Sandisk: Thrown Away Twice, Best Stock in the S&P — and the Question That Arrives at an 80% Gross Margin

Panoramic research report · longitudinal history + cross-sectional rivalry + synthesis Subject: Sandisk Corporation (NASDAQ: SNDK) Report date: 2026-07-04 · Data cutoff: fiscal Q3 FY2026 (quarter ended 2026-04-03) + latest quote and news flow Sources: Yahoo Finance (quote / fundamentals / financials / analysts), company IR releases and filings, Investing.com, TrendForce, Tom's Hardware, Futurum, industry NAND price coverage For information and research purposes only. Not investment advice.

Before we start

Sixteen months ago, you could have bought Sandisk — the company that invented the flash memory card, whose chips have probably touched every photo you took between 2001 and 2015 — for $27.89 a share, about $4 billion for the whole business. It had just been spun off by a parent that didn't want it, into a market that wanted it even less, and it celebrated its first independent quarter by writing off $1.8 billion of goodwill — an accounting funeral held, as it turned out, at the exact bottom.

Last week the stock traded at $2,354. On Thursday it closed at $1,745 — down 14% in a single session, 26% off its high, and still the best-performing stock in the S&P 500 for the first half of 2026, up roughly sixty-fold from the April 2025 low.

If that arc sounds familiar, it should. This is the second time in five weeks this research feed has opened a report with the same impossible sentence: a memory company — the textbook commodity business, the industry that broke everyone who ever ran it — is suddenly printing software margins, and the market cannot decide whether the cycle is dead or merely coiling. We wrote it about Micron and DRAM. Today we write it about Sandisk and NAND, and the numbers are, remarkably, even more violent. Gross margin went from 22.5% to 78.4% in five quarters. Revenue in the March quarter: $5.95 billion, up 251% year-over-year — and up 97% from the quarter before. Guidance for the current quarter: $7.75–8.25 billion at a 79–81% gross margin, which would mean this company grew revenue roughly 4× in a year and a half while expanding margins by nearly sixty points.

Against that: a trailing P/E of 59 and a forward P/E of 9.3. The same split personality we found at Micron, stretched further. The market is paying a premium multiple on what Sandisk just earned and a distressed-cyclical multiple on what it will earn next — a public confession that nobody believes an 80% gross margin on NAND flash can survive contact with the industry's own history. Analyst targets run from $1,000 to $3,250. And on Thursday, with no company news at all — just a strategist's warning, a report of Samsung capacity additions, and the smell of H2 rebalancing — the stock surrendered $42 billion of market value before lunch.

There is one more thread that makes this report the natural third act of a trilogy. The AI trade has been rotating down the shortage stack: first compute (Nvidia, covered), then the memory that feeds compute (Micron, covered), now the storage that AI datacenters drain. And in one of those symmetries markets occasionally gift to writers: the man who co-founded Sandisk in 1988 — Sanjay Mehrotra — is the same man who, as CEO of Micron, presided over the 84.6% gross margin we dissected five weeks ago. The two ends of this trilogy share a founder. They also share a question, and it is the only question that matters at these altitudes: when a commodity starts earning monopoly margins, is it still a commodity?

Longitudinally, we trace how a physicist's floating-gate patents became the memory card in two billion devices, how the company was nearly bought at the bottom in 2008, actually bought at the top in 2016, buried inside a hard-drive manufacturer for nine years, and finally thrown clear in a spinoff priced for irrelevance. Cross-sectionally, we map the 2026 NAND board — Samsung the swing producer, the Kioxia conjoined twin, the China wildcard — and the two demand shocks (enterprise SSDs today, High Bandwidth Flash tomorrow) that turned the worst-behaved commodity in semiconductors into the scarcest. Finally, we put the 59×-trailing/9×-forward paradox on the table and try to locate, as precisely as the data allows, where the disagreement actually lives.

Part One · Longitudinal: three decades of being early, and two of being unwanted

1. The physicist's bet: floating gates and a company named after sunshine (1988–1995)

Eli Harari earned his doctorate studying the physics of a curious structure: a transistor gate electrically isolated in oxide, where injected electrons would simply stay — for years — without power. In 1988, when a megabyte of magnetic disk was cheap and a megabyte of solid-state anything was absurd, Harari, Sanjay Mehrotra, and Jack Yuan founded SunDisk (renamed SanDisk before its 1995 IPO, reportedly because customers kept assuming a connection to Sun Microsystems) around a specific conviction: that floating-gate flash, packaged as a system — cells plus controller plus firmware that hid the technology's flaws — could become removable, rewritable, solid-state storage for devices that did not yet exist.

The systems part deserves emphasis because it is the DNA of everything that followed. Raw NAND flash is, and has always been, a terrible product: cells wear out, bits flip, blocks die. SanDisk's founding insight was that the defects were manageable in software — error correction, wear leveling, bad-block mapping — and that whoever owned the system-level patents would tax the entire industry regardless of who fabricated the silicon. For two decades, that patent estate was a river of high-margin licensing revenue that kept the company alive through cycles that killed larger rivals. At various points in the 2000s, licensing and royalties contributed a share of profits wildly out of proportion to their share of revenue — a stream that cost nearly nothing to service and that competitors paid because litigating Harari's foundational claims was costlier than writing the check. It was a fabless-brained company that happened to co-own fabs — a hybrid that made it perpetually hard to value, and perpetually underestimated.

It is also worth pausing on how contrarian the founding bet was. In 1988, flash memory cost orders of magnitude more per megabyte than magnetic disk, and every sensible roadmap said disks would stay ahead forever — they had spinning-media physics and fifty years of learning curve on their side. Harari's counter-argument was structural, not incremental: solid-state storage would win wherever power, shock-tolerance, and form factor mattered more than cost-per-bit, and the set of devices fitting that description — which in 1988 meant almost nothing — would eventually mean almost everything. The digital camera proved the thesis at miniature scale; the smartphone proved it at planetary scale; the AI datacenter, thirty-six years after founding, is proving it at industrial scale, for reasons (watts and racks, not dollars per gigabyte) that rhyme almost perfectly with the original pitch. Few companies get to be right about the same idea three times in three different decades.

2. Format wars, a fateful marriage, and the first time nobody wanted it (1995–2008)

The consumer flash era arrived exactly as Harari had bet, and SanDisk won its format wars — a competition modern readers underrate because its losers are forgotten. CompactFlash (SanDisk, 1994) outlasted a zoo of rivals; the SD card standard (co-developed with Toshiba and Panasonic in 1999) beat Sony's Memory Stick and Olympus/Fuji's xD-Picture Card so thoroughly that "SD" became a generic noun, the way "35mm cartridge" once was for film. Winning a format war in removable media is a special kind of victory: it converts a commodity chip into a royalty-bearing standard plus a retail brand, and for fifteen years SanDisk collected on both — the only NAND player whose name a consumer could recognize on a store shelf, a fact that still quietly matters in 2026 (the consumer segment's $800M+ quarters ride on it). Two structural decisions from this period shaped the next twenty-five years.

The first was the Toshiba joint venture (2000): SanDisk would co-own and co-fund NAND fabs in Yokkaichi, Japan, taking roughly half the output, while Toshiba ran the manufacturing. The economics were elegant — half the capex, guaranteed leading-edge supply, someone else's operational headache — and the constraint was equally elegant in reverse: SanDisk would never fully control its own supply decisions. That marriage, inherited through Toshiba's memory spinoff (Kioxia, 2018) and SanDisk's own changes of ownership, is still the structural spine of the company in 2026. Roughly speaking, Sandisk today is entitled to about half the output of the world's third-largest NAND manufacturing complex without bearing hyperscale capex — an arrangement whose consequences, good and bad, recur throughout this report.

The JV's mechanics reward a moment of precision, because they explain numbers that otherwise look like errors. The fabs sit inside jointly-owned entities; each parent funds its share of expansion through those entities and commits to purchase its share of output — a take-or-pay-like obligation that, in downturns, forces both parents to keep absorbing wafers they'd rather not have (a hidden pro-cyclical liability the 2023 trough exposed brutally), and in upcycles hands them contractually guaranteed supply at cost while spot prices soar (the hidden windfall 2026 is monetizing). Because the heavy equipment lives at the JV level and much expansion is funded through JV-level cash flows and leases, Sandisk's own reported capex can run absurdly low — the $83 million net print — without the underlying manufacturing actually being starved. Investors screening on "capex as % of revenue" see saintly discipline; the fuller truth is a structure that both enforces discipline and obscures where the spending happens. Both readings matter and neither alone is honest.

The second was surviving the 2008 test. As the financial crisis cratered flash prices, Samsung made a hostile run at SanDisk at $26 a share — an offer that valued the patent estate, the JV stake, and the brand at roughly one bad year's panic price. Harari's board said no. It was the first time the market concluded SanDisk was worth more dead than alive, and the first time that conclusion was wrong. The stock would be worth many multiples of Samsung's bid within five years, as smartphones detonated NAND demand.

3. The smartphone windfall and the exit at the top (2009–2016)

The iPhone era was everything the founders had waited two decades for: every phone a guaranteed multi-gigabyte NAND socket, every generation a capacity upgrade, and — unlike the camera-card era — no removable-media format war to defend, just embedded chips sold on qualification and price. SanDisk rode embedded flash hard, and made its first serious datacenter play early: the 2014 acquisition of Fusion-io, the flash-acceleration pioneer whose PCIe cards had taught Facebook and Apple what NAND could do inside a server. The strategic read was correct a decade early — enterprise flash was the future — but the timing produced little; the enterprise SSD market of 2014 was a knife-fight of narrow margins, and the acquisition is remembered mostly as expensive tuition. File that under a pattern this company repeats: SanDisk's strategic bets have tended to be right about the destination and a half-decade early on the arrival, from the 1988 founding thesis to Fusion-io to, perhaps, HBF.

By 2015 SanDisk was a fixture of the S&P with $6+ billion in revenue — and increasingly squeezed: flash prices were softening, its own fab-lite model limited its cost leverage against integrated giants, and China's capital was circling the industry (Tsinghua Unigroup's overtures to memory assets were the era's background noise). Then, in 2016, Western Digital acquired SanDisk for $19 billion — the hard-drive giant buying its way into solid-state before SSDs finished eating its lunch. For WD the deal was existential defense; for SanDisk shareholders it was a full-ish price at what proved to be a local top; for the flash franchise itself it began nine years of strategic house arrest.

Strategically, the logic was defensible. Culturally and financially, it was a burial. Mehrotra, passed over to run the combined company, left — and in 2017 became CEO of Micron, where this trilogy's paths cross. SanDisk became a division: its brand relegated to retail memory cards, its JV obligations one line in WD's capital budget, its flash roadmap perpetually negotiating against HDD priorities for investment. For nine years the most storied name in flash effectively disappeared from the investment map — present in every phone, absent from every pitch book.

The WD years also produced one near-miss that history should record, because it nearly rewrote this entire report: in 2023, WD and Kioxia negotiated a full merger of their flash businesses — which would have created a NAND producer rivaling Samsung's scale and pre-empted any spinoff. The deal reportedly collapsed in October 2023, in significant part because SK hynix — a Kioxia shareholder through the 2018 Bain consortium — withheld its blessing, unwilling to bless the creation of a giant that would dwarf its own NAND position. Read that against Section 9's news: two years later, the company that blocked Sandisk's path to scale-by-merger became its co-author on HBF, the technology bet that may matter more than scale. In memory, your blocker, your partner, and your competitor are reliably the same three companies wearing different hats to different meetings.

4. The unwanted spinoff: thrown away a second time (2022 – April 2025)

What un-buried it was the usual combination: an activist and a bad cycle. Elliott Management arrived in 2022 arguing WD's flash and HDD businesses traded at a conglomerate discount that separation would cure. The 2022–2023 NAND downturn — the worst in the industry's history, with contract prices falling below cash production cost and every producer bleeding money — delayed the surgery but proved the argument: inside WD, flash was a capital-starved hostage to a cycle WD couldn't control.

Elliott's arithmetic at the time is worth recording, because it quantifies how cheap the market's indifference ran: the activist argued WD's flash business alone was worth $17–20B+ — roughly the whole company's market value — implying the market was pricing SanDisk's franchise at approximately zero inside the conglomerate. The eventual spinoff tested that thesis in the cruelest possible way: freed from the conglomerate, the flash business promptly traded down to $4 billion — the market not merely pricing it at zero inside WD but at a fraction of Elliott's floor outside it. Activists are sometimes wrong about value; here the activist was right about value and early about timing by almost exactly one NAND price inflection.

On February 21, 2025, Sandisk (now styled with a lowercase d) began trading as an independent company, and the market's verdict was immediate and brutal. The shares opened in the mid-$30s… actually, let the tape speak: mid-$30s at the spin, a brief pop above $50, then straight down — through the April 2025 tariff panic to an intraday low of $27.89, a market capitalization near $4 billion for a company with roughly half the output of the world's third-largest NAND complex, the industry's second-best patent estate, and a brand two generations of humans could identify on sight. The spin cohort — index funds force-fed shares of a company nobody chose to own — did what spin cohorts do: they sold first and never asked questions.

Management then performed the ceremonial act that, in hindsight, marked the exact bottom: in its first quarter as an independent company (March 2025), Sandisk wrote off $1.83 billion of goodwill, formally declaring that the business was carried on its own books at more than it was worth. Revenue that quarter was $1.7 billion at a 22.5% gross margin. The company posted a $1.9 billion GAAP loss. Twenty-two analysts covered it with the enthusiasm of an obituary desk. The second time nobody wanted SanDisk was fifteen months ago.

5. The five-quarter resurrection: the steepest margin inflection in large-cap tech (mid-2025 – mid-2026)

What happened next is best shown before it is explained:

Five quarters. Gross margin up fifty-six points. Revenue up 3.5×, with the current quarter guided to roughly 4.2× the spinoff-era run rate. The March quarter alone earned more net income than the company's entire market capitalization at the April 2025 low. There is no larger company in the market that has repriced its own unit economics this fast; even Micron's inflection — 37.7% to 84.6% over five quarters, which we called the steepest we'd seen — was, at the gross-margin line, a gentler slope from a higher base.

Before the explanation, a word on what a 78% gross margin means mechanically in this industry, because the number is so far outside NAND's historical envelope that intuition fails. NAND's previous all-time-great quarters — the 2017 peak, the 2021 spot squeezes — topped out with leaders in the high-40s. A 78% print means selling bits for roughly 4.5× their fully-loaded cost, in a product whose price customers have negotiated quarterly for three decades. It requires not merely undersupply but allocation — customers bidding against each other for guaranteed volume — sustained across two consecutive quarters and guided for a third. The only precedents in memory history are... the ones this research feed keeps writing about: HBM in 2025–26, and DRAM's current cycle. Three simultaneous all-time margin records across three memory products is either the strongest possible evidence that AI demand has re-based the whole complex, or the most synchronized cycle top ever printed. There is no boring interpretation available.

Four forces converged, and it matters that they were simultaneous. First, the supply famine. The 2022–23 crash forced the entire NAND industry into the deepest capex cuts in its history; nobody added meaningful capacity for nearly three years, and Sandisk's own discipline persists to a degree that reads like a typo — net cash capex of $83 million in the March quarter, 1.4% of revenue, partly a structural feature of the JV model, partly a stated refusal to feed the next glut. Second, the demand shock arrived anyway. AI datacenters turned out to need staggering amounts of storage, not just memory: training corpora, inference context stores, vector databases, checkpoints — enterprise SSD demand that materialized in quarters, not years. Sandisk's datacenter segment grew 233% sequentially in the March quarter; management described the market as more undersupplied each quarter and began steering customers toward multi-year supply frameworks, language NAND has never in its history been able to use. Third, the substitution squeeze. DRAM makers, chasing HBM margins, shifted wafer capacity away from... rather, toward HBM and away from everything else — and some NAND suppliers redirected investment toward DRAM entirely, tightening NAND supply at precisely the moment demand inflected. Fourth, the technology cadence held. The BiCS8 node (218-layer, CMOS-bonded-to-array, co-developed with Kioxia) shipped on time into the strongest pricing in NAND history, stacking cost reduction on top of price increases — the double gear that produces fifty-point margin swings.

Pause on the sequencing, because it explains why almost everyone missed it. The margin inflection was visible in the arithmetic two quarters before it was visible in the headlines: once contract prices turn in an industry running near cash cost, every dollar of price increase falls straight to gross profit, and operating leverage in a business with $500M of quarterly opex converts a 30% gross margin into a 50% one, then a 78% one, on price alone. But the investors capable of doing that arithmetic had spent 2022–24 being incinerated by the identical arithmetic running in reverse, and the spin cohort had just finished selling at $30. The people who best understood the machine were the least psychologically capable of buying it — which is how a $258 billion repricing begins at a $4 billion market cap with twenty-two indifferent analysts watching.

The stock chart is simply that table with a lag: $28 in April 2025, $52 by late August, ignition in September as contract prices inflected ($68 → $102 in a fortnight), $186 by late October, $275 by New Year, then the January earnings detonation ($377 → $576 in three weeks), $989 by late April, $1,562 in early May, and an intraday $2,354 on June 24 — before the week that prompted this report: a 26% drawdown on no company news whatsoever, closing July 3 at $1,745. Sixty-two times the April 2025 low, at Thursday's close, after the crash.

That last clause — no company news whatsoever — is the hinge of this report. Everything in Part One explains why the stock rose. Nothing in it explains why it just fell 26% in a week. For that, you need the board.

Part Two · Cross-sectional: the NAND board in July 2026

6. NAND economics: historically the worst-behaved commodity in semiconductors

To appreciate what a 78% gross margin in NAND means, you must first appreciate that NAND has historically been a worse business than DRAM — and DRAM was bad enough to feature in our Micron report as "the commodity that broke everyone." The structural reasons are worth two minutes, because they are the bear case's foundation.

DRAM consolidated to three suppliers by 2013 — the consolidation our Micron report chronicled body by body — and that consolidation is the entire foundation of DRAM's modern pricing behavior. NAND never got its funeral procession: it kept five to six meaningful producers all the way into the 2020s — Samsung, Kioxia, SanDisk/WD, SK hynix (plus Solidigm, the Intel NAND unit it bought when even Intel gave up on the business), Micron, and, since the late 2010s, China's YMTC. The industry has repeatedly approached consolidation — Intel selling out in 2020, the WD–Kioxia merger dying in 2023, perennial rumors around every downturn — and always stopped short, in part because the JV structures and national-strategic stakes (Japan's in Kioxia, Korea's in its champions, China's in YMTC) make memory exits geopolitical events rather than commercial ones. More suppliers means faster defection from pricing discipline in every downturn. Worse, NAND's cost curve falls faster than DRAM's: 3D layer stacking (100+ layers by 2020, 218 on Sandisk's BiCS8, 300+ on roadmaps) plus density tricks like QLC deliver 20–30% annual cost-per-bit declines, which means suppliers can cut prices 20% a year and stand still on margin — so they do, whenever demand wobbles. And NAND demand itself was historically elastic consumer demand — phones, cards, cheap SSDs — that evaporates in recessions. Put together: an industry that spent roughly one year in three earning less than its cost of capital, punctuated by short, sharp booms that reliably financed the overbuilds that caused the next bust. The 2022–23 downturn was the archetype in extremis: contract prices below cash cost, every producer loss-making simultaneously, the whole industry effectively paying customers to take inventory.

The bull case for this cycle being different rests on three asserted breaks with that history. One: consolidation of behavior, if not of ownership. Post-2023, every producer emerged from near-death with religion about capex — and unlike every prior cycle, the discipline has held through eighteen months of rising prices, which has never happened before. Sandisk spending 1.4% of revenue on capex at an 80% gross margin is behavior the NAND industry has literally never exhibited at any prior cycle peak. Two: the demand mix changed class. Enterprise/datacenter demand — contracted, planned, price-insensitive relative to consumer — is displacing the elastic consumer bit as the marginal buyer. Management's push toward multi-year supply frameworks is the tell: NAND trying to borrow HBM's contractual clothing. Three: the substitution drain. With DRAM wafers migrating to HBM economics, and some NAND capex redirected toward DRAM, the industry's aggregate ability to flood the NAND market is genuinely impaired in the near term.

A concrete feel for the cost-curve physics helps calibrate both cases. A 3D NAND generation transition — say BiCS6's ~160 layers to BiCS8's 218, with CMOS-bonded-array packing logic under the memory stack — delivers on the order of 50–60% more bits per wafer for perhaps 15–25% more processing cost: cost-per-bit falls by roughly a quarter per node, and nodes arrive every 18–24 months. Add QLC (four bits per cell instead of three, another ~30% density gain where endurance permits) and the industry's deflation engine runs at 20%+ a year with no one building a single new fab. This is why NAND booms die faster than DRAM booms: the cost curve hands every producer a standing invitation to cut price and hold margin, and in every prior cycle, someone — usually the player losing share — accepted the invitation within eighteen months of the peak. The bull case does not require the physics to stop; it requires that, for the first time, nobody accepts the invitation while demand bits outrun supply bits.

The bear case reply is one sentence long, and it has been right for thirty-five years: every NAND cycle top has featured a persuasive essay about why this one is different, and the essay has always been financed by the margins that caused the overbuild. Which essay is this report? Part Three returns to that.

7. The board: a swing producer, a conjoined twin, and a wildcard

Samsung remains the share leader (roughly a third of industry bits) and, more importantly, the swing producer — the only player with the balance sheet and the strategic incentives to add capacity into strength. Every NAND cycle of the modern era has ultimately been ended by a Samsung capacity decision, and it is no accident that last week's selloff was catalyzed in part by reports of Samsung and SK hynix supply additions. Samsung is also the natural adversary to Sandisk's HBF gambit (next section): it has its own AI-flash concepts and zero interest in adopting a standard co-authored by two rivals. Watching Samsung's memory capex line is, bluntly, watching the expiration date on this cycle.

Kioxia is not a competitor so much as a conjoined twin. The joint venture fabs in Yokkaichi and Kitakami produce for both companies on the same technology (BiCS), with output split roughly down the middle. This is the single most distinctive feature of Sandisk's business model, and it cuts both ways with unusual force. The upside: Sandisk gets leading-edge manufacturing scale at half the capital intensity — the enabling condition for that absurd 1.4%-of-revenue capex print — and the JV's aggregate size (the world's #2–3 NAND complex, combined) gives both partners scale neither could afford alone. The downside: Sandisk cannot unilaterally add or withhold supply — every capacity decision is a negotiation with a partner whose own finances (Kioxia carries heavy debt from its private-equity-era buyout and has its own public listing and its own investors to please) may point the opposite direction at the worst moment. In a disciplined upcycle the twin structure is a stabilizer: neither partner can defect cheaply. In a downturn it historically meant the JV kept running wafers when independents would have cut. The JV is the reason to trust the discipline and the reason to doubt the flexibility.

SK hynix + Solidigm is the #2 complex and the most interesting strategic actor: it is simultaneously Sandisk's HBF partner, its NAND competitor, and the world's HBM leader — meaning its wafer allocation decisions (HBM vs DRAM vs NAND) move the whole memory complex. Its enterprise-SSD franchise (built on the Solidigm acquisition) is the sharpest direct competitor to Sandisk's datacenter push — Solidigm's high-capacity QLC drives practically invented the category Sandisk is now storming. But hynix is also the producer with the least incentive to flood NAND: every wafer it adds to NAND is a wafer not making HBM at margins NAND cannot match even now. Its rational posture is exactly what the market observed overnight before this report: enjoy the NAND upcycle passively while its Korean neighbor absorbs the strategic attention. Hynix +11% on a single Seoul session was the complex-wide verdict that memory's repricing is not one company's story.

Micron plays NAND as a disciplined #5, explicitly deprioritizing it for HBM — a live demonstration of the substitution drain, and a strategic gift to Sandisk: the industry's most financially capable American producer has chosen to point its capital at a different product. And YMTC, the China wildcard, remains the long-fuse threat: technologically credible (its Xtacking architecture is genuinely innovative), export-control-constrained on equipment, and pointed primarily at China's domestic market — for now. The YMTC risk is not 2026; it is that the next downturn arrives with a state-backed producer that does not require a return on capital, competing in the commodity tiers exactly when the commodity tiers are the only ones still generating cash. Handicapping the board as a whole: five of the six meaningful producers currently have rational reasons for restraint; the entire supply-side bet is that the sixth — Samsung — chooses profit over punishment. Samsung's own margin recovery argues for restraint; Samsung's forty-year institutional reflex argues otherwise; and the announced additions that triggered last week's selloff are the first data point suggesting the reflex is stirring.

8. The demand engine: what AI actually does to storage

The lazy version of the bull case says "AI needs storage." The precise version is worth spelling out, because it explains both the violence of the inflection and its concentration in Sandisk's datacenter segment (+233% sequentially) and edge segment (+118%).

AI systems consume storage at three layers. Training-side data infrastructure: corpora, curated datasets, checkpoints — petabyte-scale, but a one-time-ish build. Inference-side working sets: this is the structural surprise of 2025–26. Serving models at scale means vector databases, retrieval indices, context/KV caches spilled to flash, session state for hundreds of millions of users — storage that scales with usage, not with model count, and that must be fast, dense, and power-efficient, which excludes hard drives. The displacement bid: as hyperscalers rebuild datacenters around AI power budgets, high-capacity QLC enterprise SSDs (30–120TB class) are displacing nearline HDD in a widening band of workloads — flash's decades-old promise of killing the hard drive, finally arriving not on cost-per-bit parity but on power, density and latency per rack in facilities where every watt is contested. Sandisk's edge segment surge tells the same story from the device side: AI-capable phones and PCs carry higher NAND content per unit, and mobile OEMs — burned by 2025's shortages — are contracting forward. The edge line ($3.7B last quarter, up 118% sequentially — the largest single segment) is easy to misread as consumer exposure; it is closer to the opposite. These are OEM supply agreements with handset and PC makers, negotiated quarterly-to-annually, and the sequential doubling reflects both allocation pricing and a genuine content story: on-device AI pushes flagship phone storage toward the terabyte and AI-PC minimums upward, while inventory-scarred OEMs pre-buy against the datacenter's crowding-out of their supply. It is a second-derivative AI trade wearing a handset costume — less glamorous than the datacenter line, twice its size.

The power arithmetic behind the displacement bid deserves its own sentences, because it is the least cyclical part of the demand story. A modern AI datacenter is provisioned against a fixed power envelope — the gigawatt is the scarce input, as our CoreWeave report belabored — and within that envelope, every watt spent spinning platters is a watt not spent on accelerators. High-capacity QLC SSDs deliver several times the terabytes per watt and per rack-unit of nearline HDD; when power is the binding constraint, flash displaces disk even at a multiple of the cost per bit, because the relevant denominator changed from dollars to watts. This is the first NAND demand driver in history that strengthens as electricity tightens — a hedge, partial but real, against the very AI-capex digestion that threatens the rest of the thesis.

Two honest caveats belong here. Consumer revenue actually fell 10% sequentially last quarter — the traditional bit bucket is being crowded out by pricing, not growing — so this boom is narrower than the headline. And the HDD-displacement thesis has been declared prematurely victorious in three prior cycles; Seagate and WD's HAMR roadmaps keep hard drives on a falling cost curve too. The demand story is real, large and probably durable — but it is a datacenter capex derivative, which means it inherits the AI capex cycle's own mortality tables. The same Morningstar warning that knocked 14% off this stock was not about NAND at all; it was about the customer.

9. HBF: the option that isn't in a 9× forward multiple

The most strategically interesting thing Sandisk owns is not in this year's income statement. High Bandwidth Flash (HBF) — co-developed with SK hynix, MOU signed August 2025, global standardization effort launched February 2026 through the Open Compute Project — is an attempt to create an entirely new memory tier: NAND stacked and interfaced like HBM, targeting HBM-class bandwidth with 8–16× the capacity at similar cost, sitting between HBM and SSDs to address AI inference's "memory capacity wall." First samples are slated for H2 2026, first devices sampling early 2027, with analysts projecting meaningful volume closer to 2030.

Why it matters, in one paragraph: inference economics are increasingly bottlenecked not by compute but by how much model and context can sit close to the accelerator. HBM is the answer today, and HBM is scarce, expensive, and capacity-limited by physics (DRAM cells don't stack like NAND — a 12-high HBM stack tops out in the hundreds of gigabytes per accelerator, while models, contexts and retrieval sets want terabytes). HBF's architecture is essentially HBM's packaging playbook applied to NAND: flash dies stacked and connected with through-silicon-via-class interconnect, presenting HBM-like parallel bandwidth to the accelerator while accepting NAND's higher latency — acceptable for the read-heavy access patterns of inference (weights are read millions of times, written once). If a meaningful slice of inference working sets can live in flash at HBM-adjacent bandwidth, HBF becomes a socket — qualified, designed-in, contracted years forward — rather than a commodity. Sockets are what turned Micron's HBM into an 84% gross margin annuity and what separates "component supplier" from "roadmap partner." A Sandisk that ships HBF at volume in 2028 is a structurally different company from the one priced today at 9× forward earnings; that multiple contains approximately zero dollars for it.

The equally honest counter-paragraph: HBF is a standard-in-progress, not a product — and the graveyard of "new memory tiers" is crowded (Optane died there, expensively, a decade's cautionary tale: technically superior, architecturally orphaned, killed by the absence of exactly the ecosystem adoption HBF now chases through OCP). Samsung will fight the standard or fork it. NAND's latency and endurance physics impose real engineering constraints that marketing bandwidth numbers elide. Nvidia's architectural choices — not Sandisk's roadmap — will decide whether the tier exists. And the partner who co-owns the spec, SK hynix, is also the incumbent whose HBM franchise HBF partially cannibalizes, an alignment that could curdle. Treat HBF as a genuine, asymmetric, unpriced call option — with 2030 expiry and meaningful probability of worthlessness. That is not a criticism; unpriced options are the best kind to receive for free. It is simply not the reason to own or avoid the stock this year.

Why did hynix co-sign at all, given the cannibalization? The strategic logic is instructive: HBM's capacity ceiling is real, inference demand is growing faster than HBM supply physics can accommodate, and if a flash tier below HBM is inevitable, hynix profits more from co-owning the standard (and supplying both tiers) than from watching Sandisk standardize it with someone else — or worse, watching Samsung do it. For Sandisk, the calculus is simpler and starker: HBF is the only visible path by which a pure NAND company acquires what Micron's HBM has and Sandisk's NAND lacks — qualification friction and multi-year contractual cover. Every other element of Sandisk's future margins rests on industry behavior; HBF is the one element that would rest on a socket. That is why a technology three years from revenue belongs in the middle of a report about a stock that moved 26% last week: it is the only candidate answer to this report's central question that does not depend on the prisoner's dilemma holding.

10. The supply-side risk ledger: reading the cycle's suicide gene

Every memory upcycle carries the gene of its own destruction, and it expresses through the capex line. The ledger as of July 2026: Samsung and SK hynix have announced supply additions (the proximate catalyst for last week's selloff); Kioxia's balance-sheet repair as prices recover will eventually permit — and its investors will eventually demand — JV expansion; YMTC adds what the export-control regime allows; and every producer's margin structure now screams for more wafers. Against that: fab construction takes 18–30 months, the 2023-era capex cuts left a genuine hole in 2026–27 bit supply, the DRAM/HBM substitution drain continues, and — the genuinely novel variable — every producer's management is publicly committed to discipline and, so far, behaviorally consistent with it.

The arithmetic that decides everything: NAND demand bits are growing at perhaps 15–20% a year in this AI-inflected regime; industry cost-per-bit falls 15–25% a year on layer scaling; so prices can fall meaningfully every year while margins hold — if supply bits grow no faster than demand bits. Every historical bust came from supply bit growth in the 30%+ range meeting demand growth in the teens. The entire bull case compresses to one monitorable claim: announced and financed wafer additions through 2027 keep industry bit growth below demand bit growth. The entire bear case compresses to its negation, plus thirty-five years of base rates. This is refreshingly falsifiable, and Part Three's tracking dashboard is built around it.

It is also worth naming the game-theoretic structure underneath, because it clarifies why "discipline" is such a fragile equilibrium. At an 80% gross margin, each producer's individually rational move is to add a little capacity — their own addition won't crash the price, and the margin on incremental bits is spectacular. But every producer runs the same calculation, the additions sum, and the price crashes for all: a textbook prisoner's dilemma that the industry has historically resolved in the defecting direction every single time. What is genuinely novel in 2026 is the set of side payments changing the payoff matrix: hynix and Micron earn more redeploying wafers to HBM than defecting in NAND (defection is dominated, not resisted); Kioxia and Sandisk are structurally slowed by the JV's bilateral veto; YMTC is equipment-capped. The dilemma has, temporarily, only one free player. That is a real structural difference from 2018 or 2010 — and it is also why the entire equilibrium hangs on a single Korean boardroom, which is an uncomfortable amount of thesis concentration for any $258B market cap.

11. The Micron mirror: same paradox, purer expression, shorter leash

Set this company beside the one that opened our trilogy and the comparison does analytical work. Both are memory producers printing unprecedented margins into an AI demand shock. Both trade on the identical valuation paradox — a premium trailing multiple and a distressed forward one (MU: 22× / 6.5×; SNDK: 59× / 9.3×). Both were repriced violently upward over the same twelve months, and both just got hit by the same macro tremor (on July 3, MU fell 5.5%; SNDK fell 14%).

The differences are where the insight lives. Micron's crown jewel is contracted; Sandisk's is not. HBM sells on multi-year agreements, qualified socket by socket, sold out through 2027 — Micron's margin has a contractual floor under a meaningful revenue slice. NAND still prices quarterly-to-semiannually at best; Sandisk's "multi-year supply frameworks" are an aspiration being negotiated, not a book of signed take-or-pay. That single difference explains the beta gap in both directions — SNDK rose further on the way up and falls harder on every wobble — and it is why Burry's short went to Micron while the market's nerves show up first in Sandisk's tape. Micron is diversified across DRAM+HBM+NAND; Sandisk is pure NAND — the cleanest instrument in the market for expressing a view on exactly one commodity's cycle. And in the strangest symmetry: Sanjay Mehrotra co-founded this company, was denied its leadership in 2016, and now runs its mirror. The two best-performing large-cap memory stories of 2026 are, in a genealogical sense, the same story — separated at birth by a hard-drive acquisition, reunited by the same customer's insatiable appetite.

Part Three · Synthesis: where Sandisk stands, and what the number says

12. Where it stands: the July 2026 snapshot

Read the table's internal argument: the operating rows describe the fastest fundamental repricing in large-cap technology; the valuation rows describe a market bracing for it to reverse; and the segment row quietly warns that the boom is narrower than the headline — a datacenter capex derivative wearing a consumer brand's name.

Separate, carefully, what changed last week from what did not. Did not change: the FQ4 guide, the undersupply, the framework negotiations, the balance sheet, the HBF timeline — every company-level fact in this report predates the drawdown. Did change: a research house put a number (20–30% giveback) on AI-trade froth; Samsung and hynix supply additions moved from hypothetical to reported; and the calendar flipped to H2, the traditional moment for institutions to harvest a 535% year-to-date winner. In other words, the week repriced the probability weights across Part Three's scenarios — modestly toward the base and bear paths — while leaving the scenarios themselves untouched. A 26% drawdown for that reweighting is either the market being efficient about tail risk or momentum being mechanical about exits; the volume signature (heaviest selling into the close, no capitulation gap) reads more like the latter, but readers should hold that judgment loosely: positioning is the one variable this report's methods observe worst.

The one-year price path, annotated with this report's vocabulary: $28 (April 2025, the impairment bottom) → $52 (August, discipline holding) → $102 (September, contract-price ignition) → $186 (October) → $275 (New Year) → $576 (late January, the FQ2 earnings detonation) → $989 (April) → $1,562 (May, FQ3 guide) → $2,354 intraday June 24 (euphoria) → $1,745 (July 3: Morningstar's 20–30% giveback warning + Samsung/hynix supply-addition reports + H2 rebalancing; −14% in one session, −26% for the week, on zero company news). A stock that moves 26% in a week without news is not reacting to information about the company; it is reacting to information about its own crowd — the surest sign that positioning, not fundamentals, sets the marginal price at these altitudes.

13. Management and capital allocation: the discipline IS the product

CEO David Goeckeler — who ran Western Digital and, notably, chose to lead the spinoff rather than the remaining parent — inherited the industry's most interesting capital structure and has so far played it with unusual restraint. The choice itself was informative: in early 2025, picking the flash spinco over the cash-generating HDD parent meant voluntarily boarding what the market priced as the sinking half of the ship. Executives allocate themselves like capital; Goeckeler's self-allocation was the single loudest insider signal at the bottom, and essentially nobody read it. The numbers make the case better than adjectives: at an 80% gross margin, with customers begging for supply, Sandisk spent 1.4% of revenue on capex last quarter. Some of that is the JV's structural accounting (heavy lifting sits at the JV level), but the company's own gross number (4% of revenue) tells the same story. No dividend; no buyback spree yet; a balance sheet swung from strained to net-cash in four quarters; and pricing power exercised through allocation and framework negotiations rather than spot-market gouging of strategic accounts — the behavior of management that intends the cycle to be long rather than tall.

Two structural notes temper the applause. First, the discipline is partly involuntary: the Kioxia JV means Sandisk could not unilaterally splurge even if it wished — and symmetrically, may not be able to unilaterally restrain if Kioxia's creditors and shareholders push for volume as prices recover. The genius and the flaw of this company remain the same clause in the same 2000-era contract. Second, watch dilution and insiders as the euphoria ages: diluted share count is drifting up (145M → 157M in five quarters on spin-related equity grants), a mild but real tax on the ascent.

The capital-allocation decision that will actually define Goeckeler's tenure has not been made yet, and it is worth framing in advance. Within four quarters, this company will hold a cash pile in the tens of billions against essentially no debt, and it will face the memory executive's eternal fork: return it (buybacks/dividends — the "we know this is a cycle" signal, historically rewarded), bank it (the fortress option — Harari's old playbook, which carried SanDisk through 2008), or spend it (JV expansion, HBF industrialization — the "this time is different" bet, historically punished). Watch the August earnings call for the first hints. A management that initiates even a modest buyback while capex stays disciplined is telling you it believes mid-cycle EPS is far above what the forward multiple implies; a management that announces JV expansion is telling you the prisoner's dilemma has begun to resolve the old way. The cash pile makes the next twelve months' words unusually informative.

14. Three scenarios for the year ahead

The bull path — the frameworks sign. Undersupply persists through 2027 (the 2023 capex hole is real; Samsung's additions take two years to yield bits); the multi-year supply frameworks convert from negotiation to signed book, giving NAND its first-ever contractual demand floor; margins plateau in the 70s rather than round-tripping; FY27 EPS lands near the consensus implied by that 9.3× forward multiple (~$185+), and the multiple re-rates toward even 15× as the market concedes semi-permanence. That path revisits the highs and the $3,250 top target. HBF samples validating on schedule adds the 2030 story on top — pure optionality, priced at zero today. The under-appreciated accelerant in this path is the balance sheet: two more quarters at guided run-rate puts roughly $10B of net cash on a company with no dividend and no stated buyback — at which point capital-return announcements become their own catalyst class, the kind that re-rates cyclicals when they signal management's own confidence in mid-cycle earnings power.

The base path — the glide. Margins peak this quarter or next as Samsung/hynix bits arrive through 2027; pricing softens but the crash never comes because discipline half-holds and AI demand keeps growing; GM glides toward the 50s–60s; EPS runs $90–130 annualized rather than $130+; the stock chops violently in a $1,300–2,200 range while trailing and forward multiples converge toward each other — which is what the mean analyst target ($1,864, just 7% above spot) is really encoding: massively positive fundamentals, fully priced. The glide path's distinguishing evidence, if it is engaging, will be mix divergence: datacenter and edge holding pricing while consumer-adjacent tiers leak first — watch the segment margins, not just the blended line, because the blended line will flatter the story for two quarters after the turn has begun underneath it.

The bear path — the essay fails again. AI capex digests in 2027 exactly as the July 3 warning sketched; Samsung chooses share over price as it historically does; Kioxia's recovery unmutes JV volume; consumer demand (already −10% QoQ) provides no cushion; NAND does what NAND has always done, and margins round-trip toward the 30s within 18 months. EPS collapses toward $30–40 annualized; the $1,000 low target is reached not on panic but on arithmetic (a 10× multiple on trough-ish earnings — and note that $1,000 is still 36× the April 2025 low; even the bear case here concedes the resurrection). The tell that this path is engaging: NAND contract prices printing sequential declines while any producer's capex guidance rises.

The asymmetry note that honest analysis requires: unlike CoreWeave (levered, financing-dependent), Sandisk enters whichever scenario arrives with net cash, positive FCF ($2.3B TTM and compounding), and capex it can cut to near-zero — this company cannot be margin-called by its own cycle. The equity can fall a great deal; the company itself is in no scenario fragile. That distinction — equity risk without enterprise risk — is exactly what the 2008 Samsung bid episode teaches: the entity survives its drawdowns and compounds through them.

The near-term catalyst calendar: late July, Samsung's earnings and memory-capex commentary — the single most important exogenous data point of the quarter; early August, Sandisk's FQ4 print against the $7.75–8.25B guide, plus FY2027 framing and — the item this report cares most about — any quantification of signed multi-year supply frameworks; through the quarter, monthly NAND contract-price prints and hyperscaler capex guidance in the Q3 earnings season; H2 2026, the HBF first-samples milestone, dated and binary. Each of these is capable of producing another 15%-in-a-day session in either direction; that is the deal this stock currently offers its holders.

15. The valuation paradox, third verse, loudest rendition

Before the general law, decompose the analyst spread, because $1,000-to-$3,250 is not twenty-two people disagreeing randomly — it is three coherent models fighting. The low targets are normalized-earnings models: assume mid-cycle GM in the 40s, mid-cycle EPS somewhere near $50–70, apply a cyclical's 12–15×, land near $1,000. The mean cluster ($1,700–2,000) is decay models: current earnings, glided down over 2027–28, discounted — the "great year, mortal margin" camp. The high targets are regime-change models: hold the 70s margins, credit the frameworks, sometimes credit HBF, apply a growth multiple. Note what all three share: nearly identical revenue forecasts for the next two quarters. The disagreement is entirely about quarters five through twelve — which is exactly where NAND's history says forecasting is impossible and where this cycle's novelties (discipline, demand mix, substitution drain) say history might not bind. The spread, in other words, is not confusion; it is the honest confession that the distribution is trimodal.

59.5× trailing, 9.3× forward. We have now written this paragraph three times in five weeks — Micron (22×/6.5×), and now Sandisk at a wider spread than either — so let us state the general law the trilogy has surfaced: for an AI-era memory producer, the P/E spread between trailing and forward is the market's real-time estimate of the cycle's mortality. A widening spread means the market believes this margin structure dies faster; a narrowing spread means contractual or structural evidence (HBM's multi-year books, signed NAND frameworks) is convincing it otherwise. On that meter, the market currently judges Sandisk's 78% margin the most mortal of the trilogy — wider spread than Micron — for a defensible reason (no contract cover) and possibly by too much (net cash, capex discipline, and a demand mix upgrade the old base rates never contained).

Run the crude arithmetic yourself: at the current-quarter guide midpoint, Sandisk annualizes to roughly $32B revenue and ~$125 of EPS — about 14× annualized earnings at $1,745, for a company growing revenue triple digits with an 80% gross margin, net cash, and a free option on a new memory tier. That is either the cheapest great business in the S&P or a peak-earnings mirage at 3× book-value-appropriate pricing — and the honest answer is that both descriptions are simultaneously accurate today, and only the supply-bit line resolves them. The July 3 session was a preview of how violently the market will toggle between the two descriptions on every incremental data point. Expect more such sessions; at a 26%-in-a-week realized volatility, the market is pricing this equity like an option on the cycle — which, as with CoreWeave, is not irrationality. It is taxonomy.

History offers a specific calibration for what "the essay fails" looks like from a peak, because memory has run this experiment repeatedly. The 2017–18 memory boom — the closest analog, also demand-shock-driven (cloud datacenter build-out), also accompanied by a sincere "structural change" literature — saw the leaders' gross margins peak in late 2018 and give back roughly half their expansion within four quarters as supply caught up, with the stocks bottoming 45–60% below their peaks while earnings were still near record levels — the market front-runs the margin, not the earnings. The 2000 and 2010 episodes were harsher. Against those base rates, note what a $1,000 bear-case price already encodes: roughly a 57% peak-to-trough drawdown, in line with history's worst-but-one outcomes. The bear case for this stock is not exotic; it is median memory-cycle history applied to a bigger number. The bull case is that the demand mix and the discipline break the pattern. Base rates versus regime change — the same courtroom as Micron, with a purer defendant.

16. The bottom line: what would have to be true

To own Sandisk here, you need to believe: that industry supply discipline survives its first real temptation (80% margins) for at least another 18 months; that Samsung's announced additions are measured, not vengeful; that AI storage demand is a structural re-rating of NAND's demand curve rather than a one-time datacenter build; that the multi-year framework push converts into signed cover before the next soft quarter; and that the Kioxia twin stays financially incentive-aligned. None of these is heroic individually; jointly they ask NAND to behave, for the first time in its history, like an industry rather than a knife fight.

To avoid or short it, you need only the base rate — thirty-five years, zero exceptions, every "this time is different" essay eventually marked to zero — plus the observation that the marginal buyer of the stock this spring was momentum, and momentum's exit shows up as 26%-in-a-week airpockets regardless of fundamentals.

What you cannot do, after this trilogy, is treat these as three separate stories. Nvidia's compute, Micron's memory, Sandisk's storage — one demand shock propagating down the shortage stack, one financing complex (as our CoreWeave report mapped) underwriting the demand, and one shared mortality question mark hanging over every 2027 estimate. The AI hardware trade is a single organism wearing four tickers. Diversification within it is an illusion; sequencing within it — which layer's scarcity outlives the others — is the entire game.

On that sequencing question, the trilogy's own evidence suggests an ordering worth stating explicitly, as hypothesis rather than verdict. Compute scarcity is defended by architecture and software moats (Nvidia's, chiefly) but attacked by every customer's silicon program and, as of last week, by customers reselling their surplus. HBM scarcity is defended by qualification friction and multi-year contracts — the strongest contractual cover in the stack. NAND scarcity is defended by nothing but collective self-restraint — the weakest cover, which is precisely why its margin inflection was the most violent: uncovered scarcity prices spike hardest and die fastest. If the AI build-out plateaus, the unwinding should run in reverse order of contractual protection — NAND margins first, then GPU rental rates, with HBM's contracted book failing last. Sandisk, on this reading, is the trade's canary as well as its purest expression — the ticker to watch even for investors who never touch it, because its monthly contract prices will announce the complex's turn quarters before the contracted names confess.

17. The risk matrix and the tracking signals

As with CoreWeave, the matrix's rows interact, and the most dangerous pairing deserves naming: supply discipline breaking and AI capex digestion are correlated through the same underlying event. If hyperscaler demand plateaus, prices soften; softening prices pressure the weakest producer's cash flow; pressured producers chase volume; volume kills discipline — the two "critical" rows are one cascade observed at two points, which is exactly how 2018 unfolded (cloud digestion in Q4'18 → discipline collapse by mid-2019). Conversely, the risk pairing that would confirm the bull case is also linked: signed multi-year frameworks would simultaneously derisk the demand row and harden the discipline row, because contracted demand removes the incentive to defect. One signing announcement moves two critical rows at once — which is why the frameworks item, superficially a commercial detail, is this dashboard's highest-information signal.

The five-signal dashboard for the next two quarters: (1) NAND contract prices — monthly prints; the first sequential decline while capex guidance rises anywhere in the industry is the cycle's turn signal; (2) the FQ4 report (early August) against the $7.75–8.25B / 79–81% guide — and more than the print, whether multi-year framework signings get announced and quantified; (3) Samsung's memory capex language at its late-July earnings — measured expansion vs share reclamation; (4) hyperscaler Q3 capex guidance — the demand side's tide chart, same signal our CoreWeave dashboard watches; (5) the HBF H2'26 sample milestone — binary, dated, and the only 2026 event that can add a new decade to the story.

Everything in this report was produced by the Aya Invest research platform from public data — the platform demonstrating itself. Sources: Yahoo Finance (quote, fundamentals, financial statements, analyst data, July 4, 2026); Sandisk IR (FQ1–FQ3 FY2026 results and FQ4 guidance; HBF standardization release, Feb 25, 2026); Investing.com (FQ3 FY2026 slides); Futurum (FQ2 FY2026); TrendForce and Tom's Hardware (HBF, Feb 2026); Yahoo Finance/Motley Fool/TIKR (July 1–3, 2026 selloff coverage; H1 2026 performance). For information and research purposes only. Not investment advice.

FAQ

Why did Sandisk's stock rise ~60x from its April 2025 low?

Four forces converged: the deepest industry capex cuts in NAND history (post-2023 crash) starved supply; AI datacenters created explosive enterprise-SSD demand (Sandisk's datacenter segment grew 233% in one quarter); DRAM makers shifted wafers to HBM, further tightening NAND; and the BiCS8 node cut costs into rising prices. Gross margin went from 22.5% to 78.4% in five quarters, and revenue roughly quadrupled — the fastest unit-economics repricing in large-cap tech.

Sandisk trades at 59x trailing but 9.3x forward earnings — what does that mean?

It is the market's confession that it doesn't believe an 80% gross margin on NAND can last. For a cyclical, the P/E spread between trailing and forward is a real-time estimate of the cycle's mortality — and NAND, unlike Micron's HBM, has no multi-year contract cover: prices reset quarterly. That's why SNDK fell 14% on the same day Micron fell 5.5%. Analyst targets run $1,000–$3,250 — three coherent models (normalized earnings, decay, regime change) fighting over quarters five through twelve.

What is High Bandwidth Flash (HBF) and why does it matter for Sandisk?

HBF, co-developed with SK hynix and moving through OCP standardization, stacks NAND like HBM to deliver HBM-class bandwidth with 8–16x the capacity — aimed at AI inference's memory capacity wall. First samples are due H2 2026, devices in 2027, volume closer to 2030. If it becomes a designed-in socket, Sandisk gains what its NAND lacks: qualification friction and multi-year contract cover — the ingredients of Micron's HBM margins. Today's 9.3x forward multiple prices it at roughly zero: a free, asymmetric, long-dated option.