Micron Technology expects memory supply to fall further behind demand in 2027 than it will in 2026, the company’s president said, extending a squeeze that has already reset prices across the chip industry.
Scott DeBoer, who is Micron’s president and chief technology and product officer, said in an interview that supply and demand for memory will not come into balance for a long time.
In the foreseeable future there is no clear equilibrium point, he said. He expects 2027 to be tighter than 2026.
The forecast matters because memory has become the binding constraint on artificial intelligence hardware. Accelerators cannot ship without high-bandwidth memory stacked beside them, and every stack consumes DRAM wafers that would otherwise make conventional memory.
Micron is one of three companies that supply that memory at scale, alongside SK Hynix and Samsung Electronics. The three decide, between them, how much of the world’s AI computing can be built in any given year, and at what price the rest of the industry buys it.
DeBoer’s view runs against the reflex that has governed the industry for three decades: that a shortage carries the seed of its own end, because high prices pay for the capacity that ends them. Micron’s own history supports the reflex. The company lost $5.8 billion in its 2023 fiscal year when a post-pandemic glut crushed prices, and revenue fell by half.
What is different this time, in the company’s telling, is where the demand comes from. Memory used to be sold into phones and personal computers, markets that grew slowly and turned quickly. A growing share is now sold into data centers under contracts signed years ahead, which removes some of the volatility that used to end a cycle early.
Micron has said its high-bandwidth memory output is committed under agreements that run into 2026, with pricing fixed in advance for much of the volume. That structure protects revenue in a downturn and caps the gain in an upswing, which is one reason the company is arguing publicly that tightness will last.
Capacity is the other reason. Micron said in June 2025 that it would invest about $200 billion in U.S. manufacturing over roughly two decades, including a fab in Boise, Idaho, and a cluster of plants in Clay, New York. The Commerce Department awarded the company $6.1 billion under the CHIPS Act in late 2024 to support the build. The New York site is planned for as many as four fabs.
None of that capacity arrives quickly. A leading-edge memory fab takes about three years from groundbreaking to production, and the equipment inside it is subject to its own queues. Money committed in 2025 shows up as supply in 2028 or 2029, which is a long wait after the shortage it is meant to fix.
The demand side has its own schedule. Nvidia’s Rubin platform, the successor to Blackwell, is due to ramp through 2026 and uses a newer generation of high-bandwidth memory. Each generation carries more memory per accelerator, so the same number of chips absorbs more wafer supply than the one before it. New stacking processes also yield less usable output per wafer in their first year.
Analysts said the forecast is credible over the next eighteen months and harder to defend further out. Memory has overshot in both directions for as long as it has existed, and the industry has a record of adding capacity in a shortage and delivering it into a glut.
One variable is how the hyperscalers behave with those capital budgets. Microsoft, Amazon, Alphabet and Meta have all raised capital spending plans this year, and their orders are what the memory makers are expanding against. A single round of budget revisions at two of those companies would change the arithmetic DeBoer is describing.
Another is China. CXMT and other domestic producers are adding DRAM capacity with state support, mostly at older nodes, and their output does not yet reach the AI supply chain. It does reach the consumer markets that Micron still depends on for volume. NAND has moved with DRAM, with enterprise SSD orders rising as buyers work around long waits for hard drives.
The risk is the one that has broken every memory forecast before it. Capacity ordered in a shortage arrives in a glut, usually about two years after the orders are placed, and the price of the correction is paid by whoever expanded last.
For now the numbers run the other way. Contract prices for conventional DRAM have risen for several consecutive quarters, and buyers who once negotiated annually are signing quarterly contracts with take-or-pay terms. Micron reports results on a fiscal year that ends in late August, so its next set of guidance will be the first test of whether the company’s own customers agree with its president.


