SanDisk (NASDAQ: SNDK) arrived at Wednesday morning trading at $1,581 — more than 3,400% above where it stood 12 months ago, but 32% below the all-time high of $2,354.39 it set on June 22 — following an 8.31% drop Tuesday after Morgan Stanley published a Morgan Stanley crowding note identifying SNDK as the most institutionally crowded semiconductor stock in the S&P 500, with large funds holding approximately 2.3 percentage points more of the stock than its index weighting warrants. That positioning signal arrives one day before OCP Korea Tech Day in Seoul on Thursday, where SanDisk's Distinguished Engineer for SSD Architecture, Ross Stenfort, is scheduled to present "Storage: Past, Present and Future" — the first major public technical appearance for SanDisk since its August 13 Investor Day.

The question the crowding note cannot answer is the one that actually determines whether SNDK's stock recovers toward the $2,094–$2,164 analyst consensus or revisits its late-July lows near $1,000. That question comes down to a single data point buried inside SanDisk's fiscal fourth-quarter results: of the massive sequential revenue increase the company delivered, approximately two-thirds came from higher average selling prices — not from shipping more NAND. Only one-third came from higher volume. The history of memory semiconductors says that ratio matters more than any analyst target.

What the Earnings Beat Actually Showed — and What It Obscured

SanDisk's fourth-quarter results for the period ending July 3 were extraordinary by any conventional measure. Revenue reached $8.97 billion — up 372% year-over-year and 51% sequentially — against an analyst consensus of $8.42 billion. Non-GAAP gross margin reached 84.6%, a figure more typical of a pure-software company than a chip manufacturer. Non-GAAP earnings per share came in at $39.25, beating the $34.59 consensus estimate. Full fiscal year 2026 revenue reached $20.25 billion, up 175% year-over-year.

The number that requires specific technical understanding is the composition of that sequential jump: two-thirds of the quarter-over-quarter revenue increase came from higher average selling prices, not from higher bit shipments. A market where suppliers raise prices by two dollars for every dollar they add in output is a market where demand has outpaced supply sufficiently to sustain scarcity pricing. That is not a commodity market. A commodity market delivers pricing only when supply is insufficient — and pricing collapses the moment supply catches up, without any warning.

In NAND flash memory's history, no company has sustained a 2:1 ASP-to-volume ratio for more than two or three quarters before supply additions from any of the four main producers — Samsung, SK Hynix, Micron, and SanDisk itself — eroded the premium. The question that determines SNDK's trajectory is whether AI infrastructure demand is structurally different enough from prior demand drivers to sustain the ratio longer.

Why Morgan Stanley's Crowding Note Is Positioning, Not Fundamentals

Tuesday's 8.31% decline was driven by Morgan Stanley's finding that institutional portfolios are over-positioned in SanDisk by approximately 2.3 percentage points relative to the stock's S&P 500 index weighting. This is a positioning assessment — a measurement of how concentrated a specific stock has become in institutional portfolios — not a judgment about whether SanDisk's earnings are sustainable. When a crowded stock needs to be sold — for risk management, client redemptions, or rebalancing after a historic run — all sellers move simultaneously, amplifying the decline.

The intraday range on Wednesday of $1,570 to $1,710 reflects exactly that dynamic: a stock that has moved faster than most institutional managers can accommodate in their models, now encountering a positioning-based headwind at a moment when no new fundamental information exists. The same stock that was "cheap at 10 times next year's earnings" 10 days ago, per several analysts, is now being sold because too many funds own it relative to how much of the index it represents.

That is not a permanent condition. Crowded trades resolve when either the fundamental story deteriorates (making the crowding justified in retrospect) or the excess positioning works its way through the market as investors find alternative ways to express the same thesis.

How Does KV Cache Actually Drive NAND Demand — and Where Does It Break Down?

The AI demand thesis for NAND storage rests on a specific technical mechanism: KV cache. When a large language model generates a response, it produces key-value (KV) pairs at each step of the attention calculation. Preserving those pairs in cache — rather than recomputing them from scratch for each new token — dramatically reduces inference latency. As context windows expand from 4,000 to 128,000 tokens and beyond in modern models, the per-session KV cache grows proportionally. At millions of concurrent users, aggregate KV cache demand creates a need for fast, high-capacity storage that exceeds what DRAM can economically provide at scale.

SanDisk estimated at the Future of Memory and Storage conference that KV cache alone could drive 75 to 100 exabytes of additional NAND demand in 2027, with potential to double in 2028. JPMorgan estimates the total addressable NAND flash market could expand from roughly $70 billion in calendar 2025 to more than $300 billion in 2026 and approximately $500 billion in 2027.

This is where the technical constraint that no analyst note mentions becomes relevant. SanDisk's flagship NAND technology — BiCS10 QLC — achieves its extraordinary bit density (332 layers, more than 37 gigabits per square millimeter, 60% higher than the prior generation) by storing four bits of data in each transistor across 16 discrete voltage states. QLC NAND's advantage — high density, low cost per gigabyte — comes with a tradeoff: write endurance is lower than in TLC (triple-level cell) or SLC (single-level cell) NAND, because more voltage states mean more precision required in the charge management, and that precision degrades with repeated write cycles.

KV cache in AI inference is predominantly a read-heavy workload — the model reads cached context but does not continuously overwrite it. That makes QLC NAND well-suited for KV cache tiering specifically. The risk is architectural: if AI compute workloads evolve toward more write-intensive patterns — on-device training, federated learning, frequent model updates in inference serving — the QLC endurance assumption changes. SanDisk's pricing power depends not just on AI demand growing in aggregate, but on AI inference remaining structured in ways that favor NAND's read-access architecture over its write limitations.

High Bandwidth Flash: NAND's Next Frontier Before Thursday's Test

Beyond KV cache, the technical development that could expand SanDisk's total addressable market structurally is High Bandwidth Flash (HBF). On August 3, SanDisk and SK Hynix published the first HBF OCP technical specification through the Open Compute Project — the standards consortium that governs data center hardware interoperability. Google and Tenstorrent joined the consortium during the standardization process.

HBF applies the same vertical die-stacking principle behind High Bandwidth Memory — connecting multiple dies with through-silicon vias (TSVs) — but to NAND flash instead of DRAM. SanDisk's target configuration stacks 16 flash dies plus a base die in the same physical footprint as an HBM package, delivering read bandwidths of 1.6 terabytes per second and capacities up to 512 gigabytes per stack. For comparison, HBM4 offers approximately 64 gigabytes per stack. NAND provides approximately 30 times DRAM's bit density per unit area — the ratio that makes HBF a potentially transformative near-compute memory tier.

On Thursday, SanDisk's Ross Stenfort is scheduled to present "Storage: Past, Present and Future" at OCP Korea Tech Day in Seoul — where Stenfort leads the storage session on the server-storage-memory day specifically framed around the KV cache tiering question. That presentation will be the first public technical signal from SanDisk's engineering team since the Investor Day financial targets — and the first opportunity for attendees to probe whether the HBF specification is attracting the hyperscaler interest that would convert a technical standard into actual volume commitments.

What History Says the ASP Ratio Predicts

Motley Fool's analysis of comparable parabolic runs concludes that the ASP-to-volume composition is the decisive variable — and the history is genuinely split.

Micron's 2018 cycle is the cautionary case. The company's net income peaked at $14.1 billion in fiscal 2018 — driven in large part by ASP increases in DRAM — and then fell to $6.3 billion in fiscal 2019 and $2.7 billion in fiscal 2020, down 81% from peak over two fiscal years. The stock had peaked at around $64 per share with a forward price-to-earnings ratio of approximately 4 to 5 times at the top — and fell 56% even as earnings were still reported as rising, because markets are forward-looking. The 2022–2023 cycle was faster and deeper: Micron posted its largest-ever quarterly loss of $2.31 billion and the stock fell roughly 50%.

Nvidia and Tesla represent the recovery case. Both experienced drawdowns exceeding 30% after parabolic runs and went on to set new all-time highs. The common element: earnings either continued growing through the drawdown or dipped once and quickly recovered.

Bank of America semiconductor analyst Vivek Arya offers a framing that helps calibrate how different this cycle is from prior ones. Even under bear-case assumptions — DRAM prices falling 30% and NAND prices falling 40%, consistent with historical downturns — BofA estimates Micron could still earn close to $100 per share. That would be more than eight times Micron's $12 per share earnings peak from the 2018 cycle — a structural comparison suggesting the AI demand uplift has permanently raised the floor, even if a downturn eventually arrives.

For SanDisk specifically, the AI-era structural defense takes three forms. First, the New Business Model (NBM) contracts: as of the August 13 Investor Day, SanDisk has signed ten long-term supply agreements with eight enterprise customers, carrying combined minimum purchase commitments of approximately $93.9 billion at floor pricing, covering more than half of SanDisk's total bit shipments in fiscal 2027 and nearly two-thirds in fiscal 2028. No NAND manufacturer in the industry's history had previously contracted more than half its bit output simultaneously at floor pricing. Second, the multi-year supply constraint: Samsung and SK Hynix have been redirecting wafer capacity from NAND toward High Bandwidth Memory, reducing overall NAND supply even as AI demand grows. Third, the Q1 FY2027 guidance: $10.3 billion to $10.8 billion in revenue and $44 to $46 per share in non-GAAP earnings — both well above prior analyst consensus.

The structural defense and the historical cycle pattern are not mutually exclusive. The memory industry built its most aggressive new capacity during prior cycles' peak earnings — and that new supply arrived just as demand moderated. Micron has guided fiscal 2026 capital expenditure above $25 billion. SK Hynix has board approval for a $38.3 billion fab expansion. Samsung is expanding aggressively. History suggests that when all three major memory suppliers commit to capacity simultaneously, oversupply typically emerges within two to three years. Agreements can shape how a downturn arrives without preventing one.

What the Analyst Range Reflects

The analyst consensus on SNDK — with an average 12-month price target of approximately $2,164 across 23+ analysts covering the stock, implying roughly 37% upside from Wednesday's price — reflects the broad institutional view that the structural case is credible. Cantor Fitzgerald's CJ Muse leads the bull camp at $2,900, citing AI inference workloads sustaining premium NAND pricing into the decade. JPMorgan's Harlan Sur, who upgraded SNDK to Overweight with a $2,250 target after the Investor Day, described the NBM framework as having "structurally reset" SanDisk's margin profile. Bernstein called the HBF technology "a game changer for AI" and maintained a $3,000 target.

Morningstar Senior Equity Analyst William Kerwin holds the counterweight: a $1,000 fair value estimate and a "no economic moat" rating, based on the assessment that NAND is a commodity whose pricing is determined by industry-wide supply conditions rather than any single producer's competitive advantage. At the bearish extreme, Morningstar's argument is that the extraordinary margins are a product of temporary supply tightness, not structural differentiation — that SanDisk's 84.6% gross margins are identical to what any NAND producer would have earned in the same environment, because customers had no alternative.

The institutional departure that carries the most weight is David Tepper's Appaloosa Management, which fully exited its SNDK position — worth more than $400 million at peak value — in the second quarter of 2026, according to the fund's Q2 13F filing submitted August 14. Appaloosa simultaneously reduced its Micron position by 41% and rotated into Amazon, Meta, and Alphabet — the hyperscalers that purchase NAND storage rather than produce it. Tepper appears to have expressed his AI infrastructure conviction through the demand side of the equation rather than the supply side.

What OCP Korea Could Reveal That Investor Day Did Not

The Investor Day on August 13 delivered a financial framework — mid-to-high-teens revenue growth through fiscal 2030, gross margins near 80%, operating margins near 75%, free cash flow near 50% of revenue, 100% excess cash returned to shareholders. It quantified the NBM backlog and growth targets. What it could not do is demonstrate whether the technical standard underlying the next growth layer — HBF — is gaining traction in the engineering community that actually designs AI inference systems.

OCP Korea Tech Day is a practitioner conference: the 1,500 engineers at Seoul's COEX will be evaluating the HBF specification not as a financial thesis but as an engineering specification they might eventually implement. Stenfort's session and the adjacent KV cache tiering sessions will be the first indication of whether hyperscaler engineering teams — the ones actually signing NBM agreements — are treating HBF as a near-term roadmap commitment or a longer-horizon research direction. Those two outcomes have meaningfully different implications for the volume-side of the ASP-to-volume equation that currently sits at 1:2.

Investors watching Thursday's event should focus on whether KV cache tiering sessions attract specific architecture commitments from named hyperscalers, and whether the HBF consortium membership — which currently includes Google and Tenstorrent alongside SanDisk and SK Hynix — expands to include Microsoft or Amazon, whose AI infrastructure spending collectively represents the largest single driver of enterprise NAND demand.

What the Next 90 Days Will Actually Answer

The pricing composition question — 2:1 ASP-to-volume — will be answered by the next two quarters of reported results, not by a single conference session. SanDisk's Q1 FY2027 guidance implies continued strength through the period ending late September or early October. Micron is expected to report its own Q4 fiscal 2026 results in late September, which will provide a second data point on whether enterprise NAND ASPs are holding at current levels or beginning to compress.

The bear case does not require a catastrophic supply shock. It only requires that the rate of ASP increases decelerates — that NAND pricing rises 40% instead of 70% in the next quarter. At that deceleration, SanDisk's revenue still grows, but at a rate that makes the current forward P/E of approximately 22 times fiscal 2026 earnings look less asymmetrically cheap than it does today.

The bull case requires exactly the opposite: that AI inference demand continues to outpace even the production ramp enabled by SanDisk's own capacity expansion, Samsung's BiCS output growth, and the supply additions that high NAND prices inevitably invite. The NBM contracts provide a floor — but they cover approximately half of bit shipments, not all of them. The uncovered half still faces the commodity cycle's arithmetic.

For an investor sitting with SanDisk at $1,581 — 32% below its all-time high and 3,400% above where it was 12 months ago — the question is not whether the AI demand story is real. It is whether the pricing composition that turned a 26% gross-margin business into an 84% gross-margin business was the beginning of a sustained structural shift or the peak of a cycle that, like every prior NAND cycle, will eventually normalize.

The next earnings report will know more than any Investor Day can.

This article is for informational purposes only and does not constitute investment advice. Market data as of August 19, 2026.

Frequently Asked Questions

What does it mean that two-thirds of SanDisk's Q4 revenue growth came from higher prices rather than higher volume?

In commodity industries, revenue typically grows because companies ship more product — volume growth. When revenue grows primarily because prices are rising, it signals that supply is so constrained relative to demand that buyers have no pricing leverage. SanDisk's 2:1 ASP-to-volume ratio in Q4 FY2026 means it raised prices twice as fast as it raised output. That is structurally unusual for NAND flash, which has historically been highly competitive. It is the single number that distinguishes a cyclical pricing peak (which will reverse when supply catches up) from a structural break (which would sustain at elevated margins). The historical record offers examples of both outcomes after similar ratios, but the determining factor in every prior case was whether earnings held in the two to four quarters following the peak.

What is High Bandwidth Flash (HBF), and how is it different from the NAND used in regular SSDs?

Conventional NAND flash SSDs connect flash dies to a controller over a standard interface and are accessed sequentially rather than randomly, making them slower than DRAM for data that needs to be read repeatedly and quickly. High Bandwidth Flash (HBF) stacks multiple NAND flash dies vertically and connects them with through-silicon vias — the same technology that makes High Bandwidth Memory fast — then interfaces directly with AI accelerator packages. SanDisk's target HBF configuration delivers read bandwidths of 1.6 terabytes per second at capacities up to 512 gigabytes per stack, versus roughly 64 gigabytes per stack for HBM4. The key difference: NAND provides approximately 30 times the bit density of DRAM, meaning a GPU package combining HBF stacks with HBM stacks could carry far more total memory capacity than HBM alone at comparable cost. The HBF OCP specification, published August 3, creates the technical framework that accelerator designers need to build HBF support into their hardware.

Why did SanDisk drop 8% on a Morgan Stanley report that did not change its earnings outlook?

Morgan Stanley's note did not change SanDisk's earnings estimates — it identified that large institutional funds hold approximately 2.3 percentage points more SanDisk stock than the company's S&P 500 index weight warrants. That is a positioning signal, not a fundamental judgment. When a stock is over-owned relative to its benchmark weight, any market event that prompts institutional risk managers to reduce exposure forces all of them to sell simultaneously, amplifying the move. After SanDisk's 35% rally in the week following Investor Day, the stock had become the most over-owned position in institutional semiconductor portfolios — meaning the next catalyst for selling was amplified by the concentration itself. The August 18 drop reflects crowded positioning being unwound, not a change in the underlying NAND demand story.

Does QLC NAND's lower write endurance limit SanDisk's AI storage opportunity?

QLC (quad-level cell) NAND stores four bits per transistor, which achieves high density at lower cost per gigabyte but requires more complex error correction and degrades faster under repeated write cycles than TLC or SLC NAND. For AI inference specifically — where KV cache is read far more often than it is overwritten — QLC's endurance limitation is less severe than in consumer or write-intensive enterprise applications. The risk is workload evolution: if AI compute shifts toward architectures that require more frequent write operations (such as continuous model fine-tuning, federated learning, or on-device training at scale), QLC's endurance tradeoffs become more significant. SanDisk's current AI storage thesis is most robust under the assumption that inference remains the dominant AI workload — a reasonable assumption today, but one that could shift as AI deployment models mature.

Originally published on Tech Times