Banned Nvidia GPUs spotted powering telecom, cloud workloads in China: report - SDxCentral

Core judgment: This is not an isolated compliance lapse—it is a structural arbitrage window being systematically exploited. Technical chain: Telecom and cloud represent the highest compute-density workloads. Banned chips still in production means operators are paying a performance premium to offset compliance costs. The FP16 throughput and NVLink bandwidth of A100/H100 remain irreplaceable by compliant variants, effectively locking in CUDA migration costs and stretching replacement cycles from quarters to years. The entire stack, from HBM supply to inference optimization, is servicing this gray inventory. Compliance risk: Nvidia's China-specific SKU strategy is product fragmentation in exchange for market access, but audit costs are eroding margins. More critically, every gray-market exposure hands Washington ammunition for tighter secondary sanctions. The 2018 entity list and October 2022 controls follow a loose-tight-tighter trajectory with no reversal. Market dynamics: AMD's MI300 is filling the compliant-but-high-performance niche, yet the CUDA moat remains intact. Huawei's Ascend deployment experience in telecom is the real long-term threat. Paradoxically, the gray market delays forced domestic substitution—both a buffer and a poison for Huawei and Cambricon. Outlook: Within 12-24 months, US enforcement will pivot from chip-level controls to end-user verification and re-export tracking. Domestic GPU penetration in China's critical infrastructure will cross a tipping point by 2026. The global GPU market will complete its irreversible split from one market with two rulebooks to two markets with two ecosystems.