Illustration: AI-generated · Inside China AI
China AI Weekly — September 26, 2026
Chinese models crossed a threshold on the platforms where developers actually work. Alibaba put a chip, a cloud and a model family on one stage. Washington and Beijing agreed to pick up the phone when something goes wrong with superintelligence — and agreed on very little else.
Models & Technology · Frontier Models & Platforms
■ Chinese models now carry the majority of tokens on two developer platforms
The story: Usage data from OpenRouter and Vercel shows Chinese models moving from a minority share to a majority in under a year.
- On OpenRouter, the Chinese-model share of tokens rose from 6–13 per cent in February to 57–67 per cent in the week of 14 September. On Vercel, it went from 11 per cent in January to 55 per cent in August (CNBC).
- Open Chinese models are quoted at 60 to 90 per cent cheaper than leading Anthropic and OpenAI models.
- Alibaba’s Qwen family has overtaken Meta’s Llama in cumulative Hugging Face downloads.
- Among businesses OpenRouter classes as Global South, 67 per cent of tokens run on Chinese models.
- The counterweight belongs in the same paragraph: US frontier models still attract more total spending. Tokens and revenue are not the same measure.
Why it matters for Europe
This is what our open-weights analysis predicted, now visible in usage data: permissive licensing buys reach, and reach compounds. For a European CTO the practical question has changed. It is no longer whether Chinese models are good enough — the developers answered that — but which licence the chosen model carries, and what that licence demands at ten times today’s volume.
→ Go deeper: full story
Companies & Startups · Chips & Compute
■ Alibaba puts chip, cloud and model on one stage
The story: At its Apsara conference in Hangzhou on 22 September, Alibaba presented a full domestic AI stack — its own accelerator, a data-centre build-out and the next Qwen generation.
- T-Head, Alibaba’s chip arm, unveiled the Zhenwu V900: 216 GB of memory, 1,200 GB/s inter-chip bandwidth, three times the performance of its predecessor, scaling to clusters of up to 500,000 chips. Mass production is set for the first quarter of 2027 (TechNode).
- Alibaba is targeting more than 20 GW of data-centre capacity by 2032, with roughly 380 billion renminbi — about 53 billion dollars — committed to AI infrastructure over three years (CNBC).
- A Qwen model with up to 10 trillion parameters sits on the published roadmap (Tom’s Hardware).
- The performance claims are the vendor’s own and have not been independently benchmarked. “China’s most powerful AI chip” is Alibaba’s description, not a measured result.
Why it matters for Europe
Export controls assume a chokepoint. A company that designs its own accelerator, operates its own clouds and trains its own models has fewer places to be squeezed — and every gigawatt built domestically narrows the gap further. Whether the V900 performs as claimed is unknown and matters enormously; what is already certain is the direction of the investment.
→ Go deeper: full story
Regulation & Policy · Other
■ Washington and Beijing agree a channel for superintelligence incidents
The story: Following the Trump–Xi summit, the two governments agreed to open a bilateral communication channel for incidents involving advanced AI systems.
- The White House confirmed the channel on 26 September. A newly formed US–China Superintelligence Dialogue is to convene by November (SCMP).
- Reporting describes the summit as otherwise short on breakthroughs; both sides confirmed two further meetings this year (The Japan Times).
- The stated positions differ more than the agreement suggests. Xi framed AI as a shared responsibility to keep development “always under human control”; Trump indicated a preference for leaving regulation where it stands.
Why it matters for Europe
Two governments that regulate AI in opposite directions have nonetheless built a wire for emergencies. That is a low bar and a real one — incident channels exist precisely because the parties do not trust each other. The European question is uncomfortable: the EU has the world’s most developed AI rulebook and no seat at this particular table. Rules without a hotline are not the same as influence.
→ Go deeper: full story
Models & Technology · Other
■ China passes the US as workplace for elite AI researchers — on one measure
The story: A Carnegie China study finds China now hosts the largest share of top-tier AI researchers, reversing the position of three years ago.
- China’s share rose from 27.1 per cent in 2022 to 40.6 per cent in 2025; the US share fell from 46.4 to 34.2 per cent (SCMP).
- Returnees from US institutions include two startup founders, two technology executives and five professors, drawn by compensation and compute access.
- The qualification the headline omits: measured by publications at top conferences, US institutions still employ 59 per cent of the world’s elite AI researchers — a lead built almost entirely on foreign-born talent (Al Jazeera).
Why it matters for Europe
Both readings are defensible because they count different things — where researchers were trained versus where they now publish. What neither disputes is that America’s position rests on people who were born elsewhere, and that a growing number of them now have somewhere else to go. Europe appears in this data mainly as a source of talent, not a destination. That is the number to watch, and it is not improving.
→ Go deeper: full story
Companies & Startups · AI Video & Media
■ AI video becomes industrial policy
The story: Chinese cities are competing to attract AI film and short-drama production with subsidies, compute vouchers and dedicated clusters — the same playbook used for electric vehicles, solar and robotics.
- In the first half of 2026 the cost of producing AI short dramas fell from 5,000 yuan per minute to a few hundred, according to state broadcaster CCTV (Reuters).
- Shanghai introduced measures in May offering compute and cloud models to micro-drama producers; Beijing runs a 260-million-yuan audiovisual technology fund, and its Huairou film district issues vouchers against compute costs.
- One filmmaker reports being courted by officials from cities and industrial parks across the country.
Why it matters for Europe
The interesting part is not the technology but the method: treating generative video as an industry to be built rather than a phenomenon to be regulated. Europe’s film and advertising sectors will meet the output of this policy in their own markets, at a cost base shaped by subsidised compute. The AI Act governs how such content must be labelled here; it says nothing about who can afford to produce it.
→ Go deeper: full story
⚡ Quick hits
- Huawei used its Connect 2026 conference to set out progress on AI chips, clusters and models, pressing the same full-stack argument as Alibaba.
- SemiAnalysis published a first structured model of China’s data-centre build-out — a rare attempt to measure the capacity rather than repeat the announcements.
- EU–China talks remain deadlocked on medical devices, with discussions continuing on plug-in hybrid curbs.
- German industry associations have called for a stricter national China policy, while European officials voice concern about having no seat at the Trump–Xi table.
- Chinese home-storage supplier AlphaESS is expanding across European residential energy markets — the quieter half of the energy-and-AI build-out.
→ Deep dives this week: Open Weights, Two Meanings and Ox Alpha Was Chinese.
All figures are linked to their sources. Vendor performance claims are identified as such and have not been independently verified. Where two studies measure the same question differently, we report both. Corrections are welcome: press@insidechinaai.com · Our editorial standards.
No responses yet