The AI explosion is rewriting the underlying logic of the NAND flash industry. Long treated as a “commodity” trapped in a cycle of oversupply and price collapses, NAND has now—for the first time—become a “necessity” on the compute path, driven by the explosive growth of AI inference workloads. In 2026, data centers’ share of NAND demand historically overtook smartphones to become the largest end-market.
According to Kioxia’s projections, total data center NAND demand will grow from 286 EB in 2025 to 1,686 EB by 2031, a compound annual growth rate of 34%. Within that, AI inference-driven demand is growing five times faster than AI training demand, making it the dominant force behind this NAND supercycle.
A Fundamental Shift in Demand Structure
Before the AI boom, NAND industry demand growth relied primarily on per-device capacity upgrades in smartphones, PCs, and other consumer electronics, plus the gradual replacement of HDDs by SSDs—steady and predictable. The industry’s boom-bust cycles were driven largely by disorderly capacity expansion and abrupt cutbacks on the supply side. NAND makers could only hope to “lose less than competitors,” and technology leadership rarely translated into pricing power.
2026 marks the inflection point: data centers’ share of total NAND demand jumped from roughly 30% in 2025 to approximately 45%, while smartphones contracted to about 23% over the same period. The demand driver has shifted from price-sensitive consumer electronics to cloud service providers making decisions based on total cost of ownership (TCO) and return on invested compute (ROI).
Cloud providers’ procurement logic has fundamentally changed: from “are NAND prices cheap enough this quarter” to “can we secure sufficient high-performance enterprise SSDs on time to keep our GPU infrastructure on track for the next three years.” This shift from short-term price squeezing to long-term supply assurance provides the demand-side foundation for NAND makers to push new long-term agreement models (NBM).
KV Cache Offloading: NAND Becomes a “Token Battery”
The most profound change of the AI inference era is that NAND now stores “intermediate computational products” for the first time. The massive KV Cache generated by long-context inference is no longer static data imported from external sources—it is “working memory” produced after GPUs expend enormous compute resources. If lost, it can only be regenerated at the cost of expensive GPU compute and electricity.
SanDisk offers an apt analogy: SSDs have essentially evolved into “Token Batteries”—tokens represent work already done, and rather than discarding and recomputing them, it is far better to “charge” them into storage. Storing KV Cache is fundamentally about saving cloud providers money on power and compute.
The viability of this shift rests on three conditions. First, HBM capacity expansion cannot keep pace with KV Cache growth. Mainstream model context lengths have been doubling or more every year—from GPT-3’s 2K tokens in 2020 to million-token contexts by 2026, a 500-fold increase in six years. Take Nvidia’s Rubin NVL72 rack as an example: its roughly 15.4 TB of pooled HBM space can serve only 193 to 230 concurrent million-token sessions at the limit, which translates to extremely low compute monetization efficiency for cloud providers.
Second, the cost gap is enormous. HBM4 costs approximately $16 per GB, while enterprise SSDs cost only about $0.30–$0.40 per GB—a 40- to 50-fold difference. Keeping “warm/cold KV Cache” that may not be accessed for hours in HBM is a severe misallocation of resources.
Third, latency issues have been addressed through engineering. GPU Direct Storage (GDS) technology bypasses the CPU and DRAM, establishing a direct channel between enterprise SSDs and GPU memory. Publicly available validation shows that GDS combined with compression can improve KV Cache recovery latency by 10x, making offloading commercially viable.
Nvidia has accordingly established a new tiered memory architecture: G1 HBM → G2 system DRAM → G3 local SSD → G4 shared network storage, orchestrated at the software layer by NVIDIA Dynamo. In the Rubin platform, Nvidia has also built the ICMS/CMX context storage layer, defined as the G3.5 tier, with approximately 16 TB of capacity per tier—nearly 60 times the 288 GB of HBM on a single GPU.
The economics of KV Cache offloading depend on cache hit rates. According to disclosures from the Manus team, tokens served from cache hits are roughly 10 times cheaper than cache misses. Typical large model hit rates range from 40% to 70%, while multi-step Agent scenarios can sustain above 95% over the long term, unlocking the “write once, read a thousand times” dividend.
The Division of Labor Across Three Core Workloads
SanDisk estimates that AI data center NAND shipments will reach 1.2 ZB by 2030, roughly triple the 2026 level. The three core workloads have clearly defined roles:
Workload TypeShare of AI Data Center NAND DemandPrimary NAND TypeCore CharacteristicsStaging Layer40%High-performance TLC and pSLCHigh-intensity I/O buffer close to GPUs; heaviest write loadKV Cache Offload Layer35%TLC/QLC tieredEntirely new incremental demand in the inference era; absorbs HBM overflow as dynamic storageFast Data Lake25%High-capacity QLCRemote centralized data repository; read-dominated
Note: Data sourced from SanDisk’s breakdown forecast of AI data center NAND demand for 2030.
By NAND type, TLC accounts for 66% and QLC for 34%. Staging inherently excludes QLC due to write intensity, with TLC firmly holding the base; QLC focuses on data lakes and warm/cold KV Cache offloading. The two workload categories are evolving in opposite directions: data lakes are moving toward “cheaper,” while Staging and hot paths are moving toward “faster and more durable.”
The rise of pSLC deserves particular attention. It involves using TLC or QLC dies as SLC, sacrificing roughly two-thirds of effective capacity in exchange for more than 10x endurance and lower write latency. The core trigger scenario is high-frequency vector retrieval in RAG (Retrieval-Augmented Generation) and Agentic Search—AI agents access data at extremely fine granularity with intense concurrency, reading and writing simultaneously, which rapidly exhausts the endurance of standard TLC. Achieving the same effective capacity as a standard SSD requires 3 to 4 times the wafer output for pSLC, creating additional pull on NAND supply.
HBF: NAND Makers’ “Upward Breakthrough”
In the traditional AI storage hierarchy, there is a severe latency and bandwidth gap between HBM and NVMe SSDs. NAND makers are now pushing a new product category called HBF (High Bandwidth Flash), whose structure essentially replicates HBM: multi-layer stacking, TSV through-silicon vias, and co-packaging with XPUs—except the DRAM dies are replaced with NAND dies.
SanDisk’s HBF under development claims 16-layer NAND stacking, maintaining HBM-class bandwidth while multiplying capacity by 8x. According to disclosures, HBF Gen1, planned for launch between late 2026 and early 2027, will offer up to 512 GB per stack, compared to just 48 GB for 16-layer HBM4. Per-GB pricing, according to industry research, may be only one-fifth to one-tenth that of HBM. Read bandwidth can reach 1.6 TB/s, approximately 55% of the maximum HBM4 specification bandwidth.
But HBF’s shortcomings are equally clear: NAND inherently has high latency (approximately 1 to 10 microseconds, versus roughly 0.1 microseconds for DRAM), slow writes, block-based reads, and limited program/erase endurance—it is destined never to replace HBM. It is suited only for data that does not require frequent reads or writes, such as full model parameters, compressed context caches, and MoE expert parameters not currently in use.
SanDisk and SK Hynix are currently the only two substantive proponents of HBF. The two companies formed a standards alliance and built a pilot production line in early 2026, six months ahead of the original schedule, with plans to produce chips by the end of 2026 and push toward volume production in 2027. Google, Nvidia, AMD, and others are unlikely to adopt HBF as customers until 2028 at the earliest. Samsung holds patent reserves but has not disclosed prototypes or timelines; Kioxia is pursuing its XL-FLASH route, which does not directly compete with HBF; Micron has no public plans.
Dolphin Research analysis suggests HBF is not a disruptor of HBM—the two coexist and complement each other: HBM handles latency-sensitive hot data, while HBF handles capacity-hungry warm data. Augmented configurations (HBF + HBM hybrid) and disaggregated configurations (pooled) are the two most likely deployment forms. SK Hynix’s newly proposed H3 architecture is a hybrid of HBM and HBF, with volume production in 2027 and customer adoption in 2028—currently the path of least resistance.
Nvidia’s “Layer Reduction” Signal and Supply Chain Implications
A telling signal has recently emerged in the industry: according to SemiAnalysis, Nvidia’s next-generation Rubin Ultra may reduce HBM from the originally specified 16 layers to 12 layers, with the latest rumors even pointing to 8 layers. Per-GPU HBM total capacity would correspondingly drop from 768 GB to 192 GB.
Behind this “layer reduction” choice is the hard constraint of DRAM wafer capacity. Currently, producing the same GB capacity in HBM consumes 3 to 4 times the DRAM wafers of conventional DRAM, and stacking higher only worsens wafer consumption. Meanwhile, new DRAM capacity is being released slowly—limited growth from 2024 to 2026, with more meaningful additions only in 2027.
But layer reduction does not mean total system capacity declines. Thanks to advances in scale-up interconnect technology, a single Rubin Ultra GPU—even with only 8 stacks of 8-layer HBM (192 GB total)—can interconnect up to 576 GPUs, with total shared system capacity reaching 110 TB. That is more than five times the B300 era (72 GPUs, roughly 20 TB total).
This shift transfers pressure to two areas: first, TSMC’s CoWoS co-packaging capacity, and second, the interconnect technology required for memory pooling. Nvidia is building a NAND pooling platform through its CMX product line to host the portion of context memory that is accessed less frequently than DRAM-resident data, further reinforcing NAND’s strategic position in the AI storage stack.
From a supply chain bargaining power perspective, Dolphin Research argues that if AI demand is constrained more by capacity, storage returns to a volume-driven logic where production capacity and cost are decisive—benefiting the NAND supply chain (HBF, enterprise SSDs) more. If constrained by bandwidth, packaging and foundry segments gain leverage, with TSMC benefiting most while memory makers’ pricing power may actually be diluted. If both are simultaneously tight, memory process technology and capacity become the hardest bottleneck, giving memory makers the greatest weight in the supply chain.
As things stand, rising capacity demand is the highest-certainty direction. User counts, usage volume, and Agent scenario penetration rates are highly likely to continue climbing. NAND’s transformation from a “peripheral component” to a “necessity on the compute path” is now irreversible.