Coupang convenes Korean AI leaders with NVIDIA to discuss inference at scale - Coupang, Inc.

This gathering signals a structural shift: the center of gravity in AI compute is irreversibly migrating from training to inference. Coupang's real-time e-commerce workloads are constrained by latency and concurrent throughput, not peak FLOPS — the bottleneck has moved from compute to memory bandwidth. Upstream, SK Hynix and Samsung's HBM3E directly benefit from inference's bandwidth hunger. Midstream, quantization and model-compression toolchains are becoming a deeper moat than the GPU itself. Downstream, Korean AI startups gain their first near-production-grade inference infrastructure access. A structural compliance tension is being underpriced: US chip export controls target training-grade compute, yet inference workloads naturally drift toward edge and on-premise deployment — reshaping the jurisdictional boundary of those controls. Samsung and SK Hynix face a paradox: they supply the memory that makes inference viable while their own silicon ambitions compete with the very ecosystem they enable. On the competitive front, AMD's MI300X inference cost-performance positioning, potential counter-alliances from Naver and Kakao, and AWS/Azure competing for Korean AI workloads will form a three-way squeeze within 18 months. Verdict: within 12–24 months, inference will consume over 60% of total AI compute demand. Korea will likely lock in the 'Asia inference hub' niche. Purpose-built inference ASICs will begin eroding general-purpose GPU share in specific workloads.