AI Chip Market for Data Centers to Hit $860 Billion by 2030 - Seoul Economic Daily

The Korea Export-Import Bank's latest analysis highlights a structural shift in the AI chip landscape: the center of gravity is moving from training-focused GPUs to inference-optimized NPUs, a transit... Read original →

The inference-to-training pivot isn't a demand shift—it's a reallocation of architectural power. Nvidia's 78% share is anchored to CUDA lock-in and training-cluster economics, but distributed, low-power inference workloads are structurally eroding that moat. FuriosaAI and Rebellions are betting on a narrow window: before NPU standards harden, they can compete on energy-per-token rather than peak FLOPS in sovereign and enterprise AI deployments. The real variable is geopolitical arbitrage. Middle East data-center buildout naturally sidesteps export-control gray zones, and Korea's Export-Import Bank involvement signals state-level ecosystem positioning—creating a second supplier before Nvidia's moat calcifies, mirroring Japan's 1980s procurement-mandate playbook. Within 18 months, expect Nvidia to ship inference-dedicated SKUs to close the flank, but CUDA migration costs paradoxically become the challengers' moat. Memory bandwidth, not compute, will become the scarcest resource. Cooling architectures will shift toward chip-level thermal management. The endgame isn't who computes faster—it's whose inference cost curve is steeper.