Amidst significant target price hikes by Wall Street giant Citi and growing conviction that an unprecedented "memory chip super cycle" is far from over, U.S. NAND storage leader SanDisk Corp. (SNDK.US) has regained market spotlight. Citi Group has raised its price target by over 50% to a striking $2,025, citing sustained pricing momentum driven by artificial intelligence and seemingly limitless storage demand fueled by AI training and inference workloads.
As of last Friday's close, SanDisk shares stood at $1,407, having surged over 460% year-to-date, cementing its status as a premier "AI super stock" globally. Citi's new target implies a potential upside of approximately 51.9% from Monday's closing price. Furthermore, the latest 13F filings reveal that top-tier institutional investors on Wall Street are consistently expanding their exposure to memory chips.
In its latest research note to clients, Citi senior analyst Asiya Merchant stated: "Following the rapid global proliferation of the AI Agent super-wave and Kioxia's exceptionally strong earnings report, we have significantly raised our SanDisk [target price] from $1,300 (previously 7-8x forward P/E) to $2,025 (implying a forward P/E of approximately 9-10x for calendar year 2027). This reinforces our positive bullish view on persistently robust storage demand and a highly favorable pricing environment, all driven by exponentially increasing AI compute needs."
Analyst Merchant also reaffirmed her "Buy" rating on SanDisk. She noted that the company's numerous long-term agreements are expected to provide a baseline gross margin outlook of 80% or higher, along with a "pre-determined strong shipment trajectory" that is set to grow throughout the contract periods.
Citi's analyst team's latest forecasts indicate that the average selling price for NAND storage chips will surge 186% in 2026, with enterprise SSDs (eSSD) seeing an increase of about 265%. Analyst Merchant added, "Moreover, given the recent $6 billion authorization, we believe share buybacks could help drive significant upward revisions to SanDisk's EPS estimates."
The analyst team at Melius, led by star analyst Ben Reitzes, recently published a report suggesting the AI boom will continue to fuel strong growth in memory chip demand through the end of this decade (2030). According to market research firm Counterpoint Research, the memory market has entered a "super bull market" or "super cycle" phase, with current supply-demand dynamics and pricing conditions far surpassing the historical peaks seen during the 2018 cloud computing boom.
With the launch of Anthropic's Claude Cowork and the anticipated proliferation of autonomous super-AI agent tools like OpenClaw by 2026, this wave of AI Agents is sweeping the globe. The bottleneck in AI compute architecture is shifting from GPUs, centered on matrix multiplication throughput, to "AI Agent-driven full-stack AI systems." In this shift of the primary AI narrative, data center CPUs and memory chips are likely to emerge as the biggest winners. In other words, the AI compute bull market is expanding from "compute systems centered on AI GPU/ASIC chips" to central processors and the "data storage foundation."
AI Agents are fully igniting a "storage super cycle"! The latest 13F report shows Appaloosa Management initiated a new position in U.S. NAND storage giant SanDisk of approximately 281,250 shares, valued around $178.7 million. This move, alongside increased holdings in Micron and a South Korean chip ETF, points to a strong bullish logic centered on the "revaluation of the memory chip chain."
In its Q1 2026 earnings call, Micron's management specifically highlighted exploding demand for high-capacity data center SSDs for AI infrastructure, KV cache deployments, and PCIe Gen6 SSDs related to NVIDIA's AI compute infrastructure clusters. This indicates AI-related memory chip demand is far broader than many Wall Street analysts had anticipated.
Modern AI infrastructure not only consumes more HBM memory but also requires high-bandwidth DRAM, greater storage capacity, and high-speed SSD infrastructure to meet the growing needs of retrieval and agentic AI workloads. Emerging AI applications, including robotics, multi-agent systems, and multimodal reasoning models, are continuously creating new vectors of storage demand, suggesting AI storage intensity may continue to grow exponentially even after initial AI deployment.
As Jeremy Werner, Senior Vice President and General Manager of Micron's Data Center Business Unit, revealed in a recent interview, the underlying driver of this market trend is not simply "AI needs more compute chips." Instead, the era of AI inference dominated by AI Agents like Claude Cowork and OpenClaw is pushing memory/storage from a supporting component to a system bottleneck.
AI training engineering relies more on massively parallel computing, while inference—especially involving long context, multi-turn conversations, and Agentic AI workflows—requires persistent storage of KV Cache, context states, and intermediate results. When memory/storage space is insufficient, models must recompute historical states, leading to decreased GPU utilization and increased token generation costs.
Consequently, HBM, DDR5, LPDDR, enterprise SSDs, and even HDDs/data lakes are forming an "AI memory chain" from GPU-proximate to distant storage, determining an AI system's throughput, latency, concurrency, and per-token economics. This explains the synchronized surge in stocks like Micron, Samsung, SK Hynix, SanDisk, and Western Digital: demand is not concentrated solely on HBM but is spilling over across the entire chain of DRAM, NAND, SSD, and HDD along the AI server architecture.
More critically, AI CPUs are opening a second demand curve. While the market previously equated AI compute almost exclusively with GPU+HBM, as inference workloads become more complex, CPUs are evolving from "GPU supporting actors" to "AI coordinators" that schedule multiple Agents, manage context, and orchestrate workflows. This is expected to significantly boost demand for server DDR5 and data center-grade SSD configurations.
Simultaneously, HBM capacity is heavily locked by AI GPUs, squeezing available capacity for general-purpose DRAM. Price trends for DDR5 and DDR4 are diverging, and the shortage is spilling over from high-end HBM to the broader DRAM/NAND supply chain. TrendForce also cited Micron's CEO's latest view, stating that both traditional and AI server demand is strong but constrained by tight DRAM and NAND supply. Samsung and SK Hynix recently warned that AI-driven memory shortages could persist until 2028 or beyond.
Regarding DRAM/NAND memory chip price increases, Wall Street giant Goldman Sachs' latest assessment is that the 2026 price surge will far exceed the firm's previously optimistic forecasts. Goldman recently raised its DRAM price increase forecast from about 150% to 250%-280% and its NAND price increase forecast from about 100% to 200%-250%. In other words, Goldman believes this is not an ordinary inventory recovery cycle but a "super supply shortage cycle" caused by unprecedented demand surges driven by AI compute, HBM's increasing capacity consumption due to its complex manufacturing and packaging processes, and insufficient supply elasticity for general-purpose DRAM/NAND.
From Cyclical Stock to Core AI Infrastructure Asset: Long-Term Agreements Reshape Storage Pricing Power
Citi's bullish view on SanDisk and J.P. Morgan's revaluation of Samsung, SK Hynix, and Kioxia essentially tell the same story: the global AI compute infrastructure frenzy is shifting the storage industry from "inventory cycle trading" to "capacity scarcity trading."
Citi raised its SanDisk target from $1,300 to $2,025, citing Kioxia's strong performance, sustained robust storage demand driven by AI data centers, and an extremely favorable pricing environment for NAND/eSSD. J.P. Morgan's logic goes a step further: Long-Term Agreements (LTAs) are changing the valuation methodology for the storage industry.
Historically, DRAM/NAND were viewed as highly cyclical commodities, with valuations anchored to P/B ratios due to volatile prices, profits, and inventory cycles. However, cloud providers, seeking to secure future capacity for HBM, DRAM, NAND, and eSSD, are now proactively signing 3–5 year LTAs, making advance payments, locking in prices, or setting price floors. This is transforming storage manufacturers from a "build first, sell later" commodity model to one more akin to TSMC's "build-to-order/production-to-order" model.
Consequently, J.P. Morgan advocates shifting from a P/B to a P/E valuation framework and has raised its price targets for Samsung Electronics to 480,000 KRW, SK Hynix to 3,000,000 KRW, and Kioxia to 8,000 JPY. Korean media also reported that J.P. Morgan's target hikes for Samsung and SK Hynix are based on AI server investments driving storage into a structural growth phase and LTAs potentially reshaping storage manufacturer valuation frameworks.
The most critical data for this revaluation are the supply gap and price elasticity. Even under aggressive capacity expansion assumptions, storage chip supply for AI from 2026–2030 is projected to fall short of cloud provider demand, with a gap equivalent to approximately 450,000 wafer starts per month of capacity. Kioxia's capital expenditure as a percentage of revenue is only about 5%, significantly below its five-year average of over 20%. Samsung and SK Hynix are also expected to keep their capex ratios in the mid-single digits over the next two years.
On the demand side, the opposite is true: AI servers, Agentic AI, enterprise SSDs, and HBM demand are accelerating. Kioxia's quarterly ASP increased over 100% sequentially, and DRAM prices have tripled over the past year. SanDisk's own performance confirms this profit elasticity: its Q3 FY2026 revenue surged 97% sequentially and 251% year-over-year to $5.95 billion, gross margin rose to 78.4%, and EPS reached $23.41.
From a semiconductor engineering perspective, memory is no longer just a "server component" but a core constraint on AI system throughput and economics. The training side requires HBM to support high-bandwidth access for GPUs/ASICs, while the inference side needs DRAM and eSSD to handle KV cache, long context, retrieval-augmented generation, Agent workflow states, and large-scale enterprise data calls. The more pervasive Agentic AI becomes, the stronger the demand for data I/O, context storage, vector databases, caching layers, and enterprise SSDs.
Thus, for the first time, memory manufacturers are gaining pricing power akin to an "infrastructure bottleneck." Cloud providers are no longer solely focused on driving down procurement costs; instead, to ensure the rollout pace of future AI clusters, they are proactively using LTAs and advance payments to secure capacity. This is why the current涨价 cycle for Micron, SK Hynix, Samsung, Kioxia, and SanDisk is viewed by the market as potentially different from traditional inventory cycles.
Comments