Summary
Highlights
SK Hynix, a key supplier of high-bandwidth memory (HBM) for Nvidia's AI accelerators, announced a slowdown in the production ramp of its next-generation HBM4 chips. This news triggered a sell-off across the AI semiconductor sector, with companies like Samsung and Micron seeing their stock prices drop, as the market interpreted this as a decline in AI demand.
The market's initial reaction is largely mistaken. Firstly, SK Hynix has already sold out its entire 2026 HBM supply, indicating strong demand. Secondly, the HBM4 slowdown is a supply-side timing issue, not a demand problem. It's primarily due to delays in Nvidia's upcoming Vera Rubin GPU platform, which requires HBM4. SK Hynix is simply reallocating wafer capacity to more immediately profitable ventures until Rubin is ready.
SK Hynix is redirecting its freed-up wafer capacity towards commodity DRAM, which is experiencing a scarcity-driven market expansion. Due to the significant investment in HBM production over the past three years, conventional DRAM now faces high demand, with operating margins projected to reach up to 90%. This indicates that SK Hynix is leveraging a second premium market created by the AI boom.
The market incorrectly views SK Hynix's expansion into GDDR7 and LPDDR6 as a retreat from AI memory. In reality, these are crucial AI memory products for different phases of the AI build-out. HBM is primarily for AI training, which requires extreme bandwidth. However, AI is evolving to include inference and edge computing, which have different memory requirements. Nvidia's Reuben platform, for instance, will use GDDR7 for inference prefill due to its cost-effectiveness, and LPDDR6 is vital for AI agents and localized AI applications on devices like laptops and smartphones. This diversification reflects the expanding and evolving AI memory market, not its contraction.
The sell-off in memory stocks like Micron is a mispricing. Micron, also an HBM4 partner for Nvidia, faces the same Rubin delay but also has the opportunity to redirect production to high-margin commodity DRAM. Samsung, with stronger commodity DRAM and NAND flash operations than SK Hynix, is well-positioned to benefit from this diversification trend and the scarcity in the commodity DRAM market. The uniform negative market reaction fails to recognize these distinct opportunities and the overall bullish outlook for the AI memory sector, driven by increasing infrastructure spending and market diversification.