Huawei boss claims homegrown AI chip sales top Nvidia in China

Huawei said its Ascend AI accelerators have surpassed Nvidia in Chinese market share, as prior US restrictions on Nvidia's H20 exports prompted Beijing to push domestic data-center operators toward local chips. Nvidia has resumed limited H200 shipments to approved Chinese customers, but China revenue remains immaterial, with H200 sales contributing less than 1% of its data-center revenue in Q2. Huawei is deploying a 256,000-card Atlas 950 SuperCluster and targeting systems scaling to 1 million NPUs, while DeepSeek software tools and evolving programming models are reducing the CUDA ecosystem barrier.

Analysis

The investable issue is not near-term lost China revenue alone; it is the erosion of Nvidia’s platform lock-in inside the world’s second-largest AI-compute market. A domestically supported hardware/software stack can compound through developer tooling, model optimization and procurement mandates, reducing the probability that reopened export channels restore prior Nvidia share. This is a 6-18 month multiple risk for NVDA if investors continue to value China as a recoverable upside market rather than a structurally impaired one.

The most important uncertainty is whether domestic systems deliver usable cluster-level economics: availability, interconnect performance, power efficiency, yields and software productivity matter more than peak-chip benchmarks. Huawei’s market-share assertion is not independently auditable, so the next verification points are Chinese cloud-capex disclosures, AI-lab migration announcements, domestic foundry utilization and Nvidia’s China data-center commentary. A rapid increase in approved H200 volumes, or evidence that customers retain Nvidia for frontier training workloads, would weaken the displacement thesis within 1-3 quarters.

Contrarianly, this is unlikely to impair Nvidia’s consolidated earnings immediately given the currently minimal reported China contribution, and a China-specific competitor does not automatically translate into a global competitor. The larger second-order beneficiary may be Broadcom: hyperscalers outside China seeking supply-chain diversification can accelerate custom ASIC programs, while China’s stack becomes increasingly separated from the CUDA-centric global ecosystem. The relevant relative-value question is therefore NVDA’s China-optionality premium versus AVGO’s less export-control-sensitive AI infrastructure exposure.

Key Decisions for Investors

• Maintain an underweight in NVDA into the next earnings cycle; use any export-license-driven rally to initiate a 3-6 month NVDA put spread rather than an outright short. Thesis is multiple compression from permanently lower China optionality, not an imminent EPS collapse; exit if management demonstrates China data-center revenue recovering above 5% of segment sales with stable gross margin.
• Initiate a 6-12 month pair trade: long AVGO / short NVDA in equal beta-adjusted dollars. AVGO offers AI exposure through custom silicon and networking with lower dependence on reversible China licensing policy; target 10-15% relative outperformance, with a stop if Nvidia raises forward data-center guidance by more than 10% without corresponding AVGO order-strength.
• Do not establish a direct Huawei-supply-chain trade without independently verified evidence of domestic accelerator shipment volumes, foundry yields and HBM availability. Put alerts on SMIC utilization, Chinese cloud-provider capex guidance and disclosed domestic AI-cluster deployments; those data would determine whether this is a procurement shift or a durable supply-chain buildout.
• For existing NVDA longs, reduce position size ahead of any Chinese regulatory or procurement directive over the next 1-3 months. The asymmetric downside is a policy-led customer migration that arrives before sell-side estimates remove China recovery assumptions; the upside catalyst is a material license approval expansion or demonstrable H200 deployment at Chinese frontier labs.