Nvidia rejects near-term memory glut fears, leaves CXMT cooperation open

Jensen Huang's GTC Berlin remarks are less a supply forecast and more a strategic pressure test on the DRAM oligopoly. The real signal: Nvidia is telegraphing to Samsung and SK Hynix that buyer-side leverage is shifting. CXMT's DDR5 yield is approaching commercial viability, yet the TSV stacking gap for HBM3E remains 18 to 24 months behind the Korean incumbents—that is the actual moat, not node size. Under the current export-control architecture, "openness to cooperation" is diplomatic hedging, not a procurement roadmap. What matters is the marginal erosion of seller pricing power: when the world's largest AI accelerator buyer hints at a Plan B, HBM4 contract negotiations lose their one-sidedness. Over the next 12 to 24 months, conventional DRAM will likely digest a capacity overhang, while AI-driven HBM demand fractures the market into two parallel regimes—tight at the top, glutted below. If CXMT clears HBM2E mass production by 2026, the global DRAM pricing center of gravity shifts structurally. The Korean duopoly's three-decade pricing discipline faces its first credible asymmetric threat, and the bargaining table for next-generation memory contracts will never look the same.