Sandisk’s Optimistic Outlook Challenges the Old Memory-Cycle Playbook

Sandisk triggered a new Google News memory rally after presenting long-term targets that challenged the sector’s familiar boom-and-bust pattern. TradingKey reported that Kioxia jumped more than 8%, while SK hynix gained over 6% as investors weighed Sandisk’s outlook.

Those moves followed Sandisk’s August 13 investor day, where management described a business built around AI storage, customer commitments, and controlled capacity planning. The presentation followed fiscal fourth-quarter results that had already shown extraordinary revenue, margin, and data-center growth.

The real story is not one strong trading session. Sandisk is asking investors to believe that customer contracts and AI inference can make NAND earnings more predictable. That claim puts the company against decades of evidence that memory profits eventually attract supply, competition, and falling prices.

Kioxia and SK hynix sit on both sides of that argument. They benefit when Sandisk validates stronger flash demand, but they also compete for the same customers and capital. Their Asian share moves turned a United States investor presentation into a broader test of the memory industry’s valuation.

What the Sandisk Announcement Actually Changed

Sandisk replaced a short-term recovery story with a detailed claim about durable earnings from 2028 through 2030.

At its investor day, Sandisk projected mid-to-high-teens annual revenue growth during fiscal 2028 through fiscal 2030. It also targeted an approximately 80% non-GAAP gross margin and an approximately 75% non-GAAP operating margin.

The company expects adjusted free cash flow margin to reach approximately 50% over that period. Sandisk also said it plans to return 100% of excess cash after funding required business investment.

These are company targets, not independently verified outcomes. Sandisk acknowledged that the model depends on estimates and assumptions that remain subject to operational, competitive, and macroeconomic risks.

Still, the numbers changed the market conversation. Investors had already seen a sharp NAND recovery. The investor day argued that exceptional profitability can persist after the immediate shortage and pricing surge fade.

Sandisk’s long-term financial model rests partly on New Business Model agreements, or NBMs. These are contracts involving committed volumes, minimum financial guarantees, and structured pricing mechanisms.

Sandisk said it has signed these agreements with eight customers. The contracts represent approximately 50% of expected fiscal 2027 bit shipments and roughly two-thirds of expected fiscal 2028 bits.

That coverage matters because NAND suppliers traditionally build capacity before knowing precisely how much customers will buy. When demand disappoints, excess inventory pushes prices and margins down across the industry.

Committed volumes offer a different planning signal. Sandisk says the agreements help align customer demand with capacity decisions and provide better revenue visibility.

The contracts do not eliminate the cycle. Customers can change deployment schedules, contract enforcement can become contentious, and market prices can still influence future negotiations. However, the agreements reduce the amount of production exposed to immediate spot-market conditions.

The company paired that contractual argument with an AI infrastructure forecast. Sandisk expects the enterprise data-center flash market to reach 1.2 zettabytes by 2030.

A zettabyte equals one trillion gigabytes. The forecast reflects Sandisk’s view that AI inference will require much more storage near expensive computing systems.

Inference is the process of using a trained model to generate answers, predictions, images, or actions. As inference volume grows, systems must repeatedly access model data and previous computational results.

Sandisk pointed specifically to KV cache, which stores previously calculated attention data during model operation. Moving portions of that cache into fast flash could expand storage demand beyond conventional file and database workloads.

The combination creates Sandisk’s new pitch. AI increases the addressable market, while contracts keep suppliers from responding with uncontrolled production.

That pitch explains why a Google News headline about Asian stock gains carried more significance than a routine sympathy rally. The market was reacting to a proposed change in how NAND demand, supply, and profits connect.

Why Kioxia and SK hynix Joined the Google News Rally

Kioxia and SK hynix rose because Sandisk’s model strengthens the economic case for scarce, higher-value memory across several product categories.

TradingKey’s Google News headline reported gains exceeding 8% for Kioxia and 6% for SK hynix. Those percentages should be read as reported market moves tied to a specific trading window and local exchange session.

Different market feeds can show different changes based on closing time, currency, and security type. SK hynix also has securities trading across different markets, which can produce mismatched daily percentages.

The direction of the reaction remains understandable. A durable NAND outlook benefits other memory manufacturers because the industry sells related products into overlapping AI infrastructure budgets.

Kioxia has the closest connection. It co-develops and manufactures BiCS flash technology with Sandisk through long-running Japanese production arrangements.

Sandisk relies on that relationship for a significant portion of its flash supply. Its own risk disclosures identify dependence on Kioxia as a material operational consideration.

The partners recently introduced new BiCS10 QLC technology for data-intensive applications. QLC flash stores four bits in each memory cell, increasing capacity while requiring careful performance and endurance management.

The new design combines 332 active layers with a faster interface and higher bit density. Independent technical reporting described the companies’ 332-layer flash as targeting high-capacity data-center solid-state drives.

Sandisk also said its BiCS10 QLC node delivers a 60% bit-density increase over BiCS8. More bits per wafer can improve production economics when yields and customer qualification meet expectations.

A stronger Sandisk demand forecast therefore supports Kioxia in two ways. It improves confidence in their shared technology roadmap and increases the expected utilization of their manufacturing partnership.

Kioxia has made a similar AI inference argument. During its June investor day, the company said flash could become extended memory for graphics processors and other accelerators.

Kioxia aims to lift data-center and enterprise products above 60% of its sales mix over the medium to long term. Its AI inference strategy also emphasizes multi-year customer agreements and more stable profit quality.

The overlap makes Kioxia a natural beneficiary of Sandisk’s investor day. It also creates an important tension because each company wants investors to credit it for the value produced by shared technology.

SK hynix has a more diversified memory position. It is a major supplier of high-bandwidth memory, or HBM, which feeds data directly to AI accelerators.

HBM uses vertically stacked DRAM to deliver much more bandwidth than conventional memory modules. It has become a central component in high-end AI systems.

SK hynix also participates in NAND through Solidigm and its broader flash operations. Stronger enterprise SSD demand can improve the outlook for that business alongside its HBM franchise.

Sandisk and SK hynix are also cooperating on High Bandwidth Flash, or HBF. HBF aims to place high-capacity NAND closer to processors while delivering much greater throughput than traditional storage.

The companies released an initial specification through the Open Compute Project. The proposed design supports packages with capacities reaching 512 gigabytes and bandwidth reaching 3 terabytes per second.

Those figures describe a specification, not a commercially proven product. The HBF specification still requires hardware implementation, customer validation, and a supporting ecosystem.

Even so, the collaboration shows why SK hynix can benefit when investors take Sandisk’s AI storage thesis seriously. It expands the role of flash without directly replacing HBM’s highest-performance functions.

The Asian rally therefore reflected more than sector momentum. Sandisk supplied a common demand narrative for partners and competitors spanning NAND, enterprise SSDs, HBF, and conventional HBM.

Sandisk Says Contracts Can Tame the Memory Cycle

The decisive question is whether contractual demand can prevent the overinvestment that has repeatedly ended memory upcycles.

Memory manufacturing has high fixed costs. Suppliers spend heavily on fabrication equipment, process transitions, and production facilities before receiving revenue from finished chips.

Once installed, that capacity encourages manufacturers to keep producing because each additional unit can cover part of those fixed costs. Several suppliers making the same choice can flood the market.

Demand forecasting makes the problem harder. Consumer electronics, cloud infrastructure, and enterprise purchasing can change faster than fabrication plans.

Customers also build inventory when they expect shortages. They later reduce orders while consuming that stock, producing a sudden demand gap for manufacturers.

Sandisk argues that its NBMs provide an answer. Committed volumes and minimum financial guarantees give the company stronger evidence before assigning supply.

Structured pricing can also reduce dependence on immediate market quotations. That can protect both sides from extreme movements, depending on each contract’s undisclosed terms.

The company’s latest financial results make the argument look credible at first glance. Sandisk reported fiscal fourth-quarter revenue of $8.97 billion, up 51% sequentially and 372% from the prior-year period.

Management said approximately one-third of sequential revenue growth came from higher volumes. Roughly two-thirds came from higher pricing.

That split reveals both the strength and vulnerability of the current results. Pricing produced most of the sequential increase, so a future price reversal would have an outsized impact.

Fiscal fourth-quarter gross margin reached 84.6%. Fiscal-year revenue rose 175% to $20.25 billion, while data-center revenue grew 437%.

The company’s quarterly results also showed fiscal fourth-quarter data-center revenue of $2.98 billion. That was more than double the previous quarter’s level.

Sandisk expects fiscal first-quarter 2027 revenue between $10.3 billion and $10.8 billion. It guided for a non-GAAP gross margin between 83% and 85%.

These results show unusually strong demand and pricing. They do not yet prove that the long-term contract system can preserve comparable economics through a weaker market.

The fiscal 2028 through 2030 targets go much further than near-term guidance. Sandisk expects approximately 80% non-GAAP gross margins even after the present period’s supply pressure evolves.

That is the core reversal behind the Google News rally. The company is treating current profitability as the foundation of a new model, rather than the temporary peak of an old cycle.

Investors have heard similar “this time is different” arguments during earlier commodity upswings. New demand sources can extend a cycle without removing the incentives that eventually create excess supply.

AI does add a substantial source of demand. Training clusters need HBM, enterprise SSDs, networking, and large pools of conventional storage.

Inference broadens the opportunity because it spreads computing across more applications and locations. Each deployed model can generate new stored data, logs, embeddings, and cached computations.

However, higher demand alone does not guarantee durable margins. Supply discipline matters just as much, and Sandisk cannot control every manufacturer.

Samsung, Micron, SK hynix, Kioxia, YMTC, and other suppliers make independent investment decisions. A favorable pricing environment can motivate several of them to expand simultaneously.

Long-term agreements could reinforce discipline if customers accept meaningful commitments. They could weaken if customers resist guaranteed economics after market prices fall.

The missing information sits inside the contracts. Sandisk has not publicly disclosed detailed pricing formulas, cancellation protections, or customer remedies for significant market changes.

Without those terms, investors cannot fully measure how much risk has moved from Sandisk to its customers. They also cannot know how durable the agreements will be during a downturn.

The NBMs remain important because they cover large portions of planned output. Yet their real test will arrive when spot pricing, contract pricing, and customer deployment schedules move in different directions.

The Rally Does Not Erase Competition or Valuation Risk

Sandisk’s targets are unusually ambitious, while growing competition and volatile share prices leave little room for execution mistakes.

Memory stocks have experienced sharp reversals throughout 2026. Strong earnings have not prevented sudden declines when investors questioned AI spending, Chinese competition, or future supply.

A late-July regional selloff showed that sensitivity. Kioxia fell 18.3%, while SK hynix dropped 14.7% during a broader Asian semiconductor retreat.

The Asian chip selloff followed concerns about elevated valuations, AI financing, and Chinese memory competition. Those issues did not disappear after Sandisk’s investor day.

The later rebound instead showed how quickly sentiment can change. The same operational leverage that expands memory profits during shortages can accelerate losses when pricing turns.

Sandisk’s long-term model assumes revenue growth, sustained margins, disciplined expenses, and controlled investment. Missing one assumption can affect several others.

For example, weaker pricing would reduce revenue and gross profit without immediately lowering research or administrative expenses. Slower bit growth could also leave manufacturing commitments underused.

Faster bit growth creates another problem. Sandisk and Kioxia must introduce denser nodes without sacrificing yields, reliability, or customer qualification schedules.

BiCS9 and BiCS10 use new combinations of memory arrays and control circuitry. Sandisk says its bonded architecture enables more flexible scaling and capital-efficient product variants.

Those claims still require volume production evidence. Sampling a device and producing it economically at customer scale are different milestones.

HBF carries even more uncertainty. The concept promises capacity beyond HBM while offering bandwidth above conventional flash devices.

However, HBF must gain support from accelerator designers, server vendors, controller developers, and software platforms. A specification alone does not create demand.

The technology also occupies an awkward position. It cannot match DRAM’s latency, yet it will cost more than ordinary data-center storage.

Customers must find workloads where capacity and bandwidth matter more than the latency gap. KV cache storage is one candidate, but real deployments must establish the value.

Competition adds pressure. Kioxia is developing its own high-bandwidth SSD and flash products for AI inference.

SK hynix benefits from established HBM leadership and can integrate NAND through Solidigm. Micron sells both DRAM and NAND into data centers, while Samsung has broad manufacturing scale.

Chinese manufacturer YMTC presents another challenge, particularly in high-density NAND. Trade restrictions can limit access to some customers and equipment, but they do not eliminate its impact on global supply.

Kioxia and Sandisk also share technology while remaining separate public companies. That arrangement creates operational efficiency but complicates how investors assign value.

A successful shared process can benefit both companies. Production disruptions, investment disagreements, or uneven customer exposure can also affect them differently.

Sandisk expressly lists reliance on Kioxia among its business risks. The partnership therefore supports the optimistic model while remaining one of its important dependencies.

SK hynix faces a different balance. Its HBM position gives it direct exposure to AI accelerator demand, but that success can overshadow the economics of its NAND operations.

The reported 6% share gain does not mean investors valued every business segment equally. It reflects a market response to the combined memory demand narrative.

There is also a measurement problem around the rally itself. Financial headlines often compare the latest trade with the previous local close.

Korean, Japanese, and United States markets operate at different times. News released during one session can affect another security before the underlying home-market shares reopen.

Readers should therefore avoid treating a single headline percentage as a precise measure of corporate value creation. The moves are evidence of sentiment, not proof of Sandisk’s forecast.

The strongest skeptical interpretation is straightforward. Sandisk announced aggressive margins near the top of a pricing cycle, and investors extrapolated those conditions years forward.

The strongest favorable interpretation also deserves attention. Customer commitments and AI storage demand have altered supply planning enough to make past cycle comparisons incomplete.

Both positions can be true for a period. Contracts can soften volatility without eliminating it, while AI can extend demand without guaranteeing permanent scarcity.

Google News Is Following a Fight Over AI’s Memory Hierarchy

The wider industry question is whether flash becomes an active layer of AI computing rather than remaining a secondary storage destination.

AI systems have traditionally separated computing memory from persistent storage. HBM and DRAM hold data close to processors, while NAND SSDs retain larger datasets farther away.

That division reflects a tradeoff. DRAM is fast but expensive and loses data without power. NAND is slower but denser, nonvolatile, and more economical.

Larger models strain that arrangement. Keeping every model parameter, cache entry, and retrieval object in expensive memory becomes impractical.

Developers increasingly use tiered systems that move data among HBM, DRAM, local SSDs, and remote storage. Software decides which information needs immediate access.

Sandisk’s strategy assumes flash will move higher in this hierarchy. HBF and high-bandwidth SSDs aim to make NAND useful for frequently accessed AI data.

KV cache provides a concrete example. A model can reuse previously calculated attention states instead of recomputing the full conversation every time.

Long contexts and repeated requests can make those caches enormous. Storing less-active portions on fast flash can reduce pressure on scarce accelerator memory.

The benefit depends on latency, bandwidth, software scheduling, and request patterns. A poorly managed storage tier can slow responses enough to erase its cost advantage.

This is why the HBF ecosystem matters. Hardware specifications must connect with accelerator interfaces, memory controllers, operating systems, and inference frameworks.

The Open Compute Project route could help multiple companies align around common interfaces. It also prevents Sandisk from controlling the entire standard alone.

SK hynix’s participation adds credibility because the company understands both HBM and NAND. It can evaluate where flash complements high-bandwidth DRAM instead of competing directly.

Kioxia approaches the opportunity through high-bandwidth SSDs and its CM product family. Its strategy also treats stored KV cache as a way to improve GPU utilization.

These parallel efforts suggest the industry sees a genuine architectural problem. They do not show which implementation will win.

Conventional enterprise SSDs may capture much of the opportunity without a new memory category. Software vendors could also improve cache compression or reduce memory use.

Accelerator companies might add more HBM, change interconnects, or develop proprietary storage extensions. Each alternative changes the addressable market behind Sandisk’s forecast.

Google News interest has focused on stock gains because they provide a visible response. The more consequential competition is happening inside future AI servers.

If flash becomes essential to inference, the opportunity extends beyond unit growth. Suppliers can sell products with greater performance, integration, and reliability requirements.

Higher-value products usually support better margins than undifferentiated consumer storage. That mix shift helped Sandisk’s recent data-center growth.

Yet specialization requires investment. Suppliers must fund controller development, firmware, qualification, support, and new packaging.

Customers will demand evidence that these systems improve total computing efficiency. A denser device has limited value if it increases response time or complicates operations.

Sandisk’s long-term targets assume it can capture the economic value of that transition. Competitors intend to capture the same value.

The rally therefore represents an industry-wide vote on the size of the opportunity, not a final judgment about the winners.

Three Signals Will Test Sandisk’s Outlook

Contract performance, production execution, and customer adoption will determine whether the rally marked a structural change or another cyclical peak.

The first signal is Sandisk’s fiscal 2027 contract coverage. Management says NBMs already represent approximately half of fiscal 2027 bits and two-thirds of fiscal 2028 bits.

Investors should watch whether those percentages rise without weaker protections or unfavorable pricing. Stable margins alongside growing committed volumes would strengthen Sandisk’s argument.

Any contract delay, renegotiation, or customer concentration warning would weaken it. The company’s disclosures should also show whether cash collections match reported profitability.

The second signal is the BiCS9 and BiCS10 production ramp with Kioxia. Density improvements matter only when factories deliver acceptable yields and qualified products at scale.

A smooth ramp would support the claim that Sandisk can increase bits without proportionally increasing capital spending. It would also reinforce Kioxia’s position as a central beneficiary.

Manufacturing delays would challenge the long-term margin framework. They could force continued reliance on older nodes or increase costs during a critical demand window.

The third signal is real adoption of HBF and flash-based KV cache systems. Announced specifications and prototypes must become customer programs with defined deployment schedules.

Design wins from accelerator vendors, cloud operators, or server manufacturers would support the 1.2-zettabyte enterprise flash forecast. Production shipments would provide stronger evidence than demonstrations.

Limited ecosystem participation would weaken the thesis. It would suggest that AI storage demand remains concentrated in conventional SSDs rather than a new memory layer.

Investors should also distinguish these operating signals from daily share prices. Kioxia, SK hynix, and Sandisk can rise together even when their product exposure differs substantially.

The opposite is also true. Their stocks can fall together because of macroeconomic sentiment despite stable customer orders.

Sandisk has offered an unusually clear test. It expects long-term contracts, AI inference, and disciplined investment to sustain economics once considered temporary.

That forecast deserves attention because the latest results provide a strong starting point. It deserves skepticism because pricing created most of Sandisk’s recent sequential revenue growth.

The next one to three months should reveal more about customer commitments, production schedules, and the first commercial HBF relationships. Those facts will matter more than one Google News rally.

For developers and enterprise technology buyers, the outcome will influence SSD availability, infrastructure design, and the cost of serving long-context AI applications. Watch where vendors place KV cache, which interfaces gain support, and whether contractual supply produces steadier purchasing conditions. Those decisions will show whether flash is becoming part of the AI computing path or simply benefiting from another shortage. Sandisk has placed a measurable bet on the first outcome. Kioxia and SK hynix are participating through different products and partnerships. The useful question is no longer whether AI needs more memory. It is whether these companies can supply that memory without recreating the excess capacity that ended previous cycles.