The quarterly revenue line reads $8.97 billion. The non-GAAP gross margin reads 84.6 percent. One year earlier, the same company printed 26.4 percent. Same NAND. Same fabs in Japan. Same ARM cores inside the controllers. What changed is the price of memory, and the narrative bolted to it.
SanDisk, newly independent from Western Digital, is being called an AI chipmaker. The market does not read datasheets.
Over six months, Jefferies analyst Blayne Curtis raised the target price seven times, to $3,000. Then he cut it to $1,750. The stock lost roughly a quarter of its value within a month. Vested interest distorts the lens of analysis. Here, the lens produced a target that moved like a price rather than like an output of a model — a model in which the margin was up 58 points, the data center line was up 103 percent, and yet the revision was downward.
The margin says abundance. The stock says fear. One of them is standing on the wrong side of a cycle.
The protocol does not lie; the interface does. The interface here is the one between a storage supercycle and the analysts who describe it. It is doing what interfaces do: formatting the inconvenient truth out of the display.
Context
SanDisk is a NAND flash IDM. Vertical by design: it designs the flash, co-operates the fabs with Kioxia in Japan, assembles the drives, and writes the firmware. The latest process generation is BiCS8, roughly 218 layers of 3D charge-trap flash. That sits behind Samsung's 300-plus layers, SK Hynix's 321, Micron's 276. In HBM — the most contested memory product of the AI era — SanDisk has nothing. Zero.
Its leverage to the AI story is the enterprise SSD. Data center revenue came in at $2.98 billion for the quarter, up 103 percent year over year, roughly a third of total revenue. Every AI training cluster is built like a patient memory. GPUs compute. HBM feeds the die. Terabytes of enterprise NVMe flash hold the checkpoints, the datasets, the logs. Models crash; checkpoints get reloaded. The SSD is the bulk storage of the machine-learning era.
The market heard "storage for AI" and translated it into "AI chipmaker." What it actually bought is a cyclically positioned commodity with a temporarily favorable supply curve.
In my audit work, I learned to separate the asset from the story. The asset is a stack of silicon. 218 layers, vertically written, erased in blocks, organized into dies, packaged into drives that must survive seven years of hostile thermal cycling inside a server. The story is that this is a structural-growth company at the center of the intelligence revolution.
Both are true in the way two objects can share a coordinate at different times. The skill is in telling which truth applies now.
Core
The margin measures the past
Eighty-four-point-six. Let the number sit.
Non-GAAP gross margin. Same period last year: 26.4 percent. A swing of 58.2 points in four quarters. No production process on earth improves that fast. No product mix shifts that fast. What moves that fast is price, plus an accounting tailwind with a long half-life.
Three components.
Price. NAND contract pricing spent 2024 into 2025 in a genuine supply deficit. Hyperscalers bought aggressively; AI server builds consumed capacity that had been planned for the PC replacement cycle. Spot markets ran hot. That part of the margin is real, and it is cyclical.
Cost. This is the part the dashboard does not show. SanDisk's Japanese fabs are old friends. The wafer fabrication equipment was purchased, depreciated, and written down across prior cycles. A fully depreciated fab has a cash cost close to the marginal wafer cost — electricity, gases, wafers, labor. Depreciation, the largest fixed cost in memory manufacturing, has already been paid by the income statements of previous years. The 84.6 percent is not a measure of operational excellence. It is a measure of time passing. The load was paid; this cycle collects the receipt.
Mix. Enterprise and data center SSDs carry higher blended margins than retail flash. The 103 percent surge in data center revenue tilted the mix upward. Real, but secondary in magnitude.
Now the non-GAAP asterisk. The 84.6 percent excludes stock-based compensation, restructuring costs, and the other line items management sets aside. The GAAP number is lower. The non-GAAP number is the interface: designed for comparison, adopted for celebration.
Historical high-water marks for NAND manufacturers in genuine supercycles: 50 to 60 percent gross margin. NVIDIA prints above 70 percent on the strength of a compounding architectural moat. A storage company printing 84.6 percent is not becoming NVIDIA. It is catching a wave of spot scarcity and accounting tailwinds at the instant the market decides the wave is over.
I will be blunt. This margin is not an achievement of engineering. It is an accident of timeline. The next node — BiCS9, 300-plus layers, a new clean room — will bring a depreciation wave that washes margins back toward the 50-60 percent range. The stock will then be repriced on that number.
The margin measures the past. Depreciation schedules do not accelerate for narratives.
Counting layers
Memory companies are judged in layers, the way protocols are judged in total value secured. Layer count is the NAND distance metric. Samsung runs V9 near and beyond 300 layers. SK Hynix announced 321. Micron is at 276. SanDisk and Kioxia ship BiCS8 at about 218. The lag is approximately one generation — six to twelve months — in the critical dimension of stacking depth.
That is survivable in NAND. It has been survivable before.
What is not survivable is the HBM absence. High-bandwidth memory is where the AI memory profit pool concentrates. It sits beside the GPU, stacked with through-silicon vias, connected through CoWoS-class packaging, sold out through 2025 and beyond. SK Hynix, Samsung, and Micron harvest that margin. SanDisk watches from outside the package. You can call the company an AI memory play, but it is the AI memory play with no ticket to the most valuable tier of AI memory.
The SSDs it does sell are real. A frontier training run checkpoints enormous state across nodes; a single run can touch tens of terabytes of writes. Inference serving demands high-capacity, low-latency drives. The storage content per AI server is five to ten times a traditional server. That is the demand story, and it is true.
Here is the part the demand story omits. NAND supply is elastic. Five major manufacturers run this market — Samsung, SK Hynix with Solidigm, Micron, Kioxia with SanDisk, and a state-backed YMTC climbing from the low end. Twenty-five years of history show the same behavioral constant: high prices trigger expansion; expansion arrives late; late capacity crashes prices. The industry has never failed to overbuild. The lag between high prices and high capacity is the only reliable schedule in this market — a sixteen-to-twenty-four-month loop that functions with the regularity of a clock.
Certainty is a bug in a stochastic world. But that lag is a near-certitude. The only question in NAND, always, is not whether the glut will come. It is whether your model has priced the dotted line of new capacity.
As a technologist, I see the gap as closable in layer count by 2027 through the Kioxia partnership. A credible HBM entry would require massive capital and years of client certification. That means SanDisk's answer to the AI memory question will remain the SSD for at least two more years.
The target price as a data point
Let me treat Blayne Curtis's target arc as data, because it is the most instructive dataset in this report. Seven upgrades in six months to a $3,000 objective. Then the revision down to $1,750. The spread between those two numbers is $1,250 — a range larger than most companies' entire market value.
This pattern is not confined to sell-side equity research. I watched the same behavior inside DeFi protocol governance during the 2020 yield summer. Model outputs were updated to match the price. The mathematics had not changed; the desire had. When you observe a forecast revised in the direction of the market seven consecutive times, you are not observing discovery. You are observing a feedback loop: rising price, rising sentiment, rising target, rising price.
The cut to $1,750 is not proof the analyst found a new fact. It is proof the feedback loop broke. The stock fell 25 percent, and the model was recut to fit the new price.
The signal: the target was revised because the market moved first. Notice that non-GAAP EPS of $39.25 came in above consensus, and the target was still cut. When good news produces bearish price action, the marginal buyer has left the venue. That is a liquidity statement, not a fundamental one. The stock tells you what the interface does not: the strongest holders of the narrative have re-priced.
At the $1,750 target, the trailing multiple sits near 44.6 times earnings. Storage cyclicals historically trade between 10 and 20 times. Even the bull case requires the margin to be permanent. Margins are never permanent in storage.
I do not write this to posture from a moral high ground. Sell-side incentives are what they are. But structured professionals should read a seven-raise-one-cut arc the way auditors read a journal entry that does not tie out: someone is telling themselves a story. The company has real assets and a genuine position in a vital market. The equity, at target prices near $3,000, was pricing the margin as a perpetual property.
It is not. When a forecast moves seven times with price, the forecast has stopped being an analysis and has become an echo.
The pull-in problem
The 103 percent data center growth deserves skepticism of a mechanical kind.
In shortage cycles, hyperscalers do not just consume storage. They hoard it. Procurement teams forward-buy to lock supply and price. They build inventory cushions. The demand curve visible in the income statement is lifted, in part, by purchases pulled forward from quarters that have not yet arrived. This is exactly the pattern I identified in 2020 inside yield-farming liquidity: activity that exists because the incentive demands that it be pulled forward in time.
Every storage cycle contains a period in which apparent demand exceeds true consumption by a meaningful margin. The 84.6 percent gross margin is the fingerprint of that overshoot — spot scarcity, double-orders, and allocation discipline all at once. When the dust settles, usually twelve to eighteen months later, inventory corrections do the arithmetic the models refused to do.
Then there is the buyback. $4.5 billion repurchased in the quarter. $14 billion reauthorized. This is management standing at a cycle peak and choosing to return capital rather than to build capacity. Storage economics dictate that incumbents should spend aggressively when margins are high and the future is bright. Choosing buybacks at the top is a mature-cycle tell. It signals, as clearly as a press release could, that management views the marginal expansion dollar as lower-return than the marginal repurchase dollar. They know the yield curve of the cycle. They are closer to the end than to the beginning.
The buyback is not a bearish fact in isolation. It is bearish in the context of an 84.6 percent margin and a doubling data-center line. The people with the best data are not chasing the cycle; they are harvesting it.
The bifurcating stack
The manufacturing geography is its own signal. An American brand running its flash fabs in Japan, in partnership with Kioxia, is a supply-chain hedge in silicon form. It locates the brand inside the American alliance's industrial core while avoiding direct Chinese exposure. It answers the question of where the next wafer comes from without waiting on the slow machinery of the CHIPS Act.
The bifurcation of the global stack is not abstract. Gallium and germanium export controls on one side. Advanced-node restrictions on the other. YMTC continues to climb with state backing; its 232-layer parts are already a price-competition force in the midrange. The Chinese market is a substantial consumer of NAND. A SanDisk without full access to it grows slower than a Samsung running a mega-fab in Xi'an. That is a long-term growth differential, and the analysts who cut the target know exactly where it lives.
For those of us building decentralized compute networks, geography is no longer optional metadata. Storage location determines regulatory jurisdiction, data-residency compliance, and the verifiability of provenance. When replication checks run across enterprise SSDs in Tokyo, Frankfurt, and Virginia, the physical circuit of the network becomes part of its security model.
To own the chain is to own the history. To own the storage is to be able to prove it.
What the crypto layer absorbs
This is where my own bias sits. In 2025, I co-authored a technical specification for a decentralized compute marketplace — one designed so that AI models could not be trained on stolen data without economic penalty. We spent six months on the incentive layer. The hardest number was the price of storage.
If the network prices storage at the narrative peak, the network becomes unaffordable at the moment the narrative turns. If the network prices at replacement cost — the fully burdened cost of standing up new capacity — the network survives the cycle, and the margin distortion of a quarter like this one stays at arm's length.
The volatility of storage cost is a protocol risk. Zero-knowledge proof generation depends on witness sizes and state checkpoints that live on exactly this hardware class. Decentralized training pushes tens-of-terabytes writes across commodity nodes. Every design that consumes NAND inherits the NAND cycle. The prudent architect writes the cycle into the economics, not into the pitch deck.
Contrarian
Here is the counterintuitive part.
The bear case is not the falling stock price. The bear case is the 84.6 percent margin itself.
A gross margin resting on a fully depreciated asset base is not a signal of operating health. It is a receipt for investments made ten years ago, collected during a pricing spike. It tells you nothing about the cost of the next generation. The market's error is treating a backward-looking accounting artifact as a forward-looking competitive moat. When BiCS9 arrives with its depreciation wave, the 84.6 percent will be remembered the way investors remember models that used the 2021 spot peak to price 2023 earnings.
The HBM absence is the companion flaw. The most valuable tier of AI memory passes this company by. The label "AI chipmaker" requires marking to reality. The resilient position is not a bet against storage demand. Storage demand is strong and getting stronger. The resilient position is a bet against the margin's arithmetic, and against the interface that presented a 58-point margin swing as a stable property of a business.
The 103 percent data-center line. The seven target raises. The $3,000 screen. The $14 billion buyback. These form a coherent stack of signs. They describe a cycle in the late innings, not the first act. My old audit partner would phrase it differently: when the call auction shows everything marked high and everyone holding a ticket, begin reading the exit row.
Takeaway
When a target price rises seven times without the underlying silicon changing, the model has stopped describing the company and has started describing desire. My guidance to protocol builders and allocators alike is the same: price at replacement cost, not at the narrative peak. The fab will be built. The margin will normalize. Certainty is a bug in a stochastic world, and the only reliable constant is the lag between oversupply and the moment the market admits it.
Silence before the block confirms the truth. In this case, the block is a fully depreciated fab — and the silence that will follow its replacement will carry the actual numbers.