Memory Shortage to Persist Through 2028 as AI Demand Outpaces Supply

Micron CEO Sanjay Mehrotra told investors that demand for the company's high-bandwidth memory (HBM) and server DRAM will exceed available supply through at least 2028. Roughly 75 percent of Micron's 2027 output is already committed, and most current sales discussions center on 2028 allocations. HBM pricing for 2027 is "much higher" than 2026 levels, reflecting the scramble for AI-grade memory.

Capacity shifts toward AI, away from consumer

Manufacturing capacity is being redirected to HBM, which Samsung projects will consume nearly 30 percent of DRAM wafer starts in 2027, up from 20 percent today. That shift leaves less silicon for consumer DRAM, a trend already visible after Micron halted Crucial retail RAM sales in December 2025. Prebuilt PCs, smartphones, and consoles are seeing higher prices or reduced memory configurations.

New clean rooms slated for 2028 will not provide immediate relief. Mehrotra noted that production ramps gradually, and each node transition yields smaller productivity gains per wafer. Meanwhile, larger models, longer context windows, and agentic workloads continue to drive memory intensity higher across enterprise and consumer segments.

Why it matters for GPU / AI infrastructure: GPU clusters are only as fast as the memory feeding them. Persistent HBM shortages mean longer lead times for AI server builds, higher bill-of-materials costs, and pressure to optimize memory efficiency in model design. Infrastructure buyers should lock in supply agreements early and evaluate memory-aware architectures.