There are businesses where getting something wrong is expensive, and businesses where getting something wrong becomes expensive four years later. Memory belongs to the second group, and that delay explains almost everything that has happened to the industry over the past four decades, including the disappearance of nearly all of its participants.
Building a chip plant takes three to four years of construction alone. Then come the machines, calibration and production-line validation. By the time the factory begins producing something that can actually be sold, the market that justified building it may already have changed. That is part of the reason the industry has shrunk so dramatically: of the dozens of DRAM manufacturers that once existed, only three remain, and just one produces in the United States.
That company is Micron. It has been led since 2017 by Sanjay Mehrotra, who has spent more than forty-five years in semiconductors. He started as an engineer, co-founded SanDisk and now runs the last U.S. memory manufacturer.
Much of his management comes down to three constraints that no executive in this industry can avoid.
Concrete cannot be rushed
The first constraint is physical, and there is not much room for negotiation.
When Mehrotra is asked when the chip shortage that is making phones and computers more expensive will ease, he does not offer broad forecasts. He looks at the factories under construction and answers from there. Boise’s advanced facility will begin producing in the middle of next year. A second plant is expected toward the end of 2028. Four more are planned in New York and will be built in stages. Before 2028, he says, there is little room for supply to loosen significantly because those factories simply do not exist yet.
That leads directly to Micron’s current problem: demand is running far ahead of supply. For its largest customers, the company can provide roughly half of what they request and, in the best cases, about two-thirds. This is not a commercial decision or a pricing tactic. It is simply everything Micron can manufacture.
One detail that is often overlooked makes the constraint even tighter. For decades, each technological generation made it possible to extract more capacity from the same amount of silicon, and those improvements acted as a pressure-release valve. That valve is beginning to close: current technology transitions deliver much smaller gains than previous ones. In practical terms, innovation alone can no longer produce enough additional supply. New factories have to be built, and construction takes as long as it takes.
Someone else sets the price
The second constraint is economic, and responsibility for it is spread across the industry.
The familiar explanation for the memory cycle is straightforward: manufacturers become too optimistic, produce too much and drive prices down. Mehrotra adds another part of the story, one that is less comfortable for his customers. He argues that years of negotiating every last dollar out of chip prices left the industry with less room to invest. According to him, part of today’s shortage began long before chips actually became scarce. It began at the negotiating table.
Micron is trying something less glamorous to break that pattern. It has signed sixteen supply agreements lasting between three and five years with data-center operators and automotive companies.
The advantage is concrete. Customers know they will have product, and Micron can commit capital knowing that demand will be there. But something else changes as well: short-term pricing no longer has complete control over how much production makes sense. If those contracts hold, Mehrotra’s legacy may be less about successfully riding an extraordinary boom and more about beginning to change the cycle that repeatedly takes the industry from abundance to shortage.
Geography matters
The third constraint is where the chips are made. And here Micron is making a very large bet: it plans to invest close to US$200 billion in the United States across new factories and research. Today, roughly 10% of its DRAM memory is produced in the country. The goal is to bring that share to around 40% over the next decade.
The least glamorous part of the plan may be the one that says the most about where Micron is heading. In Virginia, the company has already begun producing long-life-cycle chips intended not for AI servers but for cars, aircraft, defense systems, industrial machinery and networks. It is unglamorous memory based on older generations of technology, but it must remain available for years because the products that depend on it can last for decades. That production used to come from Asia and is now being quadrupled on U.S. soil.
The plan is expected to support around 90,000 jobs across the ecosystem, and Mehrotra does not present that simply as good news. He treats it as another logistical problem: finding enough trained workers is nearly as difficult as obtaining the machines. That is why Micron works with universities and local education systems on training programs and has for years sent its own employees into high schools in the communities where it operates to explain what semiconductor careers actually involve.
The bet Micron did not have to make
Everything above was decided under pressure: supply is tight, customers are demanding more, and the company’s stock pushed Micron above a trillion dollars in market value at the end of May after quarterly revenue increased more than fourfold in a year.
The decision that was not driven by any immediate pressure is different. Micron plans to invest US$10 billion in a research laboratory headquartered in Boise, with branches in locations where the company already conducts development. Construction begins next year. The facility will not focus on products already in the catalog or even those coming next season. The idea is to bring together customers, universities, startups and Micron’s own scientists to work on materials, semiconductors and computing architectures with a horizon extending beyond any current commercial plan.
There was no need to announce it now. And that is precisely what reveals something about the way Mehrotra thinks.
In 2023, when the semiconductor sector was going through one of its worst periods in years and many companies were freezing investments, Micron maintained its development spending. Two years earlier, when high-bandwidth memory represented less than 1% of the market and few people were paying attention to it, the company was already warning that manufacturing capacity would eventually fall short.
Neither position was particularly comfortable at the time. Both eventually proved correct.
Announcing a ten-year investment plan immediately after the best quarter in Micron’s history reflects the same logic from the opposite side of the cycle. Maintaining investment when everything is falling requires endurance. Committing capital for a decade when everything is rising requires something less dramatic and more unusual: remembering that this moment will end too.
