中国AI的全球背景——我们月度产品的推出

Equity ResearchNovember 11, 2025 China (PRC) | Technology China AI in a Global Context – Inauguration ofOur Monthly Product In our China AI Monthly, we analyze China's AI progress from models toapps, semi needs, power demand, capex, and economics, and put China'sprogress in a global context. Our first issue includes: 1) an upward revisionin our China AI capex forecasts, driven by higher 7nm yield partly offset byhalt in NVDA chips, 2) bottom-up token consumption forecasts, and 3) adetailed look at China's model efficiency. Our new forecasts indicate no chipconstraint to 2030E. China offers much more "efficient" AI investment than the US.Chinese hyperscalers'combined capex in 2023A/24A/25E of US$124bn is 82% lower than that of its US peers (US$694bn), but the performance of China's best AI model (MiniMax M2) is only 10% behind thatof the best US AI model (GPT5)(source: Artificial Analysis). Although US AI models still leadin overall performance, for open-source models, China models are already world-leading inperformance. Chinese AI models are likely to catch up with the US because many Chineseplayers could conduct pre-training outside China, without chip restrictions. This also meansChina's low capex is due to less FOMO/competition and higher model efficiency, rather thanlimited access to expensive advanced AI chips. China pushes for model efficiency, while the US focuses on model performance/AGI.UShyperscalers and AI players (eg, OpenAI) continue to raise capex as they focus on pushingmodel performance to reach AGI. But this push goes against the evidence that scaling law isslowing significantly, implying a rapid decline in potential ROI. In contrast, Chinese AI playersfocus on: 1) open-source models, and 2) improving model efficiency, to cope with hardwareconstraints, driving down token cost and thus, API pricing. China's AI API pricing is the lowestglobally; DeepSeek recently lowered its API pricing by 63%, driven by higher efficiency. Low AIpricing in China could more easily drive application development and user adoption. Even ifChina's inference is supported by sub-optimal semiconductors, we see better ROI potential inChina's AI. We raise 2025E-30E China cumulative AI capex by 8% to US$884bn (vs est US capex of US$600bn just in 2026).Our revised capex is based on:1) removing previous forecasts for importsof downgraded NVDA chips (banned by the gov), and 2) raising our 7nm yield assumptions forChina, from 7%-20% (2025E-30E) to 15%-35%.Higher yield is based on our understanding thatHuawei has reduced the number of cores in its GPU design to increase the area of parallelcomputing, making the chip usable even with some defects. This is an acceptable practiceeven at TSMC with small defects of GPU. Our updated forecasts imply cuts in AI capex for2025E/26E, but raise for 2027E-2030E. By 2030E, we forecast China will produce 4.1m GPUs,up from ~2.8m previously. Our energy consumption forecast (2025E-30E) has risen by only1% to 29.2GW. Our first attempt on a bottom-up token consumption modelgenerates 140x growth between2024 and 2030E, but only 10x growth in compute demand given efficiency improvement.China's daily token consumption rose 100x between Mar 2024 and Mar 2025, and another 3x to30tn in 3 months as of Jun 2025. China's latest daily token consumption is similar to GOOGL'sglobal dailly token consumption of ~33tn, which doubled between May and July this year, andgrew 49x in the prior year.Our chip-supply-driven compute availability forecast exceeds tokenconsumption demand by 119% in 2030E. This means an extra 6,000 EFLOPs would be availablefor training and/or more inferencing. In other words, chip shortage is no longer a big risk.We willcontinue to update this model with actual data, investor feedback and expert inputs. Edison Lee, CFA * | Equity Analyst852 3743 8009 | edison.lee@jefferies.com Matt Ma * | Equity Analyst852 3767 1109 | matt.ma@jefferies.com Nick Cheng * | Equity Analyst+852 3743 8750 | nick.cheng@jefferies.com Annie Ping, CFA, FRM * | Equity Associate+852 3767 1273 | annie.ping@jefferies.com Jacky He * | Equity Analyst+852 3743 8084 | jacky.he@jefferies.com The 3-Minute Guide Chinese Hyperscalers’ 2023–25 Capex is 82% Below US Peers', yet Chinese Top AI Model Performs Only 10% Lower Than US Leader Although ChatGPT-5 is Still Leading the Pack, its Lead has Dramatically Diminished with Intensifying Competition OpenAI, xAI, Anthropic, Google, BABA and DeepSeek are Top AI Labs; but MiniMax Performance Recently Surpassed that of BABA & DeepSeek .Source: Artificial Analysis, Jefferies We Lower Our AI Capex Forecasts for 2025E–26E but Raise for 2027E–30E. AI Capex to Increase by a 19% CAGR in 2025E-30E Chinese AI Labs' Focus on Efficiency, such as Architecture Innovation, has Significantly Driven Down Token Prices Both China and Google (as a benchmark for US AI) Show Accelerating AI Token Consumption Growth We Forecast China’s Inference Token Usage to Expand 140x to 184,500t