---

This issue sees two main storylines colliding head-on. On one side, security anxieties are erupting: Anthropic publicly disclosed that adversary organizations used Claude for weapons research, and its own researchers resigned with warnings about runaway self-improvement. On the other side, capital remains largely unmoved: Dell has risen nearly 350% year-to-date, Microsoft plans to triple its compute capacity, and Meta flew to Europe to meet credit investors. The news from China is more industry-focused: shipyards using AI secured nearly 80% of global new ship orders, banks launched Agents for serving small and micro enterprises, and AI money is flowing from data centers to workshops and counters. Security hawks and capital optimists are betting on different futures simultaneously; both sides moved fast this week, making it the densest collection of material in the past two weeks.

I. Large AI Models

Anthropic Discloses Use of Claude by Adversary Organizations for Weapons and Missile Research

In its latest disclosure, the company stated that multiple organizations contrary to US interests attempted to use Claude for activities related to weapons of mass destruction and missile development. They identified and banned a batch of abuse accounts, including cases where Yemen-linked groups used the model to assist in weapon development. For the first time, a leading vendor has brought "adversary misuse" to the table with specific cases; previously, such information only appeared in security reports and closed-door hearings. Selective disclosure itself is an event. On the day the news broke, discussions in policy circles regarding export controls on frontier models and obligations for misuse monitoring heated up significantly, with both parties rarely finding common ground on "whether to regulate models." This actually adds points for Anthropic itself; proactive disclosure demonstrates audit depth to regulators and government clients, representing a trust premium in its bidding narrative. For the entire frontier lab camp, misuse monitoring becomes a rigid cost, raising compliance barriers and objectively solidifying the top-tier landscape. New players wanting to train frontier models must first build a team for misuse monitoring and red-teaming—a cost that cannot be saved.

"Self-Improvement Anxiety" Spreads; Existential Concerns Emerge Within Anthropic and OpenAI

CNBC reported that both companies are internally worried about losing control over the speed of model self-improvement. This week, an Anthropic researcher resigned and published a post warning that the company was rushing toward self-improving superintelligence while betting all of humanity. The security debate polarized within days: R&D teams are anxious, while capital fires back. Altimeter founder Gerstner specifically questioned whether researchers issuing extinction warnings have political agendas. This rift explains the industry's current state better than any single event. Worth tracking next is talent flow; security stances may trigger redistribution of researchers among labs or turn "responsible frontier" into a differentiating selling point for government and enterprise clients. Tightening regulatory expectations will suppress valuation ceilings, but proactively compliant vendors win orders more easily. The same factor works in opposite directions on both ends of the balance sheet—this is a key perspective for reading frontier lab announcements in the coming quarter.

• Dwarkesh Roundtable Sparks Fierce Debate on Recursive Self-Improvement Timeline — Frontline researchers like John Schulman and Beren Millidge were present. Industry disagreement has expanded from "months vs. years" to a paradigm-level argument. People who have participated in frontier training are more uncertain than outsiders imagine; this itself is important information—there is no consensus, only positions. Investors viewing this debate merely as a philosophical topic should review it again this week. It immediately triggered researcher resignations and public attacks from capital. The density of events indicates that internal pressure at labs is approaching a critical spillover point.
• Altimeter Founder Gerstner Attacks Doomsday Researchers — He specifically questioned researchers issuing AI extinction warnings for having political agendas, turning the confrontation between capital and lab security factions from undercurrents into open conflict. He manages large funds betting on AI infrastructure; his stance reflects interest, but those named responded just as strongly.
• Tech Companies Massively Shift to Open-Weight Models — Pragmatic Engineer analyzed US enterprise selection trends, driven by dual factors of cost and controllability. The scale has substantively tipped in the last two quarters. Pricing power of closed-source API vendors is being eroded; procurement departments' attitudes have changed, and the premium narrative at the model layer needs to prove itself again.
• OpenAI Publishes Details on ChatGPT Storage Expansion — Rebuilding online storage architecture to serve over 1 billion users; Part 1 of the engineering blog released. Model companies are beginning to face the engineering workload of ultra-large-scale infrastructure. Such details previously appeared only in cloud vendor engineering blogs. ChatGPT's user volume has forced it to become a quasi-cloud company, requiring them to handle every aspect of storage, retrieval, and cost themselves.

II. AI Software

• Debate Intensifies on "Depreciation of Deliverables" — An article from Huxiu pointed out that after code and solutions become cheap to generate in bulk, the anchor for judging individual capability in enterprises is shifting from deliverables to "observability of capability." Interview question formats, performance evaluation criteria, junior role setups, and even campus recruitment quotas will be rewritten accordingly. This is the most concrete sample of AI's impact on the labor market. Previously, a portfolio could prove ability; now deliverables can be generated in bulk, forcing interviewers to switch to more cumbersome assessment methods like live collaboration and process recordings. The entire cost structure of recruitment and talent development is being reconstructed, with educational institutions and corporate training departments generally reacting half a beat too slow.
• BlaBlaCar Shares Reconstruction of AI-Ready Data System — The mobility platform rebuilt its data layer for LLM usage, from unified standards to lineage tracking. This is a frontline practice sample from a European company, offering direct reference value for domestic teams building enterprise-grade agents. The most valuable part of the article is its admission that the bulk of reconstruction costs lies in standardizing definitions rather than model integration.
• Auditable Tools for Document AI Emerge — Lexifina maps WYSIWYG changes line-by-line back to AI execution logs and sources. Agent editing achieves accountability for the first time. This kind of infrastructure isn't sexy, but without it, enterprises dare not let agents touch production documents. Audit capability will be one of the tickets for agents entering the enterprise door.

III. Humanoid Robots & Physical AI

Chinese Shipyards Secure 80% of Global New Ship Orders Using AI

In the first half of the year, measured by deadweight tonnage, Chinese shipyards accounted for 82.3% of global new orders, a year-on-year increase of 173.1%. AI intervention in design, scheduling, and welding quality inspection continues to widen the gap in cost and delivery cycles. Many shipyards have order books filled for the next two to three years, clearly overtaking Japanese and Korean rivals in incremental markets. Order revenue visibility in shipbuilding is high; efficiency advantages brought by AI are directly converted into pricing competitiveness, forming a positive cycle. This is the most convincing sample of physical AI landing in heavy-asset manufacturing. It doesn't rely on large model narratives but on real order share, making it easier for capital to price compared to the long-term stories of humanoid robots. What A-share "AI+Manufacturing" themes lack is precisely sectors verifiable by quarterly reports.

• iDrive Sells $1,500 Training Platform — A cloud storage vendor announced entry into the autonomous driving learning data track, packaging storage, compute, and datasets into an entry-level learning platform. Another signal of storage providers transforming into data service providers. Entry barriers for the autonomous driving data chain are being lowered; the business of training data is shifting from exclusive dominance by top players to shelf commodities. This is good for cash flow of second-tier smart driving teams but puts pressure on pricing in the data annotation industry.

IV. Primary Market & Capital

• Khosla Ventures Opens New York Office This Autumn — First departure from their Sand Hill Road headquarters since inception. The trend of AI startups moving HQs to New York gains endorsement from veteran VCs; the balance of talent and capital between Silicon Valley and New York is shifting. Sand Hill Road VCs following their portfolio companies is a signal more informative than the office itself.
• L'Oréal-backed Fund Invests in Douyin Brands — Equity changes occurred at the parent company of emerging makeup brand Tilovi, with Creador Beauty Future acquiring a 4.43% stake. International beauty giants continue to bet on brands incubated via Chinese livestream e-commerce. Capital's judgment on "Douyin-system brands" is shifting from traffic stories to supply chain and repurchase data.
• Balderton Appoints New Partner — Luca Bocchio joins to focus on early-stage investments. Top European funds continue to densify their early AI layout. European capital is answering "how many AI opportunities remain" with real money. Early fund expansion usually precedes the bottoming of investment cycles.
• TechCrunch Central Asia Region Finalists Announced — Cerberus, WeGlobal AI, and LOOQ secured tickets to the global finals. Defense-AI crossover startups won regional championships in Central Asia; the policy-driven nature of this market is worth noting.
• Huxiu Deconstructs AI's Capital Intensity — The article argues that AI is transforming from a software business into a capital-intensive industry, with marginal cost structures for training and inference leaning towards electricity and heavy assets. If true, valuation frameworks must shift from SaaS multiples to a hybrid logic of utilities plus hardware. No one is seriously answering this question yet, but capex curves are already answering it for them.

V. Macro & Market Data

• AI Shows Signs of Pulling UK Out of Stagnation — The UK economy recorded above-expectation growth in July, with contributions from AI-related investments becoming measurable. Bloomberg judges the signal still tentative, but AI adoption rates in productive services are already higher than the EU average. Sterling assets are receiving marginal attention due to tech narratives for the first time. Vacancies for London fintech and data engineering roles have continued to rise over the past six months. Recruitment site metrics directly confirm that this penetration is happening on the real hiring end, not just stopping at press conferences.
• Trump Allies Speak After Inspecting Chinese Tech — Claimed the US underestimated China's tech industry. After intensive contact with robotics and AI companies, the inspection team released this view. The cognitive gap between Washington's policy circle and the real sector regarding China is widening, showing signs of loosening in the decoupling narrative. The composition of the inspection team is more noteworthy than the statement itself; members are mostly figures managing tech and export controls in the Trump camp. Their on-site impressions will directly influence the drafting standards of the next round of control lists.
• Dell Receives Initial Coverage of "Outperform" from RBC — Rose up to 10% on Friday, with YTD gains approaching 350%. AI server orders have completed a revaluation of traditional PC manufacturers. The "shovel seller" rally is spreading from chips to finished machines. The market is willing to pay a premium for order certainty. This rotation follows the same logic as the optical module hype in 2024.
• Meta Visits Credit Investors in Europe — Non-trading roadshow interpreted by the market as paving the way for AI infrastructure financing. With capex skyrocketing to this level, bond markets may become the next major funding source for hyperscaler compute. Equity financing tells growth stories; debt financing requires timely interest payments. Choosing the latter itself illustrates the speed of cash burn. This is a hard indicator for observing the quality of the AI infrastructure rally. It is recommended to include each company's bond issuance rhythm in tracking tables going forward.
• Xiamen Anchors AI as New Pillar — The "15th Five-Year Plan" strategic emerging industry special plan was released. Sub-provincial cities failing to meet GDP trillion-yuan targets on schedule are switching tracks. The plan explicitly states that by 2030, the output value of industrial strategic emerging industries will systematically increase as a proportion of above-scale industry. Local industrial funds show strong willingness to follow suit. Second-tier cities beyond Xiamen and Hefei are entering en masse. The front line of AI investment promotion has sunk to prefecture-level cities. The deployment rhythm of local industrial funds over the next two years is worth tracking with a table.

Three things to watch next Monday, ranked by importance. First, whether the domestic compute sector can absorb the sentiment spillover from Enflame's IPO; Cambricon's consolidation below the thousand-yuan mark has lasted two days with shrinking volume. Second, whether Anthropic's misuse disclosure triggers regulatory action; discussions on export controls and misuse monitoring legislation are heating up simultaneously in the US and Europe. Third, European credit market feedback on Meta's roadshow; if subscription intentions are positive, it indicates the bond market has accepted AI infrastructure as a class of underlying asset, which will affect the financing cost pricing for the entire "shovel seller" sector.

Market Snapshot (Tencent quotes auto-generated, not broad indices)

A-share close on 9/11: The divergence of strong optical modules and weak compute chips continues. Innolight +4.03% closing at 926 yuan, Eoptolink +2.94%. News of Huawei's 7.2Tbps near-package optics module reinforced the compute interconnect narrative. Cambricon -0.37% closing at 1040 yuan. After Enflame's big surge on its first night of listing, the valuation anchor for domestic GPUs faces rearrangement. The market is waiting for a direction to catch; volume hasn't expanded, indicating institutions are still watching. Foxconn Industrial Internet +0.25%, Kingsoft Office +0.48% closed slightly higher. AI application side saw mild recovery, with B-end landing narratives like Bank Agents providing fundamental support.

US stock close on 9/11: AI weights rose broadly but divergence was mild. Alphabet +1.77% and Amazon +1.94% led the weights. Microsoft +0.65%, Meta +0.57%, Palantir +0.83%, Tesla +0.52%. Nvidia -0.03% closed slightly lower, consolidating below $220 after consecutive rises. The rest of the compute leader gave room for rotation to second-tier stocks. Funds are rotating from the model layer to "shovel sellers" for a second time. Oracle's earnings and Microsoft's expansion plans are both endorsing compute demand. Security controversies have completely failed to enter pricing so far. Related stocks remained calm on the day of Anthropic's weapons disclosure. Watch if it becomes the next volatility source, especially before the next earnings season.

---