SanDisk Corporation

Every photograph on your phone, every file on your laptop and, increasingly, every answer an artificial-intelligence model gives you is stored on NAND flash — the memory that keeps its contents when the power goes off. SanDisk designs that memory, has it manufactured in eight factories in Japan that it co-owns with Kioxia, and sells it as solid-state drives for data centres, as chips inside phones, PCs, cars and games consoles, and under its own famous name as the memory cards and USB sticks in every electronics shop. A ten-year-old could understand it: SanDisk sells storage by the gigabyte, and the price of a gigabyte is set by the whole world's supply and demand.

That last clause is the whole report. In the autumn of 2023 the price of a gigabyte was so low that SanDisk — then still a division of Western Digital — sold its chips for less than they cost to make. In the quarter to 3 July 2026 it sold $8.97 billion of them, up 372% on a year earlier, and kept 84.6% of that as gross profit and $6.9 billion as net profit — in a single quarter, six and a half times what it earned in the whole of fiscal 2022, the best year of the previous cycle. The factories were the same. The price of a gigabyte was not.

Wall Street noticed. A share that changed hands for about $112 at the low of the past year closed at $1,787.69 on 1 October 2026, valuing SanDisk at about $265 billion. On the company's own chart, $100 put into its shares when they began trading in February 2025 was worth $4,847 sixteen months later. This report asks the oldest question in commodity investing in its newest costume: is this a different kind of business now, or the same business at the top of its cycle?

Two lessons for an owner. The first is that this company has been valued at every extreme within two years: at less than its book value in the spring of 2025, so cheaply that it had to write down goodwill, and at about $345bn at the June peak. Nothing about the factories changed between those two prices. When the same assets are priced at both ends of that range in so short a time, the market is not telling you what the business is worth; it is telling you it does not know.

The second is that the spin-off happened at the bottom, and the price at which Western Digital let go of the stake it kept reflects it. WD exchanged 21.3m SanDisk shares in June 2025, mostly to retire its own debt. At today's price those shares are worth about $38bn — more than twice what WD paid for all of SanDisk in 2016. It is a reminder that the best time to own a cyclical is when nobody wants it, and that the people closest to a business are not always the ones best placed to time it.

Read the cost line, not just the revenue line. In the fourth quarter SanDisk's cost of revenue was $1.38bn — almost exactly what it was a year earlier, $1.40bn — on revenue 4.7 times larger. Nothing about how the chips are made changed. What changed is what customers would pay for them. That is the signature of a commodity in a shortage, and it is the same signature, reversed, that produced a negative gross margin in the quarter to September 2023.

The joint venture is the amplifier. SanDisk does not own a fab. All of its flash comes from the Flash Ventures joint ventures with Kioxia, whose eight fabs in Japan make about 80% of the capacity in Kioxia's buildings. SanDisk buys about half the output at cost plus a small mark-up — and pays half the ventures' fixed costs whatever volume it takes. That is why its own balance sheet carries only $674m of property and its capital spending was a mere $177m last year: the heavy assets sit in the ventures, and SanDisk pays for them through the wafer price and through commitments — $6.6bn of them to Flash Ventures and $11.8bn in all at 3 July 2026. It makes the business look capital-light in a boom. In a bust the fixed costs do not go away; in fiscal 2023 Western Digital booked $296m of charges for factory capacity it was paying for and not using.

Buffett set the test for a commodity producer in his 1982 letter: "A few producers in such industries may consistently do well if they have a cost advantage that is both wide and sustainable. By definition such exceptions are few, and, in many industries, are non-existent." SanDisk's cost position is competitive. It is not wide, and half of it belongs to a partner who also competes with it. The 84.6% gross margin is a shortage rent, not the yield of a moat — the proof being that the cost of making the chips did not change between the quarter with a 26% margin and the quarter with 85%.

The same letter carries the line every memory investor should pin above the desk: "What finally determines levels of long-term profitability in such industries is the ratio of supply-tight to supply-ample years. Frequently that ratio is dismal." We score the moat 4: a good business in a bad industry, with the one genuinely new element — the contracts — still untested.

Management knows what every reader of this report is thinking, and it has an answer with a name: the New Business Model, or NBM. Since the start of fiscal 2026 SanDisk has signed long-term agreements with large data-centre and device customers that, in the words of its annual report, "commit us to deliver, and our customers to purchase, a stated volume of products, mostly over multi-year periods." Prices have fixed and variable components. The customers put up money in advance, refundable security deposits, and collateral held at third-party banks. The company says these agreements are expected to become "our predominant way of doing business" and will reduce "certain elements of industry cyclicality." If that is true, the price of the shares is not crazy. If it is not, nothing else in this report matters very much.

This is genuinely new, and it deserves to be taken seriously. Memory makers have signed long-term agreements before, but rarely with this much customer money attached: the trade press reported in January that SanDisk was asking for prepayments of up to 100% for long-term supply, which buyers called unprecedented. When a hyperscaler pays years ahead for gigabytes, it is telling you two things — that it fears not getting them, and that SanDisk had the leverage to ask.

But read the contract as an owner, not as a seller. Three questions decide what it is worth:

First, what happens to the variable part of the price when the spot market falls? It falls with it. Only the floor is protected, and SanDisk has not disclosed where the floors sit relative to today's prices. A floor set when gross margins were 30% protects a very different business from one set at 85%.

Second, will customers honour minimums when they can buy the same gigabyte cheaper elsewhere? In past memory downturns buyers have pushed, often successfully, to defer or renegotiate. The guarantees — $16.5bn against a contracted minimum of $93.9bn, about 18 cents on the dollar — limit SanDisk's loss; they do not make the customer want the product. The company's own risk factors say the guarantees "may not fully offset" lost revenue.

Third, the contracts cover the volume SanDisk sells, not the volume the industry makes. If Samsung, SK hynix, Micron and YMTC all build for the AI boom, the surplus lands on whoever is not under contract — and on the next round of negotiations. Buffett's textile mills had customers too. What they did not have was a way to stop everyone else standing up at the parade.

★ Our reading. The NBMs probably make the next downturn shallower for SanDisk than the last one, and they turn some of the boom's cash into the balance sheet's protection — that is real and we give it real weight. They do not turn NAND into a franchise. They have never been tested by a falling market, and the first test is the one that will tell us what they are worth. Until then we value SanDisk as a cyclical with a cushion, not as a utility with a contract.

The pattern across the rows. Every competitor on this table is reporting record profits — Samsung, SK hynix, Micron and Kioxia all grew NAND revenue by 70% to 99% in a single quarter. That is the most dangerous sentence in this report. Buffett wrote in 1987 that "when shortages exist, however, even commodity businesses flourish", and that "whenever shortages appear, the typical manager simply can't wait to expand capacity and thereby plug the hole through which money is showering upon him." The memory makers say they are disciplined this time, and so far the new spending is going mostly into DRAM and high-bandwidth memory rather than NAND. But TrendForce already expects NAND supply growth to accelerate in 2027 as denser chips and new fabs ramp, with possible oversupply in the second half; Micron says the same; and the one producer whose capacity plans are set partly by a government, YMTC, is building. SanDisk does not control any of those decisions. It controls half of one partner's.

The irony we cannot leave out: SanDisk's co-founder runs Micron; its manufacturing partner is also a competitor; and its partner in the most promising new product, high-bandwidth flash for AI, is SK hynix — the company whose veto in 2023 sent SanDisk out alone. In memory, everyone is everyone else's customer, supplier, partner and rival. It is an industry in which nobody's advantage stays private for long.

Integrity: no veto. The disclosures are unusually candid for a memory maker — the 10-K spells out the joint venture's fixed-cost burden and the limits of the guarantees in plain words. Capital allocation: good on the balance sheet, questionable on the buyback. Repaying debt in a boom is what a cyclical should do. Spending $4.5bn on shares at about $1,600 is what cyclicals usually do instead. A trading plan adopted in May is a reasonable way for an executive to diversify after a fourteen-fold rise, and we do not read it as a signal about the business. We do note that, across more than a hundred insider filings since the spin-off, no officer or director has bought a single share in the open market — not at $28 in April 2025, and not since. And the company bought at about $1,600 in May and June; in September its chief executive sold at about $1,580. A shareholder is entitled to notice which side of that price the person with the most information chose.

Two subtractions before any multiple. The first is the Nanya stake: SanDisk bought $970m of the Taiwanese DRAM maker's shares in March and marked them at $1.78bn in July, and most of that $0.8bn gain sits in the fourth quarter's GAAP profit — which is why GAAP earnings per share ($43.97) are above the company's own adjusted figure ($39.25). We use the adjusted figure. The second is cash: of the $11.5bn of reported free cash flow, about $2.5bn was customers' prepayments and deposits under the new contracts — real money, but money SanDisk owes back in product or, for deposits, in cash. The company's own "adjusted free cash flow" is $8.7bn, and that is the number we use.

Why a low P/E is the wrong signal here. In a steady business a P/E of eight is cheap. In a cyclical business at the top of its cycle, a P/E of eight is what the market charges for earnings it expects to shrink — and historically memory shares have looked cheapest on earnings precisely when it was most dangerous to buy them, and most expensive (or loss-making) when it was best. The honest way to value SanDisk is the way Graham taught for any cyclical: on average earnings across a full cycle, not on the earnings of the best year in it.

We do not know what a full cycle looks like with the new contracts in place, so we sketch one and show our working. Take fiscal 2022 to 2026 as reported, add the two boom years analysts expect, and assume two leaner years after them at about the level of fiscal 2022, cushioned by the contracts. That averages roughly $60 a share. A cyclical with a clean balance sheet deserves perhaps 15 times average earnings, which gives about $900; add some $42 a share of net cash and the Nanya stake and you are a little above it. The shares are twice that. To justify $1,788 you need to believe either that the boom lasts well beyond fiscal 2028, or that the contracts have abolished the downturn. Either may be true. Neither can be known today.