---

Today's Main Thread

The AI public opinion sphere this weekend was ignited by one event: Anthropic CEO Dario Amodei published a long-form article calling for the industry to "hit the brakes" on frontier models. He proposed a three-step plan involving opening models to third-party evaluation agencies, establishing capability threshold trigger mechanisms, and promoting international coordination. Elon Musk and Sam Altman rarely voiced support for the discussion. This marks the first time a leader of a frontier lab has written "proactive deceleration" into a systematic industry proposal, rather than scattered warnings.

However, capital and policy gave diametrically opposite answers—Oracle founder Ellison halted the sale of up to 50 million shares on the same day; the White House explicitly rejected slowing down via regulation in exchange for security, placing leadership over China at the top of the agenda; domestically, "Compute Token Loans" were pushed nationwide, using financial tools to leverage compute expansion. Initiatives cannot change the competitive structure; they only change the probability distribution of policies. Meanwhile, real money and credit issuance are casting a vote against the "deceleration theory" regarding the sustainability of the AI capex cycle.

The following source draft is filtered from an investment perspective, focusing on financing, equity actions, and industrial capital movements, comprising 4 headlines and 15 selected items; market windows cover the most recent closes for A-shares and US stocks. Two observation points on this week's schedule: September 14 pre-orders for the new WEY Gaoshan model (pricing validation for high-compute penetration into MPVs), and Q3 earnings capex guidance from major US AI weights (the first formal test of the judgment that "the spending cycle hasn't peaked").

---

White House Sets Tone: No Brakes for AI; Trump Downplays Doom Scenarios, Prioritizing Lead Over China

Anthropic CEO Amodei published a long article on Saturday proposing a three-step initiative to slow down frontier models—opening models to third-party evaluators, establishing capability thresholds, and pushing for international coordination. Trump responded publicly over the weekend: he will not slow AI development due to safety concerns, with the policy focus ensuring US AI superiority over China. This implies the US regulatory path will likely continue to lean towards deregulation and speed racing, further diverging from Europe's preference for ex-ante compliance. Transatlantic rule arbitrage opportunities will persist. At the market level, regulatory discount expectations for leading model companies and compute infrastructure providers are narrowing; conversely, third-party firms relying on compliance audit and safety assessment narratives face demand-side skepticism—their commercial prospects are more tied to EU and state-level legislation than federal levels.

Larry Ellison Hits Stop: Cancels Sale of Up to 50 Million Oracle Shares

According to Bloomberg, Oracle Founder and CTO Larry Ellison canceled his previously disclosed plan to sell up to 50 million shares. The background is that Oracle is facing cash flow pressure while heavily investing in AI data centers and has included this round of layoffs in its 2026 restructuring plan, raising total costs to approximately $2.8 billion. Founder selling is typically a dual negative signal of supply pressure above the stock price and confidence. Halting it during a weekend when debates about peaking AI capex are heating up was directly interpreted by trading desks as insider endorsement of the spending cycle's sustainability; more broadly, concerns about peaking across the entire AI infrastructure supply chain—including database clouds, compute leasing, and data center construction—received a tangible counter-evidence. Note: The founder's personal liquidity arrangements and the company's capex plans are not the same ledger. This halt is closer to a signal trade than fundamental information, and its validity awaits verification with Oracle's next guidance update.

---

I. Large AI Models

• Google Accused of Copying Open Source Code in Mobile AI Tools — According to Neowin, community members accused Google's mobile AI tools of directly appropriating open-source project code without attribution per licenses, demanding removal of infringing components. Google has not yet responded. If confirmed, this will drag down the rollout pace of the Gemini edge ecosystem and serve as the latest footnote to the controversy over boundaries between big tech self-development and open source: beyond the model layer, compliance costs for toolchains and code layers are being re-calculated into hidden liabilities of big tech AI products. License audits may even become a new category of enterprise service projects.
• Schneier's Long Article Deconstructs Concept Shells: LLMs Are Real, AI Is Fake — The core argument is that after stripping away marketing shells, verifiable model capabilities and anthropomorphized "intelligence" are two different coordinate systems; safety discussions must be built upon the former. The article was the most shared sobering reminder in the tech community over the weekend, fermenting alongside Amodei's deceleration initiative, constituting two hedging forces in industry sentiment: one side exaggerates loss of control, while the other reminds the public to clarify the subject first.
• Anthropic's Deceleration Initiative Gains Rare Cross-Camp Volume — Musk and Altman successively expressed support for discussing deceleration; Altman further conveyed to employees that OpenAI is open to slowing down. However, the White House has explicitly rejected adoption, while Europe continues to push forward ex-ante compliance frameworks. Impact on valuations is concentrated in repricing regulatory discount expectations, not any substantive change in R&D pace—the initiative itself cannot change the competitive structure, but changes the probability distribution of policies, as well as the weight of "safety narratives" in bidding and government procurement.

II. AI Software

• Tencent May Be Required to Rename WorkBuddy — The English product name faces obstacles in both mass market propagation and search within China; internal discussions are underway for a consumer-facing rename. The naming logic of efficiency tools "going overseas first, then feeding back to domestic markets" is being re-examined: as AI assistants move from geek novelty to daily use for all, brand language distance equals conversion rate. Localization at the entry point is becoming an organizational issue for big tech AI product lines.
• HarmonyOS Version of Tencent Video Debuts Edge-Side A2A Capability — A single command wakes up system-level Xiao Yi and app collaboration, scheduling drama-watching services across apps. Edge-side agent protocols are beginning to enter real scenarios of head content applications. As the entry point shifts from Apps to system-level assistants, the distribution logic and ad slot ownership of content platforms will be rewritten accordingly.
• More Managers Negotiate Salary/Performance After Rehearsing with AI — A CNBC survey reveals that managers commonly rehearse scripts with AI before delivering difficult HR decisions like performance gaps, salary refusals, or resignation advice. Algorithms have substantively entered the HR decision chain. Who is responsible for AI-generated negative feedback wording, and whether employees can request disclosure, has become a new point of contention in labor-management boundaries.
• First Horizon Call Center Introduces AI Quality Inspection and Agent Assist — Management reported significant improvements in call sentiment metrics and customer satisfaction, with synchronized shortening of agent training cycles. Regional banks are the hardest sector to scale the AI cost-reduction narrative. The value of this case lies in providing quantifiable operational metrics; if unit call costs for customer service across the industry enter a downward channel, the proportion of generative AI in bank tech spending will become a new observation metric in earnings season.
• AI Agents Scale Classical Concert Scraper to 70,000 Events — ClassicalBot used agents to expand scraping, covering 4,000 venues and 70,000+ future events. Previously, aggregating such long-tail data was an engineering burden only affordable by commercial companies. It validates the scalable operating cost curve for individual developers under agent workflows and hints that the distribution value of vertical data assets is being repriced.
• Agentica Launches Six-Agent Company Pipeline — Research, planning, building, and deployment are divided among six AI agents, going from a one-sentence requirement directly to landing page and app launch. This is a productized sample of the "one-person company" narrative, and also a direct impact on the business model of low-code platforms: when delivery itself is agenticized, pricing power for selling tools will quickly shift to selling results.
• Six-Person Team Uses AI to Produce Sci-Fi Feature 'Deviant' End-to-End — Another case of commercialization for feature-length generated content. Creators stated that a six-person team completed production stages traditionally requiring hundreds of people in an extremely short cycle. Cost collapse and copyright chain gaps remain two major question marks, but streaming and distributors' trial purchases of "AI-native films" have begun. The unit supply cost curve of the content industry deserves re-evaluation.
• AI Word Guessing Game Synapse, Health Subscription Coach Roger, Tattoo Planning Flow TattooPreview Launched Intensively — Low-cost acquisition experiments for generative micro-products are still emerging in batches: teams of three to six people, single-page websites, subscription pricing, collectively forming the standard start-up for C-end AI applications in 2026. Their retention and renewal rate data will test the true elasticity of this application narrative earlier than any framework demo.
• University AI Teaching Autumn Guide Released: Rising Failure Rates Coexist with Incompetent Passes — Scholars compiled a 2026 autumn university AI teaching guide stating the academic situation is severe: failure rates hit record highs, while more students "pass despite lacking competence." Redesigning assessment methods has no buffer period. AI evaluation and process-based assessment tools on the EdTech side may become the next scenario validated by budgets.

III. Physical AI & Compute Infrastructure

Compute Token Loans Accelerate Nationwide Rollout: Financial Liquidity Directly Supplies Data Center Construction and GPU Cluster Procurement

According to CCTV Finance, this year's CIFTIS debuted a "Finance + Tech Innovation" joint exhibition mode. New financial tools like compute loans, compute insurance, and investment-loan linkage appeared centrally. Compute Token Loans are accelerating nationwide rollout. Incorporating operational metrics like Token consumption and rack occupancy rates into credit pricing effectively opens a leveraged channel bypassing equity financing for heavy-asset compute projects, significantly lowering expansion barriers for small/medium compute service providers. Beneficiary ranking: Server OEMs, liquid cooling, and data center EPC chains benefit most directly; already heavy-invested third-party IDCs gain a handle for refinancing valuation re-rating; banking and fintech targets add a new professional asset track.

• China's Compute Capacity Expected to Account for 30% Globally by 2030 — At the 2026 China Compute Conference, Academician Wu Hequan provided industry figures: Domestic Token consumption reached a daily average of 140 trillion in March, with agents being the main cause of the sharp increase; he simultaneously emphasized that Token consumption should pride itself on efficiency, not quantity. Official narrative is shifting from stacking card scale to unit compute output, mutually confirming the logic of pricing compute credit by operational metrics: Financing looks at rack occupancy and Token throughput, not P-counts on planning charts.

IV. Smart Cars

NDRC: Promote Mergers and Restructuring of Large Auto Groups via Market-Oriented and Rule-of-Law Approaches

Clear signals were transmitted at the MIIT press conference on September 11: Actively support reforms of large auto groups, promote inter-enterprise mergers and restructuring via market-oriented and rule-of-law approaches, while noting that technology paths are still evolving and disruptive technologies may emerge. The official stance sets the tone for market clearing, implying that amidst price war attrition, integration of weak brands by head groups will gain policy channels. Referencing home appliance and smartphone industry history, relaxing merger/restructuring access is often a precursor signal for industry profit margin inflection points. Head SOEs and large private enterprises with multi-brand matrices are most likely to become integrating entities. Existing pricing logic for tail-end NEV projects in primary markets faces re-evaluation.

• Great Wall Motors WEY Gaoshan New Model Cabin Revealed — Qualcomm Snapdragon 8797 automotive chip (4nm), AI compute 640 TOPS, 29.6-inch integrated screen, front/rear dual 5-megapixel ultra-wide cameras. Pre-orders open September 14. The arms race for high vehicle-side compute continues to sink from flagship SUVs to the MPV category. Cabin compute is becoming a key advertised parameter equal to range. For upstream players, this means simultaneous upward movement in Qualcomm smart cockpit chip ASP and shipment mix.
• Stellantis CEO Filosa: Global Car Market Split into Two Parts—US, and Rest of World — In an analyst meeting, Filosa stated tariffs and demand mismatches are reshaping regional valuation weights for multinational automakers: The US market defends margins with high barriers, while other markets are mired in price war attrition. Two operating environments within the same group are forcing multinationals to split capital allocation and product cadence by region. For supply chains, following regional margins captures order flows earlier than following total vehicle sales.

V. Macro & Market Data

On the A-share side (close on 9/11, markets closed for weekend), optical module twin stars bucked the trend: Innolight closed at 926.00 RMB (+4.03%) and Eoptolink at 423.00 RMB (+2.94%), supported by consensus on 800G delivery expectations. Domestic GPU duo saw slight pullback from highs: Cambricon closed at 1040.00 RMB (-0.37%) and Hygon Information at 230.89 RMB (-0.64%), with shrinking volume and increased divergence near the 1000 RMB mark. Large-cap AI server ODM Foxconn Industrial Internet closed red at 64.07 RMB (+0.25%), while application-side Kingsoft Office showed defensive attributes closing at 228.82 RMB (+0.48%). The overall pattern is internal differentiation in compute hardware: Performance visibility of optical modules forms a scissors gap with valuation levels of domestic compute. Capital is switching from crowded domestic compute targets to export-chain names with stronger performance certainty. If nationwide rollout of Compute Token Loans lands, financing benefits for data center and GPU cluster procurement will further redistribute to infrastructure chains like liquid cooling and EPC. Key variable next week: Whether credit tool details provide a list of signable scenarios, not just display metrics at CIFTIS.

On the US stock side (close on 9/11), AI weights generally rose: Amazon closed at $256.78 (+1.94%) and Alphabet at $338.50 (+1.77%) leading (aided by AWS compute price hike transmission and Gemini ecosystem expectations respectively). Meta closed at $648.03 (+0.57%), Microsoft at $495.63 (+0.65%), Tesla at $365.44 (+0.52%), Palantir at $167.23 (+0.83%). Nvidia closed flat at $218.29 (-0.03%), resting in place at the trillion-dollar market cap level. The rotation pattern of compute and cloud platforms taking turns while absolute leaders rest continues. Ellison halting share sales and the White House opposing AI brakes landed on the same day, jointly reinforcing the market's judgment that the AI infrastructure capex cycle has not peaked. Two variables this week determine the narrative's quality: First, the Q3 earnings window's first formal test of capex sustainability; second, marginal changes in AI infrastructure budgets in hyperscalers' next guidance updates.

---