The volume available for sale could run out
The number that should worry every data center buyer isn’t a price increase — it’s an inventory figure. Samsung and SK hynix, who together control roughly 64% of global DRAM revenue, saw their finished memory stockpiles fall below ten days of supply on September 7, against a healthy-market norm of 30 to 45 days, according to a KB Securities report to Seoul Economic Daily.
“Beyond a simple recovery in demand, a situation could arise in which the volume available for sale itself is depleted,” warned Kim Dong-won, head of research at KB Securities, in the report — a starker warning than the price spikes that have dominated coverage of the memory shortage to date.
The mechanism is a striaght forward trade-off
HBM4 requires roughly three times the wafer capacity of conventional DRAM per unit of output, and TrendForce data shows Samsung and SK hynix are now running 150,000-to-200,000 HBM wafers per month between them — capacity pulled directly from DDR4 and DDR5 lines. Both suppliers ramped HBM4 ahead of schedule: SK hynix moved 12-layer HBM4 shipments to Nvidia forward to the end of June, and Samsung entered HBM4 mass production in February, months earlier than the market expected.
KB Securities projects memory’s share of total AI infrastructure investment will jump from 14% in 2025 to 40% this year and 57% next — a trajectory TrendForce puts even higher, at 68% by 2027.
Bottom line: For data center operators, this is no longer purely a GPU-procurement story. Server DRAM and enterprise SSDs — the tiers behind every training and inference cluster — are now drawing from the same starved wafer pool as consumer electronics and automotive chips, with KB Securities and TrendForce both signaling no relief before 2027 at the earliest.
