| I wrote this after Jensen Huang's Nov 7 statement that Nvidia now has "zero share" of China's AI datacenter market, down from 95%. The geopolitical split is creating three distinct compute ecosystems.
- China revenue impact: Dropped from 21.4% of Nvidia's total (FY2023) to 13.1% (FY2025), with ~$15B in annual datacenter revenue now offline - Huawei Ascend 910C performance: Delivers roughly 60% of H100 throughput for inference workloads, but training large-scale models remains challenging. The bigger gap is software CUDA's 15-year ecosystem vs Huawei's CANN framework - Google TPU positioning: While Nvidia Blackwell has 12-18 month lead times and China is locked out, Google's TPU becomes more attractive for specific workloads (Transformer training, inference at scale) especially for GCP customers - China's 30% mandate: State-funded datacenter projects under 30% completion must remove all foreign AI chips or lose funding (Reuters reporting from this week) The article traces the export control timeline from Oct 2022 initial rules through Oct 2023 tightening, with all revenue figures verified against Nvidia's public financials and multiple sources. Three parallel tracks are forming 1. Nvidia Blackwell (premium, supply-constrained, no China) 2. Huawei Ascend (domestic alternative, 60% H100 performance, ecosystem gap) 3. Google TPU (available hedge for workload-specific optimization) Happy to discuss the geopolitics, technical performance comparisons, or supply chain implications. Also interested if anyone has hands-on experience with Ascend 910C vs H100 workloads. Sources include Reuters, BIS.gov export control documents, Nvidia's FY2025 financials, and verified tech specs from multiple outlets. |