SNDK and SKHY Fall as Strong Results Raise the Bar

Sandisk and SK hynix shares moved lower before Thursday’s opening bell, despite results that reinforced the scale of the AI memory boom. The conflict facing Google News readers is unusually sharp. Revenue, margins, and data center demand remain strong, yet investors are demanding clearer proof that extraordinary pricing can last.

Sandisk entered the session after reporting fiscal fourth-quarter revenue of $8.97 billion, up 51% from the prior quarter. Revenue increased 372% from the year-ago period. Still, the initial market reaction showed that historical growth no longer settles the debate.

SK hynix faces the same test from a different position. It leads the market for high-bandwidth memory, or HBM, which feeds data to AI accelerators at very high speeds. Sandisk focuses primarily on NAND flash, the nonvolatile storage used in enterprise solid-state drives and other systems.

That distinction matters. Investors are deciding whether AI creates a durable expansion across memory and storage, or another cycle inflated by scarce supply and aggressive pricing. Micron, Samsung, and Kioxia give the market additional reference points, but SNDK and SKHY now capture the central tension.

Google News Focuses on the Selloff, but Sandisk’s Results Were Exceptional

The immediate selloff reflected elevated expectations, not weak reported demand.

Sandisk’s fiscal fourth quarter ended July 3, 2026. The company reported revenue of $8.97 billion, compared with $5.95 billion during the previous quarter. Its quarterly results also showed GAAP net income of $6.90 billion.

The revenue composition explains why the report attracted so much attention. Sandisk said approximately one-third of sequential growth came from higher volumes. About two-thirds came from higher pricing.

That split supports the bullish view in one sense. Buyers accepted substantially higher prices while Sandisk shipped more products. Demand was not limited to customers purchasing the same volume at inflated rates.

However, the same split creates the central risk. Pricing supplied most of the sequential increase, so investors need confidence that scarcity will persist. A normalization in average selling prices would quickly change the earnings profile.

Sandisk’s gross margin reached 84.6%, up 6.2 percentage points sequentially. Operating income increased 71% to $7.04 billion. Those figures show how strongly higher selling prices flow through a memory supplier’s income statement.

The company also reported a major shift toward data center customers. Quarterly data center revenue reached $2.98 billion, up 103% sequentially. Edge revenue increased 48% to $5.43 billion, while consumer revenue declined 32% to $556 million.

For the full fiscal year, data center revenue rose 437% to $5.15 billion. Total annual revenue reached $20.25 billion, a 175% increase. The results tie Sandisk’s expansion directly to enterprise infrastructure rather than consumer memory cards or portable drives.

Sandisk expects fiscal first-quarter 2027 revenue between $10.30 billion and $10.80 billion. That range implies further sequential growth. The company also expects a non-GAAP gross margin between 83% and 85%.

Yet the market had already priced in an exceptional trajectory. A strong forecast can disappoint when investors expect an even larger increase. That dynamic often appears when a stock becomes a concentrated expression of one popular investment theme.

The company added another signal through capital allocation. Sandisk’s board approved an additional $14 billion share-repurchase program. Total remaining authorization increased to $15.5 billion.

A buyback of that scale signals management’s confidence in future cash generation. It does not guarantee that current margins represent a sustainable baseline. Investors still need to separate supply-driven profitability from durable product differentiation.

Sandisk also disclosed five additional New Business Model agreements. Three involved new customers, while two expanded existing arrangements. These agreements appear intended to deepen customer commitments and improve revenue visibility.

The company has not disclosed every commercial term, so outsiders cannot fully measure their durability. Customer commitments help, but contract structures can differ in volume, duration, pricing, and cancellation protections.

That verification gap explains the premarket conflict. Google News headlines naturally emphasized falling shares. The underlying report, however, showed accelerating data center demand and continued pricing strength.

The better interpretation is not that investors rejected Sandisk’s results. They raised the standard for what counts as a positive surprise. Once revenue grows several times over, continued acceleration becomes harder to deliver.

AI Memory Demand Is Expanding Beyond HBM

The AI memory race now connects accelerator bandwidth, server capacity, and persistent storage inside one infrastructure budget.

HBM receives the most attention because modern AI accelerators need rapid access to model parameters and intermediate calculations. HBM stacks multiple DRAM dies near a processor, creating a wide interface with high data throughput.

SK hynix built its AI position around that bottleneck. The company supplies HBM for high-performance computing and has continued developing HBM4, HBM4E, and later generations. Its products compete directly with offerings from Micron and Samsung.

Sandisk addresses another part of the system. NAND flash stores data without continuous power and offers greater capacity at a lower cost per bit than DRAM. Enterprise SSDs use NAND to hold training data, model checkpoints, embeddings, and inference caches.

Those workloads are growing together. Training systems must read enormous datasets, while inference services repeatedly retrieve model components and user context. Larger clusters also generate more logs, checkpoints, and operational data.

This makes the simple “HBM versus NAND” framing incomplete. AI systems need both fast working memory and high-capacity persistent storage. Their revenue cycles can still differ because each market has distinct suppliers, inventories, and manufacturing constraints.

Micron’s fiscal third-quarter results reinforced the broader demand signal. The company reported revenue of $41.46 billion, compared with $23.86 billion one quarter earlier. Its record results covered HBM, DRAM, and NAND products.

Micron said its cloud memory business generated $13.77 billion in quarterly revenue. Its core data center unit produced another $11.52 billion. The company also guided for $50 billion in fourth-quarter revenue.

Micron reported that HBM4 was shipping in high volume for a lead customer’s platform. It had also sent qualification samples to multiple additional customers. That matters because qualification determines which suppliers enter future accelerator systems.

SK hynix remains heavily exposed to the same qualification cycle. Winning a position in a major accelerator platform can support several years of volume. Missing one can redirect demand toward Micron or Samsung.

NAND has a different route into AI infrastructure. Enterprise customers can qualify several SSD capacities and performance profiles across storage tiers. This creates a broader opportunity, but it also allows more substitution when supply conditions change.

Sandisk’s data center growth suggests that flash storage has moved closer to the center of the AI spending discussion. A 103% quarterly increase cannot be explained by consumer devices. It reflects large infrastructure deployments and a richer customer mix.

Still, investors should avoid assuming every AI capital expenditure dollar flows proportionally into memory. Accelerator design, networking, cooling, and software efficiency can change the memory required for each unit of compute.

Software teams continually reduce memory consumption through quantization, caching, compression, and more efficient model architectures. Quantization stores model weights with fewer bits, reducing capacity and bandwidth requirements. Those improvements can lower memory per task.

Usage growth can offset those gains. Cheaper inference often encourages developers to process more requests, retain longer context, and deploy larger models. Efficiency reduces unit requirements, while adoption expands total workloads.

The outcome depends on which force moves faster. Current financial results indicate workload growth is winning. They do not prove that the balance remains unchanged through 2027.

Investors should therefore watch customer purchasing behavior, not only supplier commentary. Long-term commitments, product qualifications, and sustained data center shipments offer stronger evidence than generalized statements about AI demand.

Sandisk and SK hynix Are Partners Inside the Main Contest

The defining contest is not Sandisk against SK hynix. It is durable AI demand against the memory industry’s history of oversupply.

Sandisk and SK hynix compete for investor attention, manufacturing resources, and positions within AI infrastructure. They also cooperate on technology intended to bridge the gap between NAND capacity and HBM bandwidth.

On August 3, the companies released the first technical specification for High Bandwidth Flash, or HBF. HBF stacks specialized NAND dies and connects them through a wide, high-speed interface for AI inference systems.

The specification was released through the Open Compute Project. That route matters because an open technical foundation can encourage adoption across processor, packaging, controller, and system vendors.

The first specification describes packages with up to 512GB of capacity. It supports eight-die and 16-die stacks, with projected bandwidth reaching 3 terabytes per second. Those figures describe a specification, not commercially deployed performance.

HBF does not replace HBM. NAND has different latency, endurance, and access characteristics. Instead, HBF aims to provide a larger, less costly memory tier for model data that does not require HBM’s lowest latency.

That tiered design addresses a real AI infrastructure problem. Large models can exceed the memory capacity located directly beside an accelerator. Operators then move data between HBM, system memory, storage, and networked resources.

Every movement consumes time and energy. A high-capacity flash package placed closer to the processor could reduce some traffic to conventional SSDs. It could also let systems reserve expensive HBM for the most latency-sensitive operations.

The design remains early. The specification does not establish production yields, real application performance, controller maturity, or customer adoption. Those questions will determine whether HBF becomes an important layer or a specialized product.

The partnership also complicates the stock-market narrative. Sandisk benefits if AI inference requires more NAND-based capacity. SK hynix benefits from HBM demand and could participate in the new HBF category.

Their shared work suggests that the industry expects memory hierarchies to become more complex. No single technology offers the best capacity, bandwidth, latency, endurance, and energy use at every point.

SK hynix is also trying to improve HBM’s thermal behavior. In May, it announced iHBM, which places integrated cooling elements inside the package. The company says the design reduces thermal resistance by 30%.

Thermal resistance measures how strongly a package impedes heat transfer. Lower resistance can help a memory stack operate more consistently beside processors that consume substantial power.

The company says iHBM uses its established MR-MUF packaging process. MR-MUF fills gaps between stacked chips with protective material, supporting structural stability and heat transfer. SK hynix plans to apply iHBM to future products, including HBM5.

These technical programs matter because the memory boom must become more than a pricing story. Persistent returns require products that improve system performance, capacity, or operating cost in ways customers cannot easily replace.

Sandisk has offered another technology signal through BiCS10. Developed with Kioxia, this tenth-generation 3D NAND design uses 332 active layers. Sandisk says its TLC version provides a 59% bit-density improvement compared with BiCS8.

Higher density allows manufacturers to place more storage capacity on a wafer. That can reduce production cost per bit after yields mature. It can also threaten pricing if several suppliers expand output faster than demand.

This is the old memory contradiction. Better manufacturing creates competitive products and larger margins at first. Widespread capacity growth can later push the market toward oversupply.

The AI memory race therefore contains cooperation and rivalry at the same time. Sandisk, SK hynix, Micron, Samsung, and Kioxia need shared standards and compatible systems. They also want the most valuable portion of each server.

The stock reaction reflects this deeper contest. Investors are rewarding present scarcity while preparing for the supply response scarcity eventually attracts.

What the Record Margins Do Not Prove

Record margins confirm an extraordinary upcycle, but they cannot establish how long suppliers will maintain pricing power.

Sandisk’s 84.6% quarterly gross margin offers the clearest reason for skepticism. It represents a dramatic change from 26.2% one year earlier. Such an expansion deserves attention precisely because it is so unusual.

The company attributed two-thirds of sequential revenue growth to pricing. That does not make the result low quality. It shows that limited supply and customer urgency currently give suppliers substantial leverage.

However, memory markets have repeatedly moved from shortage to surplus. Manufacturers respond to high margins by improving yields, converting production, adding equipment, and building fabrication capacity. Customers also accumulate inventory when they fear shortages.

These actions can extend an upcycle before reversing it. Inventory accumulation strengthens near-term orders but borrows demand from later periods. New capacity takes time, yet it can arrive after purchasing growth has slowed.

Sandisk’s annual filings list demand volatility, average selling prices, customer deployment timing, and competitive pricing among its material risks. Those disclosures are standard, but they closely match the present debate.

The company’s reliance on Kioxia manufacturing ventures adds another dimension. Shared production can improve scale and capital efficiency. It also means output decisions, technology transitions, and execution depend on a strategic relationship.

SK hynix faces parallel risks in DRAM and HBM. Leading suppliers must invest heavily before customer demand becomes certain. Advanced packaging equipment and leading-edge fabrication capacity cannot be added instantly.

HBM also consumes more wafer capacity than conventional DRAM because its stacks require multiple dies. Expanding HBM can tighten other DRAM categories. That interaction supported pricing across the market during the current cycle.

Eventually, capacity additions and improved yields can loosen the constraint. A slowdown in accelerator deployments would amplify the effect. Suppliers would then have more output chasing fewer urgent orders.

Competition matters as much as demand. Micron is already shipping HBM4 in volume, while Samsung continues working to strengthen its position in advanced HBM. Successful qualifications would give major customers more negotiating leverage.

Chinese memory suppliers create another source of uncertainty, particularly in conventional NAND and DRAM. Export controls limit access to some equipment, but domestic investment can still expand output in selected categories.

Neither Sandisk nor SK hynix needs the market to remain permanently undersupplied. They need demand growth, disciplined investment, and product differentiation to prevent a severe pricing correction.

The distinction is important. A cyclical decline in prices would not mean AI demand disappeared. It would mean supply grew faster than demand for a period.

Google News readers should also treat premarket trading carefully. Liquidity is thinner before the regular session, and a relatively small number of orders can produce large moves. Premarket direction does not guarantee the closing result.

Still, the initial response carries information. Investors had enough concern about expectations, guidance, or the cycle to sell despite reported growth. That reaction challenges the idea that strong earnings automatically support higher valuations.

The strongest bearish case argues that current margins already represent a peak. Under that view, customer commitments and AI demand cannot prevent supply from catching up.

The strongest bullish case argues that inference growth has changed the industry’s demand curve. Under that view, HBM, enterprise SSDs, and new tiers such as HBF expand together for several years.

Neither case is proven. Sandisk’s results support the demand argument, while its pricing contribution supports the cyclical warning. The same numbers provide evidence for both sides.

Investors also lack enough detail about the new customer agreements. Sandisk reported ten New Business Model agreements after adding five during the quarter. It has not published terms allowing outsiders to model their exact protection.

Those agreements become more persuasive when they translate into stable volumes across changing prices. Until then, they remain helpful signals rather than guarantees of durable earnings.

HBF carries a similar burden of proof. A specification provides architectural direction, but commercial adoption requires controllers, packaging, software support, qualified systems, and attractive economics.

SK hynix’s iHBM claims also await deployment. A reported 30% reduction in thermal resistance sounds meaningful. Customers will care about system-level temperature, reliability, bandwidth, power, and manufacturing yield.

This is why the market can sell two leading memory names during an apparent boom. Investors are no longer asking whether AI needs memory. They are asking which profits survive when supply and competition respond.

Three Signals Will Decide the AI Memory Race

The next phase depends on customer commitments, production discipline, and proof that new memory tiers improve deployed AI systems.

The first signal is Sandisk’s Investor Day on August 13. Management has an opportunity to explain its customer agreements, long-term operating model, technology roadmap, and capital allocation priorities.

Investors should listen for measurable details rather than broader statements about AI. Contract duration, committed volume, pricing mechanisms, and customer concentration would clarify how much revenue visibility Sandisk has secured.

More detail would strengthen the durable-demand case. Limited disclosure would not invalidate it, but markets would continue treating current earnings as heavily exposed to spot pricing and industry conditions.

The second signal is supplier capital spending and output guidance. Sandisk, Kioxia, SK hynix, Micron, and Samsung must decide how quickly to expand production across NAND, DRAM, and advanced packaging.

Disciplined spending would support prices and reduce the risk of oversupply. Aggressive expansion without matching customer commitments would weaken the current investment case.

Investors should distinguish spending on technology transitions from spending that adds total capacity. New process equipment can improve density and lower cost without producing the same supply increase as a new fabrication facility.

Yield improvements also matter. A supplier can generate more usable dies from existing wafer starts by reducing defects. That expansion may appear gradually, making it harder to track than a new factory.

The third signal is customer qualification for HBM4, enterprise SSDs, and HBF-related products. Qualification converts a promising component into a credible revenue stream.

Micron has already said its HBM4 is shipping for a lead customer. SK hynix must defend its leadership as accelerator platforms evolve. Samsung’s progress can increase competition and improve buyers’ alternatives.

For Sandisk, continued data center growth is the clearest near-term test. Another quarter of expanding enterprise shipments would show that fiscal fourth-quarter demand was not simply an inventory build.

HBF requires a longer runway. Investors should look for processor partners, controller implementations, system demonstrations, and production schedules. These steps would move the technology beyond an open specification.

The companies must also show where HBF fits in actual inference workloads. A useful deployment would identify which model data stays in HBM, which moves to HBF, and which remains on conventional SSDs.

That evidence would strengthen the view that AI creates additional memory tiers rather than shifting spending between existing products. Failure to secure system partners would weaken the mechanism behind the HBF thesis.

The larger lesson from the SNDK and SKHY decline is not that AI memory demand has failed. Reported results show the opposite. The lesson is that investors have moved from discovering the theme to auditing its durability.

Sandisk’s revenue, margins, data center growth, and outlook set a remarkably high baseline. SK hynix’s HBM position gives it strategic importance inside leading AI systems. Their joint HBF work creates another possible growth path.

Each advantage also carries a corresponding question. High prices invite capacity. Strong margins attract competitors. New specifications require commercial validation, and long-term agreements need enough detail to establish their value.

That balance should shape how readers interpret future Google News alerts. A daily share move offers one snapshot, while customer commitments and manufacturing decisions reveal the direction of the cycle.

Watch the August 13 disclosures first. Then follow production plans and product qualifications across the major suppliers. Those signals will show whether the selloff reflects temporary expectation management or an early warning about memory pricing.

For developers and enterprise technology buyers, the outcome reaches beyond semiconductor portfolios. It can affect server availability, storage planning, cloud costs, and the design of inference systems.

Track whether suppliers convert scarcity into stable contracts and better architectures. If they do, AI memory becomes a longer-lived infrastructure category. If capacity outruns deployments, today’s record margins will look more cyclical than structural.