DeepSeek and Huawei Target a Key Source of Nvidia’s A.I. Dominance - nytimes.com

Nvidia's true moat was never transistor density—it's the cognitive lock-in of CUDA across the global AI developer base. The DeepSeek-Huawei pairing strikes precisely at that vulnerability: DeepSeek compresses inference compute to roughly a third via MoE architecture and aggressive quantization, while Huawei's Ascend stack provides a credible non-CUDA hardware substrate. The assumption that H100-class silicon is mandatory for frontier models is fracturing. Ripple effects are already visible. HBM procurement is shifting toward inference-optimized chips, TSMC's CoWoS allocation faces a restructured customer mix, and AMD's MI300X alongside Intel's Gaudi 3 are positioned to capture liberated mid-market share. Paradoxically, BIS export controls' ceiling effect is being neutralized by model-efficiency breakthroughs—when equivalent tasks require half the compute, process-node gaps cease to be absolute barriers. Within 18 months, the inference segment will complete its first round of de-Nvidia validation. Nvidia's response won't be price cuts; it will be accelerated Blackwell iterations and tighter NVLink ecosystem closure. But the window is open: the emerging model-chip vertical integration paradigm is rewriting who holds pricing power in AI infrastructure.