AI reasoning pushes NAND into the "long-term contract era"! SanDisk (SNDK.US) revenue surged by 372%, and HBF is poised to take over the next storage super trend.

U.S. NAND storage chip giant SanDisk (SNDK.US) released its earnings report and future outlook after the market closed on Wednesday. The company reported a robust second-quarter performance that exceeded expectations across the board; however, its revenue and profit outlook for the next quarter fell slightly below consensus market expectations. This highlights the continuing expansion of demand for enterprise-level NAND storage chips for artificial intelligence data centers, but market expectations for SanDisk's future growth have become increasingly stringent. Even slight deviations from expectations could trigger significant fluctuations in stock prices. The company, headquartered in Milpitas, California, saw its stock price drop by more than 8% in after-hours trading. Although the midpoint guidance for revenue and profit in its current quarter (the first quarter of fiscal year 2027) was slightly above the expectations compiled by LSEG, it fell short of the average forecasts from other data providers. This decline occurred after the company's stock price had surged nearly 470% year-to-date. As NAND flash storage shifts from the long-standing classification as a traditional "cold data/capacity storage" asset to a fully upgraded "extended quasi-memory layer" for the AI inference era, Wall Street's bullish sentiment towards SanDisk seems to remain strong, leading to increasingly stringent expectations for the company's performance growth. This "expectation cliff" is also part of the rationale behind the more than 8% drop in stock price after the earnings announcements from both Western Digital Corporation and SanDisk on the same day. The "expectation cliff" indicates that even if performance and guidance surpass consensus market expectations, they may not meet the heightened thresholds required by extremely crowded positions, resulting in a collapse of confidence and a sharp decline in asset prices if expectations are not met (even if they are still showing growth). In the earnings conference call, SanDisk's management revealed that the company is transforming its highly cyclical NAND spot model into a multi-year capacity reservation model. The company has entered into several new business model agreements with a median contract duration of approximately four years, with a minimum revenue commitment of $93.9 billion. For fiscal year 2027, more than half of the company's production is covered by long-term agreements, and about two-thirds of the production for fiscal year 2028 is also secured, with some of the largest cloud computing customers requesting additional purchases for the next three to five years just months after signing. Generative AI large training sets, model weights, checkpoints, vector databases, RAG corpora, multimodal data, logs, and inference results all need to reside in high-capacity storage for long periods. During training, vast amounts of data must be continuously sent from object storage and local NVMe SSDs to GPU clusters. In the inference era, there is further demand for massive KV Cache, long context, agent states, and retrieval data. HBM handles the highest bandwidth hot data layer, DRAM serves as the system's working memory, while enterprise-grade NAND SSDs provide a much larger capacity than HBM at lower costs, serving as the "warm data and persistence layer." Therefore, NAND is not a replacement for HBM, but rather expands with HBM and DRAM in the storage hierarchy of AI servers. Meanwhile, SanDisk and SK Hynix have released the first HBF technical specification through OCP with the aim of providing a new storage layer for AI inference that offers capacity far exceeding that of HBM and bandwidth surpassing traditional NAND. HBF is better defined as a new capacity expansion layer following HBM, rather than a direct substitute. SanDisk's performance has wholly surpassed expectations! Surging demand for enterprise-grade SSDs and an increased stock buyback SanDisk CEO David Goeckler stated in a media interview that the company has shifted its NAND flash product sales model from quarterly procurement agreements to long-term procurement agreements (LTA). SanDisk indicated that the current median duration of these agreements is four years. The company has signed eight agreements with six clients, with a total contract value of no less than $93.9 billion. In the fiscal year ending July 2027, half of the company's production will be sold through these agreements; by fiscal year 2028, approximately two-thirds of production is expected to be sold under such agreements. Goeckler remarked, "This already puts us several light-years ahead compared to just three quarters ago." The company expects revenue for the first fiscal quarter to be between $10.3 billion and $10.8 billion, with the midpoint indicating a year-over-year growth of 359%. The midpoint is above the analyst average expectation of $10.47 billion derived from LSEG data; however, compared to the consensus expectation compiled by other data providers, the midpoint of approximately $10.55 billion is nearly 5.5% lower than the consensus expectations of $11.16 billion. SanDisk's management also anticipates adjusted quarterly earnings per share to be between $4.40 and $4.60, with a midpoint of $4.50, meaning slightly less than the market consensus expectation of approximately $4.558. In other words, the EPS forecast is roughly in line with expectations but does not significantly exceed them. The adjusted gross margin forecast is between 83.0% and 85.0%, maintaining a high level compared to the 85% gross margin for the company in the fourth fiscal quarter ending July 3, 2026, but showing no further significant expansion. The rapid penetration of generative artificial intelligence and AI agents has been driving up demand for SanDisk's enterprise-grade solid-state drives and NAND flash chips, as data centers require more storage capacity and AI inference computing power. The company's data center business revenue in the fourth quarter grew more than double compared to the third quarter, reaching $2.98 billion, marking a strong annual performance since the company split from Western Digital Corporation (WDC.US) in early 2025. SanDisk's total revenue in the fourth quarter was approximately $8.97 billion, exceeding the market consensus expectation of $8.39 billion. The adjusted EPS was $3.925, higher than the market consensus expectation of $3.445. Management stated on the earnings call that since April, five additional agreements have been signed under the new business model, including agreements with three new large clients, as well as expansions of two existing agreements. SanDisk's board has approved an additional $14 billion stock buyback plan, bringing the total remaining stock buyback authorization to $15.5 billion. SanDisk bets heavily on HBF, and the value of the AI storage hierarchy is set for re-evaluation SanDisk delivered a fundamental performance in this quarter that was far stronger than the response of its stock price: fourth-quarter revenue of $8.97 billion, year-over-year growth of 372%, with adjusted EPS reaching $3.925, significantly exceeding market expectations; the revenue of the data center business doubled sequentially to $2.98 billion, proving that the core of growth has shifted from traditional mobile and PC flash to AI infrastructure. The revenue guidance for the first quarter is positioned at $10.3 billion to $10.8 billion, with adjusted EPS of $4.40 to $4.60, higher than consensus according to LSEG, but lower than the higher expectations from some data sources like FactSet; adding to the fact that the stock price has climbed nearly 470% year-to-date, the significant 8% drop in after-hours trading can essentially be interpreted as "extraordinary performance failed to create a massively positive expectation difference" rather than a sudden reversal in AI storage demand. The real content that enhances the mid-to-long-term valuation during the earnings call is that SanDisk is transitioning its highly cyclical NAND spot model into a multi-year capacity reservation model. The addition of several new business model agreements, averaging about four years in duration with a minimum revenue commitment of $93.9 billion; over half of the production for fiscal year 2027 and about two-thirds for fiscal year 2028 is already covered by long-term arrangements, with some of the largest customers requesting additional future purchases just months after signing. The dynamics revealed by SanDisk's management during the earnings conference call suggest that large-scale cloud vendors no longer view NAND as a regular component that can be trimmed on a quarterly basis, but rather as a strategically scarce resource on par with GPUs, networking infrastructure, optical interconnect devices, and data center power equipment. Long-term agreements enhance demand, pricing, and cash flow visibility while reducing the probability of a repeat of the inventory collapse seen in 2023; however, they also increase customer concentration, contractual execution, and long-term pricing mechanism risks. The AI era not only urgently requires computational models but also necessitates continuous transport, storage, and low-latency access to massive state data. HBM and DRAM are responsible for the highest bandwidth working sets near GPUs, but model weights, training datasets, checkpoints, vector databases, RAG knowledge bases, inference logs, agent long-term memory, and some KV cache layers cannot all reside permanently in expensive and limited-capacity HBM; hence, enterprise-grade NVMe SSDs assume the high-performance persistent layer between HBM/DRAM and HDDs or object storage. SanDisk's TLC enterprise-grade SSDs cater to high IOPS, low latency, and highly durable training/inference workloads, while the QLC Stargate platform supports large-capacity AI data lakes at a lower per TB cost; BiCS 8 improves unit economics through higher density, performance, and power efficiency, enabling data centers to deploy more effective capacity within the same rack, electricity, and controller quantities. Global AI infrastructure cannot do without SanDisk but inevitably relies on enterprise-grade SSDs and the NAND flash layer, where SanDisk stands as one of the few core suppliers with complete capabilities in NAND design, wafer manufacturing collaboration, controllers, firmware, and enterprise-level system products. The most forward-looking growth variable for SanDisk is undoubtedly the high-bandwidth flash (HBF). SanDisk and SK Hynix have released the first OCP technical specifications, aiming to bring high-capacity, persistent NAND closer to AI accelerators to alleviate the "memory wall" during inference, thereby increasing bandwidth and reducing the overall cost of token services while providing capacity far exceeding that of HBM; it is not a short-term substitute for HBM, but rather may create a new intermediate storage layer between HBM and traditional SSDs. NAND flash appears to be undergoing a comprehensive upgrade from its long-standing identification as a conventional "cold data/capacity storage" asset to becoming an "extended quasi-memory layer" for the AI inference era, and it is likely to become one of the most important cutting-edge technology frontiers in the storage chip industry following the HBM super storage system. The significant emergence of the HBF technical path built on NAND flash (i.e., "high-bandwidth flash") further reinforces this judgment. Kioxia, SanDisk, SK Hynix, and Samsung have clearly defined high-bandwidth flash (HBF) as a new form of NAND aimed at the AI "memory wall," with the goal of providing greater capacity for AI inference and claiming that HBF can achieve performance close to "infinite capacity HBM" in relevant inference tests while significantly enhancing usable memory capacity. Wall Street financial giant Goldman Sachs Group, Inc. stated after SanDisk's earnings announcement that the market had effectively priced in almost a perfect scenario of NAND shortages, the introduction of enterprise-grade SSD designs, and rising prices prior to the earnings report; therefore, SanDisk's performance outlook merely meeting or slightly missing expectations would trigger valuation compression. Nonetheless, Goldman Sachs Group, Inc. maintains a "Buy" rating and a target price of $2,200 based on restricted supply expansion, the increasing proportion of AI data centers, and the stability of long-term agreements contributing to profitability. Goldman Sachs Group, Inc. noted that what will truly determine the central trend of stock prices moving forward is not the quarterly performance and NAND price changes but rather the performance realization rates of NBM, enterprise-grade SSD market shares, BiCS node transition costs, the pace of HBF commercialization, and whether high margins can remain resilient after gradual supply expansion.