Micron's 2026 HBM4 Supply Reportedly Sold Out
Micron is now shipping HBM4 memory chips, which are critical for AI infrastructure, and its entire 2026 supply is reportedly already committed to buyers. This development highlights the extreme demand for high-bandwidth memory driven by the AI sector. It also signals intensifying competition in the HBM market, which has been dominated by Samsung and SK Hynix.
- In the third quarter of 2025, SK Hynix led the HBM market with a 57% share, followed by Samsung at 22% and Micron at 21%. Projections suggest a stable market share distribution of approximately 5:2:2 between the three companies going into 2025. - The JEDEC standard for HBM4 doubles the memory interface width to 2,048 bits and specifies bandwidth of up to 2.048 TB/s per stack. This is a significant increase from the 1,024-bit interface and ~1.2 TB/s bandwidth of HBM3E. - Key customers like Nvidia have requested HBM4 with speeds of 10-11 Gbps, exceeding the 8 Gbps JEDEC standard. In response, Samsung announced its commercial HBM4 ships at 11.7 Gbps, while Micron has delivered samples supporting over 11 Gbps. - Micron has begun shipping 36GB 12-high HBM4 samples and started high-volume production ahead of schedule, with a full ramp planned for the 2026 calendar year. The company also claims its HBM4 is over 20% more power-efficient than its prior HBM3E products. - The overall high-bandwidth memory market is projected to experience a compound annual growth rate (CAGR) of over 26%, growing from approximately $7 billion in 2025 to as much as $71.99 billion by 2035. - Competition is accelerating as Samsung also announced it has begun mass production and commercial shipments of HBM4, claiming an industry-first. Analysts expect Nvidia's next-generation "Rubin" AI platform to incorporate all three major suppliers—Samsung, SK Hynix, and Micron—into its HBM4 supply chain. - Looking beyond HBM4, Micron is partnering with foundry TSMC to manufacture the base logic die for its HBM4E memory, which is targeted for production in 2027 and may allow for more customized memory solutions for specific AI workloads.
