Micron Sells the Memory Shortage as Five-Year Contracts

Micron Technology CEO Sanjay Mehrotra said datacenter customers want about 50% more memory than the company can commit to ship. He said it on Aug. 20 beside the Boise site where Micron is raising new fabs. Those buyers are not waiting in an open spot market. Micron already sold much of what it can make, at a ceiling set this spring, under deals that run to the end of 2030.

The leftover ask is now a rationing problem. It shows up in GPU growth caps, in leftover DRAM for phones and PCs, and in anyone who did not sign.

Data Center Customers Want 50% More Than Micron Can Commit

Mehrotra put the shortage in one line. All customers across end markets will buy everything Micron makes, he said, and datacenter buyers still want roughly half again as much as the firm will promise. That math means committed tons cover about two-thirds of the datacenter ask. The other third is demand with no Micron allocation attached.

Today there is no AI without memory. AI systems need more memory. They need higher performance memory. They need lower power memory. So, the value of memory, that equation has totally changed.

Sanjay Mehrotra, Chairman, President and CEO, Micron, speaking in Boise

He called memory the strategic infrastructure of the AI era and said customers now design it earlier, beside the processors it will sit next to. He also said Micron has signed more long-term deals since the 16 it disclosed in June. They have committed to taking the supply, he said, which gives the company assurance of demand.

THE QUARTER BEHIND THE 50% LINE

• Revenue: Micron posted $41.46 billion in the fiscal third quarter ended May 28, 2026, up 346% from $9.30 billion a year earlier.
• Margin: Non-GAAP gross margin reached 84.9%, versus 39.0% a year ago, with non-GAAP earnings of $25.11 a share.
• Datacenter: Combined cloud-memory and core datacenter sales were $25.29 billion, matching Mehrotra’s line that datacenter revenue exceeded $25 billion.
• Next quarter: Guidance is $50.0 billion of revenue, plus or minus $1.0 billion, at about 86% gross margin and $31.00 of non-GAAP earnings per share, plus or minus $1.00.

Those record third-quarter results for fiscal 2026 also showed $25.39 billion of operating cash flow, $18.3 billion of adjusted free cash flow, and $30.2 billion of cash and investments. Datacenter SSD revenue exceeded $5 billion and more than doubled from the prior quarter. Mobile and client still printed $11.52 billion. Auto and embedded added $4.63 billion. The shortage is not a single-product story. It is the whole book.

Price Ceilings Sit at Spring Market Levels

The 16 strategic customer agreements are take-or-pay. Buyers commit to set volumes over a multi-year term and pay whether they later want the bits or not. Typical terms run five years, from calendar 2026 through the end of calendar 2030. Automotive deals generally run three. The signed set covers about 20% of DRAM volume and a third of NAND volume, with four very large customers, three medium ones, and smaller auto names in the rest.

THE DEALS THAT SOLD THE SHORTAGE

Micron spelled out those take-or-pay agreements through 2030 in its June prepared remarks. The floor is the line bulls cite, and it is real. Even if prices slide to the contractual minimum, Micron says the margin on those tons would still beat every prior peak quarter. The ceiling is the line the 50% gap sits behind. If the shortage pushes transaction prices above this spring’s market, the largest signed books do not reprice all the way up. Micron sold insurance to itself and a discount to the buyers who got in.

Several smaller SCA slices use fixed prices or no bands at all, so some tons still float with the market. When the planned set is done, deals with fixed prices or ceilings near CQ2 levels are expected to be about 40% of revenue. That is a large share of the company to have already spent its upside. It is also why extra datacenter demand does not automatically become extra Micron profit at whatever the spot tape prints next week.

HBM4 12-high is ramping twice as fast as HBM3E 12-high, and Micron has already shipped more than $1 billion of HBM4. New products can still be negotiated at a premium under the SCAs. Existing products in the biggest deals cannot. The 50% Mehrotra will not commit is, in that sense, demand the contracts were never written to absorb.

Nvidia Guided 70% Because Memory Would Not Stretch

The same constraint showed up six days after Mehrotra spoke, on Nvidia’s latest quarter. Nvidia posted $96.2 billion of revenue for the period ended July 26, with $89.0 billion from datacenter hardware. Long-term purchase commitments jumped from $119 billion to $279 billion. The company said the bulk of that book is primarily tied to buying memory. Customer forecasts pointed to a doubling of demand next year. Supply, Nvidia said, supports about 70% growth.

Even though our demand is much greater than 70%, our supply allows us to confidently deliver 70%. The unconstrained would be a lot, a lot higher. We have secured a lot of supply, but we just need a lot more.

Jensen Huang, Founder and CEO, Nvidia earnings call

Huang’s 70% cap is the 50% gap seen from the other side of the tray. GPUs do not ship as systems without the stacked memory that feeds them. When Micron, SK hynix, and Samsung have already placed their near-term HBM under long-term deals, a buyer with cash still cannot pull extra wafers out of a sold calendar. Nvidia guided about $108 billion of revenue, plus or minus 2%, for its next quarter and said supply stays a bottleneck at least through fiscal 2028.

That is the second-order cost of Micron’s new model. The contracts make Micron’s own margin path more stable. They also freeze a large slice of industry output at known prices and known volumes, which is exactly what leaves an unfilled 50% hanging in the air. The people who wanted that extra memory do not vanish. They wait, they bid for leftovers, or they slow a cluster.

Who Still Waits Outside the 16 Deals?

Mehrotra would not name the four large and three medium SCA customers. He said they span datacenter, consumer, and auto. The rest of the industry is the residual claimant on whatever bits those seven plus the smaller auto names do not take.

THE SHORTAGE AFTER THE SIGNED BOOK

• Uncontracted datacenter buyers: The extra 50% Mehrotra will not promise is, by definition, demand with no Micron commitment behind it.
• Phone and PC lines: Mobile and client still sold $11.52 billion in the quarter, and HBM’s rising trade ratio pulls wafers away from ordinary DRAM.
• Shorter auto deals: Automotive SCAs generally last three years, not five, so that book rolls off sooner than the datacenter stack.
• GPU ramps without matching HBM: Nvidia’s 70% guide is the public version of a private allocation fight every accelerator vendor is in.

HBM’s wafer tax is the quiet mechanism. Each new stacked generation uses more silicon for the same bit count, which is why Micron says HBM growth further pressures non-HBM supply. The leftover for PCs and phones is not a separate cycle. It is the same factory choosing AI trays first. Module prices in that residual channel move because the wafers underneath them were reassigned, not because laptop demand suddenly doubled.

A rival stacked design that is still on a roadmap does not change 2026 output. What would change the queue is more cleanroom, and that is years of concrete.

Mid-2027 Is the Earliest Boise Can Help

Mehrotra said the first Boise fab should start producing wafers in mid-2027. The site is planned for two leading-edge memory fabs in Boise, each about the size of 10 football fields, co-located with Micron’s Idaho research work. The Idaho page puts the local job claim above 17,000 and notes up to $6.4 billion of CHIPS Act direct funding across Idaho, New York, and Virginia. Company materials put the broader U.S. manufacturing and research plan at more than $250 billion and about 90,000 American jobs.

Those fabs do not shrink the 50% gap this year. Greenfield memory plants, Mehrotra said, are large, slow, and boxed in by skilled-labor shortages, permitting, and power. Process nodes are getting harder. Bit growth per wafer is slowing. Even as industry supply is expected to improve gradually in 2028, he said Micron still has no line of sight on when supply catches rising demand. Tight conditions, in the June remarks, persist beyond calendar 2027.

THE SUPPLY CLOCK MICRON LAID OUT

1. 2023: Mehrotra has said prices fell to one-third of prior levels and that the industry then faced the most severe DRAM and NAND imbalance in 13 years, with a plan to cut about 10% of headcount.
2. June 24, 2026: Micron reports the $41.46 billion quarter and discloses 16 SCAs, about $100 billion of floor-price revenue on 14 deals, and $22 billion of deposits.
3. Aug. 20, 2026: Mehrotra says datacenter customers want about 50% more than Micron can commit and that more SCAs have been signed.
4. Aug. 26, 2026: Nvidia ties its 70% growth guide to supply, with memory the binding input.
5. Mid-2027: First Boise wafers, if the construction calendar holds.
6. End of 2030: Five-year SCA terms expire.

Micron’s next scheduled print is its fiscal fourth-quarter call on Sept. 30. That number can confirm whether the $50.0 billion guide landed. It cannot make mid-2027 arrive sooner.

Prices Fell to One-Third Only Three Years Ago

Memory has a habit of teaching this lesson. Strong demand pulls in new plants. The plants arrive late. Prices collapse. In 2023, after customers pushed pricing hard, Mehrotra said Micron’s prices dropped to one-third of what they had been, and the company went through negative gross margins just as AI demand was accelerating. He has argued that those bargains left the industry underinvested for the boom that followed.

The SCAs are Micron’s answer to that winter. Buyers who lived through it are now parking deposits to hold a place in line. That can be read as proof the cycle broke. It can also be read as the classic late-upcycle move, when the same hyperscalers who crashed the 2018 boom by destocking in 2019 lock in tons at peak-ish prices because they are frightened of idle accelerators. The contracts do not settle that argument. They freeze a slice of volume and a band of prices until 2030 and leave the other slice on the open market.

Micron is not chasing HBM share far above its overall DRAM share, Mehrotra has said, because overweighting stacked memory would starve the conventional DRAM that is also short. That allocation choice is why a 50% datacenter gap can coexist with an 84.9% company margin. The firm is choosing mix, not filling every order.

Take-or-Pay Protection Runs Out in 2030

About half or more of revenue is the SCA target, not the current mix. Until the remaining deals close, most of the company still clears in a market that can gap either way. After they close, about 40% of revenue is expected to sit at fixed prices or at ceilings near this spring. The floor on the banded deals is the crash hedge. The ceiling is the boom hedge for the customer. End of 2030 is when the five-year stack starts to roll off, unless it is extended.

Server bit shipments in calendar 2026 are expected to more than double from two years earlier, Micron said, and industry server units are now seen growing in the high teens, above a prior low-double-digit view. Some of that unit lift, the company added, comes from customers putting a bit less DRAM in each box so they can ship more servers on a tight allocation. That is the 50% gap written into a bill of materials. The box gets built. It just gets built thinner.

Boise’s first wafers are still three quarters of a year away at the earliest. The extra half that datacenter customers want is uncommitted today, and the contracts that made Micron’s margin safer also made that extra half someone else’s problem.

Disclaimer: This article is news reporting and analysis for information only. It is not investment, tax, or financial advice and is not a recommendation to buy, sell, hold, or short Micron or any other security. Memory and semiconductor stocks can move sharply with prices, contracts, and factory timing, so readers should consult a qualified financial adviser or licensed broker who can judge their own objectives and risk before acting. Figures, contract terms, and guidance here reflect company filings and public comments as of Sept. 2, 2026, and may change with later results, including Micron’s fiscal fourth-quarter report scheduled for Sept. 30, 2026.