Sandisk’s AI Storage Forecast Is Outrunning Wall Street

Sandisk lifted its forecast after data-center revenue rose 64% in one quarter, challenging Wall Street’s assumptions about where AI infrastructure spending goes next.

The company’s January results first reached many investors through Google News coverage focused on its surprisingly strong outlook. Yet the important development was not another chipmaker benefiting from AI enthusiasm. Sandisk showed that AI infrastructure demand was spreading from processors into the flash storage systems surrounding them.

That shift puts Micron, Kioxia, and other memory suppliers on notice. It also pressures cloud operators that need more storage capacity while NAND supply remains constrained. Sandisk’s opportunity rests on that collision between rising enterprise SSD demand and disciplined manufacturing capacity.

The Forecast That Reset Sandisk’s AI Story

Sandisk’s January forecast turned an improving memory cycle into a specific claim about AI storage demand.

Sandisk reported fiscal second-quarter revenue of $3.03 billion on January 29, 2026. That represented 31% sequential growth and a 61% increase from the prior-year quarter.

The company also reported GAAP net income of $803 million, compared with $112 million in the previous quarter. Its GAAP gross margin climbed from 29.8% to 50.9% during the same period.

Those improvements went well beyond a modest recovery in consumer flash products. Datacenter revenue reached $440 million, rising 64% sequentially and 76% from a year earlier.

Sandisk attributed that expansion to adoption among AI infrastructure builders, semi-custom customers, and technology companies deploying AI at scale. The company’s quarterly results connected the demand directly to enterprise solid-state drive deployments.

An enterprise SSD stores data in NAND flash rather than on spinning magnetic disks. It offers faster access, lower latency, and different power characteristics than a conventional hard drive.

The forecast carried more weight than the completed quarter. Sandisk projected fiscal third-quarter revenue between $4.40 billion and $4.80 billion.

It also expected non-GAAP diluted earnings per share between $12 and $14. Finimize noted that the midpoint exceeded the contemporary LSEG consensus by a wide margin.

That difference explains why the story spread quickly through market feeds and Google News. Analysts had modeled Sandisk as a participant in a cyclical NAND recovery. Management was describing something larger and more concentrated.

AI training initially made accelerators and high-bandwidth memory the most visible infrastructure bottlenecks. Inference creates a different mix of demands because deployed models repeatedly retrieve data while responding to users.

Inference is the process through which a trained model generates an answer from new input. Large deployments must move model data, cached information, and retrieval content through multiple storage layers.

GPUs cannot hold every useful dataset in their fastest local memory. Operators therefore need a hierarchy combining high-bandwidth memory, conventional DRAM, fast SSDs, and lower-cost archival storage.

Sandisk occupies the NAND-based portion of that hierarchy. Its January results suggested that storage was becoming a meaningful constraint rather than a secondary purchase made after servers arrived.

Management also pointed to a better product mix and stronger pricing conditions. Buyers seeking dependable supply were placing greater value on capacity allocated to enterprise and data-center products.

This combination amplified Sandisk’s financial results. Higher volumes increased revenue, while limited supply and premium products supported margins.

The result was not proof that AI had permanently changed NAND economics. It was evidence that Wall Street’s near-term models had underestimated both demand and operating leverage.

That distinction matters. Semiconductor forecasts often overshoot near a cycle’s peak, especially when customers order early to protect supply.

Still, Sandisk gave investors more than a hopeful projection. Its subsequent fiscal third-quarter results provided a demanding test of the original claim.

Google News Captured the Surprise, but the Numbers Kept Moving

The January forecast looked aggressive until Sandisk reported results that exceeded it only three months later.

Sandisk generated $5.95 billion in fiscal third-quarter revenue, according to its April earnings release. That was 97% above the preceding quarter and 251% above the prior-year period.

The result also exceeded the company’s January guidance range. It was not a narrow beat caused by rounding or a small timing benefit.

Datacenter revenue rose to $1.47 billion from $440 million. That represented 233% sequential growth and 645% year-over-year growth.

Edge revenue, which includes devices and systems closer to users, reached $3.66 billion. Consumer revenue declined 10% sequentially to $820 million, making the data-center shift easier to see.

Sandisk’s GAAP gross margin climbed to 78.4%. GAAP net income reached $3.62 billion, compared with $803 million during the second quarter.

The company then guided fiscal fourth-quarter revenue to a range between $7.75 billion and $8.25 billion. Expected non-GAAP diluted earnings per share ranged from $30 to $33.

These figures substantially strengthened the narrative first reflected in Google News headlines after January. Sandisk was not merely enjoying a general rebound in removable storage or personal computers.

Its revenue mix was moving toward higher-value data-center customers. AI infrastructure spending was becoming a direct financial driver rather than an indirect marketing theme.

The progression is striking. Datacenter revenue moved from $269 million in fiscal Q1 to $440 million in Q2, then reached $1.47 billion in Q3.

The increase also changed who faces the most pressure. Cloud companies need dependable storage supply while expanding inference services, retrieval systems, and data-processing pipelines.

Large AI deployments generate several storage workloads. Training datasets must be prepared and retained, checkpoints preserve model states, and inference systems maintain caches and indexes.

Retrieval-augmented generation adds another layer. It searches external information before a model answers, requiring accessible stores for documents, embeddings, and associated metadata.

Embeddings are numerical representations that help systems compare the meaning of different pieces of content. Searching billions of them requires capacity alongside fast data access.

Not every workload belongs on the fastest SSD. However, operators must balance latency, power use, physical space, endurance, and cost across the storage hierarchy.

That balance is becoming more important as inference expands. A data center built only around accelerator performance can leave expensive processors waiting for data.

Sandisk argues that this architecture raises the strategic value of flash storage. Its revenue trajectory supports the direction of that claim, even if it cannot isolate every AI-related purchase.

The figures also show why customers might accept long-term agreements. Storage buyers face operational risks when capacity arrives late or fails qualification inside a production system.

Switching suppliers is possible, but enterprise validation takes time. Controllers, firmware, endurance profiles, power behavior, and service requirements must all meet customer expectations.

A shortage therefore changes purchasing behavior before factories reach absolute capacity. Customers can reserve supply earlier, accept longer commitments, or prioritize availability over negotiations.

That behavior can make a favorable cycle appear more durable. It can also create future risk if buyers accumulate inventory faster than end demand grows.

For now, Sandisk’s April results validated the January forecast. The unanswered question concerns the mechanism sustaining that demand.

AI Inference Is Turning Flash Supply Into the Bottleneck

Sandisk’s advantage comes from a simple mechanism: AI systems need stored data near compute, while new NAND capacity cannot appear instantly.

NAND flash manufacturing requires specialized fabrication facilities, multiyear technology development, and close coordination across equipment and materials suppliers. Producers cannot increase output like a software company adding cloud instances.

Manufacturers also remember the previous NAND downturn. Excess capacity contributed to weak pricing, inventory corrections, and painful financial results across the storage industry.

That history encourages disciplined investment. Suppliers want to capture AI demand without recreating the oversupply that damaged earlier cycles.

Sandisk reinforced its access to production by extending its joint venture with Kioxia. Their Yokkaichi manufacturing agreement now runs through December 31, 2034.

The previous agreement was scheduled to expire at the end of 2029. The extension covers an additional five years and aligns with the companies’ Kitakami arrangement.

Under the Kioxia agreement, Sandisk agreed to provide manufacturing-related payments between 2026 and 2029. The structure gives it continued access to a major advanced-flash production base.

This arrangement matters because Sandisk does not compete through product design alone. Its ability to serve large data-center customers depends on predictable wafer supply and manufacturing execution.

The partnership has operated for more than 25 years. It combines joint NAND research with manufacturing scale at Kioxia’s Japanese facilities.

Long agreements cannot eliminate execution risk. They do, however, give Sandisk a clearer supply path when customers increasingly value allocation.

Technology improvements provide another route to additional capacity. More bits can fit into a given physical area when density rises, lowering the manufacturing resources required for each stored unit.

Sandisk began sampling its BiCS10 1-terabit triple-level-cell NAND in July 2026. Triple-level-cell NAND stores three bits in each memory cell.

The company says BiCS10 reaches density above 29 gigabits per square millimeter. It claims a 59% density improvement over BiCS8 and an interface speed of up to 4.8 gigabits per second.

Sandisk also reports 10% lower input power and 34% lower output power than BiCS8. These remain company measurements rather than independent production benchmarks.

Even so, the BiCS10 specifications show what data-center buyers increasingly demand. Capacity must rise without letting power use, rack space, or data movement overwhelm the system.

A denser flash chip can improve the economics of high-capacity SSDs. Faster interfaces can also feed information toward processors with fewer delays.

Neither improvement makes NAND equivalent to high-bandwidth memory. The two technologies occupy different positions in the system and serve different latency requirements.

High-bandwidth memory sits close to accelerators and provides extremely fast access. NAND offers far greater capacity at a lower cost per stored unit, but with higher latency.

AI infrastructure needs both. The investment story changes when spending expands from scarce accelerator memory into multiple surrounding storage layers.

This mechanism helps explain Sandisk’s stronger forecast. The company participates when data enters the data center, when models train, and when deployed services repeatedly retrieve information.

The demand also extends beyond model files. AI systems produce logs, user interactions, synthetic data, checkpoints, media, and databases that operators may retain for evaluation or reuse.

As those datasets grow, deleting everything is rarely practical. Moving all of it to slow archival systems can also undermine response times and developer workflows.

Enterprise SSDs provide an intermediate layer. They can keep frequently needed information closer to compute without consuming limited accelerator memory.

That role gives Sandisk leverage, but it does not give the company exclusive control. Micron, Kioxia, Samsung, SK hynix, and other suppliers are pursuing the same infrastructure opportunity.

Micron and Kioxia Will Test Sandisk’s Pricing Power

The contest is not Sandisk against weak demand; it is Sandisk’s allocation-led model against competitors expanding their own AI storage offerings.

Micron’s fiscal third-quarter results showed that the favorable memory environment extends beyond one company. Micron reported record revenue and described multiyear customer agreements as a source of greater predictability.

Its Core Data Center Business Unit generated $11.52 billion in fiscal Q3 2026. That was more than double the preceding quarter’s $5.69 billion.

Micron sells a broader memory portfolio than Sandisk, including DRAM and high-bandwidth memory. Direct revenue comparisons therefore do not represent equivalent businesses.

Still, Micron’s performance confirms that AI infrastructure customers are spending across memory and storage. It also shows the scale of the competitive response.

Micron is shipping a 245-terabyte data-center SSD based on quad-level-cell NAND. Quad-level-cell NAND stores four bits per cell, increasing density while introducing endurance and performance tradeoffs.

The company positions that product for large data lakes and capacity-intensive AI environments. Sandisk must compete against that approach while securing qualifications for its own enterprise drives.

Kioxia presents a different competitive relationship. It is Sandisk’s manufacturing partner, yet it also sells flash products and SSDs under its own name.

The joint venture gives both companies production scale. It does not prevent them from competing for high-value enterprise customers.

This combination makes the primary tension sharper. Sandisk needs shared manufacturing capacity while trying to capture more of the commercial value created by that capacity.

Customers benefit from this competition when multiple suppliers qualify comparable products. They gain negotiating leverage and reduce the operational danger of relying on one source.

However, tight supply weakens that leverage. A buyer facing an urgent AI deployment may prioritize guaranteed deliveries over the lowest possible unit cost.

Sandisk’s new business model responds to that behavior. By April, the company had signed three agreements under the model and added two more during the following quarter.

Management described these as multiyear customer engagements backed by firm financial commitments. Such contracts can improve visibility for both manufacturing plans and future revenue.

They can also transfer risk. Customers receive greater supply assurance, while Sandisk gains commitments that reduce its exposure to short-term purchasing changes.

This structure is attractive during scarcity. Its value becomes less certain when supply expands, technology changes, or customers revise infrastructure plans.

Competitive capacity remains the central pressure point. If rivals add output too quickly, buyers can regain negotiating power and flash pricing can retreat.

If manufacturers remain disciplined, shortages can persist longer. That would support Sandisk’s margins but raise costs for customers using NAND in PCs, phones, cameras, and other devices.

There is also a product-level challenge. High-capacity enterprise SSDs must satisfy demanding reliability and endurance requirements before hyperscalers deploy them widely.

A manufacturing technology can look promising in sampling but encounter delays during qualification. Yield, controller behavior, firmware, and thermal performance all affect the commercial ramp.

Sandisk’s BiCS10 announcement therefore represents a step, not completed market adoption. The company must translate density and power claims into qualified products shipped at scale.

Micron faces the same burden with its newer NAND and SSD platforms. Kioxia must also convert shared manufacturing advances into competitive systems.

The current shortage can hide some product differences because buyers need capacity. A looser market would expose those differences and make qualification wins more important.

That possibility explains why current margins should not be treated as permanent. Supply-constrained semiconductor profits often attract investment, substitution, and customer resistance.

Sandisk’s latest results show exceptional operating leverage. They do not repeal the memory industry’s cyclical behavior.

What the Forecast Still Cannot Prove

Sandisk has demonstrated strong demand and execution, but it has not proved that today’s pricing, margins, or purchasing behavior will survive a full cycle.

The first uncertainty concerns demand quality. Sandisk reports revenue by end market, but public figures do not reveal every customer’s inventory position or deployment schedule.

A buyer can place orders for several reasons. It may be consuming more storage, building safety stock, reserving scarce supply, or buying ahead of expected price changes.

Those motives produce similar near-term revenue. They lead to different outcomes when capacity becomes easier to obtain.

Inventory accumulation is especially difficult to measure from a supplier’s headline results. Orders can remain strong even while unused components gather elsewhere in the supply chain.

A later correction can then appear sudden. Customers pause purchases while consuming existing stock, causing supplier revenue and prices to fall together.

The second uncertainty involves AI infrastructure economics. Cloud operators are spending heavily, but each project still needs sufficient utilization and customer demand.

Inference volumes can grow while storage per task declines. Developers continually improve compression, caching, quantization, and data-retention policies.

Quantization reduces the numerical precision used by a model, lowering memory requirements and often improving execution efficiency. Better systems can therefore serve more requests with less infrastructure.

These efficiency gains do not automatically reduce total storage demand. Lower costs can increase usage enough to offset reduced requirements per query.

Still, forecasts should not assume every unit of AI activity produces a fixed amount of NAND demand. Architecture and software improvements can change that relationship.

The third uncertainty is supply discipline. Sandisk’s margins benefit from tight availability, but the company cannot control every competitor’s investment decisions.

Micron says it is investing at record levels to serve rapidly growing customer demand. Samsung, SK hynix, and Kioxia also possess substantial manufacturing capabilities.

New capacity takes time to arrive. When it does, several expansions can reach the market during the same demand window.

The fourth uncertainty comes from customer concentration. Data-center growth often depends on a limited group of hyperscalers and large technology companies.

These buyers possess technical expertise, purchasing scale, and the ability to redesign systems. A delayed deployment or supplier change can materially affect a smaller vendor.

Sandisk explicitly identifies changes in key customer relationships and customer consolidation among its business risks. Its reliance on Kioxia also creates partnership and manufacturing exposure.

The fifth uncertainty is valuation rather than operations. Strong earnings do not guarantee a rising share price when investors already expect exceptional growth.

That gap appeared across AI-related stocks during 2026. Companies could exceed current estimates and still face selling when guidance failed to outrun increasingly optimistic assumptions.

This is why an earnings surprise should not become a blanket investment conclusion. The operating evidence is strong, but the market can price that strength before it appears in financial statements.

The Google News framing of a lifted forecast captured the immediate event. It could not show whether the forecast represented sustainable demand or the steepest stage of a cycle.

Sandisk’s fiscal third-quarter performance moved the evidence toward sustained demand. It did not remove the cyclical question.

Readers should also separate revenue growth from technological leadership. Higher pricing can lift sales even without an equivalent increase in shipped bits.

A richer mix of enterprise products can further raise revenue per unit. Both are favorable for Sandisk, but neither reveals the complete volume picture.

Management’s statements deserve reporting language for the same reason. Sandisk says its new agreements create more durable earnings power.

The contracts can improve visibility, but their complete terms are not public. Investors cannot independently test every volume commitment, pricing mechanism, or cancellation protection.

The company’s outlook also remains forward-looking. Trade policy, manufacturing delays, weaker macroeconomic conditions, and competitive pricing can alter the result.

None of these risks invalidates the AI storage thesis. They define the evidence required to distinguish a structural shift from a profitable shortage.

Three Signals Will Decide What Comes After the Google News Rally

The next phase depends on contract durability, product qualification, and whether data-center growth remains strong after the easiest comparisons disappear.

The first signal is Sandisk’s conversion of multiyear agreements into delivered revenue. Investors need more than a rising count of signed contracts.

They should watch how management describes committed capacity, customer concentration, and the timing of shipments. Continued commitments beyond the current shortage would strengthen the durability argument.

A slowdown in new agreements would not automatically signal weak demand. Customers may already have secured enough supply, or negotiations may simply take longer.

However, reduced commitments combined with slower shipments would weaken Sandisk’s claim that purchasing behavior has structurally changed.

The second signal is the commercial qualification of BiCS10-based enterprise products. Sampling gives customers hardware for evaluation, but qualification determines whether it can enter production systems.

The important indicators include announced SSD products, named qualification milestones, volume-ramp timing, and evidence that density gains survive real operating requirements.

Successful qualification would strengthen Sandisk’s competitive position against Micron and Kioxia. Delays would leave the company more dependent on existing generations during a demanding capacity cycle.

Power efficiency deserves particular attention. Electricity and cooling constrain AI data-center expansion, so storage improvements must reduce total system pressure rather than add another burden.

The third signal is the direction of data-center revenue after extraordinary sequential growth. Sandisk cannot repeat a 233% quarterly increase indefinitely.

The relevant test is whether revenue settles at a materially higher base. Mix, margins, and customer commitments matter more than another eye-catching percentage.

Stable data-center sales alongside disciplined supply would support the structural thesis. Falling sales, rising inventory, and weaker pricing would point toward a conventional correction.

Competitor results provide an essential cross-check. Micron’s record quarter supports broad demand, while future inventory or margin changes could reveal a market turn.

Investors should also compare Sandisk’s progress with rival product shipments. Micron’s high-capacity SSD demonstrates that competitors are attacking the same workloads.

For developers and enterprise buyers, this contest affects more than semiconductor stocks. Storage availability influences infrastructure schedules, cloud costs, architecture choices, and data-retention policies.

Teams building retrieval systems should examine which information requires fast access. Moving every document into the most expensive storage layer wastes resources without improving every response.

Infrastructure leaders should also avoid assuming that current scarcity will persist forever. Contracts signed during a shortage can outlast the conditions that made them attractive.

At the same time, waiting for cheaper capacity carries its own cost. Delayed storage can strand accelerators, slow model deployment, or prevent teams from retaining useful operational data.

The practical question is therefore not whether AI needs storage. It is which data must remain close to compute, for how long, and under what performance guarantee.

Sandisk’s forecast made that question visible to Wall Street. Its later results showed that customers were already answering with substantial orders.

The next several quarters will reveal whether those orders define a new baseline or merely a steep memory upcycle. Watch contract-backed shipments, BiCS10 qualifications, and normalized data-center revenue in that order.

If all three hold, Sandisk’s January surprise will look less like an optimistic forecast and more like an early reading of AI infrastructure’s next bottleneck.