On November 4, SK Hynix chairman Chey Tae-won said the global memory industry is heading into its worst-ever "storage famine" in 2027, with AI customers' demand "nearly doubling," Cailian Press reported. The tight window runs longer than a quarter: SK Hynix CEO Kwak Noh-jung has separately forecast that demand will outstrip supply beyond 2030, according to overseas media.
The warning rests on order visibility. Multiple outlets report that SK Hynix has locked more than half of its 2027 and 2028 capacity with 14 customers — before new fabs come online, most of the next two years' output is already spoken for.
Expansion is moving in parallel. The company this week unveiled an investment plan of about 54 trillion won (roughly US$38 billion) for two new memory plants in South Korea, including a DRAM fab in Yongin and a NAND facility. Even so, effective relief from new capacity will take years, leaving next year's gap to be managed through allocation of existing lines.
The boom is already on the books. SK Hynix's third-quarter revenue, reported on October 28, topped 100 trillion won for the first time, a record driven chiefly by AI memory demand; the company also announced a shareholder return program of about US$75 billion.
Supply warning at a glance
2027: worst-ever memory shortage, AI demand nearly doubling (chairman Chey Tae-won, per Cailian Press)
Capacity: over half of 2027-2028 output locked with 14 customers
Capex: ~54 trillion won for two new fabs (Yongin DRAM + one NAND plant)
Results: record Q3 revenue above 100 trillion won; shortage may last beyond 2030
From HBM to conventional DRAM, memory is becoming the tightest link in AI compute expansion — after GPUs, memory pricing and lead times are the next variable moving device costs and data-center budgets. For compute builders, memory's weight in the 2027 cost sheet deserves an early estimate.