Kepler Aims to Launch Energy-Saving Replacement for HBM in 2027

Kepler's 3D ferroelectric memory, if it reaches volume production by 2027, delivers a far more structural threat than the "HBM replacement" headline suggests. The real disruption is breaking the AI compute stack's monoculture dependency on DRAM's capacitor-charge-discharge physics. Ferroelectric domain switching sidesteps that energy ceiling, and the CoWoS/TSV packaging ecosystem built around it faces a second-order repricing. Supply-chain shock: ferroelectric wafer processes are fundamentally incompatible with existing DRAM fabs. Whether Kepler builds its own line or contracts out, the foundry must open an entirely new material stack — a capital-intensive bet that directly pressures the pricing power of the three DRAM giants in China's Taiwan and South Korea. Competitive response: SK Hynix and Samsung have sunk billions into HBM4 and won't pivot overnight, but expect defensive ferroelectric IP acquisitions or internal R&D mandates within 2025–2026. Micron, positioned to leverage a "domestic memory" narrative, is the most likely strategic partner. The 12–24 month inflection point isn't the 2027 product launch — it's the Q1 2026 yield-validation data. Once third-party benchmarks confirm even half the claimed bandwidth-per-watt advantage, HBM3E pricing will face its first genuine technology-substitution discount, and AI server BOM costs could compress by 5–8%.