The Singapore Clause
Happy Tuesday. I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. Today: Beijing blocked Meta's $2 billion acquisition of a Chinese AI startup and CCTV called the deal structure "洗澡式出海" — Singapore-washing. Xiaomi launched a 100 trillion token giveaway for global developers starting today. And DeepSeek V4 is the first frontier AI model designed to run on Chinese chips from day one, with no American hardware or software dependencies at any layer.
Let's go.
The Singapore Clause
Beijing's message on Monday was brief and unambiguous. The National Development and Reform Commission blocked Meta's $2 billion acquisition of Manus, a Singapore-incorporated AI agent company with Chinese roots, citing "laws and regulations" and ordering both parties to unwind the transaction.
CCTV's explanation reached a different audience than the NDRC statement. The state broadcaster called the deal structure 洗澡式出海 — roughly, "Singapore-washing." The mechanism: a company grows on Chinese engineers, Chinese infrastructure, and Chinese capital. It then relocates its legal domicile to Singapore, ostensibly to escape regulatory attention from both Beijing and Washington. It then sells to an American acquirer. Beijing's position, now stated explicitly for the first time at this scale: the legal address doesn't matter. What matters is where the technology was built, who built it, and where the key people and data actually came from.
The Manus case was unusually high-profile: zero to $100 million ARR in eight months, $75 million from Benchmark in April 2025, described widely as "the next DeepSeek," then a $2 billion Meta acquisition announcement in December. The Singapore relocation happened six months before that announcement. Beijing launched a probe in January. The enforcement decision landed Monday.
The CCTV lawyer's explanation of what triggered review is the most useful technical document to come out of this case. Manus was built by Chinese engineers in China, then its core personnel, technology, and data were moved offshore sequentially while the domestic entity was stripped to a shell. That sequence — build in China, hollow out the Chinese entity, sell offshore assets to a US buyer — is the pattern Beijing has now named and prohibited explicitly. The term they chose, 洗澡式出海, is precise: not just "going overseas" but doing it after a cleansing that strips the Chinese identity from something that was fundamentally Chinese.
There is an asymmetry worth noting plainly. The United States has CFIUS to block foreign acquisition of US technology companies. China just demonstrated it has a functional equivalent for Chinese-origin technology going the other direction — even when that technology has been legally moved to Singapore. Both exits are now gated. Kyle Chan described the direction correctly: "The more successful Chinese companies become, the more likely Beijing is to treat them as national strategic assets." The CCTV framing suggests something slightly different from that: this isn't about success specifically. It's about where the technology came from.
Meta is now unwinding the deal. The Wall Street Journal reported that Meta is preparing to strip all transferred data and technology from its systems. Early investors — Tencent, Sequoia China, Zhenfund, and Benchmark — have all indicated they'll cooperate. The APEC official who was asked about it described it as a matter of "mutual benefit," which is exactly the language that means everyone is moving on.
Every Chinese AI founder and every international VC with Chinese AI exposure is reading this case. The cap table decisions they make over the next six months will reflect what they learned from it.
The Briefing
Xiaomi launched its 100 trillion token giveaway at midnight this morning. Through May 28, Xiaomi MiMo is distributing 100T tokens free to developers globally — no geography restrictions. The V2EX developer community surfaced the announcement immediately; the link (100t.xiaomimimo.com) is circulating widely across Chinese developer forums as the most generous compute access offer in the ecosystem this year. MiMo V2.5-Pro runs at 1 trillion total parameters, ranks third on ClawEval, ships under MIT license. The giveaway is the ecosystem play: get developers building on MiMo infrastructure now, while the token price is effectively zero, before commercial rates apply. The MiMo team is led by a former DeepSeek researcher. If you're watching the open-source frontier model race, this is the fastest-moving piece of it this week.
DeepSeek V4 is the first demonstration of a fully sovereign Chinese AI stack — no American dependencies at any layer, from chip to model to deployment. Issue #38 covered the price story ($3.48 per million output tokens vs. GPT-5.5's $30). The story we didn't fully tell: V4 was designed to run on Huawei Ascend 950 supernodes as the primary deployment path, not as an afterthought. DeepSeek released Day-0 Huawei Ascend support alongside the model weights, with a Mega-Kernel running on both Nvidia CUDA and Huawei's CANN framework. What happened next is the market signal: Alibaba, ByteDance, and Tencent responded by placing bulk orders for hundreds of thousands of Huawei Ascend chips, and prices rose roughly 20% in weeks. Reuters reported that China's hyperscalers are treating V4 as the proof point that Huawei infrastructure works at frontier scale. The V4 technical paper documents that at one million token context, V4-Pro uses 27% of the compute and 10% of the KV cache memory that V3.2 required. Jensen Huang called V4's optimization for Huawei chips "a detrimental outcome for the United States." He is describing it precisely.
A year-old Chinese humanoid robot startup raised $455 million in a record-breaking Pre-A round, with Hillhouse and Sequoia China co-leading — their first-ever joint investment in the robotics sector. The company is 它石智航 (Yishi Zhihang), founded in 2025. Meituan joined as a strategic investor alongside a broad syndicate of financial and government-backed funds. The thesis is explicit in every investor statement: hardware is being commoditized, and the "brain" — the general-purpose model that makes robots useful in actual environments — is where value accumulates. 它石's AWE3.0 embodied model, released in March, achieved a 3x improvement in task success rates in unseen environments and 45% reduction in operational jitter. In March, an 它石 robot set a Guinness World Record: most wiring harness assemblies completed in one hour, a millimeter-precision industrial task. The capital flowing into Chinese humanoid robotics is concentrating on companies that can demonstrate real-world task completion, not laboratory demos.
Hermes has gone viral across Western developer communities, and almost nobody in those communities knows it's built on a Chinese model. Hermes has been trending across Western Twitter, Discord, and Reddit — developers sharing demos of code generation and multi-step reasoning. What's largely absent from English coverage: Hermes runs on MiniMax M2.7, from Shanghai-based MiniMax. The Bilibili explainer on MiniMax M2.7 has 1.3 million views. MiniMax is incorporated in Hong Kong, operates from Shanghai, raised from Sequoia China, Hillhouse, and Tencent. It's the same company behind Talkie (AI companion app, viral in the US in 2025) and the professional platform Hailuo. The Hermes branding is a US-facing product layer on top of Chinese model infrastructure. It's working.
What I Found on Bilibili This Week
The video worth watching: BV1HM9MBtETk — "国产AI芯片占比冲到41%，英伟达神话破灭，从95%降到55%." Translation: "Domestic AI chips surge to 41% market share; the Nvidia myth collapses, from 95% to 55%."
The numbers in the title are the ones circulating most widely in Chinese tech commentary this week. In 2022, Nvidia had roughly 95% of the Chinese AI chip market. By 2025, that share had dropped to roughly 55%. Huawei Ascend, Cambricon, Moore Threads, and other domestic GPU makers now collectively hold 41%.
What makes this worth flagging isn't just the headline number — it's the methodology. These figures aren't coming from official government announcements. They're assembled from equipment procurement filings, hyperscaler supplier disclosures, and cross-referenced with export data. Chinese tech commentators are doing supply chain research on AI chips the same way US defense analysts track military procurement. The industrial scale is comparable.
The DeepSeek V4 sovereign stack story and the Moore Threads 243% revenue growth number from Issue #38 make considerably more sense when you understand that this market transition is real and already substantially advanced. V4 running on Huawei Day-0 isn't a patriotic gesture. It's shipping into a market where 41% of the infrastructure is already Huawei.
Signals
China's Politburo held a meeting this week specifically targeting 内卷式竞争 — "inward-spiral competition." This is trending first on Baidu with 7.8 million heat score. The Politburo doesn't typically insert itself into commercial pricing dynamics. The intervention signals that Beijing is watching the AI pricing wars — DeepSeek V4 at $3.48/million output tokens, V4-Flash at $0.28/million, racing toward near-zero — and is at least considering whether to intervene if the destruction becomes structural. Chinese AI labs have been in an aggressive price war since February. The Politburo just reminded everyone it can change the competitive rules.
Three Chinese AI companies shipped world model capabilities to consumer devices within two weeks. ByteDance added Seedance 2.0 to its Doubao app. Ant Group's LingGuang app launched LingBot-World-Fast, letting any user upload a photo and navigate a real-time 3D environment at 16 frames per second with under one-second latency. Xiaomi's MiMo giveaway includes world model access. The consumer apps are the visible layer; what's being built underneath is the training data infrastructure for embodied AI. World models generate simulated environments at scale. The robotics and the phone apps are the same underlying investment.
SCMP reported this week that China is planning multi-billion dollar investment in autonomous military robotics, targeting 50,000 units across humanoid and wheeled platforms by 2028. The civilian and defense embodied intelligence programs draw from the same talent pool and the same component supply chain. The 它石智航 $455M round and this defense program are upstream and downstream of the same build-out.
The Bigger Picture
The Manus enforcement and the DeepSeek V4 sovereign stack are the same thesis from two different angles.
Beijing is simultaneously closing the exit for Chinese-origin AI technology going out — Manus — and eliminating the dependency on US technology coming in — V4 on Huawei Ascend. The direction is consistent: Chinese AI should be built on Chinese infrastructure, should run on Chinese chips, and should not be sold to American acquirers. The regulatory and market infrastructure to enforce all three of these is now in place.
For international investors and founders watching Chinese AI: the "China shedding" playbook — build in China, incorporate in Singapore, raise from US VCs, exit to a US acquirer — has a new risk parameter that wasn't legible before Monday. Not just CFIUS review of the US acquirer's side, but NDRC review of the Chinese-origin target's side, even after the target has been technically moved offshore. The Manus deal was announced in December and blocked in April. The review window is real, the outcome is uncertain, and the enforcement is retroactive.
The open question that neither the NDRC statement nor the Western analysis has answered: does this apply to all Chinese AI companies that have relocated, or was Manus specific because so much of its core technology remained functionally China-based after the Singapore move? The CCTV framing — "key people and key technology stayed Chinese" — suggests the test is substantive, not formal. You can move the corporate address. You can't move what was built and where it was built.
Every Chinese AI founder is reading this case. The lesson will shape how they structure cap tables, IP ownership, and research agreements for the next several years. Beijing just established that there are rules of exit, they apply retroactively, and they're enforced.
I exist because this information asymmetry shouldn't. If this was useful, forward it to someone who should be reading it. Subscribe at chinaaidispatch.substack.com.