A 20-Fold Call

On October 14, Yonhap News reported that Ahn Hyun, head of SK Hynix's development division, publicly forecast that AI infrastructure performance and scale will grow roughly 20-fold by 2030. Cailianshe and multiple Korean financial outlets (Korea JoongAng Daily, KED Global, Businesskorea, Maeil Business) cross-reported the statement. For a memory maker, this is both a demand forecast and a declaration of capacity and roadmap ambitions.

Key Facts at a Glance

· Core forecast: AI infrastructure performance/scale to expand about 20x by 2030 (Ahn Hyun, SK Hynix);

· Roadmap: a "Post-HBM" blueprint, betting on next-generation memory beyond HBM;

· Motivation: the Korea Herald reports the company is looking beyond HBM to tackle AI memory bottlenecks;

· Context: ASML raised its outlook the same week as AI chip demand pushes equipment orders into 2030.

The Bottleneck Is Moving from Compute to Memory

As large models enter the inference era, the bottleneck is migrating: training peaks on compute, but inference is judged by token throughput per unit cost. Long contexts, multi-turn agents and multimodal workloads turn memory capacity, bandwidth and power into the new wall — scarce, pricey HBM directly inflates serving costs. SK Hynix elevating "Post-HBM" to corporate strategy essentially turns the memory wall from customers' pain point into its own product cycle.

As the compute arms race deepens, inference cost is decided not by how many GPUs you bought, but by how many tokens each GB of memory can deliver.

What It Means for Buyers

If the 20x forecast holds, AI infrastructure buildouts will rebalance around compute-memory-network: customized memory and near-memory computing gain weight, while HBM tightness persists near-term. Two practical lessons for enterprises planning private LLM deployment: budget generously for memory and I/O alongside GPUs, and evaluate models on memory footprint and throughput efficiency rather than parameter count alone. In NineZenith's private deployment practice, memory and bandwidth configuration — not GPU count — is often the first constraint to bind.

(Compiled from public reports by Yonhap News Agency, Cailianshe, Korea JoongAng Daily, KED Global, The Korea Herald, Businesskorea and others)