Sandisk

IThe business

Sandisk makes NAND flash, the memory that keeps data when the power is off. It designs the chips and controllers, writes the firmware, builds SSDs, memory cards and USB drives, and sells them to the makers of servers, PCs, phones and game consoles, to clouds, and through shops under the Sandisk brand.² It was part of Western Digital until 21 February 2025, when Western Digital spun it off to its own shareholders.² About 11,100 people work there, three quarters of them in Asia.²

One fact shapes the rest of this page. Sandisk does not own the factories that make its flash. Every wafer comes from three joint ventures with Kioxia, called Flash Ventures, in which Sandisk holds 49.9%. Kioxia owns the buildings and runs them, the ventures own or lease the tools, and each partner buys about half the output at cost plus a small markup.² Sandisk pays for half of the ventures’ fixed costs whether or not it takes the wafers, and it may not make flash anywhere else.² So the company is a design, product and sales business sitting on a share of someone else’s fabs. Its cost of goods sold was $1.65B in the latest quarter against $1.62B a year earlier, while revenue went up 4.8 times C (calculated).³ The factories did not get cheaper. The prices got higher.

Sandisk’s fiscal year ends on the Friday nearest 30 June. FY2026 ended on 3 July 2026 and had 53 weeks. Revenue was $20.2B, up 175%. Datacenter was $5.2B, up 437%; Edge, which covers PCs, phones, consoles and cars, was $12.2B, up 195%; and Consumer, the retail brand, was $2.9B, up 29%.² Net income was $11.4B, or 56% of revenue.²

The filing says how it happened, and it is one line. Total exabytes sold rose in the mid-teens percent. Datacenter revenue per gigabyte rose almost 150%, Edge almost 180% and Consumer in the low fifties.² So the volume grew at a normal pace and the price per gigabyte did nearly all the work. Revenue rose by $12.9B and gross profit by $12.3B C (calculated). Almost every extra dollar fell through to profit, because the cost of making the flash barely moved. That is the most important sentence about this company, and it also works in reverse.

The latest quarter, which ended on 2 October 2026, went further. Revenue was $11.0B, up 23% from the quarter before and 378% from a year earlier; gross margin was 85.0%; net income was $8.1B, including a $0.5B gain on shares of Nanya Technology that Sandisk holds.³⁵ Exabytes sold were up only about 10% on the year. Revenue per gigabyte was up 320% in Datacenter and more than 450% in Edge.⁵ The guide for the next quarter is $12.2B to $12.6B at a margin of about 85%, with about 151M diluted shares.³

Why storage became scarce

The company’s account is that AI inference moves a great deal of data in and out of memory, and that the cache of past computing, called the KV cache, is pushing data from expensive memory onto flash. At its August investor event Sandisk put the market for data-center flash at 1.2 zettabytes by 2030 and called AI data centers “dramatically more storage intensive.”⁸ Datacenter was 25% of revenue in FY2026 and 44% in the latest quarter, $4.8B, up from $0.3B in the quarter a year before C (calculated).²⁵ Micron names replacing hard disks as one of the uses, and Sandisk says it expects AI demand to last through calendar 2027 and beyond.¹⁷⁵

The supply side matters as much. Nobody builds NAND capacity quickly. TrendForce says suppliers are putting most new capital into DRAM and high-bandwidth memory, which limits new NAND capacity, and it expects selling prices to keep rising in the third quarter.⁶ Micron expects industry NAND bit shipments to grow in the low 20s percent in 2026 and the mid-20s in 2027 and 2028, with the industry short of supply in both years, and its new Singapore NAND plant does not start until the second half of 2028.¹⁷ Kioxia, Sandisk’s partner, said in its April to June quarter that demand would keep outrunning supply in 2027. TrendForce, reporting on that call, expects supply to loosen in the second half of 2027.¹⁸ Those are company claims and one analyst’s view. They agree on this year and 2027 and differ on what comes after.

What it sells, and to whom

No customer reached 10% of revenue in FY2026, FY2025 or FY2024.² That sounds comforting and is less so than it reads. Of FY2026 revenue, $4.5B came from China and $5.1B from Hong Kong, together 47.6%, and the United States was $3.7B, or 18% C (calculated).² Much of Asia’s share is device makers and distributors who build for the world. Before the spin-off, the Form 10 showed one customer at 15% of revenue in FY2023 and the top ten at 47%.¹⁴

The new thing is how Sandisk sells. Starting in the second half of FY2026 it signed what it calls New Business Model agreements, or NBMs, with datacenter and edge customers. Each commits Sandisk to deliver and the customer to buy a stated volume, mostly over several years, at a price with a fixed part and a variable part.² By the fourth-quarter call there were ten, with eight customers, worth a minimum of $93.9B at floor prices, lasting up to five years and more than four on average, and backed by $16.5B of financial guarantees, according to the chief financial officer as reported at second hand.⁷ The 10-Q puts the contracted minimum, the remaining performance obligation, at $88.9B on 2 October, with about 20% of it due within twelve months.⁵ Management says the agreements cover about 50% of FY2027 bits and about two thirds of FY2028 bits.⁸ I come back to what that is worth in the Valuation.

The point is the size of the guarantees against the size of the floors. $12.8B of collateral and $1.6B of deposits cover about 16% of the $88.9B minimum C (calculated), and the company says the guarantees are meant to offset a portion of lost revenue and may not offset all of it.⁵² The filings name neither the customers nor the banks.

The size of the market

TrendForce put the top five NAND makers’ revenue at $68.9B in the second calendar quarter of 2026, up 77% on the quarter. Samsung was $23.1B, SK hynix group $14.3B, Micron $11.9B, Kioxia $10.7B and Sandisk $9.0B.⁶ That makes Sandisk about 13% of the top five by revenue C (calculated), and the slowest grower of the five in the quarter, because TrendForce says its bit shipments were conservative.⁶ The same five were selling at an annualized $275B C (calculated). The question under every number on this page is what happens when the price falls toward what it costs to make.

IIThe moat

A durable business keeps competitors from taking its profits, and the test is the record. For Sandisk the record is short and in two parts. As a Western Digital division and then a standalone company, it made 33.3%, 7.1%, 16.1% and 30.1% gross margins in FY2022 to FY2025, and then 71.5%.¹⁴⁴² Before that, the old SanDisk, which made flash in the same Toshiba ventures and was bought by Western Digital in 2016, earned 1.9% in the crash year of 2008 and then between 33.3% and 46.9% in each year from 2009 to 2015, 39.0% weighted by revenue over the eight years C (calculated).¹⁰

Read the chart for what it says about pricing power. In thirteen years only one year was above 47%, and it is this one. In FY2023 Sandisk lost $2.1B, and the Form 10 explains that revenue fell 38%, nearly all from a 39% fall in the price per gigabyte.¹⁴ That is what NAND does. The same measure has since risen several times over, and a company that can only choose its volume has no say over either.

What does exist is a small group. TrendForce had Samsung at 29.3% of second-quarter NAND revenue and Kioxia at 13.6%; on its revenue figures SK hynix is 20.7%, Micron 17.2% and Sandisk 13.0% of the top five C (calculated).⁶ Entering means years of process work and tens of billions of dollars. In China, YMTC is one of the competitors Sandisk names.² A concentrated industry can hold prices while supply is short. It cannot stop its members from all adding capacity at once.

Evidence it is widening

The strongest new fact is the contract. Eight customers have signed minimums of tens of billions of dollars for up to five years, and banks hold $12.8B of collateral against them.⁵⁷ A year before, Sandisk had $126M of customer deposits. The CFO also said the company expects attractive margins even at floor prices.⁷ I can’t test that, since the contract terms are not public, but the arithmetic is that the minimum of $93.9B over a little more than four years is about $21B a year, which on its own is as large as the whole of FY2026 revenue C (calculated). The second fact is that Kioxia says it is on track to cover 50% of its calendar 2028 volume with long-term agreements as well, which suggests buyers are accepting this model across the industry.¹⁸

Evidence it is narrowing

Two things. First, the ventures are not Sandisk’s alone, and the partner makes the other half of the flash and sells it against Sandisk. Kioxia reported an 80% gross margin in the April to June quarter, close to Sandisk’s, and plans about ¥470B a year of capital spending to FY2028.¹⁸ Second, TrendForce says suppliers are putting capital into DRAM first, which helps NAND prices for now. The 10-K itself warns that sustained high prices may bring new entrants and that customers may find other technology or use less flash.² Margins this high teach buyers to look for another way out. I counted no sign of that yet in the numbers; the first would be falling exabytes in Edge, and Edge exabytes are already down about 20% on the year.⁵

My judgement: the structure of NAND is better than it was, and the contracts are real. Whether Sandisk has an advantage of its own beyond a favourable year is something I can’t yet say. One round of contracts meeting a falling market will tell us more than any argument.

IIIThe competitive advantage

Sandisk is the smallest of the five NAND makers by revenue and the least integrated. The table compares the latest quarter of each listed memory maker. The quarters end between one and three months apart, and prices were rising, so those who report later have an edge.

Sandisk trades at about 4.9 times its latest quarter’s operating income, annualized. That sits between SK hynix and Micron. The three look alike in one respect, a stock priced at a few times the profit of one quarter, which is how cyclical stocks look at a top. I would not read it as cheap.

Where Sandisk wins

Capital first. Because the ventures own the tools and Sandisk pays for the wafers it takes, its own purchases of plant were $0.18B in FY2026, 0.9% of revenue, and $0.05B in the latest quarter.²³ A company that has not spent on plant has more of its profit to keep. Contracts second: it has them in place for about half of FY2027 bits, while Kioxia says it is on track to cover 50% of its calendar 2028 volume with key customers and Micron has signed 26 agreements of its own.⁸¹⁸¹⁷ Third, it has the product line the AI buyers want most, enterprise SSDs, where Datacenter revenue rose from $0.3B to $4.8B in the quarter C (calculated),⁵ and it has the retail brand, which gives it a window into the end market that its partner lacks.

Where it loses

Scale and control. Samsung and SK hynix have their own fabs, their own DRAM and HBM businesses, and far larger balance sheets. Sandisk’s output is capped by what the ventures can make, and half of that goes to a competitor. It can choose how fast to expand only with Kioxia’s agreement, and the joint ventures end on 31 December 2034, in the middle of my forecast window, unless extended again.² The last extension, in January 2026, cost Sandisk $1.2B paid directly to Kioxia over 2026 to 2029.² And Sandisk’s revenue growth was the slowest of the five in the quarter, which TrendForce puts down to conservative bit shipments; my reading is that it is choosing price over volume.⁶ Management says it shifted shipments to higher-value customers.¹⁵ That works in a shortage. It leaves share behind if the shortage ends.

IVManagement

David Goeckeler, 64, has been chief executive since the spin-off in February 2025 and is also chairman. Before that he led Western Digital, and earlier worked at Cisco.¹² Luis Visoso, 57, the chief financial officer, came from Western Digital and before that was finance chief at Unity and Palo Alto Networks and an executive at Amazon Web Services.¹² The board has eight nominees, seven of them independent, with Richard Cassidy as lead independent director.¹²

How they are paid, and what that pays them to do

Follow the incentive first. Goeckeler’s FY2026 pay was $15.2M: a $1.35M salary, $4.7M of bonus and $9.0M of stock at grant-date value. That looks modest for a company with a $174B market value, and it is, because most of his gain is in older awards.¹²

The bonus is half profit, a quarter free cash flow and a quarter strategy. The profit target was non-GAAP operating income of $1.16B, with the maximum payout at $1.51B. Actual was $12.7B. The cash flow target was $90M, the maximum $240M, and the actual $11.2B. Both paid out at the 200% cap.¹² The targets were set from the operating plan at the start of FY2026, when the quarterly revenue was $2.3B. I do not blame anyone for that. But a target that is one eleventh of the result did not ask anyone to do anything differently, and I found no sign that the committee rescaled it.¹² The cash flow measure used for the bonus, $11.2B, is also higher than the $8.7B of adjusted free cash flow in the same proxy’s reconciliation, which takes out $2.5B of customer deposits and prepayments under the contracts C (calculated).¹²

The stock matters more. Under the longer plan three quarters of the chief executive’s stock is performance units. They turn on revenue and non-GAAP earnings per share, measured year by year and averaged over three years, and the first year also paid at the 200% maximum.¹² From FY2027 awards a relative share-return modifier of plus or minus 20% against the semiconductor index has been added.¹² Then there are the launch grants made at the spin-off, which were all performance-based and needed the share price to rise 125% above the post-separation baseline of $47.07, to $105.91, for the maximum 300% payout. The price crossed it in FY2026, and by 3 July it was $1,745.¹² The grants now vest on time in February 2028. For Goeckeler the target was 317,965 units and the maximum 953,895, which is about $1.1B at today’s price C (calculated).¹² On the proxy’s own table of what would accelerate at 3 July prices he held about 1.56M unvested units, around 1% of the company and $1.8B at $1,150.89 C (calculated).¹²

What does that pay them to do? Raise revenue and earnings per share, and keep the stock up through early 2028. Both are sensible at this point. Neither pays for what decides this business over a decade, which is return on capital through a price cycle. I found no such measure in the plan, and the new modifier adds relative share price over a few years, which a boom can fill. The strongest argument for management is the holding. Its weakest is that the bonus pays at the cap whenever prices are high.

How much of their own money is in

After the Form 4 of 17 September 2026 Goeckeler held 382,865 shares directly, about $441M at the close C (calculated).¹¹ He sold 67,679 shares that month for about $105M, at an average of about $1,551 C (calculated).¹¹ All seven insiders who sold since the spin-off took in about $134M between February 2025 and October 2026, and I found no open-market purchase by anyone C (calculated).¹¹ I do not read the sales as a verdict, because the shares went up 24 times and the grants are large. But the people closest to the cycle have been sellers, and what holds the interest is the unvested stock, not the shares sold.

What they did with capital in the first full year

Sandisk has been standalone for twenty months, so there is no full cycle to judge. What there is, is a first year, and I can say whether it was well spent.

Goodwill of $5.0B remains on the balance sheet after the two write-downs.² Western Digital still held a small number of shares at the end of the last fiscal year and has said it expects to dispose of all of them by the end of 2026.² I mention it because an overhang on the shares is small now and tends to disappear into the price.

VFinancials

Ten years of Sandisk do not exist. The business was a division of Western Digital until 21 February 2025, and the statements before that were carved out of Western Digital’s books, with costs shared with its other businesses allocated by formula.² The Form 10 gives FY2022; the FY2025 and FY2026 10-Ks give the rest.¹⁴⁴ So I have five fiscal years and one quarter, only one of them a full standalone year. There was a $0.67B goodwill impairment in FY2023 and a $1.83B one in FY2025.⁴ Anyone who tells you what Sandisk earns over a cycle from this record is guessing, and I am guessing too, from a wider base. In the Moat I used the old SanDisk’s eight years for that reason.

Five years still contain a full downturn and the beginning of a boom. Revenue fell 38% in FY2023, rose 9% in FY2024 and 10% in FY2025, and then 175% in FY2026 C (calculated).¹⁴⁴² In FY2023 the company lost $2.1B on $6.1B of revenue.

One date trap. FY2026 had 53 weeks, and the extra week fell in the first quarter, which ran 14 weeks.² The second trap is the share count. Sandisk reports about 151M diluted shares in its second-quarter guide, 142.3M were outstanding on 26 October, and the 8-for-1 split on 20 November will turn the 151M into about 1.2B. Every per-share figure on this page is before the split.³⁵⁹

What an owner could really take out

Reported earnings flatter this business, because a good deal of the cash that came in is not the owners’. In FY2026 operating cash flow was $11.7B. Of that, $3.96B was money that other people still have a claim on: $1.37B of refundable deposits from customers, $1.22B of advances for product not yet shipped, and $1.37B of tax not yet paid.² Take those off, along with the plant and the loans to Flash Ventures and the stock pay, and owner cash was about $7.0B, 61% of net income C (calculated). In the latest quarter it was $5.3B against $8.1B, 65%.

The last two lines carry the point I most want you to see. Stock pay was expensed at $0.23B in FY2026 and $0.09B in the latest quarter. But the shares that vested were worth $1.81B on the days they vested in FY2026 and $2.70B in the latest quarter alone.²⁵ The difference is that most of the awards were granted when the shares were worth $54 on average, and the cost of an award in the accounts stays at its grant-date value. The company paid $0.63B and $0.76B in cash for the taxes on vesting shares in the same periods, and it withheld about 1.0M shares in FY2026.²⁵ An owner cannot ignore that. The shares given to employees are real dilution, valued at today’s price, and about 5.7M units were still unvested at 2 October, 4% of the company.⁵ New grants of 0.4M units a quarter at about $1,650 are expensed at a far higher figure than the old ones.⁵ In my model I count stock pay at 4% of revenue, not at the $0.36B annual rate the latest quarter shows, and I use the 151M diluted shares, which already include the unvested awards.

What it takes to stand still

Sandisk’s own plant spending was $0.18B in FY2026 and $0.05B in the latest quarter, tiny for a $20B business. The reason is the joint ventures. Their tools are owned or leased by Flash Ventures, Sandisk pays half the fixed costs including depreciation reimbursement in the cost of the wafers it takes, and it guarantees half of the ventures’ leases: $923M, or ¥149B, outstanding on 3 July.² So most of the capital a NAND maker needs shows up in cost of revenue, which was $5.8B, and not in capital spending. I count it once, in cost of revenue. What sits outside is the loans Sandisk makes to the ventures to buy tools, $0.28B net in FY2026 and $0.15B in the quarter, prepayments of building depreciation, $0.84B still to be credited, and the new fixed payments to Kioxia of $1.2B over 2026 to 2029.²⁵ Flash Ventures-related commitments were $7.1B on 2 October, $2.5B in the next twelve months, and most of that is depreciation that will pass through cost of revenue.⁵ I put capital outside cost of revenue at $1.0B to $2.5B a year in the base, more than in the past because K2 started output in FY2026, and because Kioxia plans about ¥470B a year of its own spending to FY2028.²¹⁸ I can’t narrow that honestly E (estimate).

Quality of earnings

Three things to know. The first is price protection. Sales incentives and price protection were 11% of gross revenue in FY2026, down from 19%, and 9% in the latest quarter from 16%.²⁵ That is how a shortage looks from the inside. If the shortage ends, it will come back. The second is inventory, with 178 days of inventory in FY2026 against 135 the year before, and $3.0B of inventory on 2 October.²⁵ Receivables were $6.6B, up $1.9B in the quarter, after 48 days of sales in FY2026.²⁵ The auditor’s one critical matter was whether it had enough evidence for revenue, since it is recorded through highly automated systems.² The third is the tax. The rate was 12% in FY2026 and 13% in the latest quarter, helped by Malaysian tax holidays that expire between 2028 and 2031, and the company expects to pay the 15% minimum tax from FY2027.² I use 15% rising to 18% E (estimate).

The balance sheet

On 2 October Sandisk had $5.1B of cash, $2.3B in Nanya shares, and no debt.⁵ I don’t count all of that as mine. The company owes customers $1.6B in deposits and has taken $1.5B in advances, and it owes $1.9B in income tax.⁵ Taking those off leaves $2.4B of net cash C (calculated). That is small, because the company spent $5.0B on its own shares in the quarter and $0.76B on tax on vested stock. Cash was $4.8B at 3 July.⁵ The balance sheet is safe. It is not the cushion the profit might suggest.

VIValuation

First, how this goes badly wrong

Before any numbers, I asked what would most likely ruin this investment, and worked back. Three things can.

• The price falls and the floors give way. Almost all of today’s profit is price. If NAND prices fall by two thirds, the cost of making flash does not, and the loss falls on a company that pays half of fixed costs whether or not it sells the wafers.² In FY2023 revenue fell 38% and the company lost $2.1B.¹⁴ The contracts are meant to stop this, but their floors have never met a fall, their buyers are eight companies, and the filings say the guarantees may not make up the lost revenue.²⁵ I know of no memory contract of this size that has been tested by a downturn. If customers renegotiate when spot prices undercut a floor, what looks like $88.9B of backlog is worth less than it says.
• Everyone’s capacity arrives together. Micron’s Singapore NAND plant starts in the second half of 2028, Kioxia’s capital spending is rising to about ¥470B a year, and TrendForce expects looser supply in the second half of 2027.¹⁷¹⁸ Each is rational alone. Together they are how NAND downturns start. The demand may also have been partly fear of running out: buyers who are short order extra, and the 10-K says demand is hard to forecast, that many OEM customers give periodic forecasts and not firm orders, and that its products are largely interchangeable with rivals’.²
• The supply depends on one partner, until 2034. Sandisk gets all of its wafers from ventures it does not control, and may not make flash anywhere else.² The ventures end on 31 December 2034 unless extended, which is inside my forecast, and the last extension cost $1.2B.² Kioxia is both the partner and the nearest competitor. Eight fabs in two Japanese sites are one earthquake, one contamination event or one dispute from a serious problem.² I charge a renewal cost in every case and I assume the ventures continue. If they do not, the value after FY2035 is not covered by anything on this page.

None of these needs a surprise. All have happened in this industry or are in the company’s own risk factors. They set my bear case, and they are the reason I use a margin that falls to a normal level at all. China and Hong Kong, at 47.6% of revenue, is a fourth that sits outside the model.²

What the price implies

At $1,150.89 the market values Sandisk at about $174B on 151M diluted shares, $164B on the 142.3M outstanding, and 9.6 times book equity C (calculated).¹⁵³ On the face of it that is very cheap: 5.5 times the latest quarter’s GAAP earnings per share annualized, 5.5 times Nasdaq’s FY2027 mean of $210.92, and 4.5 times its FY2028 mean of $254.62, which is higher again and comes from only six estimates C (calculated).¹³³ So the Street itself does not draw a peak in FY2028. The market multiple says the earnings will not last. Low price-earnings multiples at the top are how cyclical stocks usually look, and the stock fell 13.5% on 30 October, the day after results that beat the top of the company’s guide.¹³ I do not know why, and I will not guess.

A single growth rate from a peak year means nothing here, so I turn the reverse DCF around. I take my base cash through FY2030 and ask what the business must earn after that. Those four years of cash, discounted, are $70B. The price pays for $102B more. That needs $15B of owner cash a year from FY2031, growing 3%, for good C (calculated). For comparison, the most operating cash flow in the five years before FY2026 was $1.2B, in FY2022.¹⁴²

Put another way, the price needs a gross margin near 74% held in every year from FY2029 to FY2036, with revenue staying near $35B. That is close to the roughly 80% that management targets for FY2028 to FY2030, held for a decade.⁸ The best full year in thirteen years of the old and new companies before FY2026 was 46.9%. Or the price needs about 6 more years at the FY2028 level than I allow, which is a long time for a NAND peak to last. So the market is betting that the company’s model is mostly right, not that its profit is cheap.

What I can’t judge

I can’t tell you Sandisk’s normal gross margin after this. Nobody can. The contracts are new, their terms are private, the NAND industry has never had AI in it, and the standalone record is five years of statements. What I can show is how little the answer depends on the far future. Because so much of the value falls in the next four years, moving the settled margin after FY2033 anywhere from 30% to 60% moves my base only from $516 to $735. What matters is how long the high prices last.

The contracts, honestly

The contracts are the main reason my base does not go lower, and the main thing I cannot see. The company says the minimum at floor pricing is $93.9B over contracts of up to five years, more than four on average, and the 10-Q says $88.9B of fixed amounts remained on 2 October, of which about 20%, or $17.8B, falls in the next twelve months.⁷⁵ Spread over the remaining three to four years, that is about $18B to $24B a year E (estimate). I use $21B for the table below.

Two readings. In the base, the floors are real, the customers pay them, and the floors are close to the revenue the company would earn anyway at a normal margin, so the contracts mostly delay the fall and do not prevent it. That is why my base falls anyway. In the bear, customers do not pay: they walk, or pay less and call it a renegotiation. The company holds $1.6B of deposits, $12.8B of bank collateral and says it has $16.5B of guarantees, which would cover roughly 55% to 63% of the $26B or so by which my bear revenue falls short of the floor over FY2029 to FY2031 E (estimate).⁵⁷ The CFO said the agreements earn attractive margins even at floor pricing.⁷ If I take that literally, on the contracted bits, the floors alone earn well above my 45% normal margin. But the floors run out when the agreements do, and renewals would be priced at prevailing prices. A seller who needs the next contract may take less.

The discount rate

I use 13.0%. The capital asset pricing model starts with the 4.94% ten-year Treasury yield in Q4 2026 and Damodaran’s 4.20% implied equity risk premium of 1 October.²⁵²⁶ I can’t get a beta I trust. The shares have traded for 20 months, and in that time they rose 24 times. A weekly beta against the S&P 500 comes out at 3.0 and a monthly one at 5.7, which would give 17.5% and 29%, and both are the run, not the risk C (calculated).¹²⁷ I take 13.0%, which is the 12.0% on my Micron page plus a point for a smaller company with a single source of wafers and no track record of its own. Value moves little with the rate here, as the table shows, because most of it comes soon.

Three cases

I keep the model simple. I grow bits, take cost per bit down 10% a year in the base, and then set one number a year: gross margin, or revenue for the first two years where the company’s guide and the street give me something to work from. The first quarter’s $11.0B and the $12.2B to $12.6B guide for the second give about $50B for FY2027 in my base, at a gross margin of 84.9%.³ Operating costs use the company’s non-GAAP figures, $2.2B in FY2027, and grow 3% a year. Tax is 15% in FY2027, 16% in FY2028 and FY2029 and 18% after, as the Malaysian holidays end.² Working capital is 18% of revenue, which is what it was on 2 October. Capital outside cost of revenue is loans to the ventures and own plant, $1.0B to $2.5B a year in the base. Stock pay is the larger of $1.0B and 4% of revenue, in place of the expensed figure. I charge $0.6B a year from FY2032 to Kioxia for renewing the ventures, in the base, and the terminal value leaves that out. Each year’s cash arrives at the end of the fiscal year, FY2027 only for the part after 2 October, and I take off the deposits, advances and unpaid tax from net cash.

Why does base revenue fall 61% from peak to FY2031 when the worst past fall was 38%? Because the peak is far above cost. The cost of making the flash is about $8B a year in my base, and revenue is $54B. Prices per bit in FY2028 would be more than four times FY2025’s on my numbers. To reach even a 57% gross margin, price per bit has to fall by about 75% from its FY2028 level, and I put that at three years because the floors slow it. Revenue of $21B in FY2031 is still 2.9 times FY2025 and about the size of FY2026. The base then flattens near $17B to $18B, as the price per bit keeps falling at about the rate of cost. At 45% I assume a margin the old SanDisk beat in only three of its eight years, and the allowance for contracts is what lifts it that high.

Other uses of the same dollar

Price only means something next to what else the money could do. The ten-year Treasury pays 4.94%. My base here returns −5.8% a year, so about ten points below bonds, with the bulk of the cash arriving in the first two years E (estimate).²⁵ The closest comparison on this site is Micron, which is mostly the same bet on memory prices: 7.1% a year on my base there, −3.4% on my bear and 22.2% on my bull. Sandisk’s spread is wider at both ends: −23.7% on the bear and 25.0% on the bull, and its price sits further above my base than Micron’s does, 96% against 26%. TSMC sells the same AI demand with far steadier prices: 9.4% on my base and 8.4% on my bear. NVIDIA is 9.7% and 5.5%, and Broadcom 7.5% and −4.8%. Those pages use their own prices and rates. On my numbers, Sandisk is the least attractive of these on the base and the most on the bull. Within NAND, Kioxia is the cleanest alternative, with the same fabs and the same wafers. Kioxia split its shares 3-for-1 on 1 October 2026; it had 548.0M issued on 30 June, about 1.64B after the split.³¹ I have not pulled its yen price, so I have not computed its multiple. Owning Sandisk and Micron together is not two bets. It is one, twice.

What past lows looked like

Sandisk has been listed for twenty months, so there is no earlier cycle in its own price to read. What there is, is a run: $47.15 at the end of the first fiscal year, $1,745 at the end of the second, $2,335 at the peak on 25 June, and $1,150.89 now, 51% below the peak.¹¹² The stock last closed below my margin-of-safety price of $388 on 14 January 2026.¹ The price has fallen by half in four months, after a year of 24 times. I would not call that a floor.

The margin of safety

• Price$1,150.89Close, Q4 2026 [1]
• Base value$588My estimate E (estimate)
• Margin of safety below$388set from my bear case C (calculated)

I set the margin of safety from the bear case instead of as a discount to base value, because one cycle drives this business. Below $388, 34% under my base value, my $291 bear value is at least three quarters of the price paid, and my base case would return about 34% a year and my bear case −1.8% E (estimate). That is 1.7 times my FY2027 earnings, a number that says more about how much of today’s profit is a windfall than about the stock. The price sits 197% above it, and there is no margin of safety at $1,150.89. The cushion between bear and $388 is for what the model doesn’t hold: a break in the ventures, a change in export rules, and my own error. It doesn’t fully cover the stress case: a repeat of FY2023 at this scale gives $182, which is 47% of $388.

Two different questions

How good is the business, and how good is the price? They have different answers. Today’s Sandisk earns an 85% gross margin, owes almost nothing, and has contracts of a kind the NAND industry did not have before this year. Over its history, though, it is a price-taking seller of a commodity with no wafers of its own. So I would call it a much-improved business of unproven durability. The price, at 96% above my base, already pays for that improvement to last about as long as management says, and then some. A low multiple of current earnings does not change that, because current earnings are the thing in question.

Where this could be in 5 and 10 years

For the base to be right at 5 years, the profit spike has to end the way I model it. Revenue peaks at $54B in FY2028, falls to $21B by FY2031, and gross margin drops from 85% to 57% E (estimate). The base is $156 by then because the richest cash years have been paid out and are gone: the cash of FY2027 to FY2031 comes to $626 a share. For it to be right at 10 years, margin has to settle near 45% and revenue stay near $17B to $18B, with owner cash near $2B a year. That gives $177 E (estimate).

Bear and bull differ on how long prices stay high. In the bear, customers renegotiate and revenue falls to $11B in FY2030, so the business earns almost nothing for owners and value stays near $51. In the bull, margin stays above 65% and revenue above $52B, which is the company’s own model made to last, and that leaves $2,406 in year 10. Growth, margin, capital needed and the 13% discount rate drive all three, and growth and margin matter most.

The most likely way my base is wrong is margin and its timing. Five points more in every year after FY2028 lifts year 10 to $280 and five points less cuts it to $92 E (estimate). The base also assumes that Flash Ventures continues beyond 31 December 2034 on terms like today’s; without it, year 10 has nothing under it.

These are estimates of value, not forecasts of the share price. The market price can sit above or below value for years.

VIIThe case for the price

What most people miss

Three things cut in Sandisk’s favour and don’t make the headlines. Cost of revenue was $1.65B in the latest quarter against $1.62B a year earlier, while revenue went up 4.8 times, so almost none of the boom was spent.³ Banks now hold $12.8B of collateral against customer contracts, up from $5.0B three months earlier, on top of $1.6B of cash deposits.⁵² And Sandisk has no debt, used $0.18B of cash on plant in FY2026, and has a $1.5B credit line it has not drawn.²⁵ If the contracts hold, the firm is a design and sales business with a tool bill that someone else is mostly paying.

The bull case

In plain words: AI inference has turned flash from a commodity into a constraint, and Sandisk has locked in its customers for four years or more. Management’s numbers for FY2028 to FY2030 are revenue growth in the mid-to-high teens, about 80% non-GAAP gross margin, about 75% operating margin, and about 50% of revenue as adjusted free cash flow, with 100% of excess cash returned.⁸ The chief financial officer said on the fourth-quarter call that the minimum under the contracts is $93.9B at floor pricing.⁷ At the latest results the chief executive said the company is delivering growing and durable free cash flow.³ Micron says NAND supply will stay short through 2028, and Kioxia says demand will outrun supply in 2027.¹⁷¹⁸

Holdings show who else came in, and only that. At 30 June 2026, the latest 13F date, Situational Awareness LP, the fund run by Leopold Aschenbrenner, reported 2,495,344 Sandisk shares worth $5.67B, 28.0% of its reported total, and 4,828,786 Micron shares, 27.5% of it.²⁸ A 13F lists long positions in US-listed securities on one date. It does not show short positions or holdings abroad, it is more than four months old, and it does not say why the fund held what it did or what it has done since. Index managers hold most of the rest: Vanguard 7.6%, BlackRock 6.1%, FMR 5.4% and Jane Street 5.1% at 4 September, per the proxy.¹² I have found no named investor whose published view of Sandisk I could open, so I attribute none.

Against reality

Today’s numbers support the first half. The latest quarter earned an 85.0% gross margin, the guide is for about the same, and my base has Sandisk making about $55B of owner cash in the rest of FY2027 and FY2028, 31% of the market value E (estimate).³ So near-term delivery is not where the argument is.

The argument is what comes after, and that needs several things at once. If the floors are paid, if the contracts renew at good prices, and if new supply meets demand instead of swamping it, a margin near 74% is possible, and the price with it. Today Sandisk has about three quarters of experience with the contracts, and no downturn to test them.

The bull case needs more, and it helps to multiply it out. Say the floors are paid in full and renewed at good prices, 65%. Rivals’ new supply, from Kioxia, Micron, Samsung, SK hynix and YMTC, does not cause a price slump by 2029, 45%. AI demand keeps pushing exabytes up in the mid-teens to 2030, 65%. Sandisk keeps its half of the ventures on fair terms past 2034, 85%. And buyers do not shift to other storage or design in less flash, 75%. Those are my guesses, and they aren’t independent, since good AI demand helps most of them. Multiplied, they give 12% E (estimate). Allowing for the overlap, perhaps 20% to 25%, which is the weight I give the bull. At the 25%, 50%, 25% weights my cases give $875. To get to $1,150.89 with the base at $588 and the rest of the weight on the bull, the bull needs 39% odds. Each step looks likely. All five together do not.

What a 10x would take. Ten times $1,150.89 is $11,509, about $1.74T of market value on 151M shares. The largest company I cover, NVIDIA, is about $5T.²⁹³⁰ At 20 times earnings Sandisk would need about $87B of net income, 7.6 times FY2026 and 2.7 times the latest annualized quarter C (calculated). At a 70% net margin, above anything but this year, that is about $124B of revenue, 6.2 times FY2026 and 2.8 times the latest annualized quarter C (calculated).³ At today’s 13% of the top-five NAND revenue, the five makers would need to sell about $953B a year, 3.5 times the $275B annualized rate of the second quarter.⁶ By FY2031 that means growth of 58% a year in the market value from $174B; by FY2036 it means 26% a year. The stock has already risen 24 times in sixteen months from $47.15, but that began at about $7B of value.¹ The best run I know of from this size is NVIDIA’s: about five times from its $40.11 close of May 2023, near $1T, in about three and a half years.²⁹ Sandisk would need twice that, in an industry whose five largest makers sell $275B a year. From here, tenfold is close to impossible, and I would not build a case on it.

Where I come down

The contracts moved me: without them my normal margin would be nearer 30% than 45%. The record quarters did not, because one good stretch can’t prove a change in kind, and the stock’s fall of 13.5% on its best-ever results is a reminder that record results alone have not held the price. A first downturn in which floors are paid and margins stay above 60% would raise my values most.

VIIIRisks, and what would change my view

The three ways this goes badly wrong are at the start of Valuation. These are the rest.

• The cycle itself. The same chip fell 39% in price per gigabyte in one year, FY2023, and Sandisk lost $2.1B.¹⁴ The stock has fallen 51% from its June high already.¹ Even if I’m right about value, the path can be violent in either direction.
• China and trade. China and Hong Kong were 47.6% of FY2026 revenue and Asia was 70%.² Section 232 and Section 301 investigations could change tariffs on products sold in the United States, now mostly exempt, and export rules could cut sales or raise cost.²
• Counterparty and delivery. If a customer breaks an agreement Sandisk may need to resell the volume at lower prices, and if Sandisk fails to deliver it faces damages or early termination. The guarantees may not cover the lost revenue.² The agreements may also bind a part of Sandisk’s supply and keep it from selling at better prices.²
• Partner and place. All wafers come from Kioxia’s fabs in Yokkaichi and Kitakami, Sandisk may not make flash elsewhere, and the ventures run to 31 December 2034.² Flash Ventures uses leases that Sandisk guarantees for half, and a default under a Japanese lease could be called on Sandisk, $923M outstanding.²
• Stock pay and dilution. The shares given to staff were worth $2.7B when they vested in a single quarter, 5.7M units are unvested, and the CEO’s launch grant is fixed at its maximum number of units and vests in February 2028 if he stays.⁵¹²
• Western Digital’s remaining stake. Western Digital said it expects to sell its remaining Sandisk stake by the end of 2026.²
• Governance. One person is chairman and chief executive, with a lead independent director, seven of eight nominees are independent, and the auditor’s one critical matter was evidence over revenue.¹²²
• Taxes. The 12% rate depends on Malaysian holidays that expire from 2028 to 2031, and the 15% minimum tax applies from FY2027.² I use 18% from FY2030.
• The split. The 8-for-1 split changes nothing about value, but it changes the price quoted. After 20 November a $1,150.89 close would be quoted at $143.86. Anyone comparing this page to a later quote needs to divide my per-share values by eight.⁹

Geopolitical risk, and what it means for the value

Sandisk is an American company whose flash is made in Japan, assembled in Malaysia and China, and sold mostly in Asia.² China and Hong Kong took 47.6% of FY2026 revenue by ship-to address, and the United States 18% C (calculated).² A ship-to address is where the box lands, not where the chip ends up, so the true China exposure is smaller, by an amount I cannot see. My base assumes today’s rules: China open, most US sales exempt from tariffs, and Flash Ventures renewed after 2034. The only cost already in it is the renewal payment to Kioxia. The table runs four changes through my model one at a time.

Weighted by my odds, the four take about $41 a share, or 7%, off the base C (calculated). The price is 96% above the base, so these risks are not what separates the two. They matter more to the bull, which depends on margin lasting and loses 40% if the ventures end. Japan’s government helps hold the ventures together: its trade ministry approved up to ¥150B in 2024 towards the Yokkaichi and Kitakami plants.³⁴

Two risks I do not value. Sandisk buys DRAM from Nanya in Taiwan and holds 3.9% of it, worth $1.8B at 3 July, and its controllers come from outside foundries.² A blockade of Taiwan would stop every memory maker’s customers at once, and my model would not be a useful guide in it. Korea is a competitor, not a chokepoint: Samsung and SK hynix make 54% of the top five’s NAND revenue⁶ C (calculated), so their capacity choices set the price cycle more than any ministry does.