Memory chip supercycle may stretch to 2029-2030 as AI demand outpaces supply

Deloitte and SemiAnalysis say Samsung Electronics, SK Hynix and Micron face years-long capacity lead times, with DRAM prices set to more than double in 2026 and HBM shortages possibly lasting through 2027.

Summary

The memory chip supercycle may last longer than previously expected, extending from an earlier 2028 outlook into 2029 to 2030 as AI-driven demand keeps supply tight. Samsung Electronics, SK Hynix and Micron are ramping up investment in AI-related capacity, but new fabs and equipment typically take three to five years to translate into meaningful output, limiting near-term relief. Deloitte estimates the three companies' capital spending will reach $146 billion in 2027, up 3.4 times from $43 billion in 2024, accounting for 56% of global semiconductor capital expenditure of $260 billion that year. Even so, Deloitte expects newly added supply will still be difficult to release at scale before 2029 to 2030. SemiAnalysis said memory procurement as a share of hyperscale cloud providers' capital spending rose from 8% in 2023-2024 to 30% in 2026 and could climb further to 36% in 2027. Based on projected 2027 hyperscaler capital expenditure of $1.14 trillion, about $410 billion would go toward memory purchases. The firm expects DRAM prices to rise by more than 100% in 2026 and still post double-digit gains in 2027, while structural shortages in HBM (high-bandwidth memory for AI workloads) may persist through 2027. SK Hynix CEO Kwak Noh-Jung previously said 2027 could become the tightest supply year in the industry's history, and customer demand may still exceed supply capacity after 2030. Samsung Electronics also said customers that need large-scale AI infrastructure after 2029 are continuing to seek multi-year supply contracts.

Terms & Concepts
  • HBM: High-bandwidth memory used for AI processing
  • DRAM: A common memory chip used for fast data access
  • hyperscale cloud providers: Large cloud companies operating massive data centers