Somewhere inside an artificial-intelligence data center, light is racing through a strand of glass.
That doesn't sound nearly as exciting as a chatbot writing software or an AI model generating a movie. But increasingly, it is the problem.
Thousands of specialized processors can perform extraordinary amounts of computation. They also need to talk to one another—fast. Eventually, making the processors faster doesn't help enough if the connections between them become the traffic jam.
That helps explain why Marvell agreed to acquire optical-component maker Lumentum in an all-stock deal of roughly $115 billion. Lumentum makes some of the equipment that moves information using light.
Michael Rosen wants to see the spreadsheet.
He's spent nearly 30 years in finance and lived through the dot-com crash, the housing bubble and the 2008 financial crisis. He's heard some variation of this time is different more times than he can count.
His first question isn't whether artificial intelligence will change the world. It's simpler.
“Show me why this company is worth $115 billion.”
Tara Brooks has a different problem with the question.
She works in communications in Los Angeles and spends much of her time thinking about power, inequality, and who disappears when economic stories become numbers on spreadsheets.
“You're asking whether investors make enough money,” she says. “Before we're finished, I want to know who else is paying for this.”
The Numbers Are Real
Michael warns that there’s plenty here to make an experienced investor nervous.
Lumentum's stock had risen roughly 370% during the year before the deal. The potential return from a deal could be roughly 6%, compared with a roughly 10% cost of capital. That doesn't prove Marvell overpaid. It illustrates the valuation problem.
Michael has seen this movie.
During the dot-com boom, investors correctly understood that the internet would transform civilization. They also lost fortunes.
“You can correctly predict the future,” Michael says, “and still pay too much for it.”
Then the evidence interrupts him. Lumentum isn't a company surviving primarily on AI enthusiasm.
Its fiscal-2026 revenue increased 83.2%, reaching just over $3 billion. Some optical products have reportedly been sold out years in advance.
And then comes the number that makes $115 billion look strangely smaller.
The International Energy Agency expects the world's largest data-center operators to spend more than $700 billion in capital expenditures in 2026, much of it on data centers.
Picture the problem Marvell thinks it is solving: Building more AI processors without expanding the networks connecting them is a little like building thousands of new houses while leaving the highway at two lanes.
Marvell believes optical networking can become part of AI's highway system. Michael isn't ready to surrender his calculator. But he has to concede something.
This isn't a financial boom searching for an underlying business. The underlying business is booming.
Tara Changes the Spreadsheet
Tara isn't particularly impressed that the infrastructure is real. AI doesn't exist in a cloud.
The “cloud” occupies land. It consumes electricity. It requires transmission lines, transformers, generators, cooling equipment and, in many cases, significant amounts of water.
Nearly 40% of identified AI-specific data-center capacity is in areas experiencing high or extremely high baseline water stress, according to the International Energy Agency.
If shareholders receive the upside, who pays for grid expansion? What happens to electricity customers? Who gets the construction jobs? What happens to communities competing for water? What public incentives helped bring the data center there?
For someone trained to notice who is missing from the framing, the $115 billion headline suddenly has a lot of missing people. So far, reality isn't cooperating. The International Labour Organization finds real productivity gains from generative AI, but large-scale employment displacement has remained limited.
And the physical AI boom is creating another problem: America doesn't have enough workers to build it.
The expansion of data centers and electrical infrastructure is intensifying shortages of electricians, line workers, engineers and construction workers. Goldman Sachs has estimated that adding 300 gigawatts of U.S. power capacity by 2030 could require about 507,000 additional grid-connection, manufacturing, construction and operations workers.
AI may threaten some middle-class jobs while creating demand for others.
The contradiction becomes especially visible in Nevada. In the northern part of the state, data-center construction is providing years of work for electricians, pipefitters and other skilled tradespeople.
The question isn't simply whether AI destroys jobs. It's which jobs, whose jobs, where—and who captures the productivity gains afterward.
The Companies Nobody Talks About
Something else is happening outside Silicon Valley.
Generator maker Generac is investing $250 million to manufacture larger generators for data centers. It has accumulated a $1.6 billion backlog and expects to hire roughly 1,000 workers.
The boom is reaching makers of transformers, cooling equipment, cables, pipes, construction machinery and cement. Artificial intelligence may turn out to be an industrial boom disguised as a software revolution.
That brings us back to Lumentum. Everyone can name ChatGPT. Far fewer people could explain what an optical transceiver does. Yet the spectacular AI applications people see on their screens depend upon an enormous physical ecosystem underneath them.
Chips need power. Processors need cooling. Data centers need grids. And computers need increasingly sophisticated ways to communicate with one another.
It sounds like the old gold-rush strategy: don't search for gold; sell picks and shovels. But there's a better historical analogy.
What If the Bubble Bursts?
The dot-com boom produced spectacular failures. It also produced infrastructure.
Telecommunications companies poured money into fiber-optic networks. Investors overestimated demand, companies collapsed, and fortunes disappeared. But the fiber didn't vanish.
Later businesses inherited communications infrastructure that helped support the internet economy that followed. Michael stops worrying about whether AI is a bubble long enough to notice the distinction. Investment value and technological value aren't the same thing.
Marvell could be right that optical networking becomes indispensable to AI and still discover that $115 billion was too much.
Tara encounters the mirror image. Private ownership and social value aren't the same thing either.
An AI infrastructure boom could create factories, electrical capacity, skilled jobs, and productivity while distributing the rewards and disruptions very unevenly.
Today's AI companies may fail. Valuations may collapse. Some investors may lose billions. But the data centers could remain. The electrical infrastructure could remain. The optical networks could remain. The factories could remain. The trained electricians could remain. And the computing capacity could remain.
Which leaves a different $115 billion question:
A bubble can burst, but what will it leave behind—and who will own it?