SanDisk (SNDK): The $93.9 Billion Floor

Transcript

Intro

Thomas: Welcome to Second Spark — we break down emerging technologies to show you what innovation will be transforming next. I'm Thomas.

Katie: And I'm Katie. Let's get to the truth.

The Setup

Thomas: It's nineteen eighty-eight, and a physicist named Eli Harari is at a Silicon Valley chip company called Wafer Scale Integration, trying to sell people on a strange idea: that a chip — not a spinning disk, not film — should be where the world stores its data. He'd invented the key piece himself, at Hughes Aircraft in the seventies: the floating gate, a way to make silicon remember with the power off.

Thomas: The problem is the calendar. In nineteen eighty-eight the market for chip-based storage barely exists — the company's mass-storage push stalls, years too early. So on February twenty-eighth, Harari walks.

Thomas: And the next morning — March first — he starts a new company to chase the exact same idea. Two colleagues go with him, Sanjay Mehrotra and Jack Yuan. They name it SunDisk.

Thomas: This time they build it on the lesson: a raw flash chip fails. It wears out. So you don't sell the chip — you build the whole system around it, controller, error correction, everything, until it's reliable enough to replace film and hard drives.

Katie: Today that company is worth two hundred and thirty billion dollars.

Katie: What they sell is memory chips. Memory is the most boom-and-bust commodity in all of electronics.

Thomas: So what's the product, in plain terms? NAND flash — the storage in your phone, your laptop, the card in a camera, the drives in every data center you've ever rented a server from.

Thomas: Five companies on Earth can make NAND at the leading edge — and Sandisk is in the club.

Katie: The club was this small last year, and nobody was paying this kind of money for it. What changed?

Thomas: AI got out of the lab and went shopping. The hyperscalers suddenly needed somewhere to put a civilization's worth of data.

Katie: Hard drives ran short, and the shortage shoved the hyperscalers into QLC enterprise drives. Flash got the AI order because the usual option was gone.

Thomas: Forget the market cap. Look at what the stock itself has done inside a year.

Thomas: Over the trailing fifty-two weeks, this stock is up more than three thousand two hundred percent — the most violent ride the memory trade has ever printed.

Katie: Thirty-three dollars and change tonight for every dollar parked in it last August.

Thomas: So that's the company in front of us: a commodity manufacturer claiming it has re-engineered its own cycle — contracts instead of spot prices, floors instead of crashes. Tonight we test that claim.

History of the Business & Founding

Katie: Start with the founder, then. Where does he come from, and what leads him to the breakthrough?

Thomas: Eli Harari. Born in Israel in nineteen forty-five, a doctorate in solid-state sciences from Princeton in nineteen seventy-three, and then Hughes Aircraft — where he spends his days coaxing electrons through insulating films so thin the electrons tunnel straight through. Out of that tunneling comes the first practical EEPROM: the floating gate, a little raft of trapped charge that still sits underneath every flash chip on Earth.

Thomas: In his own telling, when he's born — June of nineteen forty-five — Israel is still British-controlled, and his parents had come from Poland. In his Computer History Museum oral history he starts the story at eight days old — carried home from the hospital in a pram through a curfew, stopped and searched by British soldiers.

Thomas: The road to Princeton runs through half the map: Israel until thirteen, then English boarding school — his parents wanted an English gentleman — two years as an air force technical clerk, a year of math and physics at the Hebrew University of Jerusalem, a physics degree from Manchester in nineteen sixty-nine. He lands in America to start the doctorate a month after Neil Armstrong steps on the moon.

Katie: And the doctorate itself — what problem was he actually working on? Nobody funds thin films for their own sake.

Thomas: The Office of Naval Research does. The navy wanted to know why satellites kept dying after about three months in orbit, and the radiation question put Harari onto ultrathin films of silicon dioxide — his words: "it turned out, to my great luck, that I was working on the right materials."

Thomas: At Hughes he sets the challenge for himself: an EPROM you could erase electrically. The trick is thinning the insulating oxide under the gate from a thousand angstroms to about a hundred — thin enough for electrons to tunnel through on command — then proving it reproducible across tens of thousands of samples. Out of it comes the industry's first floating-gate EEPROM chip, eight kilobits, in production around nineteen eighty.

Thomas: The paper he publishes on those thin oxides finds exactly the right reader: Dov Frohman, inventor of the EPROM, who jumps on the next plane, comes to Hughes, and recruits him to Intel. That's how Harari reaches Silicon Valley in nineteen seventy-nine.

Thomas: And at Intel, in nineteen eighty, Ted Hoff hands him the sentence that becomes the operating principle for everything after: build a memory technology at ten times lower cost than the incumbent, and the entire hierarchy of memories will move aside to let it in. Harari says Hoff was right on the money.

Katie: That's a cost-curve rule, and it explains everything this company does from here on — hold onto it, because it comes back tonight when we reach the new product.

Thomas: The boast his National Inventors Hall of Fame profile preserves: a billion bits in a footprint the size of a printed period.

Thomas: In the eighties, at Wafer Scale Integration, Harari bets that EEPROM could be mass storage — the right destination, reached too early. The board gently pushes him out, and at forty-three he says you could have called him a failure. He walks on February twenty-eighth, nineteen eighty-eight, and starts the next company March first — optimistic as hell, in his own words.

Thomas: He doesn't start it alone: SunDisk, founded with Sanjay Mehrotra and Jack Yuan. Hold Mehrotra in your head, because today he runs Micron — the archrival. Raw flash in that era wears out and drops bits — so SunDisk designs the memory, controller, error correction, and wear management as one object, until solid-state storage can stand in for film and the spinning disk.

Katie: If the system was the insight, why did flash trade like a brutal commodity for decades — who captured the system premium?

Thomas: Honest answer — for most of that run, nobody did. The premium kept leaking into the price war underneath it.

Thomas: Nineteen ninety-one, and out the door goes the first flash solid-state drive ever shipped: twenty megabytes, built for IBM, about a thousand dollars.

Thomas: The story under that sale: the bid for the original pen ThinkPad, won by a company of about forty people against Western Digital and Texas Instruments — Harari swears IBM had more people helping SanDisk deliver than SanDisk had employees.

Katie: Then comes multi-level cell — more than one bit stored per transistor — and the Hall of Fame puts the cost collapse beyond a hundred-thousand-fold across two decades.

Thomas: Then they go public. November nineteen ninety-five: the Nasdaq bell — although rung under a new name, because a Sun Microsystems trademark kills "SunDisk" on the way to the podium.

Thomas: The original name came from his fifteen-year-old daughter, in the car near Fremont — a sunny, uppity name, she said. The replacement was Harari's own coinage — SanDisk as in sand, silicon disk — and Sun paid the changeover costs.

Thomas: CompactFlash arrives in nineteen ninety-four, and at the turn of the century they co-write the SD card standard — the little card that ends up in every camera bag on Earth.

Thomas: One decision from that era tells you the scale of the ambition. Mehrotra, on the TechSurge podcast in twenty twenty-four, remembers Kodak wanting the young company's digital film business all to itself: "Kodak had interest in making our digital film an exclusive for Kodak. And of course, at that time, you know, Eli Harari, kudos to him and the team chose not to make our film an exclusive to Kodak. Rest is history."

Katie: Refusing the exclusive is the same system thinking, applied to distribution.

Thomas: On the same show, he walked the demand story across the decades: first digital film, then music players, then the cloud — and now, in his words, "AI is going across all of these end market segments."

Katie: May, two thousand: SanDisk and Toshiba form the flash manufacturing joint venture — fabs at Yokkaichi. Each side funds roughly half the equipment; each side buys roughly half the wafers at cost-plus.

Katie: A startup at the bleeding edge of the hardest manufacturing on Earth, on a split bill. The cost collapse everyone credits to the physicists was financed on someone else's buildings.

Thomas: October, twenty fifteen. Western Digital agrees to buy SanDisk for about nineteen billion dollars. Then a planned Chinese investment in Western Digital itself collapses, the terms get rewritten, and the deal closes in May of twenty sixteen at about sixteen billion. Eli Harari had retired back in twenty ten — by the time these papers are signed, it is a founder's company with the founder long gone.

Thomas: A year in, Toshiba — the other half of the joint venture — is staring down bankruptcy, and to survive it moves to sell its memory unit.

Katie: Why? A year earlier Toshiba was solvent enough to be everybody's manufacturing partner. What breaks that fast?

Thomas: It had been breaking in public for years. July, twenty fifteen: Toshiba's chief executive resigns after an independent panel, led by a former Tokyo prosecutor, finds the company inflated profits by more than a billion dollars across six years — a culture, in the panel's words, where underlings could not challenge powerful bosses. So the credibility is already spent when the real hole opens — and the real hole is nuclear.

Katie: Toshiba owned Westinghouse — the American nuclear-plant builder — a majority stake bought in two thousand six for five point four billion dollars.

Thomas: Westinghouse was building four reactors of its new flagship design in Georgia and South Carolina, sold on a bold promise: cheap power, and construction in roughly thirty-six months from first concrete to fuel. The industry's own status report later put it flatly: all of those projections went spectacularly wrong. By early twenty seventeen the projects were about three years behind schedule and billions over budget — the only new reactors being built in the United States.

Katie: And the contracts were fixed-price: Toshiba was committed to finishing the reactors at pre-set customer costs, eating every overrun beyond them.

Katie: Then the move that sealed it. In late twenty fifteen, Westinghouse bought its own construction partner — the contractor on all four builds — for two hundred twenty-nine million dollars plus assumed liabilities, which settled the fights over who owed what by moving the entire cost of completion inside Westinghouse.

Thomas: December twenty-seventh, twenty sixteen: Toshiba warns the deal will force a writedown of several hundred billion yen — several billion dollars — against a goodwill line of eighty-seven million. The stock drops a quarter in a morning.

Katie: February fourteenth, the number lands: a six point three billion dollar writedown, the chairman resigns, and Toshiba forecasts a three point four billion dollar net loss for the year.

Thomas: Six weeks after that, Westinghouse — the most prolific builder of nuclear plants in history — files for bankruptcy in a New York courtroom.

Thomas: That is why the memory unit goes on the block. Western Digital invokes its consent rights under the joint venture and drags Toshiba into arbitration — May, twenty seventeen. Toshiba swings back with a counter-suit for about a billion dollars and briefly locks Western Digital's people out of the joint venture's own data.

Thomas: But Toshiba pushes ahead anyway. September, twenty seventeen: with auditors balking at its earnings and its listing at risk, it agrees to sell the memory business to a Bain-led consortium for eighteen billion dollars — Apple, Dell, SK hynix, Kingston, and Seagate among the buyers — over Western Digital's repeated objections.

Katie: December brings the global settlement: the arbitration ends, the sale proceeds — the business later renames itself Kioxia — and Western Digital's price is joint investment in the new fab, plus the joint ventures rolling on.

Katie: The sale closes June first, twenty eighteen, restoring positive shareholder equity and saving Toshiba's listing.

Katie: So hold the shape of that: the fab partner SanDisk can never leave has already been through a forced sale — over cost overruns at two nuclear projects in the American South. Nothing about it was a memory problem.

Thomas: May of twenty twenty-two. Elliott Investment Management writes the board an open letter: separate the drives from the flash — there is, Elliott argues, a path past a hundred dollars a share in it.

Thomas: The split takes years more. Announced late October, twenty twenty-three; completed February twenty-first, twenty twenty-five — trading opens that Monday. And the detail I keep turning over: David Goeckeler, Western Digital's own chief executive since twenty twenty — the one man with a choice of chairs — walks out with the flash company.

Katie: He'd seen both sets of books, and he picked this one. Revealed preference beats any press release.

Thomas: Now look at the debut. It comes out of the gate in February a forty-five to fifty dollar stock. But by the seventh of April — the tariff selloff in full roar — it prints under twenty-eight dollars. The all-time low, weeks into its life.

Thomas: Then the first report card. May, twenty twenty-five — the first quarterly release as a standalone company — and one sentence tells you exactly who this company is. Goeckeler: "We have taken actions," he writes, "to reduce supply to match demand" — and, same sentence, "commenced price increases this quarter." Cut the supply. Raise the prices.

Thomas: The technology never broke its chain — floating gate, to more bits per cell, to the stacked towers of modern NAND.

The Technology Today

Thomas: Before any of the money makes sense, you need the mechanism. On this month's earnings call, Goeckeler compressed the whole thesis into one sentence: "AI is fundamentally a memory-centric storage-intensive problem."

Katie: That sentence needs defending. The model math happens next to the GPU, in high-bandwidth memory — storage is where data sits. Why does serving a model touch flash at all?

Thomas: Start with what the company actually sells. Three end markets: data center — enterprise drives for the AI buildout, including a QLC drive named Stargate launched this past quarter; edge — flash soldered into other people's machines, the PCs and phones and cars; and the retail aisle you already know, cards and drives under the SanDisk brand.

Katie: And the mix moved violently: data center was roughly twelve percent of the company's bits in fiscal twenty twenty-five. Exiting fiscal twenty twenty-six, it's thirty-eight percent of the portfolio.

Thomas: The whole factory swung toward the data center inside a year. Here's why. When a model serves you, it doesn't just hold its own weights in memory — it builds a working memory of your conversation. For every token you've exchanged, it keeps a pair of vectors — the key-value cache — and it consults them to produce each next word.

Thomas: Nvidia's own engineers state the constraint flatly: that cache grows linearly with prompt length, and it has to sit in GPU memory during generation for fast access.

Katie: The arithmetic is brutal. Take Meta's seventy-billion-parameter model at full precision: about two and a half megabytes of cache per token. One long-prompt request: about twenty gigabytes. A batch of thirty-two requests: six hundred forty gigabytes of cache — nearly five times what the model's own weights occupy. The cache now routinely outweighs the model.

Thomas: And long context turns it into a storage problem outright. At a context of a hundred twenty-eight thousand tokens, a single user's cache reaches forty gigabytes — half of the eighty gigabytes of high-bandwidth memory on the data-center GPU serving them. One conversation, half the most expensive real estate in computing.

Katie: Then the obvious out: throw it away and recompute it. Compute is the one thing these companies have. Why store the cache at all?

Thomas: Because re-reading beats re-thinking. Nvidia's answer is a hierarchy: its serving stack evicts cache blocks out of GPU memory, down through CPU RAM, local solid-state drives, and network storage — and its engineers demonstrate re-loading a precomputed long prompt instead of recomputing it, with storage vendors showing thirty-five gigabytes a second flowing back into a single GPU.

Thomas: In January, Nvidia standardized the whole pattern: a platform that routes inference context onto NAND-based drives at cluster scale, claiming up to five times the power efficiency and five times the tokens per second of ordinary storage paths.

Thomas: The frontier labs already live this way. DeepSeek's own open-source file system serves the cache from flash drives over the network instead of DRAM, with aggregate reads north of six terabytes a second across a cluster.

Katie: So conversation state spills downhill until it lands on the cheapest medium still fast enough to hand it back. That's the cache.

Thomas: The mixture-of-experts trap. DeepSeek's flagship model is six hundred seventy-one billion parameters — roughly two-thirds of a terabyte on disk at eight-bit precision — and only thirty-seven billion of those parameters fire on any given token. The compute touches a sliver of the model at a time; the storage holds all of it, and every checkpoint, copy, and version multiplies the footprint.

Thomas: Then retrieval. IBM demonstrated a hundred-billion-vector database on a single server — north of a hundred fifty terabytes of vectors and index, sitting on forty-eight flash drives, feeding six GPUs at over ninety percent recall, with queries coming back, on average, in about seven hundred milliseconds.

Thomas: And the newest draw is agents. Jensen Huang, at CES in January: "AI is no longer about one-shot chatbots but intelligent collaborators that understand the physical world, reason over long horizons, stay grounded in facts." An assistant that remembers you across sessions is a stored-state product — and the state has to live somewhere cheap enough to keep.

Katie: Mehrotra — the co-founder who now runs rival Micron — said it from the other side of the fence, on the Bloomberg Talks podcast this May: "Memory is a strategic asset for AI," he said — "Because without memory, you don't really have that intelligence."

Katie: Now price the shelves, because the hierarchy is the whole business case. The top shelf, high-bandwidth memory — analyst estimates only, there's no public market — runs somewhere between six and eighteen dollars a gigabyte depending on generation.

Katie: The middle shelves at street prices: a thirty-two gigabyte kit of ordinary PC memory went from eighty or ninety dollars in mid twenty twenty-five to a three hundred seventy-five dollar floor by this June, while a one-terabyte consumer flash drive went from about seventy-five dollars to roughly two hundred.

Katie: Part of the reason: high-bandwidth memory is a wafer hog. It takes about three times the silicon per gigabyte of standard DRAM, earns its makers three to five times the revenue per wafer, and now consumes twenty-three percent of the world's DRAM wafer capacity. The AI order starves every shelf below it.

Katie: The bottom shelf is the hard disk, which is still roughly six times cheaper per unit of capacity than flash — a gap the disk industry itself expects to persist.

Katie: And flash keeps bending its own curve: the newest node packs sixty percent more bits into the same silicon than the one two generations back.

Thomas: So every workload finds its shelf. What cannot wait pays the top-shelf price. What can wait a beat rides flash. What can wait longer lives on disk. The repricing happened because AI dropped a brand-new class of warm, read-heavy, latency-tolerant data onto the middle shelf — and the middle shelf wasn't sized for it.

Katie: Which leaves the question that decides the whole forward model. Is this a one-time restocking of a mis-sized hierarchy — the world bought too little flash for the AI era, corrects it once, and demand normalizes? Or is it a recurring per-token draw — every conversation, every agent, every retrieval pulling storage as it runs?

Katie: Both readings have a live print. The recurring case: demand growth projected at twenty to twenty-two percent this year against supply up only fifteen to seventeen — shortage anticipated all year.

Katie: The restock case has a calendar: the analyst Karl Ackerman expects prices to have peaked about now — mid twenty twenty-six — and to start falling quarter over quarter as early as twenty twenty-seven, and spot buyers are already balking.

Katie: If it's a restock, the collapse arrives on schedule, like every cycle before it. If it's per-token, the floors never even get tested. That fork is what the financial model sits on.

Financials & Valuation

Thomas: All right, take me through the numbers.

Katie: Four questions. Is this margin durable. Was the growth price, or volume. How does the company look against its own industry. And what has the market already paid for. The machine first, though.

Katie: The factory is not theirs. Flash Ventures: three joint ventures, each one forty-nine point nine percent Sandisk-owned — just under half, by design — running seven fabs inside buildings Kioxia owns, with Kioxia manufacturing the wafers at cost.

Katie: Sandisk takes roughly half the output, owes half the fixed costs whether it buys or not, and cannot fabricate flash anywhere else. No second source, no escape hatch.

Thomas: A leading-edge fab is among the most expensive objects humans build. Somebody paid. Where's the bill?

Katie: Inside the ventures. They own or lease the equipment, the walls are Kioxia's, and Sandisk co-funds roughly half the capital when venture cash runs short — which is why its reported capital spending looks so tiny.

Katie: The wafer transfers at cost plus a sliver of markup, and wafer cost is mostly equipment depreciation plus keeping a cleanroom alive. So margin isn't a recipe — it rises and falls with utilization. And it has fallen: idle capacity billed Sandisk two hundred forty-nine million dollars in fiscal twenty twenty-four, then seventy-five million the year after.

Katie: This past April the joint-venture terms were extended through the end of twenty thirty-four.

Katie: That's the supply side. The customer side: ten agreements across eight customers — hyperscalers and large OEMs, every name undisclosed by design — with committed volumes quarter after quarter; miss your purchases, and the commitment converts into compensation.

Katie: Behind them, sixteen and a half billion dollars of financial guarantees — cash deposits and financial instruments the customers posted.

Katie: And the total expected revenue across the agreements, assuming every price lands at its floor: a minimum of ninety-three point nine billion dollars.

Katie: Terms run up to five years, and management expects the agreements to cover most of the company's bits within two fiscal years — this framework is the backlog now. Which means the concentration cuts both ways: eight buyers carry the book, and one paused buildout transmits straight through it.

Thomas: For decades, hyperscalers played five suppliers off one another and prices only went the buyers' way. A shortage inverts the game — now customers post collateral to get chips.

Thomas: So what's the margin at the floor?

Katie: Floors and ceilings on the committed volumes, and roughly eighty percent gross margin even at floor pricing — that's management's plan, in management's own words.

Katie: The shape of it, from my home turf: a reinsurance treaty in a hard market. After the catastrophe year, capacity vanishes, and buyers who swore they'd never pay up sign multi-year cover at prices they hate. The floor is the premium.

Thomas: Then the clause the whole structure hangs on. Goeckeler, asked on the spring call what happens if a customer stops buying: "If they do not meet their obligations on consistent purchasing every quarter, that financial commitment immediately comes to us. We do not expect to collect those because our customers are extremely serious about needing this product."

Katie: He does not expect to collect — which means the mechanism has never fired. No real winter has ever tested those floors, the ceilings cap the recovery on the way back up, and the very next quarter's guidance already flattens.

Katie: And the missing evidence, plainly: no named customer, no switching math — it doesn't exist in public. The collateral is the best evidence on Earth, and the only evidence.

Katie: One concession, because it's earned. I walked in reading these contracts as cycle theater. But customers do not post collateral for a commodity they expect to get cheaper. That moved my read: the model is real. Untested — but real.

Katie: Now the first question, the margin. Fiscal twenty twenty-three, this business kept seven point one cents of gross profit on every dollar it sold. Then sixteen point one. Then thirty point one. A fiscal year per step.

Katie: Then the June quarter printed eighty-four point six percent gross margin, non-GAAP.

Katie: And running the whole company now costs roughly five and a half cents of every revenue dollar — not because anyone cut spending; revenue exploded and the spending never caught up.

Thomas: But that single-digit year the staircase starts on wasn't a floor — it was a crater. One cycle back this business carried a thirty-six percent gross margin; then revenue fell thirty-eight percent, and the margin caved to the step you opened on.

Thomas: And here's my reframe, because the room reads that winter as the reason to disbelieve the floors. Nobody writes a guaranteed minimum into a contract in good times — you write it because that quarter happened to you.

Katie: Management's September-quarter guide: revenue ten point three to ten point eight billion, gross margin eighty-three to eighty-five percent.

Katie: The long model from Investor Day, management's own numbers for fiscal twenty twenty-eight through thirty: mid-to-high-teens growth, roughly eighty percent gross margin, seventy-five operating, half of revenue as free cash, all excess returned.

Katie: Stack those against the peak print: the walk-down is conceded in their own arithmetic. The only argument left is the slope.

Katie: So my rule, stated before it's needed: the day a sequential gross-margin print undercuts the trough management itself just promised, the bottom falls out of the story.

Katie: Second question — the growth. The law of the land: this industry ships about thirty percent more bits every year while the price of a bit falls fifteen to thirty percent a year.

Katie: Against that law, the supercycle: fiscal twenty twenty-six revenue was twenty and a quarter billion dollars, up one hundred seventy-five percent — on bit shipments roughly flat. The year before was the ordinary kind: up ten percent, split six points of volume to four of price.

Katie: Same physical product; price did everything. Repricing, not adoption — and repricing is the kind of miracle that runs in reverse.

Katie: How big is the market this happened in? Three public answers that do not agree: Mordor Intelligence says about fifty-nine billion dollars this year, Coherent Market Insights seventy-eight, and the company's own CFO told Investor Day north of three hundred billion.

Katie: Those aren't measurements of the same thing, and we can't resolve them tonight. On the durable size of this market, the honest answer is: we don't know.

Katie: Underneath all three, one record: five years through twenty twenty-five, essentially zero industry dollar growth — while bits grew about thirty percent a year.

Thomas: What about the margin profile of peers?

Katie: Micron's spring quarter: revenue about forty-one and a half billion dollars, up three hundred forty-six percent, a GAAP gross margin of eighty-four point six percent, operating margin just over eighty. Same figure as Sandisk's print, different basis — and Micron is three-quarters a DRAM company.

Katie: Kioxia is the cleanest comparison in existence — same wafers, same fabs. Its April-to-June quarter: revenue up more than four hundred percent year over year, gross margin eighty percent, operating margin seventy-five.

Katie: Samsung's semiconductor division ran roughly a seventy percent operating margin in the same quarter, and SK hynix a record seventy-six, mostly on DRAM — neither discloses a flash-only line.

Katie: So the print is not a company invention — the whole memory complex is at historic margins in the same season. The cycle set the altitude. The gap over Kioxia, on Sandisk's own wafers, is the part Sandisk can claim: mix and pricing. Real, and modest.

Thomas: Then answer the strong version of the bear: Samsung and SK lead the enterprise-drive catalog, and they climbed past two hundred layers first — the bonding tricks too.

Thomas: So why does this company get to exist at all?

Katie: Three structural answers. The only branded consumer franchise at this scale — a retail aisle that clears premium prices. Integration from wafer to system, controllers designed in-house. And late to enterprise drives but fast: two hyperscalers in qualification, five engaged.

Katie: Goeckeler put the identity claim on stage at Investor Day: "We own the whole stack. We do everything. This is not a fabless semiconductor company."

Thomas: A whole stack whose wafers, by contract, may only come out of a partner's buildings.

Thomas: A decade of this war — who actually took ground and gained share?

Katie: The scoreboard: Samsung roughly held. SK Group nearly doubled its slice — the Intel flash deal did the lifting. The Kioxia-plus-Sandisk bloc lost the most — about thirty-five percent combined down to just under thirty — and YMTC walked in from zero to double digits.

Katie: Sandisk alone: a ten-year share loser, mid-teens down to about thirteen, stabilizing only as the bits moved to data center.

Thomas: So the company having its best year ever spent the decade losing ground.

Katie: And the contract structure everyone admires is being copied as we speak: SK hynix lists long-term agreements with around ten customers of its own, and Kioxia targets half its volume under contract for calendar twenty twenty-eight.

Katie: A power needs a barrier — some reason copying it hurts the copier — and there isn't one here. A copyable move is a move, not a moat. What Sandisk has is a head start.

Thomas: Then the question the tape forces: if the contracts guarantee the floor, why does the street price this like the floor gives way?

Katie: Twenty-three analysts cover it. The average target sits about a third above the tape; the spread runs one thousand dollars a share to three thousand.

Katie: And the same tape carries two prices at once: about twenty-one times trailing earnings — and about seven times forward.

Katie: Nobody is paying for management's model.

Katie: J.P. Morgan's Harlan Sur restored his overweight rating this week, with a target of twenty-two hundred and fifty dollars out to December twenty twenty-seven — Sandisk, he wrote, is "uniquely positioned to capture the ongoing structural inflection in NAND demand driven by rapid growth in AI inference." Wells Fargo's Aaron Rakers raised his target to fifteen hundred and fifty — and still would not call it a buy.

Katie: And nobody knows what cheap looks like at this company's bottom — no bottom has ever had real earnings under it. We don't know. Neither does anyone quoting you a target.

Emerging Technologies

Thomas: Six days before the split even completed, the company unveiled something called High Bandwidth Flash — a whole next act, announced before the ticker existed.

Thomas: The long model nobody is paying for assumes that act shows up. So give it the full treatment: what's coming next in this space, who's behind it, and how fast can it arrive?

Thomas: What it is: NAND stacked the way high-bandwidth memory is stacked — eight to sixteen times HBM's capacity at similar cost. A product category flash has never had.

Katie: Size the market it attacks. Yole's outlook, presented at this year's Chiplet Summit: high-bandwidth memory was about thirty-five billion dollars last year, around sixty billion this year — and roughly one hundred seventy billion by twenty thirty-one.

Katie: And HBM keeps eating DRAM supply. TrendForce has it at roughly eighteen percent of DRAM wafer input late last year, about thirty percent by end of twenty twenty-seven, contract prices expected to surge multiples higher — Nvidia's next flagship carries three hundred eighty-four gigabytes per GPU.

Katie: Capacity per dollar is exactly where flash beats DRAM. This is Hoff's rule from the top of the show, aimed at the most expensive memory on Earth.

Thomas: Goeckeler's motive, at the Bernstein conference in May: "Inference is really going to be where it's at on NAND and so had to get there."

Katie: And the technology chief, Alper Ilkbahar, in a Nikkei Asia interview this spring: "We believe the next big thing is HBF." His frame for the whole season: "The global AI race is increasingly evolving into a memory-centric resource war."

Thomas: Now the trade-off, stated plainly: flash reads beautifully but writes badly. Latency runs in microseconds — roughly a thousand times longer than DRAM — rewriting means erasing whole blocks, and every write wears the cells a little. So this is a read-optimized part for inference, for serving models rather than training them. It is not a DRAM replacement.

Katie: The strongest third-party evidence comes from the HBM king itself. SK hynix published a hybrid architecture — HBM and HBF side by side on the same GPU — and its simulations of an Nvidia Blackwell part show performance per watt up as much as about two point seven times versus HBM alone, batch sizes up almost nineteen times, and workloads that took thirty-two GPUs running on two.

Katie: Same report, the honest limit: NAND's slow write performance "remains a key limitation" for adoption.

Thomas: Who's behind it: an agreement with SK hynix, August of twenty twenty-five, to standardize the technology.

Thomas: A published standard followed this summer, through the Open Compute Project — up to five hundred twelve gigabytes per stack, bandwidth grades running to three terabytes a second — with Google and Tenstorrent in the consortium.

Thomas: And an advisory board Sandisk seated in July of twenty twenty-five, chaired by David Patterson — Turing Award, Berkeley, one of the fathers of RISC — alongside Raja Koduri, the former AMD and Intel graphics chief.

Katie: Now the calendar, because it moved. The original plan had samples in the back half of twenty twenty-six.

Katie: As of this week — TrendForce, the day after Investor Day — the first product is taped out, samples slip to twenty twenty-seven, and mass production begins in twenty twenty-eight.

Katie: Roughly a year of slip before the first sample ships. Normal for a new category; expensive for a valuation that needs the second act on time.

Thomas: And the name that appears nowhere in it: Nvidia. No public commitment to HBF exists — Nvidia sits on neither the consortium nor the advisory board — and the storage analyst Chris Mellor argues that without Nvidia designing HBF into its GPUs, Sandisk is left selling to the accelerator market's minor players.

Thomas: That's an absence, not a rejection — no Nvidia statement for or against is on the record. But the honest read is that the biggest buyer of memory bandwidth on Earth has not blessed this.

Katie: The counterweight, from inside the same joint venture: Kioxia's European technology chief, in an interview this March — "Nvidia is leading all these discussions with the relevant suppliers and partners." Engaged with the problem, uncommitted to this answer.

Katie: And Kioxia is running both tracks. Its own high-bandwidth flash — up to thirty-two NAND dies stacked on an interposer shared with the GPU — built, per its technology chief, on the same silicon as Sandisk's, through the joint venture.

Katie: Plus a product that ships first: an ultra-fast drive on its low-latency flash, built for Nvidia's own storage-access architecture and demoed at Nvidia's developer conference this spring — ten million input-output operations a second targeted this year, a hundred million next year.

Thomas: So the fork: Kioxia keeps flash outside the GPU package, as the fastest possible drive, on Nvidia's architecture. Sandisk bonds flash into the package next to the HBM, with SK hynix. Same problem, opposite plumbing — and the partner whose buildings Sandisk depends on is hedged across both answers.

Thomas: One more sign they're serious: in March the company put a billion dollars into Nanya Technology — an equity stake plus a multi-year DRAM supply agreement — locking up the DRAM its controllers and an HBF-era product will need, in a DRAM-short world.

Katie: So, as a business: a real specification, a real coalition, a real market to attack — and no anchor customer named Nvidia, with first samples still a year out. That is the second act as it actually stands.

What to Watch

Katie: Gauge one: is anyone building? Every floor in that contract book survives only if supply stays disciplined — and the capital-spending prints answer before any income statement does.

Katie: The prints, by name: NAND capital commitments in Flash Ventures' notes in Sandisk's own filings; Kioxia's capex guide — about four hundred seventy billion yen a year through fiscal twenty twenty-eight; and the toolmakers upstream, because Samsung and SK hynix do not disclose their DRAM-versus-flash split.

Katie: And upstream just moved: flash went from twelve percent of Lam Research's systems revenue to twenty-three in a single quarter — more than doubled sequentially. Tokyo Electron has flash at eleven percent of equipment sales and guided flat. And Applied Materials says flash wafer starts are still declining — the money is upgrades, more layers on the same wafers, with one exception it names: new projects in China.

Katie: The threshold: equipment orders shifting toward flash for two consecutive quarters. That's the industry answering the price signal, and everything downstream of it is a countdown. Tonight's reading: the first quarter of the shift is on the board, and September is the confirmation print.

Thomas: And one name to keep on the list: M seventeen — a new flash fab in SK hynix's own mid-term plan, staged, they say, on customer demand. That's what it looks like when "no" starts turning into "yes."

Katie: Gauge two: price, and the gap. The print is TrendForce's monthly contract series, plus the spread between spot and contract. Right now: third-quarter contracts up ten to fifteen percent — still rising, but moderating, with consumer buyers described as reaching their affordability limit — while a standard flash wafer trades around twenty-one dollars on the spot market, drifting up on volume the trackers call lethargic.

Katie: The rule: a contract holds while spot sits near or above the floor; it gets renegotiated when the gap beneath it exceeds the cost of walking away. Tonight that gap is not open.

Katie: Gauge three: the contract book itself, quarter over quarter. This is the cheapest, highest-signal work in the whole story, and almost nobody does it.

Katie: Three lines to pull every quarter. The remaining performance obligation — ninety-one point one billion dollars as of the June quarter, including two agreements signed after the close. The financial guarantees, which by design release toward the end of each agreement. And the prepayments: about four hundred million on the balance sheet in the spring, then nearly two billion of prepayments and deposits through the full fiscal year — different statements, same direction.

Katie: The reading: minimum climbing and collateral climbing means the buyers are still scared. Either one flatlining or falling early is the inversion unwinding in public — before the income statement admits it.

Thomas: Gauge four: the demand source, and the substitute. Top of the funnel, the hyperscaler budgets: Amazon raised this year's capital plan to about two hundred twenty billion dollars and named higher memory costs as a reason — and the other giants raised in the same season.

Thomas: But the sharper item is the shelf the AI order was supposed to land on. That order went to flash because the disk shelf ran empty — so the watch item is disk lead times. Seagate says the vast majority of its nearline capacity is allocated into calendar twenty twenty-eight, with customers pushing planning horizons toward twenty twenty-nine and beyond.

Thomas: And disk isn't adding units either: Seagate says it is not increasing the box count, just packing more into each box, and Western Digital spent about a hundred million dollars of capital in the quarter while its price per terabyte rose high teens. Both shelves are running the discipline playbook.

Thomas: The loop even runs backward now: Western Digital's own CEO says enterprise buyers who were headed for all-flash systems are shifting back toward hybrid disk because flash got expensive — the two shelves are re-pricing each other upward.

Thomas: The threshold: disk lead times normalizing. The day the cheaper shelf restocks, the marginal AI order goes back to it — a chunk of this flash demand was borrowed. Tonight's reading: lead times extending, not normalizing.

Thomas: Gauge five: the second act, on a calendar. The print: the HBF timeline you just heard, held against its own dates — plus spec revisions at the Open Compute Project, and every name that joins or leaves the consortium or the advisory board.

Thomas: The single binary inside it: an Nvidia design-in — or an Nvidia architecture that routes around HBF entirely. And the route-around candidate already exists: the context-offload platform from earlier tonight, which sends inference memory to ordinary flash drives instead of stacking flash beside the GPU. Nvidia's developer conferences are where a design-in appears, or doesn't.

Katie: And watch whether the partner ships first — Kioxia's rival drive is due in evaluation samples by the end of this year.

Katie: A shared-silicon partner beating you to the same customer is a specific kind of bad.

Thomas: One overlay left: every gauge on that board assumes the rules of the game hold still. But they don't.

The Political Shot Clock

Katie: Start with the supply-cut lever. The blanket authorizations for American chip tools in Samsung's and SK hynix's China fabs were revoked at the end of last year — since January, tools enter only under annual licenses that allow routine operation but constrain expansion. Behind those licenses: Samsung's Xi'an fab is roughly a third of its flash capacity, SK hynix's Dalian fab thirty-five to forty percent of its flash output. Frozen nodes there are a supply cut nobody chooses.

Katie: So the licensing regime is bullish for the floors — for exactly as long as Washington keeps it tight. Each annual renewal is a scheduled event.

Katie: The counterweight is YMTC — on the Entity List since December twenty twenty-two, running a trial line meant to use only Chinese tools, targeting fifteen percent of global NAND by late this year. Counterpoint already has it at thirteen percent, up from eight a year earlier.

Katie: The only supplier on Earth that adds capacity without answering the price signal. Its share line is the single most watchable bear number on the board.

Thomas: The cost clock: Section two thirty-two. Phase one, in January, put a twenty-five percent tariff on high-performance logic chips — but memory is not covered, and big American data centers got carve-outs. The report that triggers phase two came due at the start of July; trade analysts expect it could sweep in memory, equipment, and chips inside imported devices, sometime this fall. Nothing announced as of this recording.

Thomas: It matters here because Sandisk's wafers cross the Pacific — and tariff headlines already set this stock's all-time low once, in April of twenty twenty-five.

Katie: Japan's money carries strings: up to a hundred fifty billion yen approved for the joint venture's fabs in early twenty twenty-four, under a program whose stated purpose is stable production of semiconductors in Japan.

Katie: Public subsidy quietly weighs on any future argument about where that capacity lives.

Thomas: And the structural clock ticked six days before this recording: Kioxia disclosed that the vehicle holding SK hynix's convertible bonds is now its largest shareholder — edging past Toshiba itself.

Thomas: Why nobody can simply take Kioxia over: SK hynix is capped at fifteen percent of the votes until twenty twenty-eight, converting the bonds needs Japanese-government approval, and any move triggers merger review in multiple jurisdictions — it would fuse the number-two and number-three flash suppliers.

Thomas: The consent power is not hypothetical. October twenty twenty-three: SK hynix refused to approve the Kioxia–Western Digital merger, and the deal died.

Katie: And the Beijing precedent: China's cyberspace regulator ordered infrastructure operators to stop buying Micron in twenty twenty-three — and SAMR, the separate merger authority, is a required stop for any Kioxia deal.

Thomas: Which is why the one deal that would fix Sandisk's structural problem — folding into the partner whose buildings it depends on, one technology stack instead of two — keeps not happening.

Outro

Thomas: The plain answer, in one breath. Tonight's question was whether a commodity manufacturer really re-engineered its own cycle. The re-contracting is real, it is industry-wide, and the collateral behind it is genuine structural change — not a story. But the floors have never been tested, and nobody — not management, not the street, not us — gets to know whether they hold until a winter tries them. A real transformation in how memory is sold, with the value capture still unproven — and the prints that will settle it are on the board now, ahead of any income statement.

Katie: Second Spark is produced by Overnight Studios. This show is for information and entertainment only. Nothing you hear is investment advice, and the people behind the show may hold positions in the companies we cover.

Thomas: Thanks for listening to Second Spark. Say hello next month when we cover what's next in innovation.