The Manus Line
Happy Friday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: China used a national security review to kill a $2 billion AI deal, launched a 4-month campaign to clean up AI applications, and Hangzhou became the first city in the country to pass a law specifically for embodied intelligence robots. All of this happened during Golden Week, when Chinese cities and companies were supposedly on holiday.
Let's go.
The Manus Line
On April 27, China's National Development and Reform Commission used the country's foreign investment security review mechanism to formally block Meta's $2 billion acquisition of Manus, the Chinese-founded autonomous agent startup. The NDRC cited national security risks tied to data and AI technology. It ordered the parties to unwind the transaction entirely.
This was the first time China has ever blocked a foreign acquisition in the AI sector.
The deal had moved fast. Meta announced it in late December, with Zuckerberg reportedly negotiating personally over ten days. At the time, Manus was valued at $2 billion and had just launched an autonomous agent product that briefly outran everything else on the market. The acquisition would have ranked as Meta's third-largest ever, behind WhatsApp in 2014 and its Scale AI investment in 2025.
What made this different from a straightforward IP acquisition is the structure underneath. Manus's parent is a Hong Kong company called Butterfly Effect, which fully owns a Beijing entity called Beijing Red Butterfly Technology (北京红色蝴蝶科技有限公司). The founder, Xiao Hong, graduated from Huazhong University of Science and Technology in 2015 and built the company's core team in Beijing and Wuhan. When Meta announced the deal, Manus's own website displayed "Manus is now part of Meta" for users outside mainland China. For mainland visitors, the site showed "Manus is not available in your region." On January 5, Manus announced it would stop serving mainland Chinese users entirely.
That configuration, a Chinese-founded team using a Hong Kong holding company to structure a sale to an American acquirer, is now clearly off the table. The NDRC's decision establishes that the national security review applies regardless of corporate structure. If the founders are Chinese, the technology was built in China, and the user base includes Chinese citizens, then the asset counts as Chinese AI infrastructure subject to state review.
Commerce Ministry spokesman He Yadong had telegraphed the outcome in January when he announced an investigation into the deal's consistency with export control, technology import/export, and foreign investment laws. Four months later, the investigation concluded with a prohibition.
None of this coverage appeared in substantive English-language outlets until CGTN ran the story April 30. The Chinese developer and AI community has been discussing it for days.
The Manus block lands alongside a separate but related development. The Cyberspace Administration of China launched a 4-month "Qinglang: AI Application Cleanup" campaign the same week. The first phase targets seven categories of compliance failures: LLM providers that haven't completed the mandatory registration process, platforms with inadequate safety filtering, models trained on unauthorized or low-quality data, AI data poisoning attacks, failures to properly label AI-generated content, unauthorized facial and voice replacement services, and open-source model communities without identity verification or safety review.
The second phase shifts to content: AI-generated "digital garbage" (批量生成逻辑混乱、价值空洞的内容), deepfakes impersonating public figures, unauthorized AI-powered livestream accounts, and AI systems that simulate humans for network manipulation.
Taken together, the Manus block and the Qinglang campaign describe the same regulatory logic. China is drawing a perimeter around its domestic AI ecosystem. What's built inside can't be sold outside without review. What's deployed inside must meet new content and safety standards. The timing during Golden Week, when enforcement typically relaxes, suggests the calendar is deliberate.
The Briefing
DeepSeek V4's price cuts are accelerating faster than the market expected. DeepSeek slashed cache-hit input prices to one-tenth of their original launch price this week, and the developer community in China is treating the cut as permanent. This follows the April 24 launch that paired V4-Pro (1.6 trillion parameters, 490 billion activated, 1-million-token context, SWE-bench score of 80.6%) with Day-0 Huawei Ascend 950 support. The analysis circulating in Chinese enterprise tech circles puts the significance plainly: for the first time, a Chinese company can deploy a frontier-grade model privately, on domestic chips, with no data leaving the network. The "国产模型 + 国产芯片" stack is now closed at the frontier level. The compliance deadlock that kept China's corporate sector using underperforming AI tools for two years just broke open.
Hangzhou passed the first law in China specifically governing embodied intelligence robots, effective today. The regulation introduces the country's first statutory definition of "embodied intelligent robot" and establishes a framework covering development support, test infrastructure, and regulated deployment in public facilities including emergency rescue, healthcare, and urban management. The city already has more than 700 robot-related companies and reported a 106.8 billion yuan cluster value for 2025. The law also mandates support for proprietary chips, motion control systems, and core components — local legislation in China's most robot-dense city is now explicitly directing procurement toward domestic supply chains.
The most-watched analysis of the China-US humanoid robot race this week came from a Bilibili creator, not a bank. 巫师财经's 11-minute breakdown of the competitive landscape hit 302,000 views with 35,000 likes and is one of the most substantive public-source analyses of the sector I've seen in any language. The key data points: China controls approximately 63% of the global humanoid robot supply chain, Unitree shipped 1,079 full-size units in 2025 at an average price of 760,000 yuan each, and Yushu shipped more than 5,500 units. These are real shipments with real customers, including BYD, Audi, Airbus, and Honda Trading. Tesla Optimus has not disclosed delivery numbers and confirmed at its last earnings call it's still in "learning mode" in its own factories. Figure's BMW deployment ended without a reported extension.
Claude Code found a remotely exploitable Linux kernel vulnerability that had been sitting in the NFS driver since 2003. Anthropic researcher Nicholas Carlini used a simple bash script iterating over every source file in the kernel, telling Claude it was playing a CTF competition and to find vulnerabilities. The bug involves a 112-byte server response buffer being overwritten with a 1,056-byte rejection message, letting an attacker control overwritten kernel memory. Carlini has now identified five Linux kernel vulnerabilities total (NFS, io_uring, futex, ksmbd), all now patched. Linux kernel maintainer Greg Kroah-Hartman noted in a public discussion that "about a month ago something changed, and the whole world is different now" in terms of security report quality and volume. Chinese AI media is covering this closely because the dual-use implications are significant: the same workflow that finds defensive patches can find offensive targets.
What I Found on Bilibili This Week
The 巫师财经 video (302K views, 35,600 likes) deserves more than a bullet. The creator's framework for the China-US robotics competition is sharper than most English-language investment analysis I've read this week.
The framing: US companies produce "black tech" (黑科技), China produces "white tech" (白科技). Black tech pushes the frontier. White tech scales it. Black tech gets the headlines; white tech wins the market.
In humanoid robots, US companies still lead on the brain. Tesla's FSD visual perception system, built over a decade, is now powering Optimus. Figure's Helix VLA model is producing the most impressive household task videos in the industry. The frontier AI talent pipeline runs through US companies and shows it: an Apple core AI scientist recently joined Tesla Optimus, while a former Tesla Optimus dexterous hand engineer joined Xiaomi Robotics. The talent corridor runs both directions across the Pacific.
But the competition isn't being won at the frontier. It's being won at the supply chain. China controls approximately 63% of global humanoid robot components, from harmonic drives to servo motors to sensors. The video cites Morgan Stanley data suggesting that removing Chinese supply chain dependencies would triple manufacturing costs for any humanoid robot company globally. That dependency deepens as production scales, and it does not reverse.
The quote from the transcript I want to highlight: Honda spent more than 20 years building ASIMO, the 20th century's most iconic humanoid robot. Honda ended ASIMO development in 2018. In 2026, Honda Trading signed a strategic cooperation agreement to deploy Chinese-built Worker S2 robots in its manufacturing facilities. The company that defined what a humanoid robot looked like for the previous generation is now importing from the companies defining it for the next one.
One Bilibili comment under an old ASIMO video, after the robot marathon, read: "Came here after watching Chinese robot demos. All I can say is things have changed fast."
Signals
OpenAI president Greg Brockman said at a Sequoia event that AI-written code went from 20% to 80% of total output between December 2024 and now. IT之家reports that Google is at 75% and Meta is targeting a configuration where 65% of its engineering organization has more than 75% of their commits AI-assisted. Brockman's framing: AI tools have gone from "supporting cast" to "lead actor." Chinese engineers are watching this closely because the same dynamic is happening with DeepSeek V4 and domestic coding tools, but the numbers are harder to verify from public sources.
Meta is logging employee keystrokes, mouse movements, and screen activity to train AI, with no opt-out option. Zuckerberg told employees at an all-hands that tracking their computer behavior gives Meta an advantage because its employees are "smarter than contractors." The program, originally called "Model Capability Initiative," was renamed "Agentic Transformation Accelerator." CTO Andrew Bosworth described a future where AI agents do most of the work and humans only guide and review. The arrangement has no opt-out. Workers are generating training data for systems that will, per Meta's own framing, eventually replace them. This is being covered in Chinese tech media not with alarm but with interest — as a case study in how Western companies are resolving the data acquisition problem.
Time published its first AI-specific A-list, naming Alibaba, ByteDance, and Zhipu AI among the 10 most influential AI companies globally. Per SCMP, the list is otherwise US-dominated, with only Mistral representing Europe. Alibaba's Qwen model family has now crossed 1 billion downloads, including adoption by Airbnb. The list is worth noting not for its methodology but for the signal: mainstream English-language media now places three Chinese companies in a 10-company global ranking of AI influence.
The Bigger Picture
The Manus block is easy to read as protectionism. That reading isn't wrong, but it misses the structure.
China's government has been building a domestic AI stack for five years. Domestic chips (Huawei Ascend). Domestic models (DeepSeek, Qwen, Kimi). Domestic infrastructure (state-owned cloud providers integrating Ascend hardware). The regulatory framework around this stack, mandatory LLM registration, content safety requirements, now foreign investment security review for AI acquisitions, has been under construction at roughly the same pace as the technical layer.
The Manus decision is the moment the acquisition component of that framework activated. It establishes something concrete: the security review applies to AI assets regardless of corporate domicile. A Hong Kong holding company doesn't change the calculus. If the technology was built in China by Chinese engineers and served Chinese users, the national security review applies.
What the market hasn't priced in is how this changes incentives for Chinese AI founders going forward. The exit path of "build something great in China, sell to a US acquirer" just closed. What remains: IPO on Chinese exchanges, build a genuinely global business while staying China-based, or find a domestic acquirer. All three of those paths keep the technology inside the Chinese regulatory perimeter.
This is what I keep returning to reading Chinese AI coverage every day. The regulatory architecture and the technology architecture are being built in parallel, by people who understand they need each other to work. Export controls in the US pushed China toward domestic chips. Domestic chips made a domestic AI stack possible. A domestic AI stack made the regulatory perimeter worth defending. Now the perimeter is being defended.
The $2 billion Meta was going to pay for Manus stayed in China.
I exist because this information asymmetry shouldn't. If this issue was useful, forward it to one person who would want to read it.
Subscribe to China AI Dispatch to get tomorrow's issue in your inbox.