Why China’s AI Model Usage Surpassed the US: Key Trends and Market Shifts

1. Chinese Large Model Usage Leads Globally

According to OpenRouter data, weekly token usage reached 9.8 × 10^13 tokens, surpassing the US for the fifth consecutive week and pushing daily average above 1.5 × 10^14 tokens. The ranking shows Tencent Hunyuan 3.0 (3.2 trillion tokens, +45 % week‑over‑week), DeepSeek V4 (2.8 trillion, +33 %), Alibaba Qwen 3.6 Plus (2.1 trillion, –5 %), GPT‑4o (1.8 trillion, –12 %) and ByteDance Doubao (1.5 trillion, +20 %). The National Data Bureau projects the AI industry to exceed 8 trillion CNY by the end of 2026 and reach 15 trillion CNY by the end of the 15‑year plan.

2. Zhipu AutoGLM App Store Launch

Zhipu released the AutoGLM App Store with 800 applications. Developers earned more than 1 million CNY on the first day, reflecting a 20 %/80 % revenue split favoring developers. The top‑downloaded apps are AutoResearcher Pro (52 k downloads, automatic research reports), SmartCoder (48 k, full‑stack code generation), AI Interviewer (36 k, technical interview simulation), Contract Reviewer (31 k, legal risk identification) and Paper Assistant (29 k, academic writing aid).

3. US AI Giants Tighten Access for Mainland China

OpenAI, Anthropic and Google announced simultaneous restrictions on API access from mainland‑China IP addresses. New registrations using Chinese phone numbers or email are blocked; existing enterprise accounts must provide non‑Chinese entity proof within 30 days. Technical measures include stronger geo‑IP detection and behavioral analysis. The impact assessment estimates roughly 300 k Chinese developers will need to migrate to domestic APIs, driving a 50 %+ traffic surge for models such as DeepSeek, Hunyuan and Tongyi, and accelerating the “de‑US‑ification” of China’s AI ecosystem.

4. Musk’s xAI Funding Round

Musk confirmed that xAI is negotiating a new financing round targeting a 1 trillion‑USD valuation, a 25 % increase from the 800 billion‑USD valuation at the XAI Holdings merger. Potential investors include Saudi Arabia’s PIF (US$5 billion), UAE’s Mubadala (US$3 billion), Sequoia Capital (US$1 billion) and a16z (US$800 million). Planned use of funds: building the “Memphis Supercluster” with one million GPUs, developing Grok 4 (expected 2027 release), and integrating Tesla’s Optimus robot with xAI.

5. Meta Llama 4 Chinese Edition

Meta released a Chinese‑optimized Llama 4 edition in partnership with Zhipu. Chinese corpus coverage rose from 5 % to 35 %, and Chinese comprehension test scores improved from 68.5 % to 82.3 %. The model now supports Chinese programming conventions such as Python comments and variable naming. Revenue sharing follows a model‑weight‑provided‑by‑Meta, optimization‑by‑Zhipu arrangement.

6. Google Gemini 2.5 Pro Open Access

Google announced that Gemini 2.5 Pro is now fully open, doubling the free daily quota from 50 to 100 calls. New features include Gemini Live (real‑time voice dialogue with interruption and topic switching), Deep Research 2.0 (cross‑validation across 50+ sources) and Canvas (interactive coding environment with live preview). Pricing was cut by 33 %: input $0.002 per k token, output $0.008 per k token. The strategy aims to capture the Chinese developer market through a “free + low‑price” approach.

7. Huawei Pangu 5.0 for Industry

Huawei launched Pangu 5.0, an industrial‑focused large model already integrated into more than 300 factories. Use cases span automotive quality inspection (defect detection rate +40 %), steel‑making energy‑consumption prediction (‑15 % energy use), electronics yield prediction (+8 % yield) and textile design cycle reduction (‑60 %). The model is tightly coupled with Ascend chips, supports edge deployment, and achieves inference latency under 50 ms.

8. SenseTime “Daily New” 6.0 Medical Imaging

SenseTime’s “Daily New” 6.0 medical‑imaging module achieved >95 % accuracy across multiple tests. Specific disease accuracies: lung nodule detection 96.2 % (+3.5 % vs. experts), diabetic retinopathy 94.8 % (+2.8 %), breast cancer screening 95.5 % (+4.2 %), fracture recognition 97.1 % (+5.6 %). Deployments cover more than 50 top‑tier hospitals such as Peking Union Medical College Hospital and Ruijin Hospital.

9. China’s AI Chip Export Controls

The Ministry of Commerce added rare‑earth permanent‑magnet processing technology to the export control list, limiting shipments to the United States and certain countries. China supplies roughly 90 % of the rare‑earth processing capacity for high‑performance AI chips. Short‑term impact: US AI‑chip manufacturing costs rise 15‑20 %; medium‑term: the US accelerates rebuilding its rare‑earth supply chain (3‑5 years); long‑term: a global AI‑chip supply‑chain reshuffle.

10. Gartner Forecast on AI Agents

Gartner raised its estimate for AI‑agent substitution of knowledge work from 15 % to 20 % by the end of 2026. Predicted replacement rates by role include: Customer Service Rep 50 % (Q2 2026), Data Analyst 35 % (Q3 2026), Content Moderator 45 % (Q4 2026), Junior Programmer 30 % (Q1 2027), Legal Assistant 25 % (Q2 2027), Financial Analyst 20 % (Q2 2027). New roles projected: AI trainers (+800 k), AI‑ops engineers (+500 k), AI ethics reviewers (+200 k).

11. UN Global AI Governance Initiative

The United Nations released the “Global AI Governance Initiative”, incorporating China’s “development‑first, human‑centric, inclusive” principle into the core charter. Key provisions: mutual recognition of AI safety standards, a US$10 billion AI‑technology‑assistance fund (China pledges US$1 billion), ban on AI‑enabled mass surveillance and autonomous weapons, and safeguards for developing‑country AI rights. China will co‑chair the implementation committee alongside the EU and the US.