SK hynix chief expects AI memory crunch to last through 2030

If SK hynix is right that AI-driven memory shortages will persist through 2030, the bottleneck in the race to AGI is not just GPUs, it is the entire memory stack. High bandwidth memory has become the scarce complement to frontier accelerators, and a multi-year structural shortage implies that access to HBM-rich systems will remain the key gating factor for very large training runs.

Kwak’s point that HBM and other AI-centric products are now custom co-designed with hyperscalers is equally important. Customization makes demand more predictable and contracts longer term, which in turn gives chipmakers confidence to keep capex high. That dynamic tends to favor incumbents like SK hynix and Samsung over upstarts, locking in an oligopoly at the heart of AI compute.

For AGI timelines, this suggests that compute will keep growing but will be preferentially allocated to customers who can sign deep, long-run supply deals: frontier labs, major clouds and defense programs. Smaller labs and open collectives may find it harder to secure enough memory bandwidth for frontier-scale experiments, reinforcing centralization even as model weights and code get more open.