After HBM, the "storage super boom" - NAND! AMD (AMD.US) decisively acquired MEXT, flash memory is dominating the "AI reasoning economy".

American PC and data center high-performance chip leader AMD (AMD.US) once again received strong bullish sentiment support from investors on Monday, after the company announced the acquisition of MEXT - a technology company that has been focusing on AI-driven data center DRAM/NAND storage optimization technology for many years. AMD's stock price surged nearly 7% to $547.26 as of Monday's US stock market close, with a market value hovering around $890 billion. Leading US NAND flash supplier SanDisk (SNDK.US) also saw its stock price rise by nearly 7%, highlighting the market's increasing recognition of NAND flash technology upgrading from "storage capacity assets" to "quasi-memory layer" in the AI computing system, likely becoming the next super trend in the storage chip field after HBM. It is understood that AMD, who rarely participates in the M&A market, completed the latest acquisition with the aim of addressing one of the biggest pain points in modern data centers - the increasing demand for AI and data-intensive workloads leading to a growing memory limitation problem. Enterprises are facing the challenge of very limited HBM/DRAM memory access capabilities, which slows down performance and significantly increases overall storage chip costs. NAND flash, which has traditionally been positioned by industry insiders as "cold data/capacity storage" assets, is now being upgraded to an "expanded quasi-memory layer" in the AI inference era. It is likely to become one of the most important cutting-edge technology trends in the storage chip industry after HBM. The signal revealed by AMD's acquisition of MEXT is crucial - AMD officially emphasizes that MEXT's AI-driven storage optimization technology aims to make flash memory behave more like DRAM, expanding available memory capacity, reducing infrastructure costs, and greatly enhancing the efficiency of expanding AI and general workloads. According to AMD's latest acquisition announcement, MEXT's proprietary technology helps improve the performance and energy efficiency of lower-cost NAND flash type storage chips to make them behave more like DRAM or even HBM systems. In simple terms, this will significantly strengthen AMD's product lineup dominated by CPU+GPU+high-performance network infrastructure in the AI data center field, and the significant cost advantage of NAND flash compared to DRAM determines that NAND market share is expected to expand further in the AI inference era. AMD said that after acquiring MEXT, it will help customers deploy massive AI training/inference workloads with higher efficiency and lower total cost of ownership (TCO). The addition of MEXT's engineering talent is also expected to support the company's future needs for accelerated expansion in enterprise AI applications and cloud computing. In its acquisition statement, AMD pointed out that as AI models, data analysis, virtualization, and high-performance computing workloads continue to grow in scale and complexity, memory has become the most critical constraint in the cloud and enterprise AI workload environments. For customers, resolving these bottlenecks is essential for improving performance per dollar, enhancing data center resource allocation efficiency, and accelerating large-scale AI deployment. AMD is addressing this challenge through the acquisition of MEXT, which is a pioneer in AI-driven storage optimization technology. The market has responded positively to this deal. AMD's stock price surged nearly 7% at the close, continuing its strong performance this year; the market also gave positive feedback to leading US NAND flash technology company SanDisk. Why are AMD, SanDisk, SK Hynix, and Samsung Electronics accelerating the update and iteration of NAND flash memory? The signal revealed by AMD's acquisition of MEXT is crucial, that MEXT's AI-driven storage optimization technology aims to make flash memory behave more like DRAM, expanding available memory capacity, reducing infrastructure costs, and significantly improving the efficiency of expanding AI and general workloads. MEXT itself positions its technology as a path to allow low-cost, high-capacity system flash memory to perform at speeds close to DRAM for the operating system, and claims to bring about 2-4 times expandable memory capacity. This also indicates that AMD is interested not only in the storage supercycle dominated by NAND itself but in pulling NAND into the AI server ecosystem led by AMD through software, controller, predictive scheduling, and memory/storage chip-level management. The underlying change in the logic of this data center construction process shifting from DRAM/HBM memory to NAND flash memory field is driven by the rapid expansion of memory capacity and cost demands in the AI inference era. When AI inference enters the stage of long context, RAG, vector database, multimodal cache, and KV cache explosion, the scarcest resource is not just the bandwidth and the AI computing power cluster dominated by NVIDIA Corporation/AMD or the TPU AI computing power cluster dominated by Alphabet Inc. Class C, but the high-capacity, low-cost, low-power, scalable memory pool. However, the supply elasticity, cost, and power consumption of DRAM/HBM cannot handle all the demands alone. TrendForce data shows that in the first quarter of 2026, the traditional DRAM contract price rose by approximately 93%-98% month-on-month, driving overall revenue growth of the DRAM industry by 81% to $97 billion; TrendForce data also shows that NAND flash prices are on an upward trajectory, but the overall growth since the start of the storage supercycle in 2025 and the price base are still significantly lower compared to DRAM/HBM memory systems. TrendForce predicts a 55%-60% month-on-month increase in NAND Flash prices in the first quarter, mainly driven by enterprise SSD orders from North American cloud service providers. Even more significant is the data calculated by Citrini Research, which shows that the cost per bit of NAND flash is about 1/55 of DRAM - QLC NAND costs about $0.05 per GB, DDR5 DRAM costs about $2.75, and HBM3E costs as high as $15 per GB. The space available for this price difference lies in the fact that the largest single memory consumer in AI inference - the KV cache (recording all previously tagged contexts in each model generation step, which can grow to hundreds of GB in long conversations) - requires much lower reading speed than the decoding path of the model weights. For such sequential read data, the speed advantage of DRAM significantly narrows, while the capacity advantage of flash memory is fully realized. The appearance of the HBF technology route built from NAND flash memory (i.e., "high-bandwidth flash") further strengthens this judgment. SanDisk, SK Hynix, and Samsung have explicitly defined high-bandwidth flash (HBF) as a new form of NAND aimed at the "memory wall" in AI, with the goal of providing larger capacity in AI inference and claiming that HBF can achieve performance close to "infinite capacity HBM" in related inference tests, while significantly increasing available memory capacity. SK Hynix and Sandisk have been jointly accelerating the standardization of HBF, with the storage industry positioning it as a new high-speed flash standard for AI inference servers. The academic community is also rapidly following suit, with papers such as HAVEN proposing HBF as a supplement to HBM inside GPU packaging for large-scale vector databases and RAG retrieval scenarios, modeling results showing that compared to GPU-DRAM and GPU-SSD systems, the highest reordering throughput can be increased by up to 20 times, and the highest delay can be improved by up to 40 times. This means that NAND is no longer just "stored externally in servers," but it may enter GPU packaging, near-memory computing, vector retrieval, long context inference, and other critical AI paths. However, it is worth noting that NAND/HBF is not immediately replacing HBM, nor is it replacing DRAM, but is becoming a new third core asset in the AI memory system, namely, NAND is not immediately replacing HBM, but is serving as a "high-capacity, low-cost, high-bandwidth" new supplementary layer for AI inference, RAG, vector database, long context, multimodal cache, and other Beijing Vastdata Technology access scenarios. HBM still handles high-bandwidth, low-latency training and high-end inference main paths; DRAM still serves as general memory and hot data cache; NAND/HBF/enterprise SSDs are more suitable for handling large model weights, KV caches, low-frequency access data, edge model parameters, and RAG vector libraries. Apple Inc.'s "LLM in a Flash" test has proven that by using windowing and row-column binding methods, it is possible to run models on devices with limited DRAM capacity that exceed the available DRAM capacity, achieving a 4-5 times boost in inference speed on CPU and GPU compared to plain loading methods. x86 chip giant AMD rushes into the rampant bull market There is no doubt that HBM is still the crown jewel of the storage system in the AI training era, while NAND/HBF may become the "capacity + cost + efficiency" crown jewel in the AI inference era. Meanwhile, SanDisk, SK Hynix, Samsung, Micron, Kioxia, as well as various enterprise SSD controllers, CXL/PCIe interconnected storage companies, and AI system platform companies like AMD and NVIDIA Corporation, are around the "AI memory wall" and "NAND/HBF" opening a new round of revaluation of the storage industry value chain. As mentioned above, the core of AMD's acquisition of MEXT is not to enter the NAND manufacturing field, but to enhance the software and AI computing system layer capabilities to make flash memory behave more like DRAM. AMD officially states that MEXT's AI-driven DRAM/NAND storage optimization technology can make NAND flash memory behave more like DRAM, expand available memory capacity, reduce infrastructure costs, and enhance the efficiency of expanding AI and general workloads. On Wall Street, top Financial Institutions, Inc. Barclays set a target price of $665 for AMD (significantly raised from the previous $500), corresponding to a potential upside of about 21.5% from the current record high stock price, implying a market value of about $1.1 trillion for AMD. The core bullish logic of Barclays is not simply "catching up with NVIDIA Corporation in the AI GPU market share," but rather the agent-driven wave of AI (i.e., the AI intelligent body wave) driving a comprehensive burst of CPU demand, narrowing the CPU-GPU ratio, and AMD with its data center server-level CPU, AI accelerator, and cloud customer base, is expected to become one of the most important beneficiaries of AI computing infrastructure alongside NVIDIA Corporation. Barclays' analyst team also believes that the scheduling, tool invocation, request routing, and multistep workflows in agent-based AI tasks will significantly boost AMD's server CPU market space in the x86 ecosystem. Another positive view comes from Citigroup with a target price of $575, mainly betting on AMD becoming a more reliable second core supplier of AI GPU architecture outside of NVIDIA Corporation and potentially gaining a larger share from mega customers like Meta. Please note that this translation may not be 100% accurate as it was generated by AI technology.