---

Rules, personnel, and money are all moving at once. OpenAI released a global AI standards proposal, and on the same day, British Columbia in Canada sued them over a failure to warn about a shooting; on the domestic front, Alibaba appointed a new head for Qwen on the eve of the Yunqi Conference, while Ant Group merged its payment, digital payment, and Sesame Credit units into one business group. Money is still pouring in: Nvidia bought another $1.5 billion worth of energy company shares, Moonshot's Kimi K3 appeared on AWS for the first time, and Meta open-sourced an internal resource scheduling tool it has used for ten years.

---

VI. Macro and Market Data

On the eve of the Yunqi Conference, Alibaba appoints a new head for Qwen

Huxiu reported that on September 21, Alibaba appointed senior AI researcher Liu Dayiheng as the project lead for the Qwen large model. He has a background as a Huawei "Genius Youth." On the official guest list for the Yunqi Conference, his photo appears after Joe Tsai and Eddie Wu, with the conference opening the next day. Placing the large model lead in the front row of conference guests is Alibaba's external signal: Qwen is shifting from a technical project to a business line responsible for revenue. Personnel changes often precede rhythm changes; update frequencies for open-source versions and pricing for enterprise editions may adjust accordingly. Qwen is a free AI assistant used by many on their phones, and a batch of developers use its open-source version for secondary development. Watch two things next: will the free quota shrink, and will commercial terms for open-source licenses tighten?

Kimi K3 lands on AWS, Chinese open-source models try charging fees for the first time

SCMP reported that Moonshot AI's Kimi K3 is now listed on AWS. These "open-weight" models feature public weights that anyone can download, copy, and modify, but this time it is being sold as a service based on call volume on overseas cloud platforms. In the past, Chinese open-source models relied mainly on download counts and reputation, gaining attention rather than income. Entering AWS means facing overseas corporate procurement processes directly. Whether they can collect payments based on usage is the real validation needed for this round of going global. If you are doing overseas business, you can now directly call models trained by Chinese teams on AWS, saving the hassle of self-deployment and adding another option for price negotiation.

OpenAI wants the US government to lead global AI standards

OpenAI released a safety proposal for frontier AI on Monday, focusing on two areas: "alignment," meaning ensuring model behavior matches human intent; and who should supervise models once they can improve themselves. The company advocates for the US government to lead the push for a set of standards shared by all countries. The proposal is not binding, but it reveals the preferences of top companies: rather than being legislated separately by various countries, it is better to establish universal rules first. For model companies planning to go global, if these rules become de facto standards, safety evaluations and documentation requirements will follow. Ordinary people will feel little short-term change, at most seeing more uniform and verbose refusal prompts for sensitive topics in AI products. What really matters is whether this becomes a threshold for cross-border compliance.

Failure to warn about shooting leads Canadian province to sue OpenAI

Bloomberg reported that the government of British Columbia, Canada, is suing OpenAI, with "failure to warn" as the core allegation: the province believes the company did not issue the necessary alerts before the shooting occurred. Specific claims and compensation amounts were not disclosed in reports. Past debates about "whether AI should report dangers when seen" stayed at the public opinion level; this time it has entered the courts. If the allegations hold, model companies will have to pay for some public safety liabilities, and the cost structure for content moderation and reporting processes will be rewritten. For users, AI assistants might pop up safety warnings more frequently or even proactively contact police when detecting dangerous content. The boundary between convenience and privacy will be pushed again; think clearly about which trade-off you can accept before using them.

• IT Home: Ant Group merges Alipay Business Group, Digital Payment Business Group, and Sesame Credit into a new Alipay Business Group; Wu Minzhi appointed President and CPO; CEO Han Xinyi sends an all-staff letter explaining the purpose — This move by Ant is about organization, not product. After merging the three business groups, payments, credit, and digital services report along a single line, eliminating internal resource competition. The announcement specifically mentions "Agent Commerce," referring to letting AI place orders, pay bills, and book tickets for users. The barrier for such businesses isn't model capability, but transaction chains and risk control—Ant's traditional strengths. Worth noting is the assessment metric: once the number of transactions completed by AI enters business KPIs, AI entry points in payment scenarios will visibly increase. Organization moves first, followed by product and price adjustments; this has been the common sequence for domestic tech giants' restructuring in recent years.
• SCMP: China's low-altitude economy faces subsidy reduction for the first time — Half of the hype around the low-altitude economy comes from local governments. Once subsidies stop, the math becomes clear: per-aircraft costs, airspace approval times, and revenue per order must be borne independently. Hangzhou's choice indicates that local finances are also tightening, making funding for demonstration projects harder to get. For logistics and inspection companies, this is actually an opportunity—operators can no longer rely on subsidies to undercut prices and grab orders; teams that honestly calculate costs can secure steadier orders. Watch whether Shenzhen, Hefei, and other cities betting on low-altitude sectors follow suit. Most orders in the low-altitude economy are held by governments and state-owned enterprises; after subsidy reductions, bidding prices will better reflect true costs.
• Bloomberg: Nvidia invests another $1.5 billion in SB Energy shares, entering before its IPO. Electricity is currently the biggest bottleneck for data centers; locking down power sources via equity means customers ask about electricity availability before buying chips — This money buys electricity, not stock returns. Data center power applications are queued in multiple states, grid expansion takes 3-5 years, and Nvidia uses equity to pre-lock power sources, solving the slowest link for customers. SB Energy is rushing to IPO; adding investment now secures a valuation discount and insures Nvidia's chip sales. To track this trend, watch two numbers: whether power infrastructure corresponding to AI server orders is secured, and whether grid interconnection schedules in US states loosen. Nvidia's capital flowing from chips to power is the most overlooked shift in this wave of AI investment.
• DevOps.com — The enforceability of open-source licenses relies on court rulings. If judgments confirm that training data usage can be more lenient, corporate legal teams will reassess whether code can be used for free, and commercial terms for open-source projects will become stricter; conversely, it will become harder for open-source authors to monetize via licenses. This news is currently only at the stage of "developer concern," with case details and verdicts not yet public. Product teams don't need to change licenses immediately but should start compiling lists of training data sources.
• Meta Engineering Blog — By open-sourcing tools like this, Meta is essentially sharing the pitfalls it encountered. Resource scheduling seems far from ordinary users but actually determines server costs: with good scheduling, the same batch of machines can run 30% more tasks. Meta has used it to manage company-wide resources for over ten years, proving its validity in ultra-large-scale scenarios. For companies building their own data centers, there is now a reference implementation that can be directly modified; for cloud providers, scheduling capability shifts from a selling point back to basic infrastructure.
• Tim Dettmers — Compressing large models to run on a single machine changes the barrier to experimentation. Previously, creating a domain-specific model required applying for budgets and queuing for GPUs; now a high-spec workstation can run through the process, accessible to small teams and students. The trade-off is accuracy and speed; overly compressed models can only handle simple tasks. This path offers additional benefits for domestic teams: data doesn't leave locally. The watershed moment is whether long, stable training runs are possible, not just one-time demos. With single-machine feasibility, the open-source model ecosystem will spread to two types of users: small teams in vertical industries and enterprises unwilling to hand over data.
• Nature — The scientific community's attitude toward AI is more complex than industry's. Tools do save time: reading literature, running simulations, and organizing data—tasks previously handled by PhD students are now taken over by software. Disagreements lie in authorship and responsibility—who is liable if a paper has issues, and how reviewers determine which steps were done by humans. Such controversies won't have unified answers soon, but new standards are already appearing in hiring and grant reviews: proficiency with AI tools is becoming a criterion parallel to research ability.
• CNBC: Options trading surges around the New York Times AI copyright lawsuit — Options data provides an early read on institutional sentiment. Increased bets around the NYT suggest the market believes the lawsuit outcome will significantly impact cash flow: winning enhances pricing power for content licensing; losing means payouts and legal fees eat into content revenue. Such lawsuits typically span years, so short-term stock prices are more influenced by quarterly subscription data. Media stock valuation logic is shifting from ads plus subscriptions to content assets plus litigation odds.
• TechCrunch — Leaving AI companies to become intermediaries in traditional industries is a common choice for a cohort of people in the last two years: as model capabilities become public goods, scarcity lies in industry relationships and process knowledge. Health benefit brokerage profits come from premium splits and administrative services, largely unrelated to model capabilities; AI's role is compressing the time for quote verification and claims processing. These companies have modest valuations and steady cash flows, suitable for long-term tracking. Risks lie in regulation: US health insurance brokerage licenses and commission rules are managed at the state level.
• TMTPost: Anew Labs hailed as experts in "distillation" — Speed is the competitive metric in AI drug discovery these days. Using small models to inherit large model capabilities lowers costs, allowing pipelines to expand faster, though originality suffers. The pharma investment world loves and hates this approach: it quickly produces decent molecules, but post-IPO faces hard tests regarding patents and clinical data. For domestic innovative drug companies, following this path is a business decision, not a technical one. The real variable is the clinical phase; AI saves time and money upfront but cannot avoid the cost of Phase III trials. Some call this "engineering innovation": use existing capabilities to build products first, then discuss originality.
• The widely circulated figure that "80% of AI projects fail" was traced back to a footnote. Author Jamie Watters followed the citation chain upward and found the original source couldn't support this ratio — This number is repeatedly cited because it's useful: writing proposals, requesting budgets, or persuading bosses to halt projects—a phrase like "80% will fail" wraps it up. Tracing to the source reveals it was just a footnote, and the original source didn't support the ratio. This serves as a practical reminder for those implementing AI: keep records of your own project data. When asked about success rates, teams with internal metrics are more persuasive than those citing industry figures.
• Bloomberg — The story of AI's electricity consumption has been told for a year; the story of copper usage is just beginning. Busbars, transformers, and cooling pipes in data centers are made of copper, and grid expansion consumes even more. Unlike chips, copper supply elasticity is low: new mines take 7-8 years from exploration to production, and short-term price hikes won't expand capacity. This is why commodity markets are refocusing on copper. Risks lie on the demand side—if data center construction slows, the copper price story will end before the electricity price story. For ordinary investors, commodity trends are now more tightly linked to AI stocks than in the past.
• Tech.eu — Selling software to banks is a slow business but a good one. Banks have long system replacement cycles; once implemented, switching is difficult, leading to high renewal rates. Valuations for such companies depend on customer retention and revenue per client, not growth slope. FintechOS raising funds via equity plus debt suggests the founding team doesn't want to dilute too much ownership at a low valuation. As more European companies emerge, Chinese banking tech suppliers face new competitors overseas.
• - Including the "right to use advanced AI" in draft legislation indicates a split in US domestic attitudes toward AI: one faction wants regulation, the other fears regulation will restrict access. The draft is far from becoming law, but its framework is worth remembering—it limits supply-side barriers while protecting individual and small business access. If such clauses land, they will directly change how capabilities are tiered between enterprise and personal versions.
• A practical guide on Medium — EU AI Act compliance requirements are tiered; higher risk levels mean heavier documentation and testing. Most guides discuss principles, but this one provides checklist-style steps suitable for self-audits. For teams going global, the trouble isn't technology but record-keeping: training data sources, test cases, and manual review processes must be retrievable. Setting up processes early is much easier than scrambling for materials when compliance is demanded. Domestic regulations have similar checklists; teams exporting products can compare both sets.
• deslop-skills on GitHub — AI-generated interfaces and copy share common traits: identical rounded corners, identical parallel structures. This repo's idea is turning "de-flavoring" into reusable rules, helping assistants bypass these clichés during generation. It costs nothing for individual developers—just copy a few rules into their workflow. It also highlights another fact: homogenization of AI output is obvious enough to require specialized cleanup tools, making aesthetic judgment and discernment more valuable. Tools can remove surface-level clichés, but deciding which copy to keep and which button to delete still requires humans.

---

Today's Market Snapshot (Self-produced quotes, Global 24 hours (US stocks close on Mon, 9/21 + A-shares close on Mon, 9/21))

US stocks closed broadly higher on Monday, with AI heavyweights leading: Meta rose over 11% in a single day, followed by gains in Nvidia and Palantir. Brent crude oil and copper prices strengthened simultaneously, indicating continued capital expectations for AI-driven electricity and copper demand. Closing points for the three major indices were not publicly confirmed at the time of publication; individual stock prices reflect the 9/21 close.

The pattern of strong hardware and weak applications remains unchanged. Two optical module stocks and domestic chip makers closed higher, while Cambricon and Kingsoft Office remained flat. Today's two major events were on the application side—Alibaba's leadership change and Ant's restructuring—but before revenue materializes, capital prefers segments that can deliver servers.

---