FuriosaAI CEO June Paik said at the FKI CEO Jeju forum that global new AI data center builds could total about 100 GW by 2030, which implies semiconductor procurement equivalent to a market worth several quadrillion won. He said generative AI is shifting from model training to application-facing inference, so future demand will concentrate on high energy-efficiency, low-power inference chips. Paik expects most newly built AI data centers will deploy inference rather than training compute and war

FuriosaAI CEO June Paik said at the FKI CEO Jeju forum that global new AI data center builds could total about 100 GW by 2030, which implies semiconductor procurement equivalent to a market worth several quadrillion won. He said generative AI is shifting from model training to application-facing inference, so future demand will concentrate on high energy-efficiency, low-power inference chips. Paik expects most newly built AI data centers will deploy inference rather than training compute and warned that energy efficiency, not raw compute, will determine competitiveness because power supply is the primary bottleneck for data center construction. He added GPUs still dominate today but achieving further efficiency gains will require task-optimized dedicated inference accelerators.