China AI Model Usage Surges as Cloud Capex Enters New Expansion Cycle
nashnova research
BofA's August tracker shows Chinese AI models closing in on US leaders in both usage and intelligence scores, with consensus capex forecasts for Tencent and Alibaba raised 25% and 29% respectively — cloud infrastructure spending is entering a new upcycle.
How close are Chinese AI models to the US frontier?
On Artificial Analysis's intelligence index, Anthropic's Claude Opus 5 leads at 63 points. Moonshot AI's Kimi K3 scores 60, ranking fifth globally — 95% of Opus 5's score.
Zhipu's GLM-5.3 also scores 60, tied sixth. On the Agentic Index — measuring a model's ability to complete multi-step tasks autonomously — GLM-5.3 ties Opus 5 for first place, both at 59.
This means → the China-US gap in raw model intelligence has narrowed to single-digit percentages, and in agentic capability China has already caught up.
Who is actually being used the most?
OpenRouter data through August 17 shows DeepSeek V4 Flash at 31.6 trillion tokens, Tencent's Hy3 at 26.2 trillion, and Xiaomi's MiMo-V2.5 at 19.1 trillion — the top three are all Chinese models.
By weekly token share, DeepSeek leads at 22%, up 4.7 percentage points from July. On Vercel, DeepSeek holds a 29.7% monthly share versus Anthropic's 24.7%.
Yet Anthropic still commands 64.8% of spending. In plain terms = Chinese models attract the most users; US models collect the most revenue — the pricing gap is the key.
How much cheaper — and why does it matter?
Claude Opus 5 charges $25 per million output tokens; GPT-5.6 Sol charges $20. The priciest Chinese model, Kimi K3, charges $15; DeepSeek V4.0 charges just $2 — less than a tenth of Opus 5.
On a per-task basis: Claude Fable 5 costs $3.14 per unit of intelligence-index task, Qwen 3.8 Max costs $0.91, and GLM-5.3 costs $0.68.
This means → similar intelligence scores at a fraction of the price. For enterprise customers, running the same workload on a Chinese model can cost one-quarter to one-fifth of the US equivalent — that is the fundamental driver behind the usage surge.
How much are cloud players spending?
Over the past 30 days, consensus 12-month capex forecasts for Tencent and Alibaba rose 25% and 29% respectively. Tencent's Q2 cloud revenue grew over 20% year-on-year; Alibaba's cloud revenue grew 45%.
BofA estimates combined capex for China's top four cloud platforms will rise 98% in 2026 to roughly RMB 726 billion, then another 44% in 2027 to exceed RMB 1 trillion.
This reflects a structural bet: cloud operators see current usage growth not as a one-off spike but as demand worth committing to for at least two years.
How big can the MaaS opportunity get?
BofA projects China's total cloud market will grow from RMB 389 billion in 2025 to RMB 1.729 trillion in 2030, a ~38% CAGR.
Within that, AI MaaS — model-as-a-service, where businesses call large models via API instead of training their own — is forecast to grow from RMB 21 billion to RMB 676 billion, an 81% CAGR. AI infrastructure cloud is projected at 51% CAGR, from RMB 45 billion to RMB 470 billion.
In plain terms = MaaS is "pay-per-call for someone else's model," growing far faster than the build-your-own segment. This means → the biggest revenue stream for cloud players will not be selling compute, but selling model access itself.
Who leads on the consumer side — and what comes next?
ByteDance's Doubao (豆包) hit 171.3 million weekly active users in the first week of August with 11.5 billion minutes of usage — a dominant lead. DeepSeek ranked second at 30.7 million WAU; Alibaba's Qwen third at 28 million.
Key dates ahead: Alibaba may unveil a new model at its Apsara Conference (Sept 22–24); MiniMax is expected to launch M3.1 and M3 Pro in September–October; DeepSeek V5 and ByteDance's upgraded Seed series are anticipated in H2 2026.
This means → the model-upgrade cadence will not slow, but the real test is whether capex expectations convert into actual revenue growth — that is the core metric for judging the durability of this upcycle.
市场有风险，内容仅供研究参考，不构成投资建议。