China weighs allowing ByteDance, Alibaba to buy new Nvidia chips, The Information reports - WTVB

This report signals a potential recalibration in the US-China semiconductor trade landscape. Rather than a wholesale easing of export controls, the reference to chips built for "high-end professional ... Read original →

This is not a reopening—it is the institutionalization of tiered control. The "high-end professional computer" framing precisely targets workstation-class GPUs (RTX 6000 Ada, Blackwell workstation variants), not H100/B200 training accelerators. China's AI compute stack is bifurcating into a dual-track architecture: large-scale pretraining on Ascend 910C and Cambricon Siyuan 590, while fine-tuning, inference, and R&D prototyping still route through Nvidia. This dilutes the urgency of full-stack domestic replacement on the inference side, yet paradoxically extends the window for training-side substitution—the most compute-intensive layer remains gated. Naming ByteDance and Alibaba specifically is a revocable administrative license, not market access. No resale, no secondary market, per-transaction approval. Nvidia's compliance overhead compounds: isolated supply-chain tracking, end-user verification, usage auditing. The 2023 A800/H800 special-edition precedent shows specs can be tightened at any regulatory cycle. AMD's MI300X, already marginal in China, loses further ground. Huawei Ascend faces short-term pressure, but its irreplaceability in training workloads secures a three-to-five-year ecosystem lock-in. Twelve-to-twenty-four-month outlook: decoupling shifts from cliff-edge severance to gradient management. Nvidia's China datacenter revenue stabilizes at five to eight percent of the segment (previously roughly twenty percent), becoming compliance revenue rather than a growth engine. China's compute landscape enters a "domestic training, mixed inference" equilibrium until Ascend and Cambricon achieve advanced-node mass production around 2027.