The storage investment craze is heating up again! Samsung is ahead in delivering the first batch of HBM4E samples in the storage field. Bandwidth, capacity, energy efficiency, and heat dissipation have all made significant leaps.

The global giant in storage chips, Samsung Electronics, headquartered in South Korea, has started shipping samples of the most advanced data center storage components to customers. In a competition for key storage components for cutting-edge AI chips manufactured by American tech giants such as Nvidia, AMD, Google, and Amazon, Samsung has gained an early advantage over competitors such as SK Hynix and Micron. Samsung's first batch of 12-layer HBM4E samples could potentially catalyze a new wave of storage investments, pushing the "artificial intelligence storage supercycle" further towards high-end, long-term, and revalued estimations. The South Korean company, with the highest market value, has started shipping its first batch of 12-layer HBM4E samples to global major customers, including Nvidia. Previously, Samsung had started mass production of HBM4 in February, marking a significant milestone in the rapid development of the high-bandwidth memory market. HBM, by vertically stacking multiple layers of dynamic random-access memory (DRAM), reduces power consumption while significantly increasing data transfer speeds. It has become a critical component in AI front-end processing systems used for training and deploying large language models such as Claude and GPT. AI large models are often created through data-intensive software and high-density matrix operations, involving tens of trillions of parameters and heavily relying on HBM storage systems. AI inference workloads involve massive parallel computing patterns, equally depending on HBM storage to provide high bandwidth, low latency, and high energy-efficient storage solutions. To avoid AI computational bottlenecks and keep expensive GPU processors working at full speed, Micron and its competitors - SK Hynix and Samsung - have developed HBM storage systems that are much faster and more efficient in communicating with other components than traditional storage speeds. HBM is a high-bandwidth, low-power storage technology used specifically in high-performance computing and graphics processing fields. HBM, through 3D stacking technology, connects stacked DRAM chips together through fine Through-Silicon Vias (TSVs) for high-speed data transfer. By stacking multiple memory chips together through 3D stacking technology, HBM not only significantly reduces the space occupation of memory systems but also greatly reduces the energy consumption of data transfers. High bandwidth can significantly improve data transfer efficiency, allowing AI large models to run more efficiently around the clock. HBM storage systems also have powerful low-latency characteristics, enabling quick response to data access requests. Generative AI large models and AI agents often need to frequently access large data sets and perform intense large model inference workloads. The strong low-latency characteristics can greatly improve the overall efficiency and response speed of AI systems. In the field of AI infrastructure, HBM storage systems are fully integrated with AI GPU servers such as Nvidia's B200 and GB200, as well as the upcoming Rubin architecture NVIDIA AI processing infrastructure. Memory chip manufacturers are competing to win large-scale design integrations for future AI systems. As the investment scale in global AI computing infrastructure grows exponentially, the demand for faster, more efficient storage components continues to rise. Samsung's early progress with 12-layer HBM4 compared to other memory chip manufacturers may enhance its competitive advantage in HBM technology over competitors like SK Hynix. These storage chip giants have been profiting significantly from the unprecedented AI boom since 2025. SK Hynix previously stated in April that their goal is to mass produce HBM4E by 2027 and supply samples to customers in the second half of this year. Therefore, if Samsung Electronics can obtain certifications and bulk orders from major clients like Nvidia before their competitors, it will help the company regain growth momentum in the high-end storage sector. Chip manufacturers are striving to meet the increasingly growing storage efficiency requirements for the next generation of AI accelerators. Samsung's advancement with 12-layer HBM4E from the roadmap to customer sampling phase signals that the high-end storage competition for next-generation AI accelerators has entered the customer certification window earlier. Performance data shows that HBM4E has a speed performance improvement of over 20% compared to Samsung's previous HBM4 product. It utilizes the sixth-generation 10-nanometer dynamic random-access memory (1c DRAM) technology and Samsung's exclusive 4-nanometer logic-based chip manufacturing process. For Samsung, this advancement is not just a technological release but a significant step towards qualifying for the next generation AI chip designs from major customers such as Nvidia, AMD, Google, and others. In terms of performance indicators, the value of 12-layer HBM4E compared to its predecessor is concentrated on bandwidth, capacity, efficiency, and heat dissipation. The stable pin transmission speed reaches 14Gbps and can be scaled up to 16Gbps. The maximum memory bandwidth per stack can reach 3.6TB/s, with a capacity of 48GB, which is a 30% increase from the previous generation. Efficiency is improved by 16%, and thermal resistance is improved by over 14%. All these data indicate that the advancements in HBM4E are not just about increasing capacity but also about addressing the bottlenecks and power constraints in AI training, inference, multimodal models, and high-performance computing. Samsung recently revealed at GTC 2026 that the next generation HBM4E can achieve a rate of 16Gbps per pin and a bandwidth of 4.0TB/s, demonstrating a firm focus on optimizing customer specifications, yield, power consumption, and thermal design in the pre-production sample stage. HBM4 has a stable transmission speed of 11.7Gbps, which can reach up to 13Gbps, showing an approximately 22% improvement over HBM3E. HBM4E further enhances speed performance by over 20% compared to HBM4. The strategic importance of HBM4E compared to HBM4 lies in transitioning the storage subsystems of AI accelerators from following AI GPU updates to defining the next generation computing platform collaboratively. For large model training, higher bandwidth can reduce the idle time of graphics processors waiting for data. For inference, higher capacity and lower power consumption can help expand the context window, increase batch processing throughput, and reduce data center unit token energy consumption. The arms race in AI storage is escalating, adding a new engine to the storage frenzy in the stock market. Samsung's initial delivery of the 12-layer HBM4E samples could potentially catalyze a new wave of storage investments. However, more accurately, it is not just igniting a trend but rather advancing the existing "artificial intelligence storage supercycle" towards high-end, long-term, and re-evaluated valuations. Samsung's delivery of the HBM4E samples is definitely not the end of the super bullish trend for storage chips but rather marks a phase where the super cycle enters a stage of "next-generation flagship product certification and long-term capacity agreements for DRAM/NAND." Mizuho, a Japanese financial giant, predicts that the global HBM market size will increase from $35.9 billion in 2025 to $246.1 billion in 2028, with a projected annual compound growth rate of around 90%. HBM4 and HBM4e are expected to be the core of the next growth cycle. Mizuho points out that the bottleneck of AI infrastructure has evolved from AI GPU/AI ASIC to a full-stack storage pressure involving central processors (CPU), data center-level dynamic random-access memory (DRAM), high-bandwidth storage, NAND, and enterprise solid-state drives. The key calculation is that the token generation rate of AI agents may be a thousand times higher than traditional generative AI, prompting Mizuho to raise its target price for Micron from $800 to $1,150 and for SanDisk from $1,625 to $1,825. The core reason is the amplification of memory and storage demands by intelligent agents. From an AI data center engineering perspective, HBM4E is a key component in the transition of AI computing systems from "stacked computing power" to "coordinated optimization of bandwidth, capacity, and efficiency." The bottlenecks in training and inference of large models are not just AI GPU/AI ASIC computing power but also the flow of data between data center GPU/ASIC chips and storage at a fast and low-latency rate. Mizuho states that the key difference between AI agents and traditional chatbots like Siasun Robot & Automation is that AI agents do not just have a question-and-answer structure but also break down tasks, access tools, search for information, read and write files, maintain long-term contexts, and enable the parallel collaboration of multiple intelligent agents. This spreads the memory pressure from high-bandwidth storage near GPUs/ASICs to LPDDR/DDR on the CPU side, inference cache, enterprise solid-state drives, and long-term data storage. This strengthens the increasingly bullish view of Wall Street analysts towards global DRAM/NAND storage chip sectors. As UBS aggressively raises the target price for Micron to $1,625, one of the memory chip manufacturers, Micron, has surpassed a trillion-dollar market value for the first time due to the tight supply and demand for AI storage. The market is shifting its focus from the dominance of GPU computing trends to include storage suppliers. Importantly, the market is not just interested in HBM but in the holistic "artificial intelligence data storage infrastructure chain" comprised of server dynamic random-access memory, data center DDR5, enterprise solid-state drives, NAND, and all-flash systems. UBS's recent aggressive upward revision of Micron's target stock price proposes a "de-cyclicalization of the storage industry" valuation proposition. UBS analyst Timothy Arcuri has raised Micron's target price from $535 to $1,625, driving Micron's stock price up by about 19% in a single day, surpassing a trillion-dollar market value. The core rationale is not just about betting on a further rise in spot prices but believing that the demand for AI storage and long-term supply agreements is changing the predictability of memory industry profits. UBS believes that long-term agreements, fixed prices, volume locks, and customer prepayments/supply guarantees will weaken the traditional logic of "price surge and fall - massive profit fluctuations - long-term valuation discounts" in the memory industry. If Micron can maintain a high-profit base during a downturn cycle, the market should not use the traditional valuation range of 5 to 8 times during cyclical stocks but may move towards a valuation framework closer to structural semiconductor growth stocks. At the NAND end, the narrative of the supercycle is further reinforced. TrendForce predicts that due to the demand from artificial intelligence and data centers, the contract price of NAND Flash in the second quarter of 2026 will rise by 70% to 75% compared to the previous quarter. Specifically, the NAND market, driven by the demand for artificial intelligence and data centers, is witnessing price increases spread across the entire product portfolio. In other words, artificial intelligence not only consumes high-bandwidth storage but also consumes enterprise solid-state drives, cold/warm data storage, model checkpoints, vector databases, log data, and backup recovery capacities. HBM is the "near-end high-bandwidth memory" for AI accelerators, and enterprise NAND/solid-state drives are the "persistent data repository" for AI data centers, together forming the dual engines of the storage investment frenzy. UBS's recent aggressive upward revision of Micron's target stock price suggests a "de-cyclicalization of the storage industry" valuation proposition. UBS analyst Timothy Arcuri has raised Micron's target price from $535 to $1,625, pushing Micron's stock price up about 19% in a single day, surpassing a trillion-dollar market value. The core rationale is not just about betting on a further rise in spot prices but believing that the demand for AI storage and long-term supply agreements is changing the predictability of memory industry profits. UBS believes that long-term agreements, fixed prices, volume locks, and customer prepayments/supply guarantees will weaken the traditional logic of "price surge and fall - massive profit fluctuations - long-term valuation discounts" in the memory industry. If Micron can maintain a high-profit base during a downturn cycle, the market should not use the traditional valuation range of 5 to 8 times during cyclical stocks but may move towards a valuation framework closer to structural semiconductor growth stocks.