AI inference is expanding the memory trade beyond GPUs and HBM. Rising demand for high-capacity flash, longer customer agreements and new technologies such as HBF could make NAND more important to AI infrastructure—and potentially make the traditional memory cycle less extreme.AI inference is expanding the memory trade beyond GPUs and HBM. Rising demand for high-capacity flash, longer customer agreements and new technologies such as HBF could make NAND more important to AI infrastructure—and potentially make the traditional memory cycle less extreme.
Is AI Inference Changing the NAND Cycle? What Sandisk Means for Memory Stocks
Key Takeaways
AI inference is expanding the memory trade beyond GPUs and HBM. Rising demand for high-capacity flash, longer customer agreements and new technologies such as HBF could make NAND more important to AI infrastructure—and potentially make the traditional memory cycle less extreme.
- AI inference requires far more than compute: large-scale deployments also need growing amounts of accessible, cost-efficient storage.
- NAND is emerging as a capacity layer alongside HBM and DRAM rather than simply a commodity storage product.
- Sandisk expects enterprise data center flash demand to reach 1.2 zettabytes by 2030 and is developing High Bandwidth Flash specifically for AI inference.
- Longer-term customer agreements at Sandisk, Samsung and other memory suppliers could improve demand visibility and reduce aggressive capacity expansion.
- AI does not eliminate the NAND cycle. The longer-term question is whether stronger bit demand and better supply discipline can make future cycles less volatile.
For most of the AI boom, the semiconductor investment story followed a familiar path: more AI models required more GPUs, GPUs required more HBM, and increasingly dense clusters required faster networking and more power. Storage was important, but it rarely received the same attention.
AI inference is beginning to change that hierarchy.
As AI moves from training large models toward serving billions of queries, enterprise agents, multimodal applications and real-time workloads, infrastructure must not only perform calculations but also store and repeatedly retrieve enormous amounts of data. That is creating a new question for the memory industry: could NAND become one of the next major capacity bottlenecks in AI infrastructure?
Recent developments at Sandisk, Samsung, Kioxia and other memory suppliers suggest the answer may increasingly be yes. The result is a NAND investment thesis that is starting to move beyond short-term flash pricing and toward a more structural argument about AI storage demand.
Why AI Inference Needs So Much More Storage
AI training and AI inference stress infrastructure differently. Training a frontier model requires massive computing clusters and extremely fast access to memory, which is why GPUs and high-bandwidth memory became the earliest bottlenecks in the AI supply chain.
Inference is what happens after those models are deployed. Each time an AI system answers a question, searches a knowledge base, analyzes an image or operates an agent, it may need to retrieve model parameters, embeddings, cached information, user context and increasingly large multimodal datasets.
As the number of queries grows, so does the amount of information that must remain economically accessible.
Keeping all of that data in HBM would be impractical. HBM offers extraordinary bandwidth, but capacity is relatively limited and the cost per bit is far higher than NAND. Traditional NAND sits at the opposite end of the spectrum: much greater capacity at substantially lower cost, but with lower bandwidth and higher latency.
That leaves a large architectural space between the two.
The emerging AI memory stack is therefore less likely to be a winner-takes-all competition between HBM and NAND. Instead, different types of memory can serve different levels of the hierarchy: HBM for the most bandwidth-sensitive workloads, DRAM for working memory, and increasingly high-performance NAND for much larger pools of data that still need to remain close to compute.
Sandisk Is Betting That AI Makes NAND a Capacity Layer
Sandisk has become one of the clearest public examples of this thesis. At its 2026 Investor Day, the company projected that the enterprise data center flash market could reach 1.2 zettabytes by 2030, while its own revenue is expected to grow at a mid-to-high-teens annual rate from FY2028 through FY2030.
That outlook follows a dramatic increase in Sandisk's data center business. The company has increasingly tied its long-term strategy to AI infrastructure rather than relying primarily on consumer devices or conventional enterprise storage.
The shift matters because AI systems do not simply need faster storage. They need more storage per system. Larger models, retrieval-augmented generation, multimodal content and persistent AI agents all increase the amount of information that needs to remain available to processors.
If inference becomes one of the dominant workloads inside future data centers, NAND demand could increasingly be driven by the growth of AI data itself rather than only by PC, smartphone and traditional cloud-storage cycles.
That is a very different demand profile from the NAND industry investors knew a decade ago.
HBF Could Fill the Gap Between NAND and HBM
One of the most interesting attempts to bridge this memory hierarchy is High Bandwidth Flash, or HBF.
Sandisk describes HBF as a NAND-based architecture designed to augment rather than replace HBM in AI inference systems. The company says the technology is intended to provide bandwidth comparable with HBM while offering up to eight times the capacity at a similar cost for certain configurations. Sandisk has also been working with SK hynix on standardization and expects early AI-inference devices using HBF to emerge as the technology develops.
The attraction is straightforward. Inference workloads often need to keep very large models and datasets close enough to accelerators that moving information from conventional storage becomes a performance bottleneck. Expanding HBM indefinitely would solve the bandwidth problem but create an enormous cost and capacity problem.
HBF is an attempt to find a middle ground: significantly more capacity than HBM, but much higher performance than conventional flash storage.
It remains an emerging architecture and should not yet be treated as a major source of industry revenue. But its existence illustrates how memory companies increasingly view NAND as something more than a backend storage medium. The industry is actively trying to move flash closer to AI compute.
QLC NAND May Matter Before HBF Does
HBF attracts more attention because it is new, but conventional NAND innovation may have a much larger near-term impact.
QLC NAND stores four bits per cell, allowing manufacturers to deliver more capacity from the same physical area than TLC NAND. That makes it particularly attractive for hyperscale data centers where customers need huge amounts of storage and cost per bit matters.
The trade-off is that increasing the number of bits stored in each cell can affect endurance and performance, which means QLC is not appropriate for every workload. But improvements in controllers, error correction and NAND architecture have steadily expanded the workloads where QLC can be used.
AI could accelerate that transition because inference generates exactly the type of demand where capacity, power efficiency and economics become increasingly important.
Sandisk is developing higher-density QLC products, while Samsung and Kioxia are also using wafer-bonding and next-generation NAND designs to increase density and performance. Samsung's latest BV-NAND architecture, for example, increases memory density by roughly 58% from its previous generation as the company targets growing AI storage requirements.
The AI storage opportunity therefore does not depend on HBF succeeding. Higher-capacity enterprise SSDs and more efficient QLC NAND can benefit from the same underlying increase in data intensity.
Why the NAND Cycle Has Historically Been So Volatile
The bigger investment question is whether stronger AI demand can actually change the notorious NAND cycle.
Historically, the problem has not simply been weak demand. It has been the interaction between demand and supply.
When NAND becomes scarce, prices rise rapidly and manufacturers generate strong margins. Those higher returns encourage additional investment in capacity. Because semiconductor fabs require long planning and construction cycles, that new supply can arrive after the original shortage has already eased. Prices fall, margins collapse and producers reduce investment again.
The industry then begins the next cycle.
AI does not automatically eliminate this mechanism. Even extremely strong end demand can eventually be overwhelmed if manufacturers add too much capacity.
What may be changing is the amount of visibility suppliers have before making those investments.
Long-Term Contracts Could Change Supply Discipline
Sandisk has increasingly moved customers toward multiyear agreements rather than relying primarily on shorter-term transactions. Reuters reported after its latest earnings that the company had eight long-term agreements with six customers valued at a combined $93.9 billion, with contracts averaging roughly four years. About half of output for FY2027 is expected to be covered by these agreements, rising toward two-thirds in FY2028.
That model matters because it gives both supplier and customer greater visibility. A hyperscaler can secure future storage capacity, while Sandisk has more information about committed demand before deciding how aggressively to invest.
Sandisk is not alone. Samsung has indicated that long-term customer agreements could eventually account for 60% to 70% of its memory sales, suggesting that a broader change may be taking place across the memory industry.
If suppliers increasingly expand capacity against committed multiyear demand rather than simply reacting to rising spot prices, the industry may become less prone to the extreme oversupply that characterized previous cycles.
That would not eliminate cyclicality. But it could change its severity.
What This Means for Sandisk, Micron and Other Memory Stocks
The shift toward AI storage does not benefit every memory company in exactly the same way.
Sandisk has one of the clearest direct exposures to NAND and enterprise flash. Its thesis increasingly combines higher AI-driven bit demand with long-term customer commitments, QLC products and HBF. This makes SNDK particularly sensitive to whether investors believe the current NAND profitability can persist.
Micron has broader memory exposure across DRAM, HBM and NAND. It therefore participates in both sides of the AI memory hierarchy: premium high-bandwidth memory close to accelerators and larger pools of flash storage. Strong AI demand has also helped improve supply visibility across Micron's product portfolio. Memory stocks rallied broadly after Micron's June results reinforced expectations that supply could remain tight through 2027.
SK hynix is best known in the AI trade for HBM, but its collaboration with Sandisk on HBF is important because it suggests a leading HBM producer also sees value in developing a NAND-based inference tier.
Kioxia offers another direct read on NAND. The company has accelerated next-generation flash production as AI workloads increase demand for higher-capacity storage, and Reuters reported in July that the AI boom has helped drive a dramatic recovery in the company's business and valuation.
Western Digital is somewhat different following the separation of Sandisk because its core exposure is now hard disk drives rather than NAND. Even so, the same explosion in AI-generated data can support demand for multiple storage tiers. Flash is better suited to performance-sensitive workloads, while HDDs remain economically attractive for enormous pools of colder data.
The broader takeaway is that AI storage is not a single-product trade. The growth of AI data potentially increases demand across HBM, DRAM, NAND, enterprise SSDs and high-capacity HDDs, with each technology serving a different part of the storage hierarchy.
AI Could Change the NAND Cycle Without Ending It
It is tempting to describe the current memory boom as a permanent break from the industry's past. That would go too far.
NAND remains capital intensive. Technology transitions can increase available bits per wafer even without building entirely new fabs, competitors can eventually respond to attractive margins, and AI spending itself will not grow at the same rate forever.
Recent stock volatility illustrates that risk. Sandisk and Western Digital both sold off sharply after strong August earnings because expectations had become so high that even upbeat forecasts were not enough to satisfy the market.
The more defensible structural argument is narrower: AI could make NAND demand stronger and more predictable, while long-term contracts and greater supply discipline could make future cycles less extreme.
That distinction matters.
A less volatile NAND cycle does not require memory prices to rise indefinitely. It requires suppliers to preserve profitability through normal pricing fluctuations instead of repeatedly moving from shortage to severe oversupply.
If that happens, the valuation framework for companies such as Sandisk may eventually change as well.
Access SNDK-Related Markets on MEXC
For eligible users following the AI storage and NAND theme through Sandisk, MEXC provides several forms of SNDK-related market access. The SNDKSTOCK_USDT perpetual futures contract allows leveraged long or short exposure to SNDK-related price movements, while Sandisk through MEXC RealStocks provides access to traditional U.S. equity exposure through the RealStocks structure.
MEXC also lists SNDKON/USDT, Ondo Finance's tokenized version of Sandisk stock. SNDKON provides tokenized economic exposure rather than direct ownership of Sandisk common shares. Readers researching Sandisk specifically can also explore MEXC's Sandisk stock guide, which covers the company's business model, products and key risks.
Different products involve different market structures and risks. Futures add leverage, funding and liquidation considerations, while tokenized assets introduce additional issuer, liquidity, blockchain and regulatory risks. Product availability may vary by jurisdiction.
FAQ
Does AI inference use NAND flash?
Yes. AI inference systems can rely on NAND-based enterprise SSDs to store model data, embeddings, cached information and other large datasets that do not need to remain permanently in expensive HBM or DRAM. As models and datasets become larger, this capacity layer may become increasingly important.
Can NAND replace HBM for AI?
Not directly. HBM offers extremely high bandwidth and low latency close to AI accelerators, while NAND provides far greater capacity at much lower cost per bit. Technologies such as HBF are being developed to narrow the performance gap, but Sandisk itself describes HBF as a complement to HBM rather than a replacement.
What is High Bandwidth Flash?
High Bandwidth Flash, or HBF, is an emerging NAND-based memory architecture designed for AI inference. Sandisk's design combines stacked flash memory with advanced bonding technology to provide much higher bandwidth than conventional NAND while retaining significantly greater capacity than HBM.
Is the NAND cycle over because of AI?
No. NAND remains a cyclical and capital-intensive industry. AI can strengthen long-term bit demand, while multiyear customer agreements may improve supply discipline, but pricing, capacity additions and technology transitions can still produce periods of oversupply and weaker margins.
The Bigger Shift: AI Needs Capacity, Not Just Compute
The first stage of the AI infrastructure boom was defined by scarcity in compute. GPUs, HBM and networking became valuable because there was not enough performance to meet demand.
Inference introduces another constraint: capacity.
Running AI at global scale requires enormous amounts of information to remain accessible without storing everything in the most expensive forms of memory. That creates room for NAND to move higher in the AI memory hierarchy through enterprise SSDs, QLC and potentially architectures such as HBF.
Sandisk is one of the companies making the strongest case that this shift can support a more durable NAND business. But the broader significance extends beyond SNDK. If AI inference continues to increase storage intensity while memory suppliers remain disciplined about capacity, the next NAND cycle may still be a cycle—just a very different one from the cycles investors have seen before.
Market Opportunity
Gensyn Price(AI)
$0.01983
$0.01983$0.01983
USD
Gensyn (AI) Live Price Chart
Popular Articles
View More
US Stock Weekly | Aug.08–Aug.13: CPI Cooling Lifts S&P 500 to Record High, AI Infrastructure Earnings Surge Across the Board
This week, the US stock market ushered in multiple catalysts: the July CPI exceeded expectations and cooled down, the S & P 500 hit a historic high, coupled with the SMCI financial report
QQQ Holdings 2026: Top Companies, AI Exposure and Nasdaq-100 Sector Breakdown
Summary Invesco QQQ is often associated with U.S. Big Tech, but understanding QQQ holdings requires looking beyond the idea that it is simply a “technology ETF.” QQQ tracks the Nasdaq-100 Index,
MEXC vs Gate KIMI Pre-IPO Futures: Fees, Leverage and Position Limits Compared
Summary Both MEXC and Gate provide Pre-IPO perpetual futures linked to market expectations surrounding Moonshot AI and Kimi. However, the compared KIMI contract settings show important differences
Hot Crypto Updates
View More
How Nvidia Plans to Finance 500 Billion Dollars of AI Infrastructure
Overview As the global development of frontier artificial intelligence models and multimodal agents expands to multi-trillion parameter scales, the capital requirements for physical infrastructure
Apple China AI Strategy Changes Again And What the Alibaba Partnership Means
Overview According to an exclusive Reuters report published on August 14, 2026, Apple Inc. has developed and trained a proprietary large language model specifically tailored for mainland China,
Why Is Cisco Stock Price Down 8% Despite Strong AI Orders?
Overview Following the release of its fiscal fourth quarter and full-year 2026 financial results on August 13, 2026, networking infrastructure giant Cisco Systems (NASDAQ: CSCO) saw its shares
Trending News
View More
SK Hynix ADR Prices at $149 as SKHY Nasdaq Debut Nears
SK Hynix has officially priced its U.S. American Depositary Receipts (ADRs) at $149 each, raising approximately $26.5 billion in a major U.S. offering. This move highlights massive investor appetite f
DeepSeek Funding Restart Puts China AI Valuations Back in Focus
DeepSeek funding talks have reportedly restarted at a 500 billion yuan valuation, raising fresh questions about China AI investment and IPO timing.
Related Articles
View More
What Is Invesco QQQ ETF? Nasdaq-100 Holdings, Strategy, Fees and Risks
Summary Invesco QQQ ETF (NASDAQ: QQQ) is an exchange-traded fund designed to track the Nasdaq-100 Index, which represents 100 of the largest non-financial companies listed on the Nasdaq Stock Market.
What Is QQQON? Ondo Tokenized Invesco QQQ ETF Explained
Summary QQQON, styled by Ondo as QQQon, is a tokenized product designed to provide economic exposure linked to the Invesco QQQ ETF (NASDAQ: QQQ). The product structure has three important layers: Nasd
Why Does QQQON Track QQQ? Tokenized ETF Premiums, Discounts and Price Differences Explained
Summary QQQON is designed to provide economic exposure linked to Invesco QQQ, while QQQ itself tracks the Nasdaq-100 Index. This creates a two-stage tracking structure: Nasdaq-100 Index ↓ QQQ ETF ↓ QQ
