Most people look at an 84.6% gross margin and reach for the buy button. That's the wrong instinct.
SanDisk just printed a quarter that makes NVIDIA look ordinary. Non-GAAP gross margin of 84.6%. Data center revenue up 103% quarter over quarter. Non-GAAP EPS of $39.25, twelve percent ahead of consensus. Then the stock dropped 25% in a month, and the same analyst who raised his target price seven times in six months — all the way to $3,000 — cut it to $1,750 in a single report.
The crowd reads that drawdown as a gift. I read it as a cycle-top signal firing in real time. And the crypto-AI complex should be watching closely, because the hardware layer underneath this whole narrative just got repriced.
Context: What SanDisk Actually Is
SanDisk isn't an AI chipmaker. It's a NAND flash manufacturer. An IDM — integrated device manufacturer — that designs and fabricates its own memory, runs fabs in Japan as a joint venture with Kioxia, and sells enterprise SSDs to every major cloud provider running the AI buildout.
The technology narrative is thinner than the price action suggests. Current-generation BiCS8 NAND is roughly 218 layers. Samsung is shipping 300+. SK Hynix is at 321. Micron at 276. SanDisk sits a full product generation behind the leaders. And in HBM — the most supply-constrained, highest-margin memory product of this cycle — SanDisk has zero exposure. The entire "AI memory" story rests on enterprise SSD volumes. Real demand. But structurally different economics than the HBM trade that's been printing for SK Hynix.
The latest quarter delivered $8.97 billion in revenue. A third of it — $2.98 billion — came from data center with 103% growth. That's the AI punch. Each AI server carries five to ten times the storage value of a conventional server. Cloud providers, rational actors in a shortage, did what they always do: they bought early, and they bought ahead of the next price increase.
The Margin Is an Accounting Artifact Wrapped in a Supply Squeeze
Let me decompose the 84.6%, because almost nobody else seems to have done it in writing. A year ago, this company printed a 26.4% gross margin. The improvement traces to three variables: NAND contract prices exploding, product mix shifting toward high-margin enterprise SSDs, and — the part every headline missed — a Japanese fab fleet that is largely fully depreciated.
Depreciation matters more than price. Old fabs running at near-zero marginal cost produce outsized margins at the peak of a supply squeeze. The same fabs produce catastrophic margin compression when the cycle turns. The 84.6% is not a technology breakthrough. It's an accounting artifact. SanDisk and Kioxia have to build next-generation fabs for BiCS9 and beyond — 300 and 400-layer NAND. That capex is coming. When it hits, depreciation returns, and the margin snaps back. History says the landing zone is 40-50%, not 84%.
The analyst whipsaw deserves its own post-mortem. Seven target raises in six months. Then a 43% cut. This isn't analysis — it's momentum with a letterhead. I've seen the same pattern in crypto for a decade. Sell-side strategists raise targets as price rises, then slash them as price falls, narrating the past with the confidence of prophecy. Liquidity doesn't care about your forward P/E. The $3,000-to-$1,750 trajectory is a chart of the analyst's own crowding, lagging the actual supply-demand data by weeks.
At the new target, SanDisk trades at roughly 44 times trailing non-GAAP earnings. Its historical range as a memory producer is 10-20 times. Micron trades near 25. Samsung near 15. Wall Street is paying AI-monopoly multiples for a commodity product with a two-year supply response — and an existing challenger, Yangtze Memory, scaling with Chinese state capital behind it.
The NAND market has a structural flaw that the AI narrative temporarily masked: product homogeneity. Memory is memory. There is no ecosystem lock-in, no application-layer switching cost, no winner-take-most network effect. The differentiation between Samsung, Kioxia, and SanDisk SSDs is minor, and procurement decisions come down to price and reliability track record. When supply catches up with demand — and it always does, with a lag — the pricing power evaporates.
There's also a geopolitical dimension the earnings press release won't mention. SanDisk's manufacturing base sits in Japan, inside the Kioxia joint venture. That keeps it inside the US-Japan alliance's equipment supply chain — secure, but politically exposed. China is a massive NAND consumer, and the export-control environment is pushing Chinese hyperscalers toward domestic suppliers. Losing China permanently caps the addressable market. Samsung runs a large fab in Xi'an. SanDisk has no equivalent. The long-term revenue ceiling is lower than the market prices in.
A Stress-Test, Not a Story
I don't allocate capital to stories I can't stress-test. That habit got me through May 2022, when Terra's algorithmic stability module depegged and the feedback loop was structurally broken. The market narrative then was "UST is a new money paradigm." The data showed an oracle failure compounding an irreversible death spiral. I hedged into a collapse most people refused to acknowledge, because the numbers didn't match the story.
This is the same discipline. The AI-storage story is not false — it's incomplete. Data center demand is genuinely compounding. But the current earnings trajectory assumes three things simultaneously: NAND prices stay elevated, supply stays constrained, and the depreciation holiday never ends. All three have known expiration dates. When even one expires, the multiple contracts.
The buyback tells you management reads the same curve. $4.5 billion executed in the quarter, $14 billion authorized. At the cycle peak, management chose shareholder returns over next-gen capex. That's not confidence — it's the capital-allocation pattern of executives who know the depreciation holiday is ending. Building a new NAND fab takes two to three years and tens of billions in up-front spend. If you believe the cycle has legs, you invest. If you know the price window is closing, you return the capital and let public shareholders carry the cycle risk. I saw the same pattern in DeFi during 2021-2022: protocols that hoarded treasuries at the peak spent the bear market defending insolvency. Capital allocation at cycle tops is a confession.
The Contrarian Read
The conventional wisdom says a 25% drawdown reset the risk. I don't buy it. The drawdown is the beginning of multiple compression, not the end. And the pull-in effect hasn't been fully felt: when data center revenue jumps 103% in one quarter, some of it is procurement teams buying ahead of price increases. That demand gets pulled forward from future quarters. When the price rolls over, those orders decelerate, and the year-over-year comparisons flip negative. It's the same feedback loop that broke Terra's algorithm — the consensus extrapolating an unsustainable input far past its shelf life.
For crypto builders, the read-through is concrete. The 2026 AI-agent wave I've been auditing — wallets executing on-chain trades, managing keys, calling DeFi protocols — all of it sits on the same enterprise SSD stack. Node operators, indexing infrastructure, training pipelines, even the cheapest RPC endpoint ultimately runs on NAND. If you're building AI-crypto infrastructure, your cost base just became a function of the NAND cycle. And right now, you're entering at the top of that cost curve. The team that buys storage during the downcycle — the one that comes after this correction plays out — will run the same product at half the hardware cost. That's a structural edge most VCs aren't modeling.
The same trap extends to crypto's AI-token complex. Most AI tokens have no relationship to actual inference, training, or storage hardware. They're tickers attached to a narrative. SanDisk at least sells a real unit with a real margin. The fact that a genuine hardware company with real revenue gets valued like a narrative token — and then corrected like one — tells you how much of today's "AI everything" trade is borrowed belief rather than measured demand.
The Takeaway
Watch three signals. NAND contract price rollovers — when spot pricing turns, the ceiling falls. A public announcement of SanDisk or Kioxia next-gen fab construction — that's when the margin narrative breaks under depreciation weight. And whether the buyback persists into the next two quarters — when management pivots from buybacks to capex, the cycle has turned.
I don't trade narratives. I trade supply and demand. The AI-storage demand is real. The price is not. A fully-depreciated fab producing an 84.6% margin isn't a moat. It's a lagging indicator of the best days already passing.