Why Fast AI Chips Still Have to Wait

Picture a $300,000 sports car stuck in rush-hour traffic.

It’s got 700 horsepower and can top 200 miles an hour. Right now, it’s doing about eleven.

Nothing’s wrong with the car. The problem is everything around it.

That’s what’s happening inside many AI data centers. Companies have spent billions on Nvidia’s (NVDA) fastest chips. But even a fast chip can be held back when information doesn’t arrive quickly enough. When it gets stuck in traffic, some of the priciest hardware on the planet sits idle.

The good news is somebody gets paid to fix it.

Other companies store the information, move it around, and keep it flowing. They’re not the car. They’re the road.

Nvidia noticed and dropped $4 billion in investments in two of these companies in March. More on this in a minute.

Let’s head back to “Earnings Desk,” the fictional research app from this series on AI. It’s a busy earnings-season morning, and we’ve asked it to compare a company’s new sales forecast with management’s promises from a year ago.

We want that answer before the opening bell. Unfortunately, thousands of other investors want theirs too.

Earnings Desk has plenty of computing muscle, but first it must find the right reports. If that’s slow, we’ll be drumming our fingers no matter how much somebody spent on chips.

(Note: This is the third article in Bob Byrne’s AI series. Read the first, “Who Gets Paid When You Ask AI a Question?” and second “Meet the Chips Behind Every AI Answer.”)

Keeping the Reports Within Reach

Earnings Desk’s cloud provider might store years of filings, transcripts, and earnings releases, including reports nobody’s opened in months.

That’s where Seagate (STX) and Western Digital (WDC) come in. Their hard drives hold huge amounts of information at a relatively low cost.

The reports in Earnings Desk are just one example of what AI businesses need to store. Developers also collect enormous amounts of text, images, and video to train their models, while customers save the answers and other material AI creates. As those collections grow, so does the need for storage.

But when everybody shows up at once, like that earnings-season rush, getting things out can take a while.

That’s where SanDisk (SNDK) helps. Its solid-state drives, or SSDs, use flash chips instead of spinning disks to deliver information much faster. They cost more, so providers use them where delays would hurt their AI service.

Think of storage like this:

Hard drives are the filing cabinets down the hall. SSDs are the drawer next to your desk, holding the files you grab every day. The computer’s memory is the desk itself, where the work gets done.

Customers can use both kinds of storage, giving investors a reason to watch Seagate, Western Digital, and SanDisk as AI spending grows.

Keeping the Traffic Moving

Storage is only part of the trip. The report may still need to reach another computer before the AI can use it. And processors sharing one big AI job are constantly passing information back and forth.

Put more cars on the highway without adding lanes, and you get a bigger traffic jam. Add processors without upgrading the network, and high-dollar equipment has to wait its turn.

Arista Networks (ANET) sells the networking gear that directs that traffic. Think traffic lights and on-ramps for data.

Arista doesn’t have the road to itself, though. Nvidia sells networking equipment alongside its processors, making it one of Arista’s toughest competitors. Its AI systems can come with built-in networking.

Even the cables are big business. Credo Technology (CRDO) makes copper cables with built-in electronics that keep information moving fast and reliably. Astera Labs (ALAB) sells chips that help server components talk to each other.

It might sound small, but a hiccup between two pieces of equipment slows both down. That’s why customers pay for it.

Nvidia Is Betting on Light

Copper works great for short hops. As connections get longer or carry more information, optical fiber helps carry that traffic as light.

Lumentum (LITE) and Coherent (COHR) make the lasers and other equipment that send and receive those signals. They’re not selling the glass. They’re selling the gear that puts the light to work.

Here’s why that matters if you’re investing in the AI stack.

Nvidia’s newest AI systems, called Vera Rubin, pair its latest processors with the equipment that connects them. Nvidia says their optical networking uses less power than the older connections it replaces.

That’s bigger than it sounds. In a giant AI operation, every watt saved on the network makes it easier to grow. As more of these systems get built, the optical suppliers stand to benefit.

In March, Nvidia announced a $2 billion investment in each of Lumentum and Coherent, plus multibillion-dollar purchase commitments to both. The investments fund research and new manufacturing, and the commitments cover optical products.

When the biggest name in AI commits that much money to its suppliers, it’s telling you where it thinks demand is headed.

Those deals cover Nvidia’s broader buildout, not just Rubin, and when the purchases turn into revenue depends on customer rollouts and supplier deliveries. But the story is moving in the right direction.

The road doesn’t stop at the building’s walls, either. Ciena (CIEN) sells optical networking systems that link entire data centers.

Back to the ‘Desk’

Back at Earnings Desk, we just want our answer before the open. We don’t care how it gets there.

Now you can see why a company spending heavily on AI chips also spends on storage, networking, and optics. A sports car still needs a clear road, and keeping those computers busy can help customers earn more from the equipment they already own.

Of course, those computers need a building with enough power and cooling to run them. Finding someone who can deliver that on time is another business worth knowing. We’ll dig into that next time.

More From Bob Byrne:

At the time of publication, Byrne had no positions in any securities mentioned.