The RAM shortage could last years

The more pointed detail buried in the SK Group warning is the 2030 horizon, which means this isn't a near-term supply blip that capex can fix quickly. Even aggressive fab investment today runs into 18-to-24-month lead times for advanced packaging and HBM tooling, so the shortage compounds precisely as AI infrastructure demand is accelerating.

Stratechery's piece from April 13 ('Mythos, Muse, and the Opportunity Cost of Compute') argued that controlling demand, not supply, becomes the leverage point when compute is constrained. A multi-year RAM shortage is exactly the condition that argument was stress-testing. Meanwhile, the funding activity we covered that same week, including Upscale AI raising at a $2B valuation just seven months in, reflects investor bets on infrastructure layers that sit above the physical memory constraint. If memory stays scarce through 2027 and beyond, those infrastructure bets get more complicated: efficiency and orchestration matter more, but raw capacity ceilings still bite.

Watch whether Samsung or SK Hynix announces accelerated HBM4 capacity commitments before end of Q3 2026. If neither does, the 60% demand-fulfillment figure for 2027 should be treated as a ceiling, not a floor, and AI infrastructure buildout timelines will need to be revised accordingly.