AI inference is reshaping the NAND cycle? What Sandisk means for memory chip stocks

Key Takeaways

AI inference is extending the storage investment thesis beyond GPUs and HBM to high-capacity storage. Growing flash demand, long-term customer agreements, and new technologies like HBF could elevate NAND's importance in AI infrastructure and potentially reduce the amplitude of traditional storage cycle fluctuations.

Takeaway

• AI inference requires more than just compute power; large-scale deployment also demands continuously growing, rapidly accessible, and cost-effective storage capacity.
• NAND is increasingly becoming an AI capacity layer alongside HBM and DRAM, rather than just a traditional commodity storage product.
• Sandisk projects enterprise data center flash demand will reach 1.2 ZB by 2030 and is developing High Bandwidth Flash specifically designed for AI inference.
• Storage makers like Sandisk and Samsung are adopting longer-term customer agreements, which could improve demand visibility and reduce the risk of over-expansion.
• AI will not eliminate the NAND cycle. What is truly worth watching is whether stronger bit demand and better supply discipline can reduce the volatility of future cycles.

Over the past few years of the AI boom, there has been a relatively clear path in the semiconductor investment thesis: more AI models require more GPUs, GPUs require more HBM, and increasingly dense compute clusters require faster networks and more power. Storage has always been important, but it has rarely received the same level of attention as these other areas.

AI inference is gradually changing this allocation of resources.

As AI shifts from training large models to serving billions of queries, enterprise AI agents, multimodal applications, and real-time workloads, infrastructure must not only perform computations but also continuously store and repeatedly access vast amounts of data.

This raises a new question for the storage industry: could NAND become the next major capacity bottleneck in AI infrastructure?

Recent developments from Sandisk, Samsung, Kioxia, and other storage makers suggest the answer is increasingly leaning toward yes. As a result, the NAND investment thesis is gradually shifting from short-term flash price increases toward more structural AI-driven storage demand.

Why Does AI Inference Require More Storage Capacity?

AI training and AI inference place different types of pressure on infrastructure.

Training frontier large models requires massive compute clusters and extremely high-speed memory access, which is why GPUs and HBM were the first to become core bottlenecks in the AI supply chain.

Inference, on the other hand, occurs after the model has been deployed. Every time an AI system answers a question, searches a knowledge base, analyzes an image, or executes an agent task, it may need to read model parameters, embeddings, cached data, user context, and increasingly large multimodal datasets.

As query volume increases, the amount of data that needs to remain accessible in a cost-effective manner grows in tandem.

Storing all data in HBM is not realistic. HBM offers extremely high bandwidth, but its capacity is relatively limited, and its per-bit cost is far higher than NAND. Traditional NAND sits at the other end, offering massive capacity at lower cost, but with lower bandwidth and higher latency.

There is therefore a significant architectural gap between the two.

The future AI memory stack is more likely to be a multi-tiered system rather than a choice between HBM and NAND: HBM handles the most bandwidth-intensive workloads, DRAM serves working memory needs, and high-performance NAND begins to handle larger data pools that still need to be close to the compute side.

Sandisk Is Betting NAND Will Become AI's Capacity Layer

Sandisk is currently one of the companies most explicitly articulating this thesis. At its 2026 Investor Day, the company projected that the enterprise data center flash memory market could reach 1.2 ZB by 2030, while also guiding for mid-to-high teens annual revenue growth from FY2028 to FY2030. These expectations can be referenced at Reuters.

This outlook builds on Sandisk's already rapidly growing data center business. The company is progressively aligning its long-term strategy with AI infrastructure, rather than relying primarily on consumer electronics or the traditional enterprise storage market.

This shift is important because AI systems require not just "faster storage," but more storage capacity per system configuration. Larger models, RAG, multimodal content, and continuously running AI agents all increase the amount of data that needs to remain accessible over the long term.

If AI inference becomes one of the dominant workloads in future data centers, the demand drivers for NAND could shift from PC, smartphone, and traditional cloud storage cycles toward the data that AI generates and uses itself.

This is already very different from the NAND demand structure the market understood a decade ago.

HBF Could Fill the Gap Between NAND and HBM

High Bandwidth Flash (HBF) is currently one of the most interesting memory tier solutions.

Sandisk positions HBF as a NAND architecture designed to complement rather than replace HBM. The company states that in certain AI inference configurations, HBF aims to offer bandwidth approaching that of HBM while providing up to roughly eight times the capacity at a similar cost. Sandisk is also working with SK hynix to drive standardization, hoping that future AI inference devices can adopt this technology. More information can be found in Sandisk's HBF technical announcement.

The appeal of this technology is straightforward. AI inference often requires keeping very large models and datasets close enough to accelerators; otherwise, moving data from traditional storage tiers could become a performance bottleneck. Expanding HBM indefinitely would improve bandwidth, but cost and capacity would become significant issues.

HBF attempts to find a middle ground: significantly higher capacity than HBM, while delivering far better performance than traditional flash memory.

HBF is still an emerging architecture and should not be viewed as a mature, large-scale revenue market. However, it illustrates a more important industry direction: storage makers are trying to move NAND beyond backend storage and bring it progressively closer to AI compute itself.

QLC NAND Could Have Real-World Impact Before HBF

HBF tends to attract more attention because it is a new technology, but continued innovation in traditional NAND could have a bigger impact in the short term.

QLC NAND stores four bits per cell, allowing it to deliver higher capacity in the same physical area compared to TLC NAND. This makes QLC particularly attractive for hyperscale data centers that require massive storage capacity and are highly focused on per-bit cost.

The trade-off is that storing more bits per cell can impact endurance and performance, making QLC unsuitable for all workloads. However, as controllers, error correction, and NAND architectures continue to advance, the range of use cases for QLC is expanding.

AI could further accelerate this transition, as inference workloads particularly value capacity, energy efficiency, and overall cost.

Sandisk is continuously developing higher-density QLC products, while Samsung and Kioxia are also improving density and performance through wafer bonding and next-generation NAND designs. For example, Samsung's latest BV-NAND architecture can deliver approximately 58% higher storage density compared to the previous generation, addressing growing AI storage demand.

Therefore, growth in the AI storage market does not need to rely entirely on HBF's success. High-capacity enterprise SSDs and more efficient QLC NAND can also directly benefit from continuously increasing data density.

Why Has the NAND Cycle Historically Been So Volatile?

The bigger investment question is whether AI demand can truly change NAND's historically volatile cycle.

Historically, the problem has not just been insufficient demand, but the interaction between demand and supply.

When NAND supply is tight, prices rise rapidly, and manufacturers enjoy high gross margins. Higher returns attract more capacity investment, but semiconductor fab planning and construction take time. By the time new capacity actually comes online, the initial supply gap has often narrowed.

Prices subsequently fall, gross margins compress quickly, and manufacturers reduce investment.

Then the cycle repeats.

AI will not automatically eliminate this mechanism. Even if end demand is very strong, if manufacturers ultimately over-expand, supply can still outpace demand.

What could genuinely change is how much demand visibility suppliers have before making investment decisions.

Long-Term Agreements Could Change Supply Discipline

Sandisk is progressively transitioning customers to multi-year agreements rather than relying primarily on spot transactions.

Reuters reported following Sandisk's latest earnings that the company has signed eight long-term agreements with six customers, with a total contract value of approximately $93.9 billion and an average contract duration of about four years. Roughly half of FY2027 output is expected to be covered by these agreements, increasing to about two-thirds by FY2028. The relevant data can be found at Reuters.

The importance of this model is that it improves visibility for both suppliers and customers simultaneously.

Hyperscalers can lock in future storage capacity in advance, while Sandisk can also gain more clarity on committed demand before deciding on the scale of expansion.

Sandisk is not the only company adopting this model. Samsung has also indicated that long-term customer agreements could account for 60% to 70% of storage sales in the future, suggesting that a broader business model shift may be underway across the storage industry.

If suppliers increasingly base expansion decisions on multi-year committed demand rather than immediately and aggressively adding capacity when spot prices rise, the storage industry may be able to reduce the risk of extreme oversupply it has seen in the past.

This will not eliminate the cycle, but it could change its intensity.

What Does This Mean for Sandisk, Micron, and Other Memory Chip Stocks?

Growing AI storage demand will not benefit all storage companies in exactly the same way.

Sandisk is one of the most direct pure plays on NAND and enterprise flash. The company's investment thesis is now combining AI-driven bit demand, multi-year customer commitments, QLC, and HBF, making SNDK particularly sensitive to whether the market believes current NAND profitability can be sustained.

Micron's product portfolio spans DRAM, HBM, and NAND, exposing it to both ends of the AI memory hierarchy: high-priced, high-bandwidth memory close to accelerators, and higher-capacity flash. Strong AI demand also improves supply visibility across multiple Micron product lines.

SK hynix's primary role in the AI investment narrative has been HBM, but its collaboration with Sandisk on HBF is worth noting, as it indicates that a leading HBM maker also believes there is room for a NAND-based high-capacity memory tier in future AI inference architectures.

Kioxia offers another direct window into NAND. As AI workloads boost high-capacity storage demand, the company is accelerating the production ramp of next-generation flash, and the AI boom has significantly improved its operations and valuation.

Western Digital's core business has shifted to hard disk drives rather than NAND following the Sandisk spinoff. But if the volume of AI-generated data continues to grow rapidly, it could support multiple storage tiers simultaneously. Flash is better suited for high-performance, low-latency workloads, while HDD retains a cost advantage in large-scale cold data storage.

The broader conclusion is that AI storage is not a single-product trade.

Continued growth in AI data could simultaneously drive demand for HBM, DRAM, NAND, enterprise SSDs, and high-capacity HDDs, with different technologies occupying different positions in the storage hierarchy.

AI Could Change the NAND Cycle, But Won't Eliminate It

Describing the current storage rally as a permanent escape from the industry's cyclicality would be overly optimistic.

NAND remains a highly capital-intensive industry. Even without building entirely new fabs, technology upgrades can increase the number of bits per wafer; competitors may also add supply due to high gross margins, and AI capital expenditure itself cannot grow at the same pace forever.

Recent stock price volatility reflects this risk. Both Sandisk and Western Digital fell sharply after reporting strong earnings in August, because market expectations were already so high that even optimistic guidance was not enough to satisfy investors.

A more reasonable structural view, therefore, should converge on this:

AI can make NAND's long-term bit demand stronger and more predictable, while multi-year customer agreements and tighter supply discipline could reduce the severity of future storage cycles.

This distinction is crucial.

Reducing NAND cycle volatility does not require storage prices to rise forever. What is truly needed is that even when prices return to normal