner might be a company you barely know.

This month, at COMPUTEX (the big chip conference in Taipei), Marvell's CEO gave a keynote about why AI now lives or dies on connectivity.

Halfway through, Nvidia's Jensen Huang walked out to back him up.

A few weeks earlier, Nvidia put $2 billion into Marvell $MRVL ( ▼ 9.84% ), a check it could cover with about three days of its own data center sales.

Here's the weird part.

Marvell helps Amazon, Microsoft, and Google design their own AI chips, the ones meant to make those giants less dependent on Nvidia.

So Nvidia just funded the company arming its rivals.

On purpose.

The reason is a piece of Nvidia tech called NVLink Fusion.

Think of it as a private highway that used to carry Nvidia traffic only.

Nvidia just opened the on-ramps.

Now a custom chip from Amazon can pull onto the same highway as an Nvidia chip and the two talk at full speed, no slowdown.

Marvell builds the on-ramps and a lot of the road.

And here's the kicker: Marvell is one of only two companies on Earth that design these custom AI chips (silicon built for one customer) at real scale.

The other is Broadcom $AVGO ( ▼ 2.41% ).

No one else.

So whichever cloud giant "wins," Marvell probably helped build the chip, and now it gets paid on the wiring too.

The numbers are catching up fast:

Ten years ago, Marvell's standout product was the Wi-Fi chip inside Mattel's talking, voice-controlled Barbie Dreamhouse.

The CEO killed the toy business, spent years buying connectivity companies, and walked it onto a stage next to Jensen Huang.

But that is not even the wildest part.

Because most people look at the AI chip race and see Nvidia versus everyone else.

That misses where the money is actually pooling.

The biggest winner of the custom-chip era might be the one company that gets paid whether Nvidia wins or loses.

Here's the logic.

As AI systems grow, the hard part becomes getting tens of thousands of chips to behave like one machine.

That is a wiring problem.

Marvell's wiring business is growing faster than the chips themselves, up more than 70% in a year.

Designing custom AI silicon takes years of trust and factory relationships.

That's why only two names ever show up: Marvell and Broadcom.

A buyer can switch which cloud chip wins.

It's much harder to route around the two companies that design and connect them.

So here's the takeaway for your portfolio.

If your only AI bet is Nvidia, you own the chip everyone is crowding into and trying to disrupt.

The connectivity layer underneath is the quieter bet.

Only two players and a demand that rises no matter which chip ends up on top.

It fits a question worth taping to your monitor: what does AI need, no matter who wins?

Right now, a big answer is the wiring.

We called the wiring story three weeks ago in NVDA Is Building the Toll Road for AI.

Same road. This time, from the company that paves it.

While Nvidia gets the headlines, Marvell builds the part that makes the headlines possible.

We put together a one-pager on the cleanest way to get exposure to the connectivity layer:

The two names, how they differ, and what could break the thesis.

Reply "WIRES" and we'll send it your way.

(And tell us: chip, wiring, or cloud giant? We read every reply.)