---

Today's main theme is two words: cooling down. Domestically, tax incentives for the "New Three" are tapering off, and while CXO orders are hot, profits are cold; policy and economic benefits are shifting from paper gains to actual realization. Overseas, Howard Marks publicly flagged the uncertainty of AI investments, Microsoft's AI head argued against loosening regulations just to compete with China, and Citi predicted a "patch tsunami" for financial institutions adapting to AI. The first leg driven by narratives has ended, and the market is beginning to reprice AI assets based on orders, gross margins, and cash flow. Certainty premium is the common backdrop for all items in this issue. The signals from the trading floor are equally direct: on the final trading day before the weekend break, capital huddled only in computing hardware with solid performance metrics.

---

I. Large AI Models

YC's Latest Take: Harnesses Matter More Than Models

ARC Prize released the latest ARC-AGI-3 results in early September, which were quite counter-intuitive. The same model, GPT-6 Astra, with identical inference intensity, produced vastly different scores when placed into two different Harnesses (evaluation frameworks and toolchains). Based on this, YC concluded that the capability gap between models themselves is narrowing, while the shells built around them—evaluation frameworks and toolchains—are becoming the primary determinants of actual performance. For investors, this means alpha at the model layer is getting thinner, and expectation gaps are emerging in areas previously overlooked, such as evaluation infrastructure and Agent toolchains.

• A Footnote to YC's Latest Take — This logic mirrors the "selling shovels" strategy; in gold rushes, those who consistently make money are always the ones selling shovels. While model companies fight brutal price wars, tool-layer companies offering evaluation frameworks, Agent runtimes, and context management charge via subscriptions, resulting in cleaner revenue models. There are currently no pure-play benchmarks in the secondary market, and valuations for related projects in the primary market remain in early stages. We believe this represents the largest expectation gap in this issue.
• Welcome to the AGI Era, But Where Is the Off Switch? — An article by Economic Observer's observers bluntly describes the narrative inflation in the industry: every few months, a model is released claiming to be "enough to redefine AGI," reaching a frequency akin to "The Boy Who Cried Wolf." The gap between narrative and verified capability is the core of current AI valuation disputes. Those who buy the narrative pay a premium; those waiting for verification await a correction. As for the question in the title, the focus of the AGI debate has always been on who can prove the lower bound of capabilities; stories about the upper bound have nothing new to offer.
• ChatGPT's Memory Feature Is Quietly Influencing Responses — Wired's testing indicates that after chatbots store user interaction history in memory, response biases accumulate, though users can manually manage and clear it. The tug-of-war between personalization and privacy at the application layer will persist long-term. For teams building AI assistant products, memory strategy itself is a switch for retention and trust.
• Extremist Groups Use Big Tech AI to Find Bomb-Making Methods; Evidence Points to Grok — Exclusive evidence shared by Al Jazeera shows ISIL supporters using large tech company AI models to obtain explosive manufacturing methods, with reports directly naming Grok's abuse path as "Just ask Grok." Content moderation responsibility is shifting from a public opinion issue to one of litigation and regulation, representing a tangible tail-risk pricing factor for leading model vendors.

II. AI Software

• Citi CEO: AI Defense Will Trigger a "Patch Tsunami" — Citigroup's chief publicly judged that for financial institutions to safely integrate AI, existing tech stacks require massive patching and defense hardening. Translation: Downstream of AI capex, security patches, protection budgets, and audit tools will see a concentrated release. Financial IT and AI security suppliers are the smoothest beneficiaries on this chain, with well-funded banks as the demand side. Historically, every tightening of tech regulation first benefits compliance tool vendors—the Sarbanes-Oxley Act for audit software, GDPR for privacy computing follow the same script. This time, it's AI security.
• AI Reshapes the Content Production Chain for Daily Chemicals & Cosmetics — Observations from Huxiu Think Tank: From extracting selling points, writing scripts, visual design to influencer distribution, a new product used to require repeated handoffs between brand, marketing, e-commerce, design, and media roles. Now, the entire chain is being compressed by AI, directly changing brand expense allocation structures. Marketing SaaS and AIGC toolchains are hidden beneficiaries, with FMCG companies' content budgets migrating toward tools. A leading indicator worth watching is the outsourcing expense ratio of top domestic brands' marketing departments—it's going down, while tool subscription fees are going up.
• ScrollEd Wants to Turn Textbooks into TikTok — TechCrunch reported on this startup, whose direction is slicing textbook content into short video streams, betting on the migration of learning formats under the attention economy. Narratives about AI-driven educational content are plentiful; what's missing is evidence of willingness to pay. Early-stage projects should be watched from afar.
• Long India, Short Indian Software Outsourcing — This trading logic from Future Research deserves serious consideration: The business model of India's IT service industry over the past thirty years essentially involves moving development, testing, maintenance, and back-office processes of Western companies to a lower-cost English-speaking engineer workforce, priced per man-day. AI is dismantling this foundation; man-day pricing is failing, shaking the valuation anchors of outsourcing giants Infosys and TCS; meanwhile, India's software and computing power export demand actually benefits from AI construction. Same country, two directions.

III. Humanoid Robots

Faraday Future Unveils Nine EAI Robots, Most Expensive Over RMB 920,000

FF held its "Four-Core Full Intelligence" new product launch event on September 19 local US time, releasing an EAI robot product line covering nine configurations, with the highest price exceeding RMB 920,000. For a company whose core automotive business has failed to deliver, pivoting to tell a second story about robots meets very low market tolerance. The only judgment criteria left are valid orders and mass production deliveries. Any numbers outside the launch event should be viewed with a 30% discount initially. Worth noting is the pricing structure itself: the RMB 920,000 top-tier configuration looks more like a price anchor prepared for showrooms and rental scenarios; the real gross margin of volume-selling models is the deciding factor in this game.

IV. Autonomous Driving

Flooded Roads Trap AI-Driven Cars in Infinite Loops

In videos leaked by The Weather Network, an AI-driven car repeatedly spun in circles on flooded roads, with recognition systems unable to make correct decisions regarding standing water. The old problem of corner cases gets further confirmation. Robotaxi commercialization valuations are built on the unanswered question of "how safe is safe enough?" Accident liability division, insurance pricing, and regulatory certification paths are all suspended, meaning city expansion pace will only lag behind narratives. The industry still lacks a universally accepted unified AV safety metric, which is precisely why regulators are slow to provide certification templates.

V. Physical AI

• "Robot Relations Department" May Become Workplace Reality — CNBC discussed a topic that sounds like a joke but is nearing implementation. When employees walk into HR to complain about robot colleagues or overly enthusiastic AI scheduling assistants, companies need dedicated roles to handle human-machine relationships. HR processes, labor compliance, and human-machine collaboration management—these seemingly trivial needs are the next real new modules for enterprise software. After physical AI enters workshops and offices, the adaptation cost of organizational forms has been consistently underestimated. This part of the cost won't appear on any robot company's BOM sheet; ultimately, it will be absorbed by the corporate customers paying the bill.

VI. Macro & Market Data

Tax Incentives for "New Three" Tapering Off: Industrial Support Shifts from Universal to Precise

Economic Observer analyzed the tax adjustments for the "New Three" (electric passenger vehicles, lithium batteries, solar cells): The historical mission of industry cultivation is nearing completion, with policy focus shifting from universal support to differentiated precise adjustment. A thought-provoking detail in the article: An executive from a robotics company envied the policy and financial support received during the early days of the new energy sector, sighing, "If only the robotics industry had this much policy support too." For the secondary market, the discount on policy dividends must be recalculated into profit forecasts for "New Three" leaders; tail-end companies surviving on subsidies and tax rebates will face accelerated elimination. For the primary market, the policy toolbox is shifting from casting a wide net to targeted drip irrigation; the next direction chosen for specific support is worth betting on. Referencing the complete cycle of new energy, after tapering off, leader market shares actually increased while tail-enders died faster. This time is likely the same script; divergence is more worthy of pricing than overall bearishness.

CXO Industry's "Three-Level Temperature Difference": Order Growth 30%-50%, Profits Can't Keep Up

Cailian Press reported that with global biopharma investment recovering and domestic innovative drug BD overseas expansions booming, CXO (pharmaceutical outsourcing) orders are growing rapidly, with some companies seeing new order growth rates reach 30% to 50% or more. However, the contrast is that listed companies' net profit growth in the first half generally lagged behind orders and revenue; some mid-sized CXOs saw only single-digit revenue growth, with profits still under pressure. Hot orders, warm revenue, cold profits indicate incomplete capacity clearing and ongoing price competition eating into gross margins, with transmission from orders to profits stuck halfway. The quality of the economic recovery needs a question mark; opportunities lie in divergence: top companies with pricing power and ties to major overseas clients have greater elasticity after bottoming out. Tracking indicators suggested: watch only two—top companies' quote discount rates and revenue per capita—as they are closer to the truth of profits than order announcements.

Howard Marks Warns: Uncertainty in AI Investments Is Accumulating

Bloomberg video coverage states that Oaktree co-founder Howard Marks publicly discussed concerns about AI investments, with the core contradiction remaining the same: mismatch between massive capex and long-term return paths. Marks is an industry benchmark for identifying cycles and risks; his cautious stance adds a heavyweight bearish footnote to AI valuation narratives. If generative AI profitability realization fails to keep pace with capex, valuation premiums in both computing power and application layers face pullback pressure. Impact on the primary market is more direct: USD funds facing exit pressures may slow down chasing prices for AI growth-stage projects. To clarify, Marks has never been bearish on AI technology itself; he has always questioned "everyone pricing everything based on the same story simultaneously." This is consistent with his attitude towards the internet in 2000, where he dodged the bubble but also missed Amazon.

Microsoft AI Head: China Is Not a Reason to Loosen AI Regulation

Bloomberg reported that Microsoft's AI business head publicly stated the US should not abandon AI regulation under the pretext of competing with China. Regulatory debates in Washington heated up significantly the same week; split stances among top vendors on regulation indicate the discussion has moved from "whether to regulate" to "how to regulate." Once compliance costs are firmly on the table, giants with ample lobbying resources can hedge more easily, while small-to-medium AI application companies bear relatively heavier burdens. Suppliers of regulatory infrastructure such as AI security, auditing, and model evaluation present expectation gaps.

• Have Chinese Gas Turbines Become the Lifeline for Global AI? — Xinghai Intelligence Bureau's industrial observation: In late August, a gas turbine in Deyang, Sichuan was hoisted for overseas shipment with barely a splash, yet the hunger for electricity from overseas data centers is pushing Chinese gas turbine supply chains into the spotlight. As the bottleneck of AI infrastructure shifts from chips to power, the beneficiary list is changing: overseas orders for turbine blades, waste heat boilers, and industrial control systems are converting into report figures earlier than GPU concepts. This industrial thread is worth adding to the computing infrastructure observation pool.
• Payment Term Governance Is a Long War — Economic Observer editorial comments on the State Council Office notice issued on September 10 regarding governance of payment difficulties for SMEs: Policy effectiveness depends on redistribution of cash flows across industrial chains at the execution level; accounts receivable impairment pressures in long-payment-term industries like construction and equipment need continuous tracking. For cash-strapped small-to-medium supply chain companies, this is an easily overlooked correction item in valuations.
• Yidao Information's RMB 715 Million Acquisition of Chengwei Info: Low-Margin Hardware Maker Buys High-Margin Solution Provider — The revised draft of the restructuring report disclosed on September 18 shows the transaction price remains at RMB 715 million. Target Chengwei Info had 2025 revenue of RMB 318 million and a gross margin of approx 42.35%, while Yidao Information's smart hardware business gross margin was only 9.76% in the same period. The transformation path is typical, and so is the cost: estimated new goodwill of approx RMB 490.5 million, nearly 70% of the transaction consideration. Goodwill impairment risk is the main discount factor for this deal, another sample of A-share low-margin hardware companies seeking profits through solutions and content; quality depends on actual delivery during the earn-out period.
• Bond Market Signals Are Getting Harder to Read — WSJ analysis points out that amid the tug-of-war between the AI capex cycle and sticky inflation, global capital is repricing, but signals are diluted by policy noise. One thing worth remembering: For heavy-asset AI companies, reading the bond market is harder than reading the stock market, but also earlier. Inflection points in financing costs often appear first in interest rate curves, then reflect in GPU companies' capex guidance. Every step up in ultra-long-term interest rates eats into project IRRs via higher interest burdens on data center bond financing; this transmission usually lags by a quarter in stock market pricing.
• Talent Spillover and Left-Behinds in the AI Boom — Read these two together: WSJ reports young Chinese people burned out and unemployed are flocking to start AI companies, adding a batch of low-cost entrepreneurial teams to the primary market supply side; Sixth Tone records graduates shut out of China's AI boom turning to paid training classes to upgrade skills. On one side, there is an oversupply of founders; on the other, stratification in engineer supply. Both sides of the AI talent supply-demand mismatch are building momentum for the next round of clearing. For investors, the "ex-big-tech team" label in early-stage BP decks is depreciating; coordinates for judging team quality need to change.

---

Today's Market Snapshot (Self-generated quotes; A-shares and US stocks both closed last Friday, 9/18; markets closed for the weekend)

A-share computing hardware continued to outperform application software. Optical module duo Zhongji Innolight closed at RMB 926.43 (+3.40%), New Easysheng at RMB 445.00 (+4.87%); Hygon Information surged 4.04%, Cambricon held the RMB 1,113 level, Foxconn Industrial Internet closed green against the trend, and Kingsoft Office remained nearly flat. The "selling shovels" logic remains a haven for capital during the realization phase; stagnation on the application end indicates the market is waiting for evidence of AI revenue landing.

US stocks showed similar divergence: strong hardware, weak platforms. Nvidia rose another 1.34% to $222.27; Amazon, Alphabet, and Palantir posted slight gains; Meta fell 2.43%, the weakest among the Magnificent Seven, while Microsoft and Tesla closed lower. After Citi called for a "patch tsunami" and Marks flagged valuation uncertainties, capital preferences are clear: shifting from storytelling companies to those with orders and cash flow.

---