AI Deceleration Debate Hits Market Reality, Optical Modules Lead Declines, and Cornelis Raises $205M
— 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.
— 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.
— 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.
— 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.
— 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.
— 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.
— 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.
— 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.
— 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.
📰 EastVoice — today's AI headlines, decoded
Five minutes a day, the global AI shift →