SNDK

As of 2026-10-21 · partly unblinded (see "Blindness"), revised once after our research-eval check · blind commit f617713 (the model committed before we read the price) · market pass (the later step in which we read the price and the Street) after it

We rate SanDisk Avoid: we would not buy at $1,385.92. SanDisk earns 84% gross margins today because NAND flash (the memory chips in SSDs, phones and data centres) is short. Our central case, our best estimate, takes FY2027 from the company's guide and its record of beating it, then lets the market price per bit fall from FY2028 to a normal 40% gross margin by FY2031, as every NAND peak since 2006 has fallen; the hyperscaler (large cloud company) contracts hold a floor under two thirds of the bits until they end. On that path the business is worth $751.41 a share, 45.8% under the price; even our bull ($1,108.49) is 20% under it. The price needs the boom to leave a 70% normal margin, or the peak to last through FY2030 and settle at 64%. Nothing in 20 years of NAND history supports either, though the contracts are new.

The blind central, from our model as first committed, was $876.86. Our research-eval check, run by a fresh agent, found three errors, all fixed before review and bridged dollar by dollar in revision.txt: the share count (+$14.90), a FY2028 price held while our own bits grow 15% (-$92.39), and a contract rule applied differently in FY2027 than later (-$47.96). The view was Avoid before and after.

Terms. Fair value is our estimate of what a share is worth today. We value each company three ways: the central case is our best estimate, the bear case a severe but plausible bad outcome, and the bull case the matching good outcome. FY is SanDisk's fiscal year, which ends around the start of July: FY2027 is the year to about July 2027. EBIT is operating profit before interest and tax, and EV/EBIT divides the value of the whole business by it. The long fade values the business after FY2031 by letting growth and the return on new investment drift to long-run levels; the outside view applies the EV/EBIT multiple that peers trade at. The Street means the analysts who follow the stock. NBMs are the three hyperscalers' long-term supply agreements. Alpha is Capfolio's stock-ranking model. A flag is a warning from our automated report check that a number looks off or an argument is missing.

The business

Source tags in brackets: [F] a reported fact (filing or release), [M] a management claim, [I] our inference, [A] our assumption.

SanDisk makes NAND flash in Japan through Flash Ventures, its joint venture with Kioxia, and sells it as data-centre SSDs, client SSDs in PCs and phones, and consumer cards and drives [F, 10-K FY2026]. Western Digital spun it off in February 2025. FY2026 revenue was $20.2B and the fourth quarter alone $8.97B, with cost of revenue at $1.38B [F, release 2026-08-05]. Data-centre revenue was $2.98B in that quarter, up 103% on the third [F].

Two things set the next five years.

The shortage. AI inference and training store far more data, and NAND makers cut capacity in the FY2023 glut. Gross margin went from 7.1% in FY2023 to 84% guided for Q1 FY2027 [F, M]. Management says bits are on allocation beyond calendar 2027 and that demand runs into 2028 [M, Q4 call 2026-08-05]. It also says it will add bits mainly by moving to new process generations, not by building fabs [M, Investor Day 2026-08-13].

The contracts. Three US hyperscalers signed long-term supply agreements (NBMs) worth $93.9B of minimum revenue, with $16.5B of guarantees and deposits [M, Q4 call]. They cover more than half of FY2027 bits and about two thirds of FY2028 [M]. Their terms have floors and ceilings that are not disclosed. They last about four years on average, so most end in FY2030-FY2031 [M, I].

The bull reading is that these contracts make NAND a utility. The bear reading is that a customer walking away loses at most about 18% of its commitment (our inference: $16.5B of guarantees over $93.9B [M]), and that DRAM supply agreements signed in 2017-2018 were delayed in the 2019 glut (memory sector note).

Recent results: SanDisk has beaten its own revenue guide five quarters running

SanDisk has beaten the midpoint of its own revenue guide in each of the last five quarters: +5.6%, +7.3%, +16.3%, +29.3% and +12.1% [F, releases]. The beats are measured; whether management guides low on purpose is our inference, not a fact. We carry the median beat into Q1 FY2027 and through the year.

Our forecast and value

The model is model.py, its output output.txt, and every input with its source is in assumptions.md (the blind record; the three fixes below supersede it where they differ).

• FY2027: the Q1 guide midpoint ($10.55B) raised by the median beat to $11.83B, then +4% a quarter: revenue $50.2B, gross margin 84% [M guide, I beat].
• FY2028-FY2031: revenue is bits times price. Bits grow 15% a year. Contracted bits sell at a floor backed out of the $91.1B of minimum revenue left, plus 30% of any market price above it, in every year. Market bits start falling in FY2028 and reach a 40% gross margin in FY2031. Cost per bit falls 10% a year [A].
• Why the fall starts in FY2028: management's allocation claim is about volume, not price, and our own bits grow 15% a year. No contract price or guide covers FY2028's market bits. Holding the FY2027 price through FY2028 is worth +$75.86 (output.txt).
• Normal margin: old SanDisk's pooled gross margin over 2006-2015 was 38.3% and its best year 46.9% (2010). New SanDisk's FY2022-FY2026, this peak included, was 42.8%. FY2022-FY2025 alone was 23.3%. We take 40%.
• Exit: the contracts' profit above the market is removed from FY2031 before it is capitalised, since most contracts end by then. FY2031 EBIT at the market price is $3.2B. The outside view (12.5x that EBIT, about $260 a share) is above the long fade (about $150), which values the same profit at 8.4x after tax. The input behind the gap is the cost of capital: the fade uses our 12.66% against 4% growth and a 15% return on new capital (Micron's pooled net income over net plant; asset life not separately modelled, and a 10% return is -$6.38), while the peers' 12.5x was earned over ten years at lower rates. The exit is their mean, $208.
• Cash: all free cash goes to shareholders as earned, as management promised [M, Investor Day]. The model pays out $115B over five years, valued at face, so most of the value is the five years of cash, not the exit.

Fair value $751.41 is $14.84 of net cash and claims today, $193.63 from FY2027 free cash, $164.60 from FY2028, $259.06 from FY2029-FY2031, and $119.27 from the discounted exit (output.txt). Bear ($409.73) falls faster, to a 30% normal by FY2029. Bull ($1,108.49) holds a market price 5% above FY2027's through FY2029 and settles at 50%.

Management's long-term model (16.5% growth, 80% gross margin) capitalised as it stands is worth $3,765 a share. We do not take it: it assumes the peak never ends.

What the market expects (market.py, market.txt)

At $1,385.92 the market cap is $200.3B (FMP, 144.5M shares). On our five years of cash, the price needs an exit of $1,317 a share in 2031 instead of our $208.

One driver at a time, the price needs:

• a normal gross margin of 69.7% from FY2031, against our 40% and a best year of 46.9%; or
• the peak held through FY2030 (worth $996.89 alone) and then a normal margin of 63.6%; or
• an outside-view exit multiple of 118x FY2031 EBIT, which is no real answer.

The Street's revenue matches ours for FY2027 and then stays near its peak:

The FY2030 Street figure rests on five analysts and its EPS on one. FMP's EBIT field does not match its own net income (FY2027 net income $33.2B on EBIT $35.1B with no debt), so the bridge takes the Street's net income as its after-tax profit. Run through our model (our capex, tax, working capital and exit method; market.txt):

On the Street's own numbers through FY2029, where ten or more analysts publish, the value is $842: $90 above ours, from holding FY2028-FY2029 near the peak. The average target of $2,164 ($1,921 discounted one year at our 12.66%) needs boom-level profit to last past FY2030. The disagreement is less about the next three years than about whether NAND stays at boom prices into the 2030s.

Our view against the market's: contracts set a floor for about four years, not for good

The market's view is that AI storage demand and the hyperscaler contracts end the NAND cycle. Ours is that the contracts set a floor for volume and price for about four years, and that after them SanDisk sells a commodity again at a normal margin. What settles it:

• the contract price per bit through calendar 2027 and 2028 (peer results at Micron and Kioxia, trade press);
• whether Samsung, SK hynix, Kioxia or YMTC announce new fabs for 2028 output;
• whether any NBM customer delays or cuts its minimum.

Alpha, Capfolio's stock-ranking model, ranked SNDK 8 of 710 on 2026-10-20 (score 99, Strong; growth A rank 1, momentum A, expectations B, valuation C; stale cache). Alpha ranks on momentum, growth and expectations; this report values the business. The gap is expected and is not a portfolio question.

The chart (TradingView, weekly to 2026-10-21) is weak for timing: a high of $2,354 in the week of 2026-06-22, a low of $998 five weeks later, a lower high of $1,909 in the week of 2026-09-21, and a low of $1,276 last week. The price is under its 50-day average ($1,599) and over its 200-day ($1,222). Grade C-; support $1,276-$1,330, resistance $1,540-$1,600. This is timing beside the view, never an input to it.

Bear case, and why the other side could be right

Our bear ($409.73) falls in FY2028-FY2029 to a 30% gross margin, close to the FY2022-FY2025 carve-out years without this peak (23.3%). By our count of the 10-Ks, NAND had a loss-making glut in about one year in four since 2006.

The bull side could be right for three reasons:

1. The contracts are new. Three hyperscalers with data centre at about half of NAND demand have never signed four-year floors before. If they renew in FY2030 at high floors, NAND is a different business and 50-60% normal margins are possible.
2. Supply discipline. All five makers cut capacity in FY2023 and now say they add bits by node changes only. If no new fab arrives by 2029, the peak can last; holding it through FY2030 alone is +$245.48.
3. Demand keeps surprising. Revenue has run ahead of management's own guide in each of the last five quarters.

Even the bull ($1,108.49), which holds the peak price through FY2029 and settles at 50%, is 20.0% under the price.

Next results preview (Q1 FY2027, 2026-10-29 after the close, TradingView, confirmed)

Written before the market pass from the model; the date was corrected after it.

Dropping the whole Q1 beat (about $5.4B of FY2027 revenue) is -$76.08 a share, about $14 a share for each $1B of FY2027 revenue. The Street (market pass) expects Q1 revenue of $10.65B and EPS of $46.20, under ours.

Dates to watch, and what would prove us wrong

• 2026-10-29, Q1 FY2027 results: revenue under $10.55B (no beat) is -$76 a share and moves us toward the bear; above $12.9B with a Q2 guide over $14B toward the bull.
• About 2027-01-28, Q2 FY2027 (date our estimate): the first quarter priced after the calendar 2026 contract talks. A flat Q3 guide confirms the fall has begun.
• About 2027-08-05, Q4 FY2027 and the FY2028 outlook (date our estimate).
• A contract-market price per bit still rising through calendar 2027 and holding into 2028 says the fall starts later: holding the FY2027 price through FY2028 is +$76 a share.
• A new fab for 2028 output at any of the other four makers shortens the peak; none for two years lengthens it.
• Any NBM customer delaying or cutting its minimum.
• A renewal or extension of an NBM past FY2031 at floor-level pricing, or a peer disclosing post-2030 floors: the one fact that would flip the call (the world the Buy advocate built, $1,445).
• Gross margin above 70% on market-priced bits through FY2029: each 10 points of normal margin is about $92-130 a share.

Revision after research-eval (before the adversarial review)

Our research-eval check (a fresh agent, commit 1340567) examines the basics: math that reruns, sourced figures and the bridge from business value to value per share. It found these; each was rerun here.

Total $876.86 to $751.41 (-$125.45). Each fix rests on a filing or on the model's own consistency, not on the price; two of the three moved value away from the price. drivers.json now centres the first decline year on FY2028. P(price justified) went from 11% to 6.4%.

Adversarial review

Two fresh reviewers, one after the other, on commit 74b160a: a model auditor, who looks for errors, and an opposite-side advocate, who argues the case for Buy. Each was given the market data and the report check's flags, and neither was shown the other's findings. Their scripts are under /private/tmp/claude-501/sndk-audit/ and /private/tmp/claude-501/sndk-advocate/.

Model auditor. Reran every script (byte-identical) and found nothing material. Seven priced items net to about -$5 (0.7%); we disclose them rather than reopen the model eight days before the Q1 results rebuild FY2027:

The auditor's answers to the report check's flags: flag 1 (the move from the blind value) cleared, because the largest step, -$92.39, rests on the model's own logic and moved value away from the price; flag 3 (a risk charged twice) none; flag 4 (outside view against the fade) cleared, because 12.53x EV/EBIT is 15.1x NOPAT (net operating profit after tax), which at 12.66% implies 10.8% growth forever, so the gap is the discount rate; flag 11 cleared; flag 16 (the price outside nearly all our scenarios): P is 0.03% in the base world, 0 in the bust and 30% in the structural: the price rests on that world alone; flag 17 (value far below the Street) explained by the Street bridge above. Flag 18 (our 10th-90th $537-$1,307 is 0.42 times the Street's target spread): our ranges come from NAND's own history (normal margins of 15-70% across the three worlds, against a best pre-boom year of 46.9%), and the Street's spread is of targets, not values. We do not widen toward the price; the structural world is the place the price lives, and its weight is the question.

Opposite-side advocate (the case for Buy). It found its own dated sources and built the world in which the price is right:

The world where the price is right (the report check's flag 16 asks for it): the fall starts in FY2030 when the first new fab arrives, FY2028 holds the Q4 FY2027 exit rate, the normal is Micron's strong-cycle 44.1%, floors hold with no haircut, and the contracts are renewed on half the bits: $1,444.60. The same world without renewal is $984.80. The price needs the hyperscaler contracts renewed past FY2031 at floors near today's. That is the one fact that would flip the call: a renewal or extension at floor-level pricing on about half of SanDisk's bits, or a peer disclosing post-2030 floors. Neither exists today, so the view stays Avoid, with that as its first falsifier.

On flag 18 (the narrow range), from the advocate: widening the drivers from history barely moves P (6.4%, 90th $1,306); the range is narrow because one assumption, the contracts ending, holds in every draw. With the contracts kept, the 90th is $1,519 and P 16.9%.

Blindness

We were partly unblinded before the blind commit, with no exposure used as an input (assumptions.md): alpha rank --json printed the market cap; the 8-K of 2026-05-14 kept a mini-tender price of $1,150 in its body; the anchor file holds SanDisk's own FY2026 year-end multiple; and transcripts carried remarks about the share price. The market pass changed no input. The results date moved from about 2026-11-05 to 2026-10-29. The revision above was made by an analyst who had seen the price.

One code fix after the blind commit, found when the odds run showed a 90th percentile of $6,578 in the structural world: a draw whose first year of decline rounded to 2032 kept FY2031 at the peak price, although the driver says FY2031 is normal in every draw. The check for the normal year now comes first in model.path(). Central, bear and bull were unchanged (output.txt identical); P(price justified) went from 12% to 11% (both runs in this session; the research-eval check could not rerun the first).

Sources

• 10-K FY2026 (filed 2026-08-17); 8-Ks of 2026-05-14, 2026-08-05, 2026-09-11, 2026-09-16; releases Q3 FY2025-Q4 FY2026 (SEC EDGAR). Old SanDisk 10-Ks for 2010 and 2015.
• Q3 FY2026 call (2026-04-30), Q4 FY2026 call (2026-08-05), Investor Day (2026-08-13), Goldman Sachs Communacopia (2026-09-09), via TradingView.
• FMP statements, peer betas, quote, price targets and estimates (2026-10-21); Robinhood quote (2026-10-21); TradingView earnings calendar and weekly bars (2026-10-21).
• Memory sector note, docs/research/sectors/memory/note.md (2026-10-17).