Monday 5 October 2026 537 stories on file Full archive
Daily Edition
newscms

Volume III Edition Daily

Micron Won $73 Billion in Cash and Spent $10.8 Billion: The Memory Business Is Choosing Scarcity Over Share

There is a number buried in Micron Technology's fiscal fourth-quarter release that tells you more about the semiconductor industry than the headline does. The headline is that revenue hit $54.23 billion in the quarter…

Semiconductors 2,330 words 11 min read By Tech Desk

Micron Won $73 Billion in Cash and Spent $10.8 Billion: The Memory Business Is Choosing Scarcity Over Share — Semiconductors No Image Semiconductors
Lead image · Filed 5 October 2026, 02:49

Micron Won $73 Billion in Cash and Spent $10.8 Billion: The Memory Business Is Choosing Scarcity Over Share

Introduction

There is a number buried in Micron Technology's fiscal fourth-quarter release that tells you more about the semiconductor industry than the headline does. The headline is that revenue hit $54.23 billion in the quarter ended September 3, 2026, up 379 percent from a year earlier, with net income of $37.70 billion and an 86.8 percent gross margin. That is the kind of number that ends up on a ticker tape and gets forgotten. The buried number is capital expenditure: $10.77 billion for the quarter, against a cash and marketable investments position of $73.48 billion at year end.

Read those two figures side by side and the industry's most important strategic decision of 2026 comes into focus. Micron could, on paper, fund nearly four complete new fabs with the money it is sitting on. It spent the equivalent of about one. Its peers are not straining either. The memory industry, after four decades of building capacity into the next downturn, has decided that the current one is more valuable left unfinished.

This is the most consequential industrial-policy shift in semiconductors since export controls became a live commercial variable, and almost nobody is covering it as a strategy rather than as a balance sheet. Most reporting frames Micron's results as evidence of AI demand. That is true and it is the least interesting reading. The more durable story is what a set of competitors with collectively unprecedented capital and a collective instinct for self-restraint are now doing with it.

What the Quarter Actually Reported

Micron's own fiscal fourth-quarter and full-year 2026 release is unusually explicit about the shape of the business, and it is worth taking the segment numbers at face value rather than reaching for the usual AI framing.

Revenue for the full fiscal year reached $133.19 billion, against $37.38 billion in fiscal 2025. GAAP net income was $84.97 billion, or $74.33 per diluted share, on operating cash flow of $89.68 billion. Six consecutive quarterly records.

Within the fourth quarter, DRAM accounted for $39.77 billion, or 73 percent of total revenue. NAND flash contributed $14.10 billion, or 26 percent. Average selling prices for DRAM rose in the high-teens percentage range sequentially; NAND prices rose approximately 30 percent. Bits shipped rose only a single digit percentage. In other words, this quarter's result was overwhelmingly a price event, not a volume event, and the company said so.

The business-unit breakdown makes the concentration starker. The Core Data Center unit — conventional server DRAM and flash — produced $18.0 billion, up eleven-fold year on year, at a 90 percent gross margin and an 85 percent operating margin. The Cloud Memory unit, dominated by high-bandwidth memory, produced $16.28 billion, up 3.6-fold, at an 83 percent gross margin. Even the marginal segments are extraordinary: Mobile and Client at $13.1 billion, Automotive and Embedded at $6.8 billion, both at 84–90 percent gross margins. A 90 percent gross margin on commodity server DRAM is not a market condition. It is a supply condition.

Guidance for the first quarter of fiscal 2027 is $61.5 billion plus or minus $1.5 billion, with non-GAAP diluted earnings per share of $38.15 — against consensus of roughly $57 billion and $35.40. Micron expects fiscal 2027 to be another record year with sequential growth every quarter. Chairman and CEO Sanjay Mehrotra's framing in the release was that "AI is becoming Super Intelligence, and memory enhances this intelligence and the competitiveness of our customers' platforms."

The Capital Number Is the Real Story

Against that revenue, Micron spent $10.77 billion on capital expenditure in the quarter and $27.37 billion across the full year. It ended the year with $73.48 billion in cash, marketable investments and restricted cash. Operating cash flow in the quarter alone was $43.97 billion.

The arithmetic is stark: quarterly operating cash flow covered capex more than four times over. The company is generating, in one quarter, roughly what it would cost to build a leading-edge DRAM fab, and it is banking most of it.

That is not a statement about confidence in demand. It is a statement about confidence in the shape of demand. Micron is not convinced that the current pricing environment extends forever, and it is not willing to add the capacity that would end it. CNBC's coverage of the release notes Micron is investing $250 billion to build two new campuses dedicated to HBM, with the larger breaking ground in Clay, New York in January and a first new Boise fab scheduled online next year — but that build is HBM-specific and phased, not a DRAM volume expansion.

Mehrotra's guidance on the call confirms the restraint was intentional rather than a symptom of constraint. Micron expects DRAM bit shipments to grow in the mid-20s percentage range in calendar 2026, then only the low-20s percentage range in calendar 2027 and 2028. Flash runs the opposite way — low-20s percent this year, mid-20s percent after — precisely because flash fabs are cheaper and faster to add than DRAM fabs. The company is telling its customers it will not solve this shortage by flooding it.

Set against that, the pricing leverage is being converted into contracts rather than left to the spot market. Micron has signed 26 supply agreements covering roughly 35 percent of revenue through fiscal 2030, with a stated intention to reach around half of sales. Customer prepayments have reached $32 billion, and remaining performance obligations stand at $150 billion — and that figure counts only deals with fixed price floors and committed volumes. On HBM specifically, Mehrotra said the company has completed agreements for the vast majority of its calendar 2027 HBM bit supply at significant year-over-year price increases, explicitly to narrow the gross-margin gap between HBM and conventional DRAM.

The Yield Wall Nobody Built

The reason flash can be added faster than DRAM is worth stating plainly, because it explains the asymmetry in the whole industry. HBM capacity is getting harder to expand for a structural reason that has nothing to do with equipment supply or construction costs: taller stacks at higher speeds yield worse.

As HBM progresses from HBM3E to HBM4 to HBM4E and HBM5, the die stacks get taller and the data rates go up. Both trends raise the probability that a stack fails its own electrical test and gets scrapped. Every dud is DRAM capacity already spent — wafers pulled, packages built, test time burned — that generates no saleable bits. The industry's yield loss per unit of raw capacity rises with every generation.

This creates a structural asymmetry that management will not spell out as bluntly, but which the numbers imply. Capacity added to flash produces sellable product. Capacity added to HBM consumes input at a rising rate per output bit. So the industry's rational move in a demand-driven price spike is to add flash, add HBM selectively and slowly, and leave conventional server DRAM tight. Which is exactly the allocation pattern in the fiscal 2026 results.

Micron is also working with Nvidia on what the company describes as the industry's first custom HBM implementation — NVHBM, targeted at next-generation GPUs and NVLink Fusion platforms. Custom HBM is a margin structure, not just a product: it lets a memory maker and a logic maker jointly define the interface, which changes who captures the value of a scarce stack. If HBM is the highest-value product in the history of commercial information technology, as the economics increasingly suggest, then the argument over who designs it matters as much as the argument over who fabricates it.

The Pricing Claim Everyone Repeats

There is a widely circulated arithmetic in this space about HBM being more valuable than gold by weight, and it is worth examining rather than repeating at face value. The claim goes roughly: a twelve-high HBM stack weighs about a quarter of a gram and sells for roughly $500 to $600 depending on customer and volume. At that weight, roughly 125 stacks make a troy ounce, putting the implied value at $62,500 to $75,000 per ounce against gold trading in the low four-thousands. That yields a multiple somewhere between 15 and 18 times gold.

The order of magnitude is not crazy, and it is a useful signal about where margin sits. But the inputs are soft in ways that matter. Stack pricing varies by contract and by customer, the quarter-gram figure is an estimate rather than a weighed measurement, and the comparison implicitly treats a semi-finished component that still has to be assembled into a package, tested, and mounted on an interposer as equivalent to a refined commodity. It is a striking framing rather than a valuation. Used as a headline claim about the industry, it outruns what the underlying data supports; used as an indication that scarcity has transferred decisively to the memory layer, it is accurate.

The defensible version of the same point comes straight from the filings. Conventional server DRAM — the least exotic, most substitutable memory in the portfolio — is selling at a 90 percent gross margin. When the most easily replaced product in the stack carries that margin, the constraint is not demand. It is wafers.

What This Means for Everyone Else

For the buyers, the immediate consequence is that the memory shortage that began with AI accelerators is now reaching consumer products, and the pass-through is visible in retail pricing. The substitution effect runs through the whole supply chain: LPDDR5X for flagship phones, SOCAMM server modules for AI servers, high-capacity RDIMMs for general data center work. Micron's own product highlights for the quarter give a sense of where the engineering effort is going — sampling of the industry's first 512GB ultra-dense DDR5 RDIMM at up to 9,200 MT/s, multiple customer qualifications of 8,800 MT/s 1-gamma server modules, and a 6600 ION SSD that completed a strong volume ramp.

For competitors, the constraint is uncomfortable. SK Hynix and Samsung are in the midst of large HBM buildouts at home, and Micron holds the smallest share of the three HBM market despite a market capitalization above $1.2 trillion. The advantage the smaller player now has is that it built the least speculative capacity relative to the demand. A company that has spent its cash on fabs has less pricing power in the next spot market than a company that spent its cash on customer prepayments.

For the Chinese domestic memory effort, the arithmetic is a headwind rather than a reprieve. Micron's chief technology officer has said publicly that the company is two generations ahead of ChangXin Memory Technologies and intends to stay two generations ahead. CXMT's Shanghai listing raised $8.6 billion — real money, and roughly a fifth of a single quarter's operating cash flow at Micron. That is enough to fund a fab and not enough to fund parity, and in a market where the incumbent's pricing power is set by its own restraint rather than by a shortage of Chinese supply, the gap is unlikely to close on capital alone.

The uncomfortable summary is that the industry's most visible supply problem is now a policy, not a bottleneck. Foundries, lithography tools and construction labour were the binding constraints of the last four years. In 2026 the binding constraint is the decision not to build. That decision is rational, defensible and profitable for everyone making it. It is also the kind of decision that only works while a small number of companies hold most of the world's memory capacity and none of them has an incentive to break the agreement first.

Conclusion

Micron's fiscal 2026 was not a quarter that broke the industry. It was a quarter that confirmed the industry had rearranged itself around an unwritten agreement to hold capacity down, and the agreement is working better than any of the formal ones the industry has signed.

The number to watch is not next quarter's revenue. It is capital expenditure relative to operating cash flow. If that ratio stays near four-to-one through 2027, the memory shortage is a durable feature of the AI buildout and every forecast that assumes supply will catch up is working from a broken premise. If the ratio climbs toward one-to-one, the shortage ends — and so does a gross margin that no commodity has ever held.

For everyone downstream, the practical reading is straightforward. Assume memory stays expensive, assume allocation matters more than price for the next several quarters, and assume the bottleneck keeps moving outward from accelerators into the ordinary servers, laptops and phones that everyone else has to buy. The semiconductor coverage on this site tracks the capital and policy decisions behind those constraints rather than the product announcements that sit on top of them.

Images

DDR5 memory modules in their various form factors, shown by SK Hynix at Computex 2025 in Taipei — illustrative of the memory products whose pricing now drives the AI buildout, not Micron hardware. Photo by Wikimedia Commons contributor 4300streetcar, CC BY 4.0.

Caption: SK Hynix DDR5 modules in several form factors, photographed at Computex Taipei on 23 May 2025. SK Hynix is a Micron competitor; the modules are shown as an illustration of the DRAM market in which Micron reported those figures, not as Micron product.

CAMM2 laptop memory modules and an NVMe M.2 SSD on a Teamgroup display at Computex 2025 — the mobile and client side of the same market, where memory price increases are reaching consumer devices. Photo by Wikimedia Commons contributor 4300streetcar, CC BY 4.0.

Caption: Teamgroup CAMM2 laptop memory modules with an NVMe M.2 Gen 5x4 SSD, displayed at Computex Taipei on 23 May 2025. These are a third-party competitor's products, included to illustrate the client-memory side of the same price cycle.

The internals of a disassembled Samsung 860 Pro solid-state drive — a NAND flash product of the type that made up 26 percent of Micron's quarterly revenue. Photo by Wikimedia Commons contributor, CC BY-SA 3.0 DE.

Caption: The interior of a disassembled Samsung 860 Pro 512GB SSD, photographed in 2018. An older consumer drive from a Micron competitor, shown to illustrate the NAND flash side of the market, which was the faster-growing half of Micron's quarter at 6.2-fold year-over-year.

References