For most of the AI boom, the semiconductor trade had a fairly clean sequence. More AI models meant more GPUs, which pulled in high-bandwidth memory, faster networking and more power. Storage mattered the whole time, but it rarely drew the same level of market attention.
AI inference is starting to change that.
As AI shifts from training frontier models to serving billions of queries, enterprise AI agents, multimodal applications and real-time workloads, infrastructure has to do more than compute. It also has to store and repeatedly retrieve large amounts of data. That raises a new question for the storage industry: could NAND become the next major capacity bottleneck in AI infrastructure.
Recent moves by Sandisk, Samsung, Kioxia and other storage makers suggest the answer is moving closer to yes. The investment case for NAND is also beginning to shift, away from a short-term flash pricing story and toward a more structural AI storage demand thesis.
Why AI inference needs more storage capacity
AI training and AI inference do not pressure infrastructure in the same way.
Training frontier large models requires massive compute clusters and extremely fast memory access. That is why GPUs and HBM became the earliest choke points in the AI supply chain. Inference comes after deployment. Each time an AI system answers a question, searches a knowledge base, analyzes an image or runs an agent task, it may need to read model parameters, embeddings, cached data, user context and increasingly large multimodal datasets.
As query volume rises, the amount of data that must remain accessible in an economical way rises with it.
Keeping all of that data in HBM is not realistic. HBM offers very high bandwidth, but capacity is relatively limited and the cost per bit is far above NAND. Traditional NAND sits at the other end of the stack. It delivers large capacity at lower cost, but with lower bandwidth and higher latency.
That leaves a large architectural gap in between. The future AI memory stack is less likely to be an HBM-versus-NAND choice and more likely to split work across layers: HBM for the most bandwidth-sensitive workloads, DRAM for working memory, and higher-performance NAND for larger pools of data that still need to sit closer to the compute side.
Sandisk is betting NAND becomes AI’s capacity layer
Sandisk is one of the clearest companies publicly arguing for that framework. At its 2026 investor day, the company said enterprise data center flash demand could reach 1.2 ZB by 2030 and projected mid-to-high double-digit annual revenue growth from FY2028 to FY2030. The source notes that related expectations can be referenced through Reuters.
That outlook rests on the company’s already fast-growing data center business. Sandisk is gradually tying its long-term strategy to AI infrastructure instead of relying mainly on consumer electronics or the traditional enterprise storage market.
The shift matters because AI systems do not just need faster storage. They need more storage capacity per system. Larger models, retrieval-augmented generation, multimodal content and always-on AI agents all increase the amount of data that has to remain available over time.
If AI inference becomes one of the main workloads in future data centers, NAND demand drivers could shift away from PC, smartphone and traditional cloud storage cycles and move toward data generated and used by AI itself. That is very different from how the market understood NAND demand a decade ago.
HBF may fill the gap between NAND and HBM
High Bandwidth Flash, or HBF, is one of the more interesting proposals in the memory hierarchy.
Sandisk positions HBF as a NAND architecture designed to complement, not replace, HBM. The company says that in some AI inference configurations, HBF is intended to offer bandwidth close to HBM while delivering up to roughly 8x the capacity at similar cost. Sandisk is also working with SK hynix on standardization so the technology could eventually be adopted in AI inference devices. The source points readers to Sandisk’s HBF technical materials for more detail.
The appeal is straightforward. AI inference often requires very large models and datasets to remain close enough to accelerators. If data has to move back and forth from a conventional storage layer, that movement can become a performance bottleneck. Expanding HBM indefinitely may help bandwidth, but the cost and capacity tradeoffs become severe.
HBF is trying to occupy the middle ground: meaningfully more capacity than HBM, with materially better performance than conventional flash.
For now, HBF remains an emerging architecture and should not be treated as a mature revenue market. Still, it points to a broader industry direction. Storage vendors are trying to move NAND beyond a pure back-end storage role and closer to AI compute itself.
QLC NAND may have a more immediate impact
Because HBF is new, it naturally gets more attention. Nearer-term change may come from continued improvements in conventional NAND.
QLC NAND stores 4 bits per cell, which allows more capacity in the same physical area than TLC NAND. That makes QLC especially attractive to hyperscale data centers that need large amounts of storage and care deeply about cost per bit.
The tradeoff is that storing more bits per cell can hurt endurance and performance, so QLC is not suited to every workload. Even so, advances in controllers, error correction and NAND architecture are widening the set of applications where QLC can be used.
The source argues that AI could speed up that transition because inference workloads place unusual weight on capacity, energy efficiency and total system cost.
Sandisk continues to work on denser QLC products. Samsung and Kioxia are also increasing density and performance through wafer bonding and new NAND designs. The article cites Samsung’s latest BV-NAND architecture as an example, saying it can improve storage density by about 58% versus the prior generation to address growing AI storage demand.
That means the AI storage market does not need to depend entirely on HBF to grow. High-capacity enterprise SSDs and more efficient QLC NAND can also benefit directly as data density keeps climbing.
Why NAND cycles have historically been so volatile
The bigger investment question is whether AI demand can really change the NAND industry’s long-standing cyclicality.
Historically, the issue was not simply weak demand. It was the interaction between demand and supply.
When NAND supply tightened, prices rose quickly and manufacturers enjoyed high gross margins. Those returns encouraged more capacity investment, but semiconductor fabs take time to plan and build. By the time new output actually came online, the original supply gap often had narrowed.
Prices then fell. Margins compressed. Producers cut investment, and the next cycle started.
AI does not remove that mechanism on its own. Even if end demand is strong, supply can still overwhelm demand if manufacturers expand too aggressively.
What may change is how much demand visibility suppliers have before they commit to investment.
Long-term agreements could reshape supply discipline
Sandisk is moving customers toward multi-year agreements instead of depending mainly on short-term transactions.
After the company’s latest earnings report, Reuters, as cited in the source, said Sandisk had signed 8 long-term agreements with 6 customers, worth about $93.9 billion in total, with an average term of about 4 years. Roughly half of FY2027 output is expected to be covered by those agreements, rising to about two-thirds in FY2028.
The importance of that model is that it improves visibility for both suppliers and customers. Hyperscalers can lock in future storage capacity earlier, while Sandisk can make expansion decisions with a larger base of already committed demand.
Sandisk is not alone. The source says Samsung has also indicated that long-term customer agreements could eventually account for 60% to 70% of memory sales, pointing to a broader shift in storage industry business models.
If suppliers expand more on the back of committed multi-year demand and less on spot-price signals, the industry may be able to reduce the kind of extreme oversupply that has defined prior cycles.
The cycle would remain. The amplitude could look different.
What this could mean for Sandisk, Micron and other storage names
Rising AI storage demand will not benefit every memory company in exactly the same way.
Sandisk is one of the most direct public proxies for NAND and enterprise flash. Its equity story is now tied to AI-driven bit demand, multi-year customer commitments, QLC and HBF. That leaves SNDK especially sensitive to whether the market believes current NAND profitability can persist.
Micron spans DRAM, HBM and NAND, so it sits at both ends of the AI memory hierarchy: high-priced, high-bandwidth memory close to the accelerator and larger-capacity flash further out in the stack. The article says strong AI demand has also improved supply visibility across several of Micron’s product lines.
SK hynix remains primarily an HBM story in AI, but its work with Sandisk on HBF stands out. It suggests that a leading HBM supplier also sees room for a NAND-type high-capacity memory layer in future AI inference architectures.
Kioxia offers another direct window into NAND. As AI workloads increase demand for high-capacity storage, the company is accelerating next-generation flash mass production, and the AI boom has materially improved its operations and valuation, according to the source.
Western Digital has shifted its core focus toward hard drives rather than NAND after the Sandisk separation. Even so, if AI-generated data keeps growing quickly, multiple storage layers could benefit at once. Flash is better suited to high-performance, low-latency workloads, while HDDs still hold a cost advantage in large-scale cold data storage.
The broader point is that AI storage is not a one-product trade. Continued growth in AI data could lift demand for HBM, DRAM, NAND, enterprise SSDs and high-capacity HDDs at the same time, with each technology occupying a different layer in the stack.
AI may change the NAND cycle, but not erase it
The article cautions against describing the current storage rally as a permanent break from cyclicality.
NAND remains a highly capital-intensive business. Even without building entirely new fabs, technology upgrades can increase the number of bits produced per wafer. Competitors can also add supply again when margins are rich, and AI capital spending itself is unlikely to grow at the same rate forever.
Recent share-price swings reflect that risk. The source says Sandisk and Western Digital both fell sharply after reporting strong August earnings because expectations had already become so high that even upbeat outlooks were not enough for investors.
That leads to a narrower structural conclusion. AI can make long-term NAND bit demand stronger and more predictable, while multi-year customer agreements and tighter supply discipline may reduce the violence of future storage cycles.
That is different from saying prices need to rise forever. What matters is whether suppliers can still sustain reasonable profitability when prices normalize, rather than repeatedly moving from shortage to severe oversupply.
If that shift does take hold, valuation frameworks for companies such as Sandisk could change with it.
MEXC listed several SNDK-linked market routes in the source
The article also mentions that MEXC offers several market routes for eligible users who want exposure to the AI storage and NAND theme through Sandisk.
SNDKSTOCK_USDT perpetual futures provide leveraged long and short exposure tied to SNDK-related price moves. Users interested in real U.S. stock trading services can also review related access through MEXC RealStocks.
MEXC has also listed SNDKON/USDT, which the source identifies as a tokenized Sandisk stock product provided by Ondo Finance. The article says SNDKON represents tokenized economic exposure to Sandisk rather than direct ownership of Sandisk common stock.
Readers who want to study Sandisk further can also refer to MEXC’s Sandisk stock guide for information on the company’s business model, products and main risks.
The source adds that different products carry different market structures and risks. Perpetual futures involve leverage, funding rates and liquidation risk. Tokenized assets also introduce issuer, liquidity, blockchain and regulatory risks, and product availability may vary by region.
Key questions raised in the source FAQ
Does AI inference use NAND flash
Yes. The article says AI inference systems can use NAND-based enterprise SSDs to store model data, embeddings, cached information and other large datasets, especially data that does not need to remain in expensive HBM or DRAM at all times. As models and datasets continue to expand, that capacity layer may become more important.
Can NAND replace HBM in AI systems
No, not directly. HBM delivers extremely high bandwidth and low latency close to AI accelerators, while NAND provides much larger capacity at far lower cost per bit. Newer ideas such as HBF are meant to narrow the performance gap, but Sandisk itself positions HBF as a complement to HBM rather than a replacement.
What is High Bandwidth Flash
The source describes HBF as a new NAND memory architecture built for AI inference. In Sandisk’s design, stacked flash and advanced bonding techniques are intended to provide higher bandwidth than conventional NAND while preserving far more capacity than HBM.
Does AI mean the NAND cycle is over
No. NAND is still a cyclical and capital-intensive industry. AI can improve long-term bit demand, and multi-year customer agreements may help supply discipline, but pricing, capacity additions and technology upgrades can still create oversupply and margin pressure.
The bigger shift: AI needs capacity, not only compute
The first phase of the AI infrastructure boom was dominated by a shortage of compute. GPUs, HBM and high-speed networking rose in value quickly because the market did not have enough performance to meet demand.
Inference introduces a different limit. Capacity.
If AI is going to run at global scale, large pools of data need to remain continuously accessible, but not all of that data can live in the most expensive memory tiers. That creates room for NAND to move higher in the AI memory hierarchy through enterprise SSDs, QLC and potentially HBF over time.
Among the companies discussed in the source, Sandisk is one of the most aggressive in arguing that this change could support a more stable NAND business model. The significance of the trend, though, goes beyond SNDK alone.
If AI inference keeps increasing storage density requirements while suppliers preserve capacity discipline, the next NAND cycle will still be a cycle. It may just look very different from the NAND cycles investors were used to in the past.