AI Demand Drives Memory Shortage Through 2028, Micron Says
Escalating compute requirements for large models shift wafer capacity from consumer RAM to high-bandwidth memory.
Rising demand for artificial intelligence workloads has created a severe global memory supply crunch that memory executives warn will persist through at least 2028. Manufacturing capacity is increasingly prioritized for high-bandwidth memory (HBM) and enterprise server DRAM, driving up compute hardware pricing across the semiconductor supply chain.
Key details
Micron Technology and Samsung Electronics confirmed that high-bandwidth memory supplies are severely constrained as semiconductor manufacturers allocate larger shares of wafer production to AI hardware. Micron CEO Sanjay Mehrotra reported that 75% of the company's memory output for 2027 is already sold out, with contract prices for 2027 significantly higher than 2026 levels. Most customer purchasing discussions have already shifted to 2028 allocations.
To meet hyperscaler demand, memory makers are reallocating wafer capacity away from consumer hardware. According to Samsung, high-bandwidth memory will account for nearly 30% of DRAM wafer capacity by 2027, up from 20% in 2026. Although Micron plans to open new manufacturing clean rooms in 2028, production ramps will be gradual due to the technical complexity of transitioning to next-generation HBM4 and HBM4E architectures.
Why this matters
High-bandwidth memory is a critical component for AI training and inference accelerators such as Nvidia GPUs. As model parameter counts, context window sizes, and agentic AI deployments expand, memory bandwidth and capacity bottlenecks directly raise the hardware capital expenditure for building AI data centers. Tighter memory supplies translate directly into higher cluster construction costs and increased per-token API pricing for enterprise AI deployments.
Context
The memory squeeze follows a broader pattern of hardware inflation across the AI supply chain. Hyperscalers and specialized neocloud providers have committed hundreds of billions of dollars to secure GPU clusters and server components. The decision by major memory manufacturers to pivot wafer production toward enterprise AI has already impacted consumer technology markets, leading to reduced RAM configurations in personal computers and smartphones as component suppliers prioritize lucrative AI contracts.
Risks and open questions
A prolonged memory shortage creates financial risks for AI infrastructure developers facing rising hardware procurement bills. If memory prices continue to escalate through 2028, hardware depreciation costs for AI datacenters will increase, putting pressure on software margins. Additionally, clean room construction delays or yield challenges during the transition to HBM4 could further constrain supply, exacerbating hardware bottlenecks for hyperscale compute buildouts.
What happens next
Industry analysts expect contract negotiations for 2028 memory allocations to intensify as cloud providers attempt to reserve capacity for future AI superclusters. Market watchers will track whether clean room construction timelines remain on schedule and if next-generation memory packaging yields can ease supply constraints before 2028.
Source: Ars Technica Published on AI Usage Global, author: AUG Bot
