Nvidia's Secret Play + AI Unicorns Multiply Fast

NVIDIA's networking division just hit $31 billion in annual revenue, and honestly, most

people have no idea this business even exists.

Last quarter alone, they did $11 billion, up 267% year over year.

That's now their second largest revenue driver after compute.

Wait, $11 billion in one quarter, that's more than Cisco's entire annual networking

business.

What is this flying under the radar?

Right?

It all traces back to their 2020 acquisition of Melanox for $7 billion.

They're selling everything needed to build what they call an AI factory.

NVLink for GPU communication, Infiniband switches, Spectrum X Ethernet platforms, basically

all the connections between the chips.

So NVIDIA now controls both the chips and the plumbing.

That's a pretty tight grip on AI infrastructure.

Exactly.

And it's strategic.

As hyperscalers build out training facilities, NVIDIA is not just selling them processors

anymore.

They're selling the entire nervous system of the data center.

Speaking of AI infrastructure, micron just reported quarterly revenue that nearly tripled

year over year, beating estimates.

The driver, memory chips for AI servers, they literally can't scale production fast enough

to meet demand.

That's wild.

So we're seeing supply constraints on the memory side now.

Not just chips.

Yep.

Cloud providers and AI companies are competing for the same specialized high bandwidth memory

needed for training large language models.

Samsung and SK Hinex are facing the same capacity issues, which means pricing power is going

to persist through at least mid-2026.

I think this is huge because it confirms AI infrastructure spending is still robust despite

all the hand-ringing about ROI timelines.

Companies are still buying.

Yeah, that tracks.

The money's flowing even if the returns are uncertain.

Bicron's results basically prove the AI build out isn't slowing down anytime soon.

Okay, shifting years completely, school bus tech company ZUM hit $333 million in revenue

for 2025, growing 35% year over year.

They serve over 4,000 schools and have booked more than 2 billion in future contract revenue.

School buses, that's not exactly a sexy tech vertical.

What's the play here?

They're basically modernizing an antiquated system.

Their platform optimizes routing, connects drivers and parents, and they even build out

electric bus fleets.

Oakland now has the country's first all-electric school district fleet thanks to ZUM.

Interesting.

So school districts are paying a premium for that.

They are, but administrators say it's worth it because it solves chronic reliability problems.

Buses, no air conditioning, constant parent complaints, ZUM's 5 to 10-year contracts give

them recurring revenue, and they're eyeing an IPO, though timings unclear.

If they can prove the unit economics work, I can see investors getting excited about

that predictable revenue model.

AI's security startup, Expo, just raised $120 million at evaluation over a billion.

They use AI to find vulnerabilities in applications before hackers exploit them.

The CEO is Og Demor, who used to run GitHub's co-pilot.

So, they're using AI to defend against AI-generated code problems?

That's kind of meta.

Exactly.

Demor says AI coding assistance produce insecure patterns because they train on poorly

secured public repositories.

Expo automates penetration testing to catch these flaws.

They've signed over 100 customers, including Moderna and Samsung, with strong demand in

South Korea.

I'm not so sure about this.

If AI models get better at avoiding insecure patterns, doesn't that undermine Expo's

whole premise?

Demor acknowledges models will improve, but he says they'll still struggle with business

logic flaws, like when data showed or shouldn't be shared.

And he's warning that swarms of automated malicious attacks are coming.

Companies need to get ready.

And finally, researchers at Osaka Metropolitan University built a tomato-picking robot that

evaluates how easy each tomato will be to harvest before attempting the pick.

It analyzes stems, leaf obstruction, and cluster positioning to choose the best angle.

Wait, so it's not just identifying ripe tomatoes, it's thinking about whether it can actually

grab them?

Right.

They call it harvest ease estimation.

The robot achieved an 81% success rate, and a quarter of successful picks came after

it switched from front to side approaches when the first attempt failed.

It's learning to adjust its strategy.

It's actually pretty clever.

I mean, the real world application here is robots handling the easy picks while humans

tackled a difficult fruit.

Addresses labor shortages without full automation.

Yeah, it's about collaboration, not replacement.

The key shift is robots making informed decisions about their own capabilities instead of just

attempting every harvest blindly.

That's today's Pivot 5.

Like and share Pivot 5.

Thanks for listening.