Micron’s $22 Billion Deposits Cover Only a Fifth of DRAM

Micron locked $22 billion of customer deposits across 16 take-or-pay deals that still cover only about 20% of its DRAM. The rest of the book still prices like a memory cycle, and the shares have already slipped from a three-month high of $1,213.37 to a September 1 close of $933.44.

Fiscal fourth-quarter results, due after the close on September 30, will test an 86% gross-margin guide. They will not settle whether those contracts turned a chipmaker into a software company.

Micron’s Contracts Cover a Fifth of Its DRAM

The fiscal third quarter ended May 28 with revenue of $41.46 billion, up from $23.86 billion in the prior quarter and $9.30 billion a year earlier. Non-GAAP earnings were $25.11 a share on $28.86 billion of net income. GAAP net income was $28.24 billion, or $24.67 a share. Operating cash flow hit $25.39 billion, and adjusted free cash flow was $18.3 billion after $7.1 billion of net capital spending.

Data-center revenue topped $25 billion in the quarter. Data-center SSD revenue passed $5 billion and more than doubled from the prior period. High-bandwidth memory of the HBM4 generation already contributed more than $1 billion of revenue, with the 12-high ramp running twice as fast as HBM3E 12-high, according to the company’s June 24 slides.

The structural disclosure sat beside those figures. Micron said it had signed 16 strategic customer agreements with data-center, consumer, and auto buyers, typically running from calendar 2026 through the end of calendar 2030, with automotive pacts generally lasting three years. Those 16 deals represent roughly 20% of DRAM volume and a third of NAND volume over the term, or about 25% of revenue. Management’s target is approximately half or more of company revenue once the remaining pacts are done.

SCA TERMS DISCLOSED JUNE 24

Fourteen of the 16 pacts carry about $100 billion of remaining performance obligations at the contractual floor. That number is a minimum, not a forecast. Chief Financial Officer Mark Murphy said actual revenue should run well above associated RPOs over the life of the deals. When all planned agreements are executed, pacts with fixed prices or ceilings near those mid-2026 levels are expected to be about 40% of revenue.

Sumit Sadana, Micron’s chief business officer, said the five-year take-or-pay structure is something the industry has not done before. The floor, management said, supports a gross margin well above peak quarterly margins in any past cycle. The ceiling is the other half of the bargain. Existing products cannot reprice above the calendar-second-quarter 2026 market, even if the shortage deepens.

The $22 Billion Lands as Customer Deposits

The $22 billion is real money, and it is not free cash flow. About $18 billion is cash deposits and about $4 billion is letters of credit. Roughly $500 million arrived in fiscal Q3. About $10 billion is expected in fiscal Q4. Murphy said the cash will show up on the balance sheet more in the fourth quarter, and that the associated flows appear in financing cash flows and will not affect free cash flow.

The deposits are held during the performance period and returned over time, weighted toward 2029 and 2030, the back half of the term. They function as a customer-funded working-capital loan that comes due as the first greenfield lines from this cycle are supposed to start producing. Micron ended the third quarter with $30.2 billion of cash, marketable investments, and restricted cash.

CASH THAT IS NOT EARNINGS

• Q3 cash from operations: $25.39 billion, against $11.90 billion in the prior quarter and $4.61 billion a year earlier.
• Adjusted free cash flow: $18.3 billion in the third quarter, after $7.1 billion of net capex.
• Q4 capex guide: around $10 billion, as the company builds toward a $250 billion U.S. manufacturing program across four sites.
• Buybacks in Q3: none, with the next step-up in capital return gated to December 9, after the second anniversary of the CHIPS Act agreements.

From fiscal 2023 through the third quarter of 2026, Micron returned $3.4 billion to shareholders, $1.4 billion as buybacks of 14 million shares and $2.0 billion as dividends. The board declared a $0.15 dividend payable July 21. Over time, management said, it expects to return 100% of excess cash. The CHIPS funding terms still sit in front of that promise.

Chairman, President, and Chief Executive Sanjay Mehrotra has been blunt about the other side of the ledger. Data-center customers want about 50% more memory than Micron can commit to deliver. HBM output for 2026 is sold out. He has said the company still has no line of sight to when supply catches demand, and that calendar 2027 looks tighter than 2026.

Four Large Customers Now Set the Mix

The 16 pacts include four very large customers and three medium-sized customers. The rest are smaller automotive accounts. Micron has named some of those auto counterparties in separate notices, including a supply pact with General Motors for LPDRAM, NOR, and UFS NAND, plus automotive Tier 1 supply deals with Qualcomm, Visteon, HARMAN, JOYNEXT, DENSO, Astemo, and Hyundai Mobis. The four very large accounts have not been named.

That silence is the point of leverage. The buyers who can write $18 billion of cash deposits are the same buyers who now hold contracted volume, a price ceiling on current parts, and a say in which wafers become HBM instead of commodity DRAM. New-product premiums still have to be negotiated, the slides say, as bit costs rise through LP6, DDR6, and later HBM generations.

WHAT WE KNOW

• Binding form: The pacts are take-or-pay, with committed volumes by year and deposits Micron holds on the balance sheet.
• Named counterparties: Auto and industrial names have been disclosed; the four very large accounts have not.
• RPO timing: Remaining performance obligations were about $5 billion as of May 28, including $422 million of contract liabilities, before post-quarter signings lifted the floor figure to about $100 billion.

WHAT IS UNCONFIRMED

• The four names: Micron has not identified the very large customers on the 16-deal list.
• 50% coverage: Half or more of revenue under SCAs is a target, not a signed book.
• Downturn behavior: No public test yet of whether these floors are renegotiated if AI capex stalls.

Shay Boloor, chief market strategist at Futurum Equities, put the bull case in a single line after the print: customers putting down deposits and accepting minimum pricing are taking on part of the cyclical risk Micron used to carry alone, which in his view supports a higher multiple than a commodity memory name.

A memory company with customer deposits, take-or-pay commitments, minimum pricing and multi-year supply agreements deserves more credit for earnings durability.

Shay Boloor, Chief Market Strategist, Futurum Equities

Ron Insana, a market commentator, answered that take-or-pay was blown to bits in the 1980s natural gas market when prices plunged, and that it was not pretty for pipelines or producers. The history does not prove these DRAM pacts will break. It does prove that a floor is only as hard as a customer’s willingness to keep writing checks after the shortage ends.

CXMT Is Filling the Bits Micron Converted Away

Every HBM wafer pulled out of a DRAM line subtracts more bits than it adds, because a gigabyte of high-bandwidth memory eats far more silicon than a gigabyte of commodity DRAM. Micron’s own slides call out an increasing trade ratio with every new HBM generation, which pressures non-HBM supply. That is the opening ChangXin Memory Technologies has been walking through.

Counterpoint Research’s Q1 DRAM and HBM market share tables put CXMT at 8% of DRAM in the first quarter of 2026, up from 3% in the first quarter of 2025. Samsung led at 38%, SK hynix held 29%, and Micron held 22%. The DRAM market grew 80% quarter on quarter and 260% year on year in that quarter, so CXMT’s share gain came inside a price boom, not a share fight in a flat market.

DRAM REVENUE SHARE, Q1 2026

SemiAnalysis models CXMT wafer capacity additions of about 85,000 wafer starts a month in 2026, 70,000 in 2027, and 80,000 in 2028, against Micron additions of 30,000, 90,000, and 115,000 in those years. Samsung’s modeled adds are 15,000, then 50,000, then 110,000. SK hynix is modeled at 60,000, 60,000, and 90,000. CXMT is still a trailing-node producer of DDR and LPDDR for Chinese and, increasingly, global device makers. It is not yet a fourth HBM peer at scale.

That distinction got noisier in early September, when CXMT moved into risk production of HBM3E and began sampling Chinese accelerator customers. Yields on early 8-high stacks have been described in research notes as far below mature lines at the three incumbents, and CXMT remains one to two generations behind SK hynix on the HBM roadmap. The conventional DRAM it does ship is the product the Big Three have been starving to feed HBM. If you want the merchant-market valve on the 80% of Micron bits that are not under an SCA, that is where it sits.

Where Micron Stands on High-Bandwidth Memory

High-bandwidth memory is the product bulls treat as the toll. Counterpoint’s first-quarter HBM revenue share still has SK hynix at 58%, down from 69% a year earlier, with Samsung and Micron tied at 21% each. Most of that revenue is still HBM3E. HBM4 shipments were only starting to show up in the second half of the year on Counterpoint’s June read, even as Micron said HBM4 built on 1-beta DRAM was already in high-volume shipments for its lead customer’s platform.

HBM4E, on 1-gamma DRAM, is slated for volume production in calendar 2027. That is also the year Mehrotra has flagged as tighter than 2026, and the year the first wafers are due from the new Boise shells. Each of those Idaho fabs, he said, will be the size of 10 football fields. Production ramps in 2028. The second Boise fab starts late that year. Clay, New York, ramps on a similar multi-year clock. None of that capacity is in the 2026 sold-out book.

Daniel Newman, CEO of Futurum Group, said after the June print that the AI build-out has been underestimated at every turn and that memory will keep commanding premium pricing on supply constraints. The premium is not evenly spread. SK hynix still takes more than half of HBM revenue. Micron is a credible third source into Nvidia’s Vera Rubin platform, not the gatekeeper. The SCA floor sits under a mix that includes a lot of non-HBM DRAM the company is deliberately shrinking as a share of wafers.

The 2018 Hyperscaler Pause Still Prices Into Shares

Memory has heard “this time the cycle is different” before. The 2017-2018 upcycle was sold as cloud and NAND tightness that would not mean-revert. Industry DRAM revenue reached $101 billion in 2018. Then datacenter buying stopped in mid-fourth quarter. Quarterly DRAM revenue fell more than 20% to $22.9 billion. Combined DRAM and NAND revenue was $35.8 billion, down 19% from the third quarter. Hyperscalers had spent 2017 and early 2018 filling racks, then shifted to hardware and software optimization. The supply that had been added into the boom arrived into that pause.

THE DATES THAT STILL MATTER

1. Late 2018: Hyperscaler DRAM demand drops abruptly after a two-year build, and contract prices follow spot lower.
2. June 24, 2026: Micron reports $41.46 billion of quarterly revenue and discloses the 16 SCAs, $22 billion of deposits, and the $100 billion floor.
3. September 30, 2026: Fiscal Q4 results are due, against a $50.0 billion ± $1.0 billion revenue guide and an 86% gross-margin guide.
4. December 9, 2026: The CHIPS Act repurchase restriction’s second anniversary, the date Micron has tied to a larger capital-return step-up.
5. Calendar 2027: HBM4E volume production and, in management’s telling, a tighter year than 2026, with first wafers from new U.S. shells.
6. End of calendar 2030: The five-year SCA window closes, and the weighted deposit returns are designed to be largely complete.

The 2026 setup is not a carbon copy. Take-or-pay with cash held on Micron’s balance sheet is a harder instrument than the loose 2018 order books. Greenfield fabs take years, and HBM’s wafer hunger makes it harder to flood bits the way the industry did after 2018. The stock is also no longer priced as if the print cannot be believed. A one-year low of $118.29 sits under that $933.44 close. The three-month high of $1,213.37 is 23% above it. Late August already showed the tape can drop on China-supply headlines even while 2026 HBM is sold out. That is a multiple argument, not an order-book argument.

September 30 Is a Gross-Margin Print

Guidance for the quarter ending in late August is $50.0 billion of revenue, plus or minus $1.0 billion, with gross margin of approximately 86%, non-GAAP operating expenses of about $1.65 billion, and non-GAAP earnings of $31.00 a share, plus or minus $1.00, on about 1.15 billion diluted shares. A 15% tax rate was the working assumption on the last call. The $10 billion capex figure is the cash going out the door while the first large slice of SCA deposits is supposed to be coming in.

Mehrotra’s line on the June call was that the quarter’s records, and an even stronger fourth-quarter outlook, reflect memory’s strategic value in the AI era, and that the multi-year agreements will enhance the durability and predictability of results. The sentence that will be tested on September 30 is the 86% margin, because that is where the ceiling, the floor, and the uncontracted 80% of DRAM meet in one number.

THE PRINT AND THE GATE

• Gross margin: An 86% result would show that calendar-Q2 price levels are still holding into late summer, including on bits outside the SCA floor.
• Deposit cash: About $10 billion of the $18 billion cash pile is expected in Q4; a miss would be a contract-timing story, not a demand story.
• SCA count: Mehrotra has said more pacts have been signed since June; the coverage ratio against the 50% target is the number that changes the mix.
• December 9: Buybacks restarting in size, or not, will show whether CHIPS cash and the deposit liability leave any excess to return at a $900-plus share price.

The contracts did change the cash cycle. Customers prepaid, accepted floors, and accepted take-or-pay language the industry had not used at this scale. They also took a ceiling on what they will pay for the parts already shipping, left most of Micron’s DRAM volume off the page, and scheduled the deposit returns for the same years the new fabs are meant to start. CXMT is using those years to add conventional wafers the three incumbents are converting away. September 30 will say whether 86% gross margin survived the summer. It will not say the other four-fifths of the bits have been repealed.

Disclaimer: This article is news reporting and analysis of Micron Technology’s disclosed results, contracts, and industry share figures, and it is for information only. It is not investment advice, a recommendation to buy or sell MU or any other security, and it is not a forecast of future earnings, margins, or share price. Readers should consult a licensed financial adviser or other qualified professional who can consider their own objectives, risk tolerance, and tax situation before making any investment decision. Figures, contract terms, ratings, and market prices reflect the company filings, investor materials, and market data cited here as of the dates of those sources and can change with the September 30 report, later SCA disclosures, and subsequent trading.