The bear case reduces to one decomposition and one forecast. Fiscal 2026 revenue grew 175% while bit shipments grew in the mid teens. Roughly two thirds of last quarter’s sequential revenue growth came from price rather than volume. NAND contract price increases are decelerating toward low single digits by the December quarter, and TrendForce sees downward pressure in the second half of 2027. The decomposition is accurate. What it implies is close to the reverse of what it’s being asked to prove.
Start with the bit number, because it’s carrying the argument and it’s being read upside down. Every NAND collapse on record ran the same sequence. Industry bit supply grew thirty or forty percent, sourced from wafer capacity committed eighteen to twenty-four months earlier. It landed into demand that had stopped accelerating. Depreciation on that new capacity then converted a business with variable economics into one with a fixed-cost base, and margins fell faster than prices did. In 2008, in 2015, in 2018, in 2022, price fell because bits showed up. Mid-teens bit growth, in the tightest NAND market the industry has recorded, is direct evidence that the mechanism has not been armed. A company earning an 84.6% gross margin without having built the capacity that historically ends the cycle isn’t standing on thin ice. It’s the reason the ice is thick.
Next, the deceleration. TrendForce put third-quarter NAND contract prices at 10% to 15% quarter over quarter, slower than the 70% to 75% of the second quarter and the 55% to 60% of the first. The sell side extends the curve toward low single digits by December. Read the level, not the second derivative. Contract NAND has risen roughly four and a half times in four quarters. A December quarter that adds three percent to that base is an extraordinary pricing quarter that happens to produce a small number on a percentage-change screen. Deceleration of an increase is not a decline, and an ASP that stops rising at four times its 2025 level still earns four times the 2025 margin.
The composition of that deceleration matters more. TrendForce attributes the moderation to consumer affordability limits: PC and smartphone buyers have reached the ceiling of what they’ll absorb. In the same survey it states that NAND demand continues to be driven primarily by AI inference and large-scale data center deployment. So the slowing is being sourced from the segment SanDisk is walking away from, while the segment it’s walking toward is the one still bidding. Data center revenue roughly doubled sequentially to about $3 billion in the June quarter. Applying a consumer-sourced price ceiling to a business mix that is rotating out of consumer is a category error.
That leaves the supply wave, which is the load-bearing piece of the bear case and is mostly misdated. Samsung will produce NAND in some cleanrooms of P5 Phase 1 and 2 at Pyeongtaek, ramping sequentially from 2027 through 2030; a four-year staged ramp is not a wave. SK hynix has only just fixed the investment scale for Cheongju M17, with construction beginning in February 2027 and the first cleanroom scheduled to open in December 2028, which puts qualified wafers somewhere in 2029. Kioxia’s new K3 fab at Kitakami, funded 60/40 with SanDisk out of the ¥5 trillion program, targets production within fiscal 2029, and the FY27 to FY29 tranche was left unchanged at ¥1.4 trillion. The three incumbents who would need to break the market are, between them, scheduled to add meaningful NAND wafer capacity starting roughly two years after the bear’s price break is supposed to happen.
YMTC is the one genuine 2027 addition, and it’s the wrong kind of bits. It reached third place in global shipments in the second quarter at around 14% of industry bits, but finished fifth on revenue, because its mix sits in consumer applications with very little high-priced data center eSSD. Counterpoint’s own conclusion from that quarter is the relevant one: profitability through 2027 will be settled by who ships the right mix, not by who ships the most bits. Server eSSDs are on track to absorb more than half of all NAND bits. A Chinese supplier scaling consumer-grade output relieves the SD card shortage. It does not qualify into a hyperscaler storage tier, and it has an export-controlled tooling problem standing between its wafer targets and its wafer output.
Then there’s the part of the bear argument that contradicts itself. The recommendation attached to it is a preference for Micron, SK hynix and Samsung on an expected rotation into HBM names. But the reason NAND is short is that those three companies are putting their wafers, their cleanrooms, their capital budgets and their best process engineers into HBM and server DRAM. Samsung’s NAND output has been capped by its own preference for high-margin DRAM. SK hynix has redirected a large share of 2026 investment toward HBM3E and HBM4. The $518 billion Korean program that makes the HBM trade work is the same program that keeps NAND wafer starts flat. You can own the HBM rotation or you can forecast the NAND supply response, and both positions cannot be right unless cleanrooms are free. Every dollar that validates the first leg starves the second.
The contracts are being read as a ceiling when they are a floor with a call attached. Eight NBM agreements carry $93.9 billion of minimum expected revenue, covering more than half of fiscal 2027 bits and about two thirds of fiscal 2028, at estimated floors near $0.29 per gigabyte, broadly where second-quarter ASPs sat. Bernstein’s read is that even in a price collapse worse than 2010, these agreements substantially mute earnings downside into 2029 and 2030. The bear objection is that a good portion of revenue remains exposed to spot. That’s correct, and it’s the asymmetry. Two thirds of fiscal 2028 is floored at something close to today’s price, and the residual third is unhedged upside that nobody is charging for. The distribution has been truncated on the left and left open on the right, which is the opposite of how a commodity producer usually enters a supply cycle.
Underneath that sits the balance sheet, which is what actually separates this from every prior NAND cycle. Net property, plant and equipment of roughly $674 million against $20.2 billion of annual revenue. The fab assets live in the Kioxia joint venture. The classic memory downturn kills companies through depreciation on a fab commissioned at the top, and SanDisk doesn’t have one on its books to depreciate. Incremental price runs almost straight to margin, and the absence of that fixed-cost base is why management could stand up at the August 13 investor day and model mid-to-high-teens revenue growth with roughly 80% adjusted gross margin out to 2030 without being laughed at.
The demand curve has also changed shape in a way the cyclical framework doesn’t capture. NAND demand used to be handset and PC units, saturated and replacement-driven. An NVL72-class rack now calls for something on the order of a petabyte of NAND, and a single high-end accelerator around 16TB. UBS raised its Kioxia bit growth estimate specifically because AI context lengths keep expanding, which ties bit consumption to tokens served rather than to devices sold. Commercial Times has 2026 NAND demand growing 20% to 22% against supply at 15% to 17%. That gap doesn’t close on a schedule anyone has published.
And there’s a free option on top. SanDisk has taped out its first High Bandwidth Flash part, with samples in 2027 and production targeted for 2028, backed by a consortium that now includes Google, Meta, SK hynix and Tenstorrent. Goldman’s read of the company’s own simulations is that an HBF-only architecture can deliver equivalent token output on half the GPU count of an HBM-only system. If any of that survives contact with a real inference deployment, NAND bits start being consumed by a workload currently served by DRAM, which would be the largest single demand shock available to the category. Nobody is paying for it at the current multiple.
Which is the last point. At roughly $1,633, the stock sits about 31% below its June record and trades near seven times fiscal 2027 earnings, with a forward PEG around 0.11 against a sector median of 1.31. That multiple is not an oversight. It is the market already pricing the party ending, and pricing it as though the ending were imminent and total. Consensus sits around $2,125 with a high near $3,600, and the S&P 100 addition at the open on September 21 adds mechanical demand on top. At seven times, the thesis doesn’t require the cycle to run forever. It requires the cycle to end later than a single-digit multiple already assumes, and the construction schedules at Pyeongtaek, Cheongju and Kitakami say it will.
The number to watch isn’t the December pricing delta. It’s fiscal 2028 NBM coverage and whether a ninth agreement gets signed, because each one converts spot exposure into floor exposure ahead of the supply that hasn’t been poured yet.