DeepSeek opens tools to help Huawei chips supplant Nvidia in AI - South China Morning Post

DeepSeek's tooling release is not a "Huawei assist"—it is a surgical strike on the software layer that has been Nvidia's true moat since 2007. The silicon was never the barrier; CUDA was. By open-sourcing the translation and optimization stack, the single most expensive entry cost in AI infrastructure is being systematically dismantled. Technical cascade: Ascend's CANN framework gains external validation for the first time. HBM supply chains bifurcate—Huawei's in-house roadmap versus SK Hynix's Nvidia-dedicated capacity. Advanced packaging (CoWoS alternatives) becomes the new chokepoint. Compliance: BIS rules constrain hardware exports, but tooling hosted on domestic Chinese cloud infrastructure effectively sidesteps the chip ban at the application layer. This is more structurally damaging to US policy than any single restricted die. Market response: Expect Nvidia to accelerate Blackwell shipments and tighten CUDA licensing exclusivity. AMD's ROCm gains a credible benchmark. The real casualties are mid-tier cloud providers who hedged across both ecosystems. 12–24 month trajectory: A bifurcated global AI stack becomes the base case. China's training workloads default to Ascend+DeepSeek. Nvidia's China revenue (~20% of total) faces irreversible erosion. The "CUDA tax" underpinning its 50%+ gross margins compresses globally as the alternative matures.