In this issue:

1. Xi’s first in-person WAIC keynote launches a 29-country AI alliance with no G7 members

2. Moonshot’s Kimi K3 triggers a “Kimi moment” — a global semiconductor selloff

3. DeepSeek reprices itself 40% higher, six weeks after its first outside round

4. Beijing’s emotional AI rules complete week one without a confirmed enforcement action

POLICY

Xi’s First In-Person WAIC Keynote Launches a 29-Country AI Alliance With No G7 Members

Xi Jinping delivered his first-ever in-person keynote at China’s flagship AI summit, a day after 29 countries founded a rival global AI governance body headquartered in Shanghai.

WAICO’s founding members — Russia, Cuba, Brazil, Venezuela, Pakistan, Indonesia, Kazakhstan, Cambodia, Thailand — represent a combined population dwarfing the G7, with no Western democracy among them. Xi pledged 5,000 AI training slots for developing countries over five years, calling AI development “a symphony of international cooperation.” Huawei debuted its 8,192-chip Atlas 950 supercomputer at the same event.

Ma’s read: The BBC segment above wasn’t a preview — Belt and Road’s decade-long lead on Global South energy infrastructure is exactly the foundation WAICO now governs. This isn’t a new stack; it’s Beijing formalizing control of one it’s financed for years - see my 2018 Book “China’s Digital Silkroad”.

Source: Xinhua · SCMP · Reuters · CGTN

CAPITAL

Moonshot’s Kimi K3 Triggers a “Kimi Moment” — a Global Semiconductor Selloff

Moonshot AI’s 2.8-trillion-parameter open-weight model, benchmarked near Anthropic’s Claude Fable 5, sent global chip stocks into their worst week in over a year.

The Philadelphia Semiconductor Index fell as much as 5.7% in a session, down 11% for the week and over 20% from its late-June peak — bear-market territory. Nvidia, AMD, Intel, and ARM all fell sharply; Chinese AI names dropped too in Hong Kong trading (Zhipu -28%, MiniMax -16%), suggesting the fear is compressed AI economics broadly, not a clean China-versus-US story.

Ma’s read: In The Hunt for Unicorns book, I’ve argued traditional VC can’t sustain AI infrastructure spending alone, forcing sovereign capital onto cap tables. This week stress-tests that thesis — sovereign funds underwriting multi-year infrastructure bets need this kind of shock priced in from day one, not treated as a tail risk.

Source: Bloomberg · Fortune · Yahoo Finance

CAPITAL

DeepSeek Reprices Itself 40% Higher, Six Weeks After Its First Outside Round

DeepSeek is in talks for a new funding round at a $71 billion pre-money valuation, up from June’s roughly $50–52 billion close, while separately preparing a mainland China IPO targeting 2027.

Founder Liang Wenfeng, who personally committed roughly $3 billion to the June round, has seen his 78% stake approach $36 billion. Separately, Reuters reported July 7 that DeepSeek has spent the past year quietly building its own AI inference chip capability — the same self-sufficiency logic behind last week’s Hubble/EDA ecosystem story.

Ma’s read: Same pattern [I traced through Hubble, at last week’s China AI Insider issue] at the corporate-VC layer, now at the model layer — pricing in infrastructure independence before the market forces the question. Six weeks to a 40% repricing isn’t normal velocity anywhere.

Source: Financial Times · Bloomberg · TechCrunch

REGULATION

Beijing’s Emotional AI Rules Complete Week One Without a Confirmed Enforcement Action

The Interim Measures for AI Anthropomorphic Interactive Services have been in force for five days, with no confirmed enforcement action against any platform yet.

Doubao, Qwen, Tencent Yuanbao, and NetEase all disabled companion-agent features ahead of the deadline, as covered here last week. The absence of enforcement tracks a pattern worth naming: voluntary compliance often carries a new measure through its first quarter before real enforcement tests begin.

Ma’s read: I flagged this rule at the draft stage on CNBC back in December — “a leap from content safety to emotional safety.” A quiet first week isn’t a sign the rule is toothless; it’s the normal on-ramp. Friday’s China AI Insider goes deep on what a real test would actually look like.

Source: 新华社 (Xinhua) · 国家网信办 (CAC)

The Signal: Beijing argued its AI model to the world this week, and its own companies argued it to the market — a governance bloc built from the Global South, a semiconductor selloff nobody outside China triggered, and a model company repricing itself faster than any Western comparable.

Coming Next: Friday’s China AI Insider goes deep on Beijing’s emotional AI companion rules — a full week of compliance data, and what enforcement (or its continued absence) signals about how Beijing tests a new measure before it actually enforces it. [Subscribe for Friday’s Deep Dive →]

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal work together:

Beijing Signal stays free, always, every Monday. Each Friday’s China AI Insider takes that week’s same stories, documents them in full with Chinese primary sources, adds new developments that emerged after the Monday dispatch, and delivers a 2,000-word institutional deep dive.

Capital Signal (paid, bi-weekly) tracks how sovereign capital actually deploys into AI infrastructure across five geographic blocs — the capital-side complement to the Insider’s policy-side depth.

Subscribe now

🔵 Beijing Signal (free, every Monday) 🔴 China AI Insider (paid flagship, every Friday) 🟡 Capital Signal (paid, bi-weekly) ⚫ Sovereign AI Outlook (paid, quarterly) · first issue August 2026

See details of all newsletters at my Substack’s website — About.

All past issues before the paywall are available at the Archive of my Substack Website, always.

Winston Ma (马文彦) is a family office AI investor; former Managing Director — Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, Global Public Investment Funds Forum; Adjunct Professor, NYU School of Law; Columnist, JIC (Japan Investment Corporation); and author of The Hunt for Unicorns (Wiley, forewords by Margaret Franklin, Russell Read, and Ajay Royan), The Digital War (Wiley), and Blockchain and Web3 (Wiley, foreword by Anthony Scaramucci), and among 2024 “six must-read blockchain books” by TechTarget.

In this issue:

1. Huawei’s τ Scaling Law returns with production data instead of a roadmap

2. Beijing weighs closing the export door on its own best AI models

3. CXMT’s ¥29.5B IPO turns a DRAM deal into a public listing

4. A two-track compliance floor arrives for emotional AI companions

HARDWARE

Huawei’s τ Scaling Law Returns With Production Data Instead of a Roadmap

Huawei’s semiconductor chief republished her post-Moore chip framework this week — this time backed by mass-production measurements, not projections.

Dr. He Tingbo’s July 3 ChinaXiv update turns May’s framework paper into an engineering-verified one — real Kirin 2026 measurements (transistor density +53.5%, CPU clock back to 3.1 GHz) replace modeled projections, alongside a concrete Ascend 950/990 roadmap. No Western outlet has parsed the delta from the May version yet.

Ma’s read: I flagged the real trigger to watch back in May (reference to previous AI Insider issue) — not the 2031 endpoint, but whether Kirin silicon hits its disclosed clock speeds on schedule. V2 just moved that verification up by two quarters. Deep Dive at Friday’s China AI Insider.

Source: ChinaXiv · 观察者网 (Guancha) · 新浪财经 (Sina Finance)

REGULATION

Beijing Weighs Closing the Export Door on Its Own Best AI Models

Chinese authorities are exploring curbs on overseas access to the country’s most advanced AI models — including some not yet released.

MOFCOM-led talks with Alibaba, ByteDance, and Z.ai over the past month reportedly floated a tiered regime — filing for basic open-source tools, security review for advanced ones, domestic-only for frontier systems — plus possible national-security penalties for leaks. Nothing is decided; the talks trace to a May roundtable of Chinese legal scholars on tiered open-source AI governance.

Ma’s read: As I told The Diplomat’s Mercy Kuo on SpaceX’s ITAR exclusion, both capitals are now drawing the same line in opposite directions — access, not just ownership, is the new chokepoint.

Source: Reuters (exclusive)

CAPITAL

CXMT’s ¥29.5B IPO Turns a DRAM Deal Into a Public Listing

China’s largest DRAM maker formally launched its STAR Market IPO prospectus, targeting the exchange’s second-largest listing ever.

CXMT set July 16 for subscriptions, CICC as lead underwriter, with a greenshoe option that could lift the offering to 7.69B shares — half the deal reserved for cornerstone investors. This is the capital-markets event; the ¥200B Tencent DRAM contract was covered here in June.

Ma’s read: Having priced state-backed strategic listings at CIC, this is the first market test of how outside capital values a domestic memory champion the state has already backed directly.

Source: Global Times · Caixin Global · Bloomberg

REGULATION

A Two-Track Compliance Floor Arrives for Emotional AI Companions

China’s dedicated rules for anthropomorphic AI take effect Wednesday, days after regulators closed the books on a four-month app sweep.

The CAC’s “清朗” (clean-up) campaign wrapped Phase 1 on July 6 — 14,000+ AI products removed, 26,000+ accounts penalized — with companion-agent abuse named as a Phase 2 target. Separately, the Interim Measures for AI Anthropomorphic Interactive Services take effect July 15; Doubao, Qwen, Tencent Yuanbao, and now NetEase have all disabled companion-agent features ahead of the deadline.

Ma’s read: I flagged this rule at the draft stage on CNBC and via 参考消息 back in December — “a leap from content safety to emotional safety.” That framing lands for real AI products this week, not just on paper.

Source: 新华社 (Xinhua) · CAC (国家网信办) · 经济观察网 (Economic Observer)

The Signal: Four different instruments, one direction — Beijing is no longer sequencing its AI buildout one layer at a time. Chips, capital markets, model access, and consumer compliance are all moving in parallel now, converging inside a single week (CXMT prices July 16, the anthropomorphic rules take effect July 15).

Coming Next: Friday’s China AI Insider goes deep on Huawei’s τ Scaling Law V2 — mapping the EDA and advanced-packaging plays it unlocks against the Kirin roadmap I flagged back in Issue 6. [Subscribe for Friday’s Deep Dive →]

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal work together:

Beijing Signal stays free, always, every Monday. Each Friday’s China AI Insider takes that week’s same stories, documents them in full with Chinese primary sources, adds new developments that emerged after the Monday dispatch, and delivers a 2,000-word institutional deep dive.

Capital Signal (paid, bi-weekly) tracks how sovereign capital actually deploys into AI infrastructure across five geographic blocs — the capital-side complement to the Insider’s policy-side depth.

🔵 Beijing Signal (free, every Monday) 🔴 China AI Insider (paid flagship, every Friday) 🟡 Capital Signal (paid, bi-weekly) ⚫ Sovereign AI Outlook (paid, quarterly) · first issue August 2026

Subscribe now

See details of all newsletters at my Substack’s website — About.

All past issues before the paywall are available at the Archive of my Substack Website, always.

Winston Ma (马文彦) is a family office AI investor; former Managing Director — Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, Global Public Investment Funds Forum; Adjunct Professor, NYU School of Law; Columnist, JIC (Japan Investment Corporation); and author of The Hunt for Unicorns (Wiley, forewords by Margaret Franklin, Russell Read, and Ajay Royan), The Digital War (Wiley), and Blockchain and Web3 (Wiley, foreword by Anthony Scaramucci), and among 2024 “six must-read blockchain books” by TechTarget.

Four items this week, one thread: Beijing keeps handing out the papers — a registered identity, a legal template, a supply contract, a public price — well before anyone settles the harder verdict underneath: who’s liable, who’s really in control, what something is actually worth.

In this issue:

1. China’s new agent interconnection standard gives AI agents a digital ID — the liability question stays open

2. State Council Order 837 turns the Meta-Manus block into a standing rule, not a one-off

3. Tencent’s ¥200B CXMT order secures a chip layer nobody forced China to secure

4. Zhipu’s public listing prices China’s AI sector twice as high as DeepSeek’s private round

POLICY

China’s new agent interconnection standard gives AI agents a digital ID — the liability question stays open

China released national interconnection standards for AI agents on June 30 — an identity, capability-declaration, discovery, and tool-invocation framework effectively functioning as a digital ID card for agents — with 100+ companies co-building it and 50+ already piloting.

The standard, alongside China’s own Agent Trust Handshake protocol from 中国信通院 (CAICT), answers “which agent is this,” not “who’s liable when it fails.” The STAR Market Daily (财联社’s 科创板日报) frames 2026 as agentic AI’s “ChatGPT moment.” Western coverage has focused on rival protocols — Anthropic’s MCP, Google’s A2A — and largely missed China building parallel identity infrastructure.

Ma’s read: In my Bloomberg Law piece two days after this standard published, I flagged exactly this gap — agentic one-person companies raise a “know your agent” liability question no jurisdiction has answered yet. Beijing is building the technical identity layer before anyone answers the legal one. Deep Dive at Friday’s China AI Insider.

Source: 财联社/科创板日报 (June 30); 国家网信办

REGULATION

State Council Order 837 makes the Meta-Manus block the template, not the exception

State Council Order No. 837 took effect July 1, formalizing the outbound-investment review logic that unwound Beijing’s block of the Meta-Manus deal into a standing rule for every cross-border AI transaction.

A companion revision to China’s Trade Secret Regulation newly classifies AI model weights, algorithms, and training data as protectable state assets. Together they close the “Singapore washing” route — relocate, reincorporate, sell to a U.S. buyer — a path Order 837 now makes formally dead. 21世纪经济报道 has covered the implementation window; Western financial press has largely missed the Trade Secret angle. The mirror image showed up the same week: SpaceX’s underwriters (Goldman Sachs, Morgan Stanley) voluntarily excluded mainland China and Hong Kong investors from its IPO book, citing ITAR — a preemptive self-restriction, not a government mandate, but the same substance-over-form logic running in the opposite direction.

Ma’s read: As I told The Diplomat’s Mercy Kuo, this is Beijing’s own “CFIUS moment” — the same test Washington has long used inbound, now deployed outbound.

Source: The Diplomat (Mercy Kuo Q&A, July 1); 21世纪经济报道; Reuters

HARDWARE

Tencent’s ¥200 billion CXMT order secures a chip layer nobody forced China to secure

Tencent signed a 3-to-5-year, ¥200-billion-plus supply agreement with 长鑫存储 (ChangXin Memory Technologies, CXMT) for server DRAM, days ahead of CXMT’s record ¥29.5 billion STAR Market IPO.

DRAM isn’t the layer export controls target — HBM and advanced logic are. Chinese hyperscalers are voluntarily de-risking a component Washington never restricted, timed to underwrite a state-directed listing. ByteDance and other major domestic internet firms are reportedly in similar talks, per 路透社 and 电子工程专辑.

Ma’s read: CXMT’s Shanghai listing is worth noting, as I commented in Reuters earlier on CATL’s HK Listing. I’d flag this as substitution nobody mandated — a different signal than sanctions-driven scrambling.

Source: Reuters; 电子工程专辑 (EE Times China)

CAPITAL

Zhipu’s public listing prices China’s AI sector twice as high as DeepSeek’s state-backed round

Zhipu AI’s market capitalization has approached $127 billion since listing, more than double DeepSeek’s $50–59 billion valuation in its National AI Fund (国家人工智能产业投资基金)-anchored private round.

The Big Fund holds direct DeepSeek equity with voting rights and no lockup — commercial co-investors get neither. That’s a policy instrument, not a growth investor. Zhipu’s market pricing is the first real-world test of how much more the public market will pay than the state does.

Ma’s read: In my recent FT Opinion piece, I noted the National AI Fund’s DeepSeek stake as the clearest China-side proof that both Washington and Beijing are shifting from regulating AI to taking direct equity in it — Zhipu’s number just measured the gap.

Source: 甲子光年 (Jazzyear); Financial Times

The Signal: Four stories, one pattern — Beijing issues the papers (a digital ID, a legal template, a supply contract, a public price) well ahead of the verdict underneath (liability, control, true valuation). That sequencing is a choice, not an accident: infrastructure first, judgment later, and by the time the harder questions get answered, the facts on the ground are already Beijing’s. As I wrote at a NYU DRI piece on the AI stack: "China is empowering — and gradually locking in — the layers that matter most.

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor (Dragon Global); former Managing Director and Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal Work Together

Each Monday’s Beijing Signal (free) surfaces four observations from the week’s Chinese-language AI news cycle — what happened, what Western coverage missed, and why it matters for investors. Each Friday’s China AI Insider takes these same four stories, documents them in full with Chinese primary sources, adds new developments that emerged after this dispatch, and delivers a 2,000-word institutional deep dive.

Capital Signal (paid, bi-weekly) tracks sovereign capital deployment into AI infrastructure across five geographic blocs.

Subscribe to the Insider Track →

See details of all newsletters at my Substack’s website — About.

In this issue:

1. China publishes 7 national standards giving AI agents a formal identity

2. Microsoft is routing American AI east and Chinese AI west — and taking the margin on both legs

3. China launches industrial 5G standalone network pilot for real-time robotics

4. STAR Market opens its doors to unprofitable AI large-model companies

REGULATION

China Publishes 7 National Standards Giving AI Agents a Formal Identity for the First Time

市场监管总局 (SAMR) released China’s first national standards for AI agent interoperability — identity codes, interaction protocols, and tool-calling rules — solving the “information island” problem that blocked cross-platform agent deployment.

The seven standards cover a full chain — identity codes (身份码), identity management, agent description, discovery, interaction, and tool-calling — from “who is this agent” to “what can it do” to “how does it transact.” Published as guidance documents to accommodate multiple technology paths, with the identity-code standard confirmed for later upgrade to mandatory status. Haidian district (Zhongguancun) is the first pilot zone.

Ma’s read: As I argued in a recent World Economic Forum piece, AI agents are developing explosively in China — and these seven standards are the governance layer arriving before the procurement wave. Friday's flex deep dive on the one-person-company framework shows the corporate-law side of the same bet.

Source: 市场监管总局 (SAMR) press conference, June 26, 2026; 第一财经 (Yicai); 新京报 (Beijing News)

TECHNOLOGY

Microsoft Is Routing American AI East and Chinese AI West — and Taking the Margin on Both Legs

While Washington blocks SpaceX from Chinese capital and Beijing restricts outbound tech transfers, Microsoft sells GPT models to ByteDance through Azure and is considering hosting DeepSeek for Western clients.

OpenAI and Anthropic refuse to sell in China. Microsoft, through its OpenAI partnership, sells GPT models anyway — to ByteDance, now its biggest AI customer on Azure. Simultaneously, Microsoft is weighing hosting DeepSeek on its Copilot platform for Western enterprise clients. American models east, Chinese models west, Microsoft takes the margin on both legs.

Ma’s read: In 2021, I told CNN that China remained part of Microsoft’s game plan despite the big hack — that its cloud strategy was structurally tied to the Chinese market in ways competitors couldn’t replicate. Five years later, ByteDance is Microsoft’s biggest AI customer, and Microsoft is the clearest proof of why both US and China concluded, as I argued in a June 2026 Financial Times Op-Ed, that “AI is too strategically important to leave entirely to private markets.”

Source: The Next Web, WSJ (Nadella interview, June 21)

HARDWARE

China Launches Industrial 5G Standalone Network Pilot — the Connectivity Layer for Real-Time Robotics

工信部 (MIIT) and 国务院国资委 (SASAC) jointly launched an industrial 5G standalone network pilot with five government departments, enabling real-time edge computing for factory-floor AI and embodied intelligence.

The pilot targets 5G-enabled real-time robotics, autonomous inspection, and machine-to-machine coordination in key industrial corridors — the connectivity layer that turns the humanoid deployment mandate (provincial plans due tomorrow, June 30) from a hardware procurement program into an operational data network.

Ma’s read: At MWC Shanghai in 2023, I told CNBC that “Chinese companies are probably more ready, are more willing to test new 5G applications.” Three years later, the pilot program is the evidence: China is connecting the factory floor to the AI stack at the network layer, not just the hardware layer.

Source: 工信部 (MIIT), June 24, 2026

CAPITAL

STAR Market Opens Its Doors to Unprofitable AI Large-Model Companies

CSRC (the SEC of China) Chairman Wu Qing announced at the Lujiazui Forum that the STAR Market’s fifth listing standard — originally built for unprofitable biomedical firms — now applies to AI large-model companies. Zhipu AI is already filing.

Shanghai Stock Exchange 上交所 (SSE) released Guideline No. 10 the same day (June 17), requiring at least one large model product launched with “scaled application” at filing, plus government approval and clear commercialization plans. AI startups like Zhipu AI (智谱AI) has filed for a ¥15 billion raise; MiniMax has also submitted backup documentation. Both already listed on HKEX, now pursuing “A+H” dual listings.

Ma’s read: Not surprising if you know the history. Shenzhen positioned itself as China’s Nasdaq; Shanghai created the STAR Market specifically for tech innovation companies. What the fifth standard does is calling the domestic market to become what it was meant to be originally — plus the new urgency from AI. I’ve commented on this dynamic across multiple Reuters HK IPO stories this year: companies like Zhipu AI and Kunlunxin are pursuing A-share listings after Hong Kong to tap deep domestic liquidity and align more closely with national priorities. Beijing wants both tracks running.

Source: 中国证监会 (CSRC), 上交所 (SSE) Guideline No. 10, June 17, 2026

The Signal: SC Order 837 activates in 48 hours (July 1 enforcement of the State Council’s outbound investment regulation — detailed analysis in earlier China AI Insider issues). When it does, every story in this issue — agent identity standards, model-routing platforms, industrial networks, listing pathways — sits inside a legal architecture that did not exist six months ago. The architecture is complete. Tuesday is the day it turns on.

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor (Dragon Global); former Managing Director and Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal Work Together

Each Monday’s Beijing Signal (free) surfaces four observations from the week’s Chinese-language AI news cycle — what happened, what Western coverage missed, and why it matters for investors. Each Friday’s China AI Insider takes these same four stories, documents them in full with Chinese primary sources, adds new developments that emerged after this dispatch, and delivers a 2,000-word institutional deep dive.

Capital Signal (paid, bi-weekly) provides global Sovereign AI coverage, which tracks sovereign capital deployment into AI infrastructure across five geographic blocs.

Subscribe to the Insider Track →

Subscribe now

See details of all newsletters at my Substack’s website — About.

COMING NEXT

China AI Insider (holiday issue, Friday, July 3 ) — a flex deep dive edition of the China AI Insider: the one-person-company framework as the legal architecture for AI-era founders, expanding the argument from Winston Ma’s recent publications in the World Economic Forum, Financial Times, and Bloomberg Law. Holiday week, focused format.

In this issue:

1. Beijing’s five-year plan treats AI jobs and job protection as one policy

2. DeepSeek’s governance instrument isn’t the headline investor

3. Beijing’s quietest AI policy just moved AI stocks

4. Humanoid mandate paperwork is due in nine days

SOCIETY

China’s new Employment-First Strategy treats AI job creation and job-loss protection as one policy, not a trade-off

China’s State Council just set a non-negotiable floor against AI-driven unemployment — while simultaneously funding AI as a job creator.

The State Council’s Employment-First Strategy for the 15th Five-Year Plan (《实施就业优先战略”十五五”规划》) names 不发生规模性失业风险 (”no large-scale unemployment risk”) as a 底线 — a bottom line, the strongest commitment level in Chinese policy language. It pairs an early-warning system for AI-driven job losses with a mandate to “adapt to AI development to promote employment and entrepreneurship” (适应人工智能发展促进就业创业) — treating AI adoption and job protection as one problem, not two.

Ma’s read: In a recent World Economic Forum piece, I argued “the state actively built the infrastructure for the agentic era” through China’s one-person-company reform. This plan is that same instinct, economy-wide. Friday’s Insider takes it as the deep dive.

Source: 国务院 (State Council), 《实施就业优先战略”十五五”规划》

CAPITAL

DeepSeek’s ¥50 billion round reveals Beijing’s real governance instrument — and it isn’t the headline investor

DeepSeek’s closed funding round shows Beijing keeping voting control through a fund most coverage hasn’t traced back far enough.

Among DeepSeek’s ¥50B+ round, only the National AI Fund (国家人工智能产业投资基金) holds direct equity with full voting rights and no lock-up — every other investor holds an indirect stake through founder Liang Wenfeng’s LP vehicle. 新浪财经 (Sina Finance) reports the National AI Fund was co-founded in 2025 by 国智投 (Guozhi Investment — literally “National Intelligence Investment”) and — structurally — Big Fund III, whose own DeepSeek involvement has been rumored but never confirmed.

Ma’s read: As I argued in my Financial Times Op-Ed last Friday, the National AI Fund was “the only one to receive direct corporate ownership and voting rights” in DeepSeek’s round — Beijing’s clearest proof yet that ownership, not just regulation, is the preferred instrument of AI-era control.

Source: 新浪财经 (Sina Finance), 21世纪经济报道 (21st Century Business Herald)

TECHNOLOGY

Beijing’s quietest AI policy this quarter just moved AI stocks

A National Data Bureau dataset plan most Western outlets missed entirely triggered a same-day rally in China’s AI-themed ETFs.

The National Data Bureau’s Action Plan for High-Quality Industry Dataset Construction (国数科基〔2026〕25号) sets a 2028 target for state-validated training datasets built specifically for agents, embodied intelligence, and world models — and shifts data labeling from human-led to human-machine collaborative annotation. China’s AI-themed ETFs rallied same-day on the announcement.

Ma’s read: In my book The Digital War, I wrote: “If data is the new oil, then China is the new OPEC.” This plan turns that thesis into procurement policy — data quality is now a state-funded input, not a private cost.

Source: 国家数据局 (National Data Bureau)

HARDWARE

China’s humanoid mandate moves from directive to paperwork — and the paperwork is due in nine days

Provinces and central state enterprises must file real-world humanoid deployment plans with MIIT and SASAC by June 30.

The June 9 joint notice’s implementation phase is active: every province and central SOE must submit a Real-Scene Training Work Plan naming deployment scenarios, partner suppliers, and validation benchmarks to 工信部 (MIIT) and 国务院国资委 (SASAC). The notice’s own model — “validate one, deploy a batch, drive a region” — converts a successful pilot directly into standardized rollout.

Ma’s read: At Fudan, I studied the materials science underneath exactly this kind of hardware scaling — the real signal isn’t the robot, it’s the procurement mechanism turning pilots into guaranteed order volume.

Source: 工业和信息化部 (MIIT) and 国务院国资委 (SASAC) joint notice

The Signal: None of this week’s four signals required new ownership — an employment guarantee, a fund structure, a procurement mandate, a dataset standard. Beijing’s preferred instrument for AI-era control isn’t acquisition; it’s architecture. Friday’s Insider asks whether that same architecture can keep a promise no other AI-deploying economy has dared make: full speed and full employment, at once.

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor (Dragon Global); former Managing Director, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal Work Together

Beijing Signal stays free, always. Each Friday’s China AI Insider takes these same four stories, documents them in full with Chinese primary sources, adds new developments that emerged after this dispatch, and delivers a 2,000-word institutional deep dive. Capital Signal (paid, bi-weekly) tracks sovereign capital deployment into AI infrastructure across five geographic blocs.

Subscribe to the Insider Track →

See details of all newsletters at my Substack’s website — About.

In this issue:

• National planners draft a ¥2 trillion framework to operationalize compute routing at sovereign scale

• 工信部 (MIIT) and 国务院国资委 (SASAC) order 10,000 certified humanoids into active public-sector operations

• Sovereign and corporate syndicates finalize a ¥50 billion convergence behind DeepSeek

• The National Data Bureau expands its data pilot to turn model weights into tradeable units

POLICY

National Planners Draft a ¥2 Trillion Data Center Framework to Operationalize Compute Routing

The 发改委 (NDRC) is drafting a ¥2 trillion ($295B) five-year plan for a nationwide interconnected AI data center network, funded by ultra-long sovereign bonds, operated by China Mobile and China Telecom, with 80% domestic chip content.

The plan directly extends 东数西算 (Eastern Data, Western Computing, 2022) — the strategy to route AI compute demand to western China’s cheaper renewable energy — into a binding financial pipeline. The 80% domestic chip mandate turns Huawei Ascend into the de facto standard for a $295 billion market, routing out Nvidia and AMD. Western coverage led with the headline figure; what it missed is the financing mechanism: ultra-long sovereign bonds immunize this commitment from any administration or market cycle. The plan remains a draft — details may change.

Ma’s read: As I commented at Financial Times – Sustainability Views last week, “With sovereign capital and state co-ordination driving [energy and computing] layers, this integration has a strong chance of creating a true sovereign AI energy advantage. This blueprint reveals Beijing treats raw compute capacity as a permanently subsidized state utility. Friday’s Insider deep-dives the long-term chip beneficiaries of this $295 billion market.

Source: 发改委 (NDRC) · Bloomberg (June 9) · Tom’s Hardware

HARDWARE

工信部 and 国务院国资委 Order 10,000 Humanoid Robots Into Active Public-Sector Operations

工信部 (MIIT) and 国务院国资委 (SASAC) jointly issued the “2026 Humanoid Robot and Embodied Intelligence Real-Scene Training Special Action,” mandating routine deployment across 100+ high-value scenarios in 10 provinces and central SOEs by end-2026.

This is the execution step the HEIS digital ID registry was always pointing toward. Beijing’s governance-precedes-procurement pattern runs exactly here: the HEIS 29-character registry (May 2026) created the compliance infrastructure; the MIIT/SASAC mandate now activates the deployment pipeline. Western coverage focused on the volume target; what it missed is the institutional mechanism — the humanoid registry has transitioned from a passive inventory log into a live state-directed procurement channel.

Ma’s read: Having studied electronic materials at Fudan, I read mandatory SOE deployment as the procurement trigger that precedes a supply-chain wave — the same pattern that preceded NEV dominance. HEIS-compliant suppliers like Unitree and AgiBot are now moving from compliance checks into live government revenue rails.

Source: 工信部 (MIIT) · 国务院国资委 (SASAC) · 新华社 (Xinhua) · Pandaily

CAPITAL

Sovereign and Corporate Syndicates Finalize a ¥50 Billion Convergence Behind DeepSeek

DeepSeek is finalizing ~¥50 billion ($7.4B) in external funding at a $52–59B valuation — with founder Liang Wenfeng committing ¥20B, Tencent ¥10B, CATL ¥5B, and the National AI Fund, NetEase, and JD.com in final talks.

Reporting by 36氪 (36Kr) confirms Liang Wenfeng secured ~84% founder control before admitting external state balance sheets — deliberate sequencing. The investor coalition spans sovereign strategic capital (National AI Fund), industrial energy infrastructure (CATL), cloud distribution (Tencent), and consumer platforms (NetEase, JD.com). Western coverage led with the valuation; what it missed is the institutional span — this is China’s full AI ecosystem making a single coordinated commitment behind one lab. The round remains in negotiation; terms may change.

Ma’s read: In The Hunt for Unicorns, I documented how sovereign syndicates reclassify high-growth technology into long-term national infrastructure assets. When a 13-year sovereign fund horizon enters the same cap table as an energy infrastructure giant, the model layer is no longer valued on commercial revenue multiples.

Source: 36氪 (36Kr) · 第一财经 (Yicai) · Reuters (June 3)

TECHNOLOGY

China’s National Data Bureau Expands the Data Element × Pilot to Turn Model Weights Into Tradeable Units

The 国家数据局 (National Data Bureau) expanded its “Data Element ×” (数据要素✕) pilot to cloud-hosted model parameters, legalizing fine-tuning rights trading inside special industrial zones under state-controlled clearing houses.

Western markets treat model weights as proprietary IP. China’s planners treat them as 数据要素 (data factors) — production inputs that can be priced, traded, and cleared like any other economic factor. This directly complements the Trade Secret Regulation (which protects weights as sovereign assets outside the inner loop) and SC Order 837 (which controls how they cross borders): inside the inner loop, tradeable; outside it, protected.

Ma’s read: As I argued in the deep-dive “Data Multiply” of a recent weekly AI Insider issue , China constructs and defends its data advantages as instruments of strategic competition. This pilot is the mirror image of Western algorithmic classification — a state-supervised exchange financializing model capability boundaries inside the inner loop.

Source: 国家数据局 (National Data Bureau) · 晚点 (LatePost)

The Signal: Beijing has moved the AI stack from conceptual architecture into operational field execution — ¥2 trillion committed to the compute substrate, 10,000 humanoid robots mandated into the field, the broadest capital coalition ever assembled behind one model lab, and model parameters now treated as a traded financial instrument. The inner loop is operational. The drawbridge goes live in 16 days (the July 1 enforcement of the State Council order) — and on Friday, the Insider decodes the full investor implications of this week’s deployment cascade.

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor; former Managing Director and Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

This Friday’s China AI Insider deep dive: China’s ¥2 trillion NDRC data center plan — Layer 1 sovereign infrastructure, domestic chip beneficiaries, and what a $295 billion market closed to Nvidia means for the global semiconductor competition.

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal Work Together

Each Monday’s Beijing Signal (free) surfaces four observations from the week’s Chinese-language news cycle. Each Friday’s China AI Insider takes the same four items, expands them with full primary-source documentation, adds one or two further items that emerged after the Beijing Signal dispatch, and delivers a 2,000-word institutional deep dive on the week’s most consequential development. More stories. Deeper analysis. One product to work from.

Capital Signal covers five geographic blocs tracking how sovereign wealth funds deploy capital into national AI infrastructure.

See details of all newsletters at my Substack’s website - About.

• Beijing’s AI agent security crowd-test is an enforcement dress rehearsal, not a bug bounty

• Shanghai plans compute futures — China moves to treat AI computing power as a traded asset class

• China’s first trade secret update since 1998 classifies model weights and training data as protected assets

• China’s humanoid robot ID registry moves from governance to procurement — the tenders are live

  

REGULATION

Beijing’s AI agent security crowd-test is an enforcement dress rehearsal, not a bug bounty

CNCERT’s crowd-testing of AI agent (智能体) vulnerabilities runs through June 30 — one day before SC Order 837 takes effect — offering up to ¥100,000 per structural defect.

Western analysts read this as a software bug bounty. The two agencies organizing the program are the same agencies that will administer the State Council Order 837 enforcement from July 1 (reference to the China AI Insider issue last Friday). Every model and agent application that participates will have a state-verified security baseline on record before the new legal architecture activates.

Ma’s read: The sequencing is the tell — a state security audit that closes the day before SC Order 837 activates is not coincidence. What CNCERT is building this month is the verified-compliance record that the new legal architecture assumes exists.

Source: 国家计算机网络应急技术处理协调中心 (CNCERT) · cert.org.cn

CAPITAL

Shanghai plans compute futures as China moves to treat AI computing power as a traded asset class

Shanghai’s municipal government has explicitly named compute futures in an official planning document for the first time — mirroring CME Group’s announcement three weeks earlier and signaling China’s intent to link financial markets to AI infrastructure.

The CME announcement got the Western headlines. The Shanghai move is the structural signal: computing power (算力) is being positioned as a financial asset class — hedgeable, tradeable, and investable — not just industrial infrastructure. The Shanghai Futures Exchange is China’s largest commodity bourse. Unitree Robotics passed its Shanghai IPO committee hearing the same week, underscoring the city’s deliberate push to become the financial hub for China’s AI stack.

Ma’s read: At CIC, we watched commodity futures markets as early indicators of where sovereign capital was routing long-term bets (for example, gold). Compute futures in Shanghai would do for AI infrastructure what LNG futures did for energy transition — create a price-discovery mechanism that draws institutional capital into the sector at scale. The CME did it first; the question is whether Shanghai’s version becomes a cross-border hedging instrument or stays domestic.

Source: SCMP · 上海市政府 (Shanghai Municipal Government)

REGULATION

China’s first trade secret update since 1998 classifies model weights and training data as protected assets

The State Administration for Market Regulation’s new Regulations on Trade Secret Protection took effect June 2 — the same day as SC Order 837 was released — explicitly classifying algorithms, datasets, model weights, and computer code as trade secrets for the first time.

SC Order 837 controls the movement of technology and people across borders. The Trade Secret Regulation defines what counts as a protectable asset in the first place. Two instruments, one legal architecture, one day. The fuller capital implications will be mapped in detail on Friday in the China AI Insider series.

Ma’s read: In The Digital War, I argued China, the “Land of Big Data” would gradually codify its data advantages in law — this regulation is that argument in statute. As a lawyer admitted in both China and the US, every cross-border AI collaboration now needs to be repriced: your model weights and training data sit under different legal regimes simultaneously.

Source: 国家市场监督管理总局 (SAMR) · Bloomberg · samr.gov.cn

HARDWARE

China’s humanoid robot ID registry moves from governance to procurement — the tenders are live

MIIT-affiliated procurement platforms have begun issuing tenders requiring HEIS compliance for government and SOE humanoid robot deployments — the execution step after the late-May launch of the 29-character digital ID system.

Last week’s Signal covered the governance layer: 28,000 units, 100 manufacturers, 29-character IDs. This week is the procurement layer. Chinese firms accounted for nearly 90% of global humanoid robot shipments in 2025 (IDC). HEIS (Humanoids and Embodied Intelligence Standardization) -compliant suppliers now have a direct channel into the state procurement pipeline that guidance-fund disbursements will flow through.

Ma’s read: Having studied electronic materials at Fudan, I read the transition from registry to live tenders exactly as intended — governance as procurement enabler, the same pattern that preceded China’s EV supply chain dominance. H2 2026 order volumes will be the real commercialization signal to watch.

Source: MIIT/HEIS · 智东西 (Zhidx) · 机器之心 (Synced)

The Signal: On June 2, Beijing published the State Council Order 837 and the Trade Secret Regulation simultaneously — one controlling the movement of technology across borders, the other defining what counts as a protectable asset. This week added a financial layer on top: Shanghai positioning compute power as a tradeable asset class, CNCERT building a compliance baseline ahead of July 1, and the first HEIS tenders opening the humanoid supply chain to state capital. The sovereign stack is moving into executable contracts, auditable specifications, and now futures markets. The drawbridge goes live in 22 days — and on Friday, the Insider decodes the Politburo blueprint that designed all of it.

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor; former Managing Director and Head of North America, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

This Friday’s China AI Insider deep dive: Xi Jinping’s January 30 Politburo Study Session speech on future industries — published in the Party Magazine求是 (Qiushi) four months after implementation of its directives had already begun. The seven future industries Beijing has officially canonized. The patient capital directive that authorized the Big Fund’s move into DeepSeek. The new nationwide mobilization system (新型举国体制) that coordinates every policy move covered in Issues 1–7. The blueprint that designed the dual circulation.

Subscribe now

How Beijing Signal, China AI Insider, and Capital Signal Work Together

Each Monday’s Beijing Signal (free) surfaces four observations from the week’s Chinese-language news cycle. Each Friday’s China AI Insider (paid, currently free during the 60-day launch window) takes the same four items, expands them with full primary-source documentation, adds further items that emerged after the Beijing Signal dispatch, and delivers a 2,000-word institutional deep dive on the week’s most consequential development. More stories. Deeper analysis. One product to work from.

Capital Signal covers five geographic blocs — Gulf & Middle East, North America, Asia-Pacific, Global, and a dedicated China Sovereign AI Capital section — tracking how sovereign wealth funds deploy capital into national AI infrastructure and how regulation shapes where that capital can go.

Subscribe now

All paid publications are currently free during the 60-day launch window. Pledge your Insider Track subscription now — no charge until the official launch in mid-June.

1. Big Fund’s first model-lab investment values DeepSeek at $50 billion — state capital crosses from chips to models

2. MIIT gives every humanoid robot a digital ID — the governance architecture for scaled deployment

3. Four ministries fuse AI compute with the energy grid — Layer 1 gets its rulebook

4. Beijing extends travel pre-approval to private-firm AI researchers — and some are surrendering passports

CAPITAL

Big Fund’s first model-lab investment values DeepSeek at $50 billion — state capital crosses from chips to models

China’s semiconductor state fund enters foundation model financing for the first time, driving DeepSeek’s valuation a reported fivefold increase in six weeks.

The identity of the lead matters more than the valuation. The Big Fund (国家集成电路产业投资基金) has deployed over $100B into semiconductors. It has never invested in a model lab — until now. As 第一财经 (Yicai) reported, Founder-CEO Liang Wenfeng restructured DeepSeek’s cap table in April, consolidating founder control to ~84% before state money entered.

Ma’s read: I documented in The Hunt for Unicorns how sovereign capital follows strategic logic VC cannot (and won’t). The Big Fund crossing from chips to models tells you Beijing now treats the entire stack as a single sovereign asset class. What that means for foreign capital access is Friday’s deep dive.

Source: FT · Bloomberg · SCMP · 第一财经 (Yicai)

HARDWARE

MIIT gives every humanoid robot a digital ID — the governance architecture for scaled deployment

China launched a 29-character lifecycle ID for every bipedal humanoid, with 28,000 robots from 100 companies already registered.

This is the EV/battery traceability platform applied to embodied intelligence (具身智能) — enabling liability assignment, standardized procurement specs, and guidance-fund deployment at scale. Chinese firms hold ~90% global humanoid market share.

Ma’s read: Having studied electronic materials at Fudan, I read a component-level registry as pre-clearance for industrial procurement, not regulation. In China, the governance layer precedes deployment.

Source: SCMP · MIIT/HEIS · 新华社 (Xinhua)

POLICY

Four ministries fuse AI compute with the energy grid — Jensen Huang’s Layer 1 gets its policy rulebook

NEA, NDRC, MIIT, and National Data Administration issued a 29-task AI-Energy Action Plan, with 51 scenarios released at a May 26 Shenzhen conference.

The four-ministry stamp is the signal. 国能发科技〔2026〕34号 sets 2027/2030 targets for clean-energy supply to AI compute — mapping to the fundamental layer of Jensen Huang’s AI stack cake. Twenty-five energy companies signed an open-scenarios commitment.

Ma’s read: This new measure is an execution-level extension of the “东数西算"(Eastern Data, Western Computing) strategy launched years ago — now supercharged with AI optimization to match clean energy supply with surging AI compute demand.

Source: 国家能源局 (NEA) · 新华社 (Xinhua) · 发改委 (NDRC)

POLICY

Beijing extends travel pre-approval to private-firm AI researchers — and some are surrendering passports

China now requires pre-approval for overseas travel by select AI personnel at Alibaba and DeepSeek — escalating the Manus founder restrictions.

Bloomberg led with talent flight. Chinese coverage from 科技新報 (TechNews) adds what Bloomberg omitted: affected personnel are surrendering passports to employers. Selection is by “individual strategic importance” (個人戰略重要性) — category-based, not named-list — moving from enterprise self-management to state-level supervision.

Ma’s read: When the Meta-Manus deal was ordered to unwind, I commented in CNBC report that “China is most concerned about whether China-origin strategically sensitive technologies — and the data and talent behind them — are effectively transferred offshore by corporate restructuring in Singapore.”

Source: Bloomberg · 科技新報 (TechNews) · 電腦王阿達

The Signal: In one month, the state published a chip-architecture doctrine, led the largest AI fundraise in its history, registered its robots, fused its energy grid to its compute grid, and catalogued its strategic engineers. One closed loop, tightening. Expect smaller private labs to consolidate toward state-backed champions as resource priority hardens. The question for foreign capital through year-end is not whether this loop works, but what will be the investment channel to get in.

Subscribe now

Beijing Signal is a free weekly dispatch by Winston Ma (马文彦) — US family office AI investor; former Managing Director and Head of North America Office, China Investment Corporation ($1.57 trillion); Executive Director, GPIFF; Adjunct Professor, NYU School of Law; author of The Hunt for Unicorns and The Digital War.

Each Monday’s Beijing Signal (free) surfaces four observations from the week’s Chinese-language news cycle. Each Friday’s China AI Insider (paid, currently free during the 60-day launch window) takes the same four items, expands them with full primary-source documentation, adds further items that emerged after the Beijing Signal dispatch, and delivers a 2,000-3000 word institutional deep dive on the week’s most consequential development. More stories. Deeper analysis. One product to work from.

In this issue:

1. Huawei unveils τ scaling (韬定律) as a post-Moore paradigm and a 2031 roadmap to 1.4nm-class chips — a claim deeper than 3D stacking, with 381 shipped chips behind the methodology

2. Alibaba’s Zhenwu M890 launches the same day Nvidia reports zero China revenue — the immediate-term evidence

3. The Trump-Xi summit’s AI deliverables — dialogue without agenda, TikTok without details — are absences, not announcements

4. CAC, NDRC, and MIIT issue China’s first dedicated AI agent rules — the regulatory architecture that retrofits the Meta-Manus unwinding

HARDWARE

1. Huawei unveils τ scaling (韬定律) as a post-Moore paradigm and a 2031 roadmap to 1.4nm-class chips — a claim deeper than 3D stacking, with 381 shipped chips behind the methodology

On May 25, Huawei publicly disclosed a multi-year chip roadmap built around a new architectural framework it calls “LogicFolding,” claiming a path to 1.4nm-class production by 2031 with 55 percent higher transistor density than current Western-leading nodes — and explicitly without EUV lithography access.

The technical paper, authored by Huawei Semiconductor head Tingbo He on ChinaXiv, introduces what the company calls a “τ scaling law” (韬定律) as an alternative to Moore’s Law — proposing that time, not transistor area, becomes the primary metric of progress across the entire computing stack. CNBC’s report led on the central skepticism: five-year roadmaps from Chinese chip designers have a mixed delivery record. That skepticism is appropriate for the 2031 endpoint. It is less applicable to the credibility anchor underneath it.

Three things in the paper deserve more attention than the 2031 headline. First, 381 chips brought to volume production between May 2020 and May 2026 — the operating record Huawei is asking readers to evaluate the methodology against, not the long-horizon roadmap. Second, the Kirin 2026 SoC is already in mass production at a 7nm-class node, with a +55% transistor density step (155 → 238 MTr/mm²), +41% performance-core power efficiency, and CPU frequency returning to 3.1 GHz — gains the paper explicitly notes “previously required three years of geometric scaling,” achieved here at a fixed device node through 3D integration. Third, the Kirin CPU frequency roadmap is already in silicon for 2026 and 2027 (3.1 GHz and 3.39 GHz respectively) and pre-silicon for 2028 and 2029 (3.71 GHz and 4 GHz). The 2031 endpoint is debatable. The 12- and 24-month validation points are not.

One misconception worth naming before Friday’s deep dive: τ scaling (韬定律) is not 3D chip stacking. LogicFolding folds individual logic gates within a single chip’s critical paths at the design stage — cell-level folding, not packaging-stage die stacking as in TSMC’s SoIC — and the broader τ framework extends to Unified Bus (a 500× system-latency reduction through fabric redesign, no 3D required) and Hi-ONE optical I/O. The paradigm shift is the optimization target: from geometry to time.

Winston Ma’s read: As a former electronic materials major undergrad, I agree that long-horizon chip manufacturing technology roadmaps deserve calibrated skepticism — including this one. But for institutional investors, the question isn’t whether the 2031 date holds. It’s whether the export-control thesis — that Chinese chip progress requires EUV access — survives the 381-chip operating record and the 2027-Kirin verification point. If LogicFolding is even partially viable, that thesis is structurally weaker than it looked before the announcement.

And if τ scaling is the right diagnosis of where the industry is going, the strategic capital question shifts from “who owns the leading-edge logic node” to “who owns the advanced packaging, optical interconnect, memory-interface, and 3D-native EDA stack underneath every post-Moore system.” Full technical and capital implications in this Friday’s China AI Insider deep dive.

Source: IEEE ISCAS 2026 keynote (Shanghai, May 25, 2026) · ChinaXiv (202605.00224v1, Tingbo He) · CNBC · Tom’s Hardware · Huawei official disclosure

HARDWARE

2. Alibaba’s Zhenwu M890 launches the same day Nvidia reports zero China revenue — the immediate-term answer to the chip-pivot question

On May 20, Alibaba’s T-Head (平头哥) unveiled the Zhenwu M890 (真武 M890) AI chip at its Cloud Summit in Hangzhou — the same trading day Nvidia reported zero Data Center Hopper revenue from China for Q1 FY27 and guided zero China revenue for Q2.

T-Head disclosed that its Zhenwu series has now shipped 560,000 cumulative units to over 400 customers, with a multi-year roadmap including Zhenwu V900 in Q3 2027 and J900 in Q3 2028. The M890 itself runs three times the previous generation, with 144GB HBM and 800GB/s interchip bandwidth, paired with an in-house ICN Switch fabric that lets 128 cards behave as one computer for agentic-AI workloads. Western coverage framed this as another Nvidia-rivalry story. 钛媒体 (TMTPost), in its May 21 post-Summit analysis, read it differently: as the moment T-Head crossed from internal Alibaba supplier to standalone IPO candidate.

Winston Ma’s read: Items 1 and 2 are the same story at different time horizons. Huawei’s roadmap is the long-horizon claim that the chip-stack debate has been waiting for; T-Head’s launch is the present-tense evidence that the operating layer doesn’t wait for the long-horizon resolution. When Nvidia confirms zero China revenue on the same trading day a Chinese cloud publishes a three-year chip roadmap, the question isn’t whether the stack decouples — it’s which layer prints a public-market valuation first.

Source: IT之家 (IT Home) · 钛媒体 (TMTPost) · 电子工程专辑 (EET China) · Nvidia 10-Q (May 20, 2026)

POLICY

3. The Trump-Xi summit’s AI deliverables — a dialogue mechanism without an agenda and a TikTok track without new details — are absences, not announcements

Following the May 14–15 Beijing summit, China’s Foreign Ministry confirmed agreement to begin “discussions about safe development of AI” — without specifying timing, leadership, agenda, or whether export controls were in scope. The TikTok restructuring track produced no new details.

USTR Jamieson Greer separately said there had been no discussions of chip export controls during the visit. Trump told reporters on Air Force One that the administration was “considering establishing some kind of mechanism on guardrails with Beijing.” CSIS analysts noted the absence of a stated timeframe is unusually thin even by post-summit standards. The 2024 Biden-era dialogue produced one session in which China sent diplomats to complain about export controls while the US sent technical experts to discuss shared risks. The 2026 mechanism is the same empty vessel.

Winston Ma’s read: Having watched bilateral mechanisms form and atrophy across the table at CIC, I read these two absences as deliberate — both sides preserving optionality on the two most economically loaded AI items. The mechanism exists on paper. Whether it produces deliverables depends on which Washington agency actually leads it, and that has not been announced.

Source: 外交部 (Ministry of Foreign Affairs) · 新华社 (Xinhua) · CSIS · 第一财经 (Yicai)

REGULATION

4. CAC, NDRC, and MIIT issue China’s first dedicated AI agent rules — the regulatory architecture that retrofits the Meta-Manus unwinding

On May 8, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology jointly issued the “Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents” (智能体规范应用与创新发展实施意见) — China’s first dedicated agent-level rules, targeting 70% adoption of intelligent agents in smart terminals by 2027.

The document defines AI agents as a distinct regulatory category — separate from generative AI and foundation models — with the capabilities of autonomous perception, memory, decision-making, interaction, and execution. It lays out 19 specific application scenarios across science, industry, consumer services, public welfare, and governance, anchored to a “safety first, innovation second” (安全为先、创新为重) principle. 财新 (Caixin) connected the rule’s timing to April’s Meta-Manus unwinding: the foreign-investment-review architecture for agentic AI was missing when Manus was blocked. Now it has been installed.

Winston Ma’s read: When we covered the Manus unwinding last month, the open question was whether the case was sui generis or the start of a regime. The May 8 rules appear to answer it — the timing suggests the regulators were waiting for the legal architecture to catch up to the case, and now it has.

Source: 财新 (Caixin) · 新华社 (Xinhua) · 人民日报 (People’s Daily) · CAC official text

The Signal: This week’s four items — Huawei’s LogicFolding roadmap, Alibaba’s M890 launch alongside Nvidia’s zero-China-revenue confirmation, the summit’s two AI absences, and the agent rules retrofitting the Manus block — read as four observations of one structural shift. The chip layer is moving on two time horizons simultaneously; the regulatory layer is installing the architecture for operating decisions already made; the diplomatic layer is producing paper. For institutional investors, the question for the next 90 days is no longer whether a closed-loop Chinese AI stack exists. It is whether the export-control thesis underneath US policy survives Huawei’s May 25 announcement — and what that means for SWFs holding Nvidia, TSMC, and ASML, and for those building alternative stacks in the Gulf.

Beijing Signal is a free weekly dispatch from SWF Sovereign AI by Winston Ma (马文彦) — former Managing Director and Head of North America, China Investment Corporation ($1.3 trillion); Executive Director, Global Public Investment Funds Forum; Adjunct Professor, NYU School of Law; and author of The Hunt for Unicorns and The Digital War.

This week’s SWF Sovereign AI publication arc:

• Today’s free Beijing Signal (Monday weekly) surfaces the four observations.

• Tomorrow’s Capital Signal (Wednesday weekly) places them alongside parallel SWF capital flows in four other regions.

• Friday’s China AI Insider goes deep on Huawei’s LogicFolding announcement — the technical claim, the credibility gap, and the institutional-investor question if even partially viable.

Next week’s coverage shifts to the Big Fund’s $45B investment in DeepSeek (Friday Insider, June 5) and its SWF capital architecture (Capital Signal flex deep dive, June 10) — the last free issue of Capital Signal before the Substack paywall activates the week of June 15.

Subscribe now

All paid publications are currently free during the 60-day launch window. Pledge your Insider Track subscription now — no charge until the official launch in mid-June.

In this issue:

1. HARDWARE — Beijing Auto Show closes as the world’s largest — it was a tech show disguised as a car show, and the AI giants stole it from the automakers

2. REGULATION — China finalizes the world’s first AI emotional safety rules — five months after I called them “a leap from content safety to emotional safety”

3. REGULATION — A Hangzhou court rules that replacing a worker with AI is unlawful dismissal — a landmark for the world’s fastest AI-adopting economy

4. TECHNOLOGY — Four ministries publish 193 metaverse deployment cases — while the West moved on, China is building

(A quick programming note: I am heading to Asia later this week for on-the-ground meetings with industry leaders and sovereign investors. As a result, Beijing Signal will pause for the next two weeks. I look forward to returning to your inbox the week of May 24th with fresh, primary-source intelligence directly from the region.)

HARDWARE

Beijing Auto Show Closes as the World’s Largest — and the AI Giants, Not the Automakers, Defined It

Auto China 2026 closed after drawing 1.28 million visitors across 380,000 square meters — the world’s largest auto show by every measure — with “Physical AI” as the defining thesis.

The numbers are staggering: 1,451 vehicles, 181 global debuts, 2,000+ exhibitors. But the real story is who stole the show. ByteDance’s Doubao AI is now in 145 car models and over 7 million vehicles. Huawei’s Qiankun ADS 5.0 powers autonomous driving across multiple brands — Audi now uses Huawei driving systems. Xiaomi unveiled a hypercar concept while XPeng’s CEO announced humanoid robots targeting 10,000 units by 2027, displayed alongside its IRON robot and flying car at the same booth. Tesla skipped its third consecutive Beijing Auto Show. As McKinsey’s China team noted: “Chinese OEM booths told a sharply different story — drones, flying cars, humanoid robots, SoC chips, and in-car LLMs.”

Ma’s read: The Beijing Auto Show has become the CES of Physical AI — the convergence the Five-Year Plan’s AI+ mandate was designed to produce. When Audi uses Huawei’s driving systems and VW co-develops EVs with XPeng, the technology transfer direction has reversed. XPeng’s 10,000-unit humanoid robot target sits squarely in the embodied intelligence capital cycle we mapped in Issue 1’s deep dive — the same supply chain thesis, the same Shenzhen manufacturing base, now extending from standalone robots into automotive AI.

→ I commented on the broader AI integration trend in China at CGTN TV, March 2026

Source: AFP (April 24) · CGTN (May 3) · McKinsey Greater China (May 2) · Electrek (April 26) · CNBC (May 1)

REGULATION

China Finalizes the World’s First AI Emotional Safety Rules — Western Media Still Hasn’t Covered Them

On April 10, the CAC published final Interim Measures for AI Anthropomorphic Interactive Services (《人工智能拟人化互动服务管理暂行办法》), effective July 15, 2026.

This is the regulation Western media treated as a curiosity when it was drafted in December 2025. The finalized version has received virtually no English-language coverage — yet it reshapes how every AI chatbot, companion app, and digital human operates in China. Key requirements: mandatory alerts that users are interacting with AI, two-hour continuous session limits, emotional state assessment with intervention protocols, and guardian consent for minors. Platforms with over one million users must undergo security assessments. To watch: AI startups like Minimax (Talkie) and Zhipu (Z.ai) who had Hong Kong IPOs recently — both directly covered by these measures.

Ma’s read: When the draft was released in December, I told CNBC that these rules mark “the world’s first attempt to regulate AI with human or anthropomorphic characteristics” and that this version “highlights a leap from content safety to emotional safety.” Five months later, the final version is stronger than the draft — and the US still has no equivalent framework.

→ I commented at CNBC, December 29, 2025

Source: CAC 政策法规 (April 10, 2026) · CAC 网信要闻 (April 10–17) · CNBC (December 29, 2025)

REGULATION

A Hangzhou Court Rules That Replacing a Worker with AI Is Unlawful Dismissal — a Landmark for Labor Rights in the AI Era

The Hangzhou Intermediate People’s Court ruled that a tech firm’s dismissal of a quality assurance supervisor — after his role was automated by AI — was unlawful, upholding a lower court decision and arbitration ruling.

The case was published on April 28, ahead of International Workers’ Day. The worker, identified as Zhou, earned 300,000 yuan ($43,900) annually verifying AI-generated answers. When the company automated his role, it offered a reassignment with a salary cut from 25,000 to 15,000 yuan per month. Zhou refused. The court ruled that AI-driven job replacement does not constitute a “major change in objective circumstances” under China’s Labor Contract Law — and that the pay cut was unreasonable. A separate Beijing case last year reached the same conclusion: a data mapping worker replaced by AI also won his arbitration. The pattern is now a precedent.

Ma’s read: As a licensed attorney in both China and New York, I read this as Beijing managing the social contract alongside the AI mandate. The Five-Year Plan demands 70% AI integration by 2027 — but courts are ruling that companies cannot shift the cost of technological transformation onto individual workers. This is the same pattern China used during the manufacturing automation wave of 2015–2018: retrain first, then restructure. In the US, by contrast, Meta just cut 8,000 jobs to cover AI infrastructure costs — with no equivalent legal constraint.

Source: Hangzhou Intermediate People’s Court (April 28) · NPR (May 1) · Bloomberg (May 2) · Caixin (April 30) · Xinhua

TECHNOLOGY

Four Ministries Publish 193 Metaverse Deployment Cases — While the West Moved On, China Is Building

On April 30, MIIT, the Ministry of Education, the Ministry of Culture and Tourism, and the National Radio and Television Administration jointly published 193 “typical metaverse cases” for 2025.

In the US, “metaverse” is a word most investors associate with Meta’s $46 billion Reality Labs loss. In China, the concept never had that association. The Chinese “元宇宙” framework is industrial: manufacturing digital twins, medical training, cultural heritage preservation, education platforms, smart city management. The 193 cases — up from 150 in 2024 — were selected through expert review from provincial submissions. The four-ministry joint notice instructs local authorities to “increase support for selected cases in technology innovation, application deployment, and government services.”

Ma’s read: Western investors wrote off the metaverse after Meta’s losses. Beijing never agreed. Four ministries co-endorsing 193 deployment cases is procurement pipeline, not policy theatre. Each case creates a reference architecture that state enterprises and local governments will replicate. In my book Blockchain and Web3: Building the Cryptocurrency, Privacy, and Security Foundations of the Metaverse (Wiley, 2022), I argued that the metaverse’s durable value would be industrial, not consumer. China is proving that thesis at national scale.

Source: MIIT 最新政策 (April 30, 2026) · 工信厅联科函〔2026〕164号

The Signal: This week’s four items share a thread that Western coverage consistently misses: China is not just building AI — it is governing AI. Emotional safety rules for chatbots. Labor rights protection alongside AI adoption mandates. Industrial metaverse deployment at ministerial scale. And an auto show where the tech companies — not the carmakers — defined the future. The governance layer is what converts AI capability into durable industrial advantage. The US has more frontier models. China has more rules, more standards, more deployment cases, and a court system already adjudicating the social contract of AI adoption. For SWF funds and institutional investors: the governance layer is where the long-term moat is being built.

Subscribe now

Beijing Signal is a free weekly dispatch from SWF Sovereign AI by Winston Ma (马文彦) — family office AI investor and former Managing Director - Head of North America Office, China Investment Corporation ($1.5 trillion); Executive Director of Global Public Investment Funds Forum and Adjunct Professor, NYU School of Law; and author of The Hunt for Unicorns and The Digital War.

The weekly China AI Insider goes further: it covers the same four stories here fully documented — Chinese primary sources translated, capital implications mapped, forward indicators identified — plus additional high-value items that emerged after this dispatch, and a 2,000-word institutional deep dive on the week’s most consequential development. More stories. Deeper analysis. One product to work from — See previous issue here.

In this issue:

1. DeepSeek V4 on Huawei Ascend answers the hardware question this newsletter asked last week

2. Beijing blocks Meta’s $2 billion Manus acquisition — the NDRC draws the line on AI technology exports

3. A US diplomatic cable weaponises “distillation” as a pre-summit negotiating instrument

4. MIIT extends China’s AI compute ambitions to orbit — and links them to green power

TECHNOLOGY

1. DeepSeek V4 on Huawei Ascend answers the hardware question this newsletter asked last week

DeepSeek released preview versions of V4-Pro (1.6 trillion parameters) and V4-Flash (285 billion parameters) on Friday — its first major model optimised for Huawei’s Ascend chips rather than Nvidia hardware.

Last week’s Beijing Signal framed V4’s delay as a hardware certification test, not a capability story. Friday confirmed the frame: Huawei announced day-zero SuperNode compatibility across its Ascend 950 series, and SMIC the chipmaker surged 10% in Hong Kong as investors repriced the domestic chip value chain. The model trails frontier closed-source systems by three to six months on DeepSeek’s own benchmarks — but it runs production inference on domestic silicon.

Winston Ma’s read: DeepSeek V4 on Ascend (probably to be manufactured by SMIC) is the first full-stack validation of “科技自立自强” (S&T self-reliance) at frontier model scale — and for fund managers still asking whether Chinese AI can perform without Nvidia, Friday was the answer.

Source: DeepSeek V4 technical report · Huawei SuperNode press release · 机器之心 (Synced) · MIT Technology Review

REGULATION

2. Beijing blocks Meta’s $2 billion Manus acquisition — the NDRC draws the line on AI technology exports

China’s NDRC today issued a one-line order prohibiting foreign investment in the Manus acquisition and requiring all parties to withdraw the deal — killing Meta’s $2 billion purchase of the Chinese-founded AI agent startup.

The decision was expected but its bluntness was not — no explanation, no named parties, just a prohibition citing “laws and regulations.” Manus is not an isolated case: the Washington Post reported last week that Beijing has also directly warned MiroMind, Shanda Group’s AGI lab, against sending talent and research abroad — a broader crackdown on what the Chinese tech sector now calls “China shedding.” For every AI startup weighing a Singapore exit, the compliance threshold just rose sharply — weeks before the Trump-Xi summit.

Winston Ma’s read: As a NYU School of Law Adjunct Professor, I commented at the WSJ report on the deal announcement day that if the Meta-Manus deal closed smoothly (without US “reverse-CFIUS” interference), “it creates a new path for young AI startups in China.” Now Beijing just closed that path — China’s CFIUS.

Source: 国家发展改革委 (NDRC) security review decision (27 April) · Reuters · Bloomberg · Washington Post

POLICY

3. A US diplomatic cable highlights “distillation” as a pre-summit negotiating instrument

On Friday, the US State Department sent a cable to embassies worldwide instructing staff to warn foreign governments about “concerns over adversaries’ extraction and distillation of US AI models” — naming DeepSeek, Moonshot AI, and MiniMax.

The cable’s timing — same day as V4’s launch, three weeks before Trump visits Xi in Beijing on May 14–15 — transforms an industry grievance into a diplomatic instrument. The White House OSTP memo two days earlier accused Chinese entities of “industrial-scale” distillation campaigns, and the cable instructs embassies to “lay the groundwork for potential follow-up and outreach” — language for building a multilateral coalition. China’s foreign ministry called the accusations “groundless.”

Winston Ma’s read: As a lawyer admitted in both the US and China, I read the cable’s legal language carefully: it targets the process of distillation, not the product — V4’s open-weight release under MIT licence is perfectly legal. The real question is whether the Trump administration uses the summit to demand Beijing regulate distillation as a condition for chip sanction relief.

Source: Reuters (exclusive) · CNBC · White House OSTP memo (via Financial Times)

HARDWARE

4. MIIT extends China’s AI compute ambitions to orbit — and links them to green power

MIIT Vice Minister Zhang Yunming announced support for forward-looking research on space-based computing (太空算力) and coordinated deployment of green electricity and AI compute infrastructure.

Most coverage of China’s AI stack focuses on chips and models. MIIT is extending the frame in both directions — downward to energy, and upward to orbit. The 东数西算 (Eastern Data, Western Computing) programme already links western China’s renewable resources to eastern AI demand; ADAspace’s “Star Compute” constellation of 2,800 satellites launched its first group in May 2025 and has since demonstrated the world’s first satellite-controlled ground robot, with second and third batches scheduled for 2026.

Winston Ma’s read: When I wrote about the 5-Layer AI Cake — where state capital meets national capability — I focused on chips, models, and applications. MIIT is adding a new dimension: orbital infrastructure (and its energy solution). China has 339 GW of wind and solar under construction, roughly two-thirds of the world total — energy is not only an input cost factor (to watch: token exports), but also a strategic advantage.

Source: 工信部 (MIIT) press conference via 证券时报 (Securities Times) · 新华社 (Xinhua) “From ground to orbit” · Global Times

The Signal: V4 on domestic chips. A $2 billion US acquisition killed by a single NDRC sentence. A State Department cable weaponising “distillation” as a diplomatic instrument. And an MIIT vice minister quietly extending China’s compute ambitions to orbit. These are not separate developments — they are the four walls of the same room: a national AI sovereignty architecture being built, defended, and extended simultaneously. The Trump-Xi summit on May 14 walks into this room. The agenda is already written.

Beijing Signal is a free weekly dispatch from SWF Sovereign AI by Winston Ma (马文彦) — family office AI investor and former Managing Director - Head of North America Office, China Investment Corporation ($1.3 trillion); Executive Director of Global Public Investment Funds Forum and Adjunct Professor, NYU School of Law; and author of The Hunt for Unicorns and The Digital War.

The weekly China AI Insider goes further: it covers the same four stories here fully documented — Chinese primary sources translated, capital implications mapped, forward indicators identified — plus additional high-value items that emerged after this dispatch, and a 2,000-word institutional deep dive on the week’s most consequential development. More stories. Deeper analysis. One product to work from — See previous issue here.

Currently free during our 60-day launch window. 👇 Pledge your China AI Insider subscription now — no charge until the official launch in June.

As the global AI race accelerates, tracking Silicon Valley is no longer enough. You have to track the sovereign capital, the state-backed infrastructure, and the regulatory mandates shaping the other half of the digital world.

Beijing Signal is the free, distilled layer of this SWF Sovereign AI Substack. It is designed for speed, clarity, and signal over noise.

Behind it sits our flagship publication: China AI Insider. 🔥

If Beijing Signal operates as your macro headline feed, China AI Insider is your private research desk. Each cycle expands into:

• A broader set of curated intelligence.

• Deeper analytical frameworks mapping the Sovereign AI ecosystem.

• A dedicated deep-dive on one critical theme per issue.

We are currently in the 60-day launch window, meaning all China AI Insider deep-dives are fully accessible. Following this period, China AI Insider will shift behind the paywall as the primary venue for full analysis.

If Beijing Signal captures the headlines, China AI Insider explains what they mean—and what comes next.

In this issue:

1. A humanoid robot beat the human half-marathon world record — here is what Western coverage missed

2. A $10 billion DeepSeek valuation prices an option, not a cash flow

3. DeepSeek V4’s delay is a hardware certification test, not a capability story

4. The 15th Five-Year Plan’s AI+ mandate sets the policy architecture for the rest of this decade

HARDWARE

1. A humanoid robot beat the human half-marathon world record — here is what Western coverage missed

On April 19, Honor’s “Lightning” (闪电) robot completed Beijing’s E-Town Half Marathon in 50 minutes and 26 seconds — faster than any human has ever run the distance.

Western outlets led with the spectacle — robots racing humans, a smartphone brand making androids. What they missed was the investment thesis embedded in the result. One year ago, the winner of the inaugural Beijing humanoid robot marathon finished in 2 hours 40 minutes. This year: 50 minutes 26 seconds. That is not incremental progress — it is the kind of performance compression that, in semiconductors, preceded commercial deployment by roughly 18 months.

Honor, which spun off from Huawei in 2020, ran its robot through a course ending at Beijing’s National IT Innovation Park (国家信创园) — a location not chosen by accident. Over 300 robots from 100+ teams competed; 40% navigated autonomously. The event was organised by Beijing Yizhuang, the state-backed zone at the centre of China’s embodied intelligence (具身智能) investment cluster.

Winston Ma’s read: China’s $138 billion National Venture Capital Guidance Fund directed 38% of its 2025 AI allocation to robotics — and yesterday’s race is what that capital is producing. The sovereign fund parallel I keep returning to: when Gulf SWFs funded data centres three years ago, Western analysts called it speculative infrastructure. Today those centres run AI at scale. China’s robotics investment is at the same inflection point. The question is not whether Chinese humanoid robots will become commercially significant. It is when — and whether you have exposure before that answer is obvious.

Source: 新浪财经 (Sina Finance) · 界面新闻 (Jiemian) · Beijing E-Town WeChat

CAPITAL

2. A $10 billion DeepSeek valuation prices an option, not a cash flow

DeepSeek is reportedly raising fresh capital at a $10 billion valuation — despite no significant commercial revenue and largely open-weight models distributed without charge.

The figure circulated this week on 暗涌 (Waves) and 36氪 (36Kr) without confirmed deal terms. Standard valuation frameworks produce the wrong answer here: DeepSeek has no meaningful revenue base, its models are open-weight, and its commercialisation strategy remains undefined.

The $10 billion reflects option value — specifically, the option that DeepSeek’s V4 becomes the reference architecture for China’s full-stack AI independence if it launches successfully on Huawei’s Ascend chips. Priced this way, the question is not “what is DeepSeek worth today?” but “what is it worth if it proves Chinese AI doesn’t need Nvidia?” That option is not zero. The more important data point than the valuation itself: who is pricing it. The composition of the investor base will tell you whether Beijing has decided DeepSeek is a national strategic asset.

Winston Ma’s read: At CIC and Wall Street, I structured option investments constantly — modest base case, transformative tail scenario. DeepSeek at $10 billion is exactly that architecture. I’d be less focused on the number and more focused on the cap table when it’s disclosed. That will tell you everything about what this valuation actually means.

Source: 暗涌 (Waves) · 36氪 (36Kr)

TECHNOLOGY

3. DeepSeek V4’s delay is a hardware certification test, not a capability story

DeepSeek’s anticipated V4 model remains unlaunched weeks after expected release, with credible reporting indicating it will run on Huawei’s Ascend chips — and that China’s major platforms have already placed large preparatory orders.

The Information reported that V4 can run on Huawei’s latest Ascend chips; five sources with direct knowledge described pre-launch orders from Alibaba, ByteDance, and Tencent. Coverage has focused almost entirely on whether V4 will be competitive with frontier Western models. The more consequential frame is the hardware substrate.

If V4 launches at production scale on domestic chips, Beijing’s 科技自立自强 (”S&T self-reliance”) directive — written into every major policy document since 2020 — moves from aspiration to demonstrated engineering fact. The delay itself may signal the Huawei path is harder than advertised, or that DeepSeek is choosing its announcement moment deliberately.

Winston Ma’s read: Fund managers tracking US export controls typically ask whether China is still acquiring Nvidia chips. V4’s launch reframes the question entirely: can China perform without them at scale? The hardware conditions of this launch — not the benchmark scores — are what the next round of US export control policy (the Governance layer) will be written around.

Source: 机器之心 (Synced) · 量子位 (QbitAI) · The Information

POLICY

4. The 15th Five-Year Plan’s AI+ mandate sets the policy architecture for the rest of this decade

China’s 15th Five-Year Plan (2026–2030), submitted to the National People’s Congress in March, elevates “AI+” as a core national strategy alongside explicit mandates for compute sovereignty, embodied intelligence, and semiconductor self-reliance.

Beijing is deliberately running a different AI race from the United States. Where Washington bets on closed frontier labs and massive private compute clusters, the Five-Year Plan builds an integrated stack of policy, standards, state capital, and hardware localisation that prioritises resilience and industrial application over benchmark leadership.

The State Council’s AI+ directive (August 2025) targets 70% AI integration across six key domains by 2027. CNNIC’s latest census in the Spring 2026 records Chinese industrial enterprise AI penetration already at 47.5% in 2025 — up from 9.6% the year before — suggesting the deployment base is ahead of the plan’s own targets. The divergence in innovation models between the US and China is now sharp enough to be a portfolio-level variable — and the gap is widening..

Winston Ma’s read: I’ve watched China execute Five-Year Plans for twenty years. The 15th is the first one where AI is not a sub-theme of something else — it is the primary organising logic of the entire document. When Beijing makes that architectural decision, state capital, procurement mandates, and regulatory alignment follow at speed. For sovereign investors, the question is no longer whether to have China AI exposure. It is which layer of the “5-Layer AI Cake” benefits most from decade-scale state coordination.

Source: 国务院 (State Council) AI+ directive · 新华社 (Xinhua) 15th Five-Year Plan coverage · CNNIC 第57次互联网发展统计报告

The Signal: Yesterday’s robot marathon, this week’s DeepSeek valuation, the V4 hardware question, and the Five-Year Plan are not four separate stories. They are four readings of the same instrument: a coordinated national AI infrastructure build, funded by sovereign capital, governed by a Five-Year Plan, and now physically visible on a Beijing racetrack. The compression from 2h40m to 50m26s in a single year is the most concrete data point yet for how fast this build is moving. Model the hardware scenario now, before DeepSeek V4’s launch conditions make it undeniable.

Beijing Signal is a free weekly dispatch from SWF Sovereign AI by Winston Ma (马文彦) — former Managing Director and Head of North America, China Investment Corporation ($1.3 trillion); Executive Director of Global Public Investment Funds Forum and Adjunct Professor, NYU School of Law; and Author of The Hunt for Unicorns and The Digital War.

The weekly China AI Insider goes further: it covers the same four stories here fully documented — Chinese primary sources translated, capital implications mapped, forward indicators identified — plus additional high-value items that emerged after this dispatch, and a 2,000-word institutional deep dive on the week's most consequential development. More stories. Deeper analysis. One product to work from.

Subscribe now

Currently free during our 60-day launch window. Pledge your China AI Insider subscription now — no charge until the official launch in June. Thank you!