[REVISION HISTORY]

Global AI memory and NAND flash shortage intensifies

Updated 20 times since CLSTR started tracking revisions of this situation.

What changed

2026-09-05 14:12 UTC → 2026-09-07 06:46 UTC · added ~~removed~~

The global memory shortage, driven by AI demand, is intensifying across DRAM, HBM, and NAND flash sectors. Major manufacturers, including Samsung, SK hynix, and Micron Technology, have reportedly exhausted their DRAM and HBM production capacity through 2027. Micron has confirmed its entire 2026 HBM capacity is fully committed through fixed-price contracts and is currently scaling 12-layer HBM4 delivery for Nvidia’s Vera-Rubin GPU platform. To meet this demand, Micron is reportedly planning a significant expansion of its High Bandwidth Memory (HBM) production, aiming to reach approximately 100,000 wafers per month by the end of 2026. This would nearly double its previous capacity, which was estimated at 40,000 to 50,000 wafers per month last year. While Micron has not officially confirmed these specific targets, it has commenced mass production for its 36 GB HBM4 12H memory and reported that cumulative HBM4 revenue has already exceeded $1 billion. The company is also providing customers with samples of 48 GB HBM4 16H memory. Despite this expansion, Micron remains behind competitors Samsung and SK hynix, who are estimated to maintain capacities of approximately 150,000 to 200,000 HBM wafers per month. As the industry shifts from AI model training to AI inference, demand for high-capacity, high-speed storage has surged. In the second quarter of 2026, enterprise solid-state drives (eSSD) accounted for 48% of total global NAND flash memory shipments. To address evolving requirements, Samsung is pivoting toward ‘memory-centric computing,’ prioritizing bandwidth to improve metrics like tokens per second. This includes the ‘zHBM’ concept, which stacks HBM directly on top of AI accelerators to minimize latency. Samsung claims zHBM could offer up to eight times the raw performance of HBM4E chips with three times better performance-per-watt. To overcome the ‘memory wall,’ manufacturers are advancing 3D architectures. Samsung has introduced a ‘CUBE’ strategy focusing on vertical transistor architectures, while SK hynix is pursuing monolithic 3D DRAM integration. Samsung also unveiled a roadmap for HBM5, utilizing a proprietary 2nm process node with mass production projected for 2028. Additionally, Intel and Innodisk have developed a reference platform for Edge AI using Intel Core Ultra Series 3 processors and heterogeneous computing to optimize performance in robotics and cameras. Recent data highlights the scale of this procurement surge. Nvidia reported that supplier commitments rose to $279 billion from $119 billion in the previous quarter. Nvidia CFO Colette Kress noted that the $160 billion increase was ‘primarily driven by memory procurement.’

Versions

Only revisions since CLSTR began indexing content versions appear here. Select a version to see what changed compared to the one before it.