Goldman Sachs: China AI Compute > China’s national computing power network is part of a "six major networks" infrastructure project projected to attract Rmb7tn investments in 2026. Bloomberg reports data center investments will reach approximately Rmb2tn ($300bn) over the next 5 years. > Capital and technology are heavily shifting toward Western China hubs. Meanwhile, tier-1 city data centers are transitioning to focus on ultra-low latency hot compute, edge nodes, and AI inference. > Operational gigawatt (GW)-scale clusters with 100k+ chips remain scarce in China compared to the US. However, Range Intelligence successfully operationalized a 200MW, 100k chip-scale building year-to-date (YTD) as part of a larger GW-scale cluster layout. > A typical GW-level campus workload profile consists of 50%+ Inference, 20–30% Training, and 10–20% Full-stack R&D. > Domestic AI accelerator chip shipments are expected to exceed 50% market share in 2026, with Huawei (20%) and Alibaba's T-Head (7%) leading the domestic segment (though Nvidia still commands 55% overall). > Capital expenditure per IT power for domestic chips is 40–50% lower than imported chips. However, their performance lags significantly: capex per computing power is 2–4x higher, and computing power per IT power is only 10–30% of what imported chips achieve. > Huawei's 910B/910C servers produce an average daily token output volume that is just $1/6$ to $1/3$ of an Nvidia H800 server. Consequently, API profit margins based on Huawei 910B hardware heavily lag behind Nvidia counterparts.