Memory chips are becoming the most expensive “water-sellers” of the AI era

Korean storage “two champions” win AI mega-deals on the same day: SK hynix locks in two-thirds of orders for NVIDIA’s Vera Rubin platform, while Samsung secures an exclusive order for OpenAI’s Titan chip—pricing power in the storage industry is shifting from “seller-quoted on demand” to “buyer queues for reservations.”

I. Background

In Q2 2026, global HBM enters a phase of rigid supply. SK hynix mass-produces HBM4 this February, then immediately captures roughly two-thirds of the order share for NVIDIA’s Vera Rubin platform. Samsung completes HBM4 mass production in the same period and ships 12-layer HBM4E samples ahead of schedule, winning the exclusive supply order for OpenAI’s in-house Titan chip and breaking SK hynix’s prior de facto monopoly in high-end HBM.

II. Data support

In Q1 2026, revenue was 5.258 trillion won, year on year +198%; operating profit was 3.761 trillion won, year on year +400% or more; operating margin exceeded 72%; HBM market share surpassed 55%.

Samsung and SK hynix’s combined market share in premium HBM exceeded 80%, and HBM revenue grew more than 200% year on year.

Supply side: SK hynix has publicly said that its HBM, DRAM, and NAND capacity for 2026 is fully sold out. Samsung warned in its April 30 earnings report that the storage shortage will continue through 2027. Micron CEO said that over the medium term it can only meet 50%-two-thirds of demand from key customers.

Expansion side: Samsung and SK hynix announced a combined domestic investment plan of 48 trillion won on June 29. In five years, DRAM capacity will double; SK Group chairman Choi Tae-won judged at NVIDIA’s GTC that the memory shortage may last until 2030.

III. Impact

1）Order mix changes the competitive landscape: In the past, SK hynix dominated the HBM market. This time, Samsung winning an exclusive order from OpenAI means that leading AI companies are beginning to intentionally cultivate a second supplier to avoid single-source dependence. This is a mid-term variable for SK hynix’s bargaining power, but in the short term it does not change its status as NVIDIA’s main supplier.

2）Industry-chain transmission: Wafer-fab equipment costs account for over 70% of the total. SEMI projects that the global semiconductor equipment market size in 2026 will grow 23.5% to $152.2 billion. Orders for equipment, materials, and OSAT are already booked through 2027, and China’s related industrial chain may see knock-on effects.

3）End-market cost pressure: Regular DRAM contract prices rose 45%-50% in Q4 2025, while the merged-contract price for HBM rose 50%-55%. Storage cost increases are gradually transmitting to servers and consumer-electronics end markets.

IV. Outlook and risk warnings

The core logic of this storage upcycle is a structural reassessment of memory capacity driven by AI inference demand, not just a cyclical price increase. This is a key difference from historical storage cycles. However, three risks need a clear-eyed view: First, capacity expansion has a lag. If the huge capital expenditures for 2026-2027 concentrate in release after 2028, a period of supply oversupply remains possible. Second, HBM orders are highly concentrated among a small number of major customers such as NVIDIA and OpenAI. If the growth rate of AI capital expenditures slows, demand elasticity will directly show up in the storage vendors’ performance. Third, Samsung and SK hynix’s current valuation already partially prices in optimistic expectations for the next 2-3 years. The market trades mid-term pricing power rather than any single-quarter earnings, so volatility will be amplified.

The above is only an industry logic and publicly available data整理. DYOR.