China has started cautiously reopening access to Nvidia’s advanced AI processors, granting approvals to a select group of its largest technology companies. These companies include ByteDance, Alibaba, and Tencent, which are anticipated to collectively purchase over 400, 000 of Nvidia’s H200 accelerators, according to Reuters. Additional technology firms might seek approvals in future rounds. [ Related: More Nvidia news and insights ] This move by Beijing indicates a shift as it attempts to balance immediate AI computing demands with its long-term goal of developing a domestic semiconductor ecosystem. “This unlocks a predictable supply of high-performance accelerators, allowing large Chinese enterprises to plan for denser GPU clusters, higher-bandwidth fabrics, and faster model-training roadmaps, ” said Charlie Dai, VP and principal analyst at Forrester. “Hyperscalers can further accelerate scaling AI infrastructure pipelines and optimize for multi-petaflop training capacity, rather than relying solely on inconsistent domestic alternatives. ” Impact on AI workloads These approvals might compel Chinese hyperscalers to reconsider the distribution of AI workloads. “We expect leading Chinese enterprises to adopt a dual-track deployment strategy: prioritizing H200s for core large-scale model training workloads, while mainly using domestic chips for inference and smaller-scale training tasks, ” said Galen Zeng, senior research manager for semiconductor research at IDC Asia/Pacific. China is also intensifying efforts to boost domestic training chip design and manufacturing capabilities, aiming to reduce long-term reliance on foreign suppliers, Zeng added. Complexities could arise if authorities require imported chips to be used alongside domestically produced accelerators.

Reuters has reported that such a mandate is possible. “A mandated bundling requirement would create a heterogeneous computing environment, significantly increasing system complexity, ” Zeng explained. “Performance inconsistencies and communication protocol mismatches across different chip architectures would raise operational and maintenance overhead and add network latency. ” Nevertheless, the approvals are unlikely to narrow the gap with US hyperscalers, Zeng noted, pointing out that the H200 lags one generation behind Nvidia’s Blackwell architecture and approved volumes are far below China’s total demand. Implications for global enterprises For global enterprise IT and network leaders, this development adds another factor to long-term AI infrastructure planning. Increased sales of Nvidia’s H200 chips could help the company scale up production, potentially allowing price reductions for Western enterprises deploying H200-based AI infrastructure, said Neil Shah, VP for research at Counterpoint Research.