---

Today's main storyline is singular: The AI "deceleration" debate has moved from the lab to the trading floor. Dario Amodei-led deceleration initiative completed its transmission from a tech-circle topic to an asset price shock on 9/14, with chips and compute chains falling while software platforms held steady; meanwhile, primary markets and IPO queues remain completely unaffected—funding continues as usual, hearings proceed, and guarantees are issued. The coexistence of cooling narratives and heating capital is the entire footnote of this issue's Alpha. Watch two things in the coming week: whether the deceleration initiative gets written into any capex guidance, and whether tech companies queued for IPOs will change their timelines.

I. Macro & Market Data

AI Deceleration Debate Hits Market Reality: Executives Call Out Investors, Chip Stocks Under Pressure — Following Anthropic CEO Dario Amodei's piece "We Must Slow Down at the Frontier," OpenAI and xAI leaders rarely aligned, calling for safety guardrails for continuously enhancing models. Subsequently, multiple tech executives turned to communicating with capital markets: Safety statements must be made, but they shouldn't scare investors (Bloomberg). This "deceleration" narrative is shifting from a technical issue to a valuation issue—chip and AI concept stocks fell pre-market on 9/14, reigniting concerns about the return cycle of compute capex. For secondary market investors, what to watch next isn't whose wording is more alarmist, but whether capex guidance in earnings reports loosens alongside the safety narrative.

AI Giants Put Regulatory Demands on the Table: Is "Pace the Frontier" Regulatory Capture? — The Register sharply points out that top labs' deceleration initiatives are essentially defining regulatory boundaries for themselves: By having the few players mastering frontier models define the "safety" standard, the closer the rules are to their capability circle, the higher the entry cost for latecomers. The open-source camp and mid-tier vendors have publicly expressed dissatisfaction. For investors, another side of the deceleration vs. acceleration debate is the existing premium under a licensing-based landscape—every point the regulatory threshold rises deepens the moat of the top four firms.

Loss-of-Control Anxiety Begins Pricing Sectors — A Bloomberg long-read reviews "Are we losing control of AI?": From rumors of OpenAI agents autonomously launching cyberattacks to model capabilities spilling over to security teams, safety-related events are transmitting into US AI sectors via volatility. Beyond technical debates, a "fear premium" has become a new market variable—every rise in loss-of-control narratives translates into implied volatility in the options market.

• Tech Giants Learn "The Art of Statement Calibration" — Bloomberg: Executives are practicing expressing concern about AI risks without unsettling the market; the wording, timing, and linkage of safety statements with earnings windows are highly coordinated, blurring the lines between PR and IR departments.
• Anthropic Leader's Risk of Direct Clash with Wall Street Rises — Bloomberg: AI executives' deceleration statements collide head-on with Wall Street's growth expectations. If the initiative continues to ferment and impacts capex expectations, the volatility center of AI sectors may passively rise; longs need to leave hedging space for "narrative-driven valuation kills."

Market Confirmation: Optical Modules Lead Declines, Compute Chain Pulls Back Collectively — US stock close on 9/14 and Asia-Pacific trading directly priced this debate: A-share compute hardware weakened across the board, Innolight down 5.72%, Eoptolink down 4.99%, Foxconn Industrial Internet down 3.90%, Hygon Information down 3.12%, Cambricon flat (1040 CNY); On the US side, Nvidia fell 3.36% to $210.96, Tesla down 1.77%, Amazon down 1.26%, while software and platform sides remained relatively resilient—Microsoft up 1.97%, Alphabet up 3.22%, Meta up 2.71%, Palantir up 3.64%. The divergence path of capital is clear: "Deceleration" rhetoric hits the heavy-asset compute expansion narrative, while cash-flow-determined software platforms absorbed safe-haven buying. The high valuations of optical modules and domestic compute chains essentially wrote perpetual capex growth into prices; now that those pricing it are changing their tune, holding logic needs recalculation.

II. Primary Markets

AI Network Infrastructure Firm Cornelis Raises $205M, Targeting Nvidia Interconnect Lane — Officially announced on Monday, 9/14, funds will expand production of networking tech helping AI chips communicate efficiently. Amid cooling compute narratives, primary market capital is still betting along the path of "avoiding single-point dependency": Interconnects, cooling, and power supply outside GPUs continue to gain premiums, with Interconnect becoming the most undervalued entry point in the Nvidia ecosystem. Choosing to announce during the hottest Monday of the deceleration debate is itself a vote of confidence.

European Satellite Firm Open Cosmos Secures €300M to Expand Satellite Manufacturing & Intelligence Connectivity Services — Funds directed toward capacity expansion and in-orbit constellation deployment. The EU's "sovereign space" narrative continues to receive industrial capital endorsement; defense and commercial aerospace valuation logics are merging in Europe. For global LPs, this is one of the few directions in "hard tech beyond AI" still supported by sovereign-level orders.

• Italian Insurtech Mama Insurance Completes New Funding Round — Led by Fastweb+Vodafone, followed by Founders Factory. Telecom operators are starting to use insurance scenarios to digest their user data assets. The buyer structure of European insurtech is shifting from funds to industry players: In industries where financial investors are exiting, giants holding channels are acquiring assets at discounts.
• Swedish Serverpod Does AI Coding in Reverse: App Builder Without Built-in Models — While dev tool companies rush to build coding agents, it launched an AI-free App Studio, betting on deterministic infrastructure after the "vibe coding" tide recedes. A rare contrarian position in the dev tools track: The stronger the model capabilities, the thinner the middle layer around them; it's betting on a return to engineering certainty. If this judgment holds, early-overvalued AI coding agent tracks will see M&A discount opportunities.
• Cold Thoughts Amid Math Fever Hit by AI — iFanr: The arms race of models topping math conjecture leaderboards looks increasingly like a new outlet for big tech marketing budgets; the scissors gap between capability narratives and commercial monetization is widening. Investors should discount benchmark battle reports from valuation narratives.
• Valuation Anchors Forming in Interconnect Track — Cornelis's single round of $205M sets a reference price for the AI networking segment, expected to drive repricing of a batch of optical interconnect, CPO, and silicon photonics targets in primary markets. For the secondary market's optical module sector, this is both demand-side endorsement and a loosening of the "there are players outside the Nvidia system" narrative.

III. AI Large Models

DeepSeek Prepares for IPO: Domestic LLMs Start Doing the Math — After releasing V4.1 Flash, officials stated its HBM demand dropped to a tier lower than the previous generation: Stronger models, less cache usage, lower prices—the capability table is on the table; another unpublished profitability table is worth watching closely for investors—under inference price cuts and free API strategies, how exactly does it make money? An open-source flagship going public will become the first public sample testing Chinese LLM unit economics, and the watershed for domestic compute narratives switching from "burning cash for share" to "earning profit for valuation." Finding the pricing anchor (benchmarking against OpenAI private valuations or cloud vendor AI businesses) is the biggest suspense of this IPO.

• Anthropic Names Kimi Again — Accusing it of routing some customer requests to Claude backends. "Choosing sides" pressure in the model supply chain is beginning to transmit to enterprise procurement decisions: Calling whose model is no longer just a performance issue, but also a compliance issue. For domestic model vendors, trust costs for overseas clients are rising; for Anthropic, repeatedly "catching relays" is also marketing to set rules for its own API.

• Cohere CEO Warns Against Turning AI Safety into a "Cartel" — A rebuttal from the younger faction against deceleration initiatives: If safety standards are monopolized by top labs, it will lock down the catch-up space for open ecosystems. Mid-tier vendors' regulatory lobbying is officially on the table; the route war between open-weight camps and closed-source frontiers now has an added layer of "safety discourse power."
• Industry Shifts Overall to "Doomsday" Tone — MIT Technology Review: After the OpenAI agent cyberattack uproar, top lab leaders collectively changed their tune to discuss risk management; the switch from "acceleration" to "deceleration" rhetoric means the industry treats regulation as a strategic asset rather than a cost burden.

IV. AI Software & Applications

• Testing Meta Muse Agent: Useful, But Uncomfortable Handing Over Digital Life — TechRadar experience article: Trust costs of agentic AI are discussed as a core selling point by mainstream media for the first time. Willingness to surrender personal data will determine the payment conversion slope of consumer agents; if "useful but afraid to use" becomes mainstream feedback, Meta's ad data flywheel narrative will take a hit before revenue realization.
• Money Spent on AI Coding Wasn't Wasted, But Duplicate Code Rose 81% — New Stack data: Same budget bought 25% more output, at the cost of soaring code duplication rates. "AI R&D efficiency" enters the final accounting phase: After one-time generation costs drop, long-term maintenance and deduplication costs will lag in eroding book profits; CFO perspectives on AI ROI metrics are being rewritten.
• Enterprise Agent Procurement Enters "Model Provenance" Review — From Anthropic's relay accusations to various safety declarations, clients are starting to require suppliers to disclose inference chains; model routing transparency will become a new bidding clause.
• Can "Open Data Infrastructure" Bridge the Gap Between AI Investment and Implementation? — TechRadar analysis: The gap between enterprise AI spending and quantifiable output is being attributed to insufficient data interoperability; open data infrastructure is upgrading from an engineering issue to a procurement budget line item, signaling a narrative turning point for the data middleware track.

V. Robotics & Autonomous Driving Hardware

HuanChuang Tech, World #1 in Robot Vacuum LiDAR Market Share, Passes HKEX Listing Hearing — Passed hearing on 9/13, joint sponsors CICC and Guosen Securities (HK). Annual net margin of only 0.4% for a core upstream component supplier for cleaning robots: The story of world #1 shipment volume rings loud, but profits are thin as blades. Ecovacs, Dreame, Roborock, Narwal feast on whole-machine brands, while upstream suppliers sip broth—supply chain bargaining power is obvious. This is also a valuation reminder for all "selling shovels" stories: When shovels themselves become a red ocean, diseconomies of scale will chase market share. Humanoid robot industry chain investors should self-check similarly—gross margin structures in sensors, joints, and dexterous hands determine the quality of this listing wave.

• Shenzhen Youth Spend 20k CNY on AI Training Classes, Don't Get "Guaranteed Jobs" — AI talent anxiety business replicates in batches, mixing training loans with direct hiring pitches. Another side of demand-side bubble squeezing: True AI job premiums remain concentrated at the top; the mid-layer training market clears first. The sold narrative of "anyone can learn AI" versus the reality of "jobs converging to the top" constitutes an expectation gap in the labor market.
• Amid Deceleration Voices, Robotics Companies' IPO Clock Doesn't Stop — From HK Main Board to STAR Market, embodied intelligence and cleaning robot supply chains are still queuing for sprints; primary market money and secondary market mouths have already split on pricing direction for this AI narrative, which is the largest arbitrage observation spot for Alpha layer.

VI. Industrial Capital & M&A

Guoen Shares Plans 15 Billion CNY Guarantee to Bet on Compute, What Are the Odds? — A modified plastics main business company just issued H-shares on HKEX in Feb this year, raising net proceeds of approx. 918 million CNY, holding 2.018 billion CNY cash at end of H1 (up 63.53% YoY), yet plans to issue 15 billion CNY guarantee quota for compute business. Traditional manufacturing crossing over to compute, leverage arrives before returns; guarantee scale exceeds cash on hand by seven times, a question no board can bypass before signing. Market stock price reaction to such announcements is a slice for observing A-share compute concept quality—if capital still blindly pumps it up, it indicates this narrative hasn't reached the point of restoring pricing discipline.

• Valeo Reported Exploring Sale of Car Smart Tech Business Clarion — Bloomberg: Auto parts giant shrinking lines, cockpit electronics assets entering buyer's market. The window for Tier1s divesting non-core electronic businesses is a picking-up time for Chinese industrial capital and PE; M&A valuation anchors for automotive smartification are being recalibrated. Once sellers list, the target's EBITDA multiple becomes the pricing reference for the entire auto electronics sector.
• Nigeria's Dangote Refinery IPO Subscription Enthusiasm Crashes Trading App — Bamboo overwhelmed by requests. The heat of emerging market energy assets' debut proves: Global liquidity's chase for "scarce physical assets + local leaders" combinations rivals AI; capital isn't flowing unidirectionally into tech, but rearranging based on scarcity.
• Capital Divergence Between Decelerators and Expander Camps Already Visible in M&A Market — At the other end of rising safety narratives, asset disposal windows open simultaneously: Giants shrink non-core lines (Valeo), cross-border players enter with leverage (Guoen Shares). Industrial capital turnover speed often reflects cycle positions ahead of stock prices.

---