SK Group chairman Chey Tae‑won warned that without a surge in DRAM output, customers will trim memory usage by tightening software and infrastructure. He highlighted that hyperscalers and AI accelerators—AMD, NVIDIA, and others—currently hoard memory to support billion‑parameter models that demand hundreds of gigabytes per inference across global data centers today and in the near future.

Manufacturers such as Samsung, SK hynix, and Micron are caught between selling all available DRAM months ahead and hesitating to build new fabs, a process that can take years. SK hynix has ordered 20 Low‑NA EUV machines from ASML, but the capacity will arrive only after a long ramp‑up in the next few years for future production.

Meanwhile, AI workloads are adopting lower‑precision data formats and smarter inference tricks that slash memory footprints. Yet the classic efficiency paradox means every gain spurs higher token throughput, keeping demand high. The current supply‑chain squeeze, coupled with the lag in fab expansion, tightens the market for the next year and for industry players today in the.

With the industry poised to double AI model sizes, the tension between supply and optimization will sharpen. SK hynix’s investment in EUV tools signals a long‑term commitment, but customers already shift to leaner architectures. The short‑term result? A persistent memory crunch that forces data centers to rethink capacity planning and pricing strategies for the next six months in the industry today.