Kimi K3 makes China's open-source AI the global benchmark

China

SIGNALS

 

In this issue

01|Kimi K3 makes China's open-source AI a global benchmark

02|China's DUV lithography still four generations behind ASML

03|CXMT closes on LPDDR6 mass production with 12,800 Mbps mobile DRAM

04|Unitree Robotics clears key hurdle for Shanghai STAR Market IPO

05|Xiaomi opens pre-orders for N90 Max SUV at ~€38,400

06|AI infrastructure software emerges as new lever against GPU crunch

Our Take

The world's largest open-source AI model just came out of Beijing — and crashed under the weight of its own success. Europe was not part of the decision.

Moonshot AI released Kimi K3, an open-weight model with 2.8 trillion parameters. Within two days, new consumer subscriptions were suspended. Overseas paid users quadrupled, and API revenue rose 400%. Chinese open-source models now account for 41% of global downloads; the top six most-used models all come from Chinese teams.

These are real numbers, but volume is not neutrality. Open-weight Chinese models run on Chinese chips, Chinese standards, and Chinese legal oversight. For a European manufacturer, 'free' API access means feeding proprietary data into a stack Beijing can regulate. And leading in downloads does not equal leading in frontier capability — closed US models still define the highest end.

The non-obvious implication: The real race is no longer model quality. It is who controls the default AI infrastructure in Europe. China is using open source to become the default choice for anyone who wants cheap AI without asking a US cloud provider. That is a strategic play for ecosystem lock-in.

European companies can and should use open-weight models to cut costs and gain leverage — but only through a European governance layer. Host on EU infrastructure, run audited and frozen versions, and refuse API dependencies on either US or Chinese providers. Policymakers should fund a European open-weight flagship and require transparency from every AI provider selling into the EU. Waiting for a European 'ChatGPT' is the wrong answer; building open-weight sovereignty is the right one.

Open weights from Beijing look free. The bill comes due in dependency.

Key Takeaway

Reports of China's plans to produce its own immersion DUV lithography machines have triggered a sharp sell-off in European chip-equipment stocks. Chinese pundits, however, say the machines remain four generations behind ASML and far from competitive.

Implication

European equipment makers should treat the sell-off as sentiment rather than substance. ASML's technological lead remains intact, but sustained state-backed Chinese R&D warrants monitoring.

Details

Key Takeaway

CXMT has nearly completed development validation for its first LPDDR6 memory and is past the halfway point toward commercialization. The company expects H1 2026 revenue of roughly €14-15 billion, up more than 600 percent year-on-year.

Implication

CXMT is building a credible third force in mobile DRAM alongside SK Hynix, Samsung and Micron. European device makers and suppliers gain potential supply-chain diversification, but should monitor qualification and technology risks.

Details

Key Takeaway

China's securities regulator approved Unitree Robotics' registration application for a STAR Market IPO on July 1. The Hangzhou-based robot maker seeks to raise about RMB4.2 billion (~€540 million), according to its prospectus.

Implication

The listing would give Unitree fresh capital for robot AI and manufacturing, intensifying competition in embodied AI. European robotics and automation players should track its product roadmap and use of proceeds.

Details

Key Takeaway

Xiaomi has opened pre-orders for the N90 Max, an extended-range SUV under its Pengcheng line, priced at RMB299,900 (~€38,400). The official launch is scheduled for September.

Implication

With a CLTC combined range of 1,705 kilometers and deep AI integration, the N90 Max raises the competitive bar in China's SUV market. European carmakers need to accelerate software and ecosystem strategies.

Details

Key Takeaway

As Google and OpenAI build custom AI chips, a fast-growing market is forming for software that squeezes more work out of existing GPUs. Baseten's valuation reached ~€11.1 billion, Fireworks ~€15 billion, and open-source inference engines vLLM and SGLang are commercializing with seed rounds above ~€85 million each.

Implication

European enterprises can cut AI costs through better GPU utilization instead of new hardware purchases. The emerging dependency on AI-infrastructure platforms, mostly US- or China-originated, is becoming a strategic factor.

Details

Your feedback

Did you learn something useful in this issue?

One click is enough — your answer is sent directly.