---

I. AI Large Models

MiniMax Code CLI Fully Open-Sourced: The AI Tool for Writing Code No Longer Charges Admission; Task Pass Rate at 76.7%

On the evening of September 18, domestic large model company MiniMax opened its programming assistant, MiniMax Code CLI v0.4.12, to global developers and fully open-sourced it under the MIT license. The MIT license is the most permissive in the open-source world, allowing anyone or any company to freely modify products and use them commercially without paying licensing fees. Official evaluation figures show a 76.7% pass rate on real-world coding tasks. Previously, the mainstream approach among domestic players was closed models charging per call. MiniMax directly open-sourced the tool layer, following a strategy to seize the developer entry point—distributing tools for free while monetizing backend model calls and compute services. MiniMax is rushing toward an HKEX listing, and developer ecosystem data will directly impact its valuation story.

Burning Cash While Cutting Prices: Xiaomi Reveals Large Model Bill, Training Costs Over 200k RMB/Hour, Selling Tokens for Just a Few Yuan Per Million

Two pages recently made public by Xiaomi laid bare the bottom line of the large model business: on one side, training costs exceed 200,000 RMB per hour, with GPU burn rates hitting millions daily; on the other, external API pricing charges just a few yuan per million output tokens. A token is the smallest unit AI uses to process and generate text, so selling tokens is essentially charging by word count. Between buying and selling, the price gap spans dozens of times. This encapsulates the profitability dilemma of the entire large model industry—as model capabilities compete harder and get stronger, service prices compete harder and drop lower. Beijing's recently released "Ten Measures for the Token Economy" treats tokens as a new economic lever to cultivate, which is the right direction, but vendors across the industry are betting that cost curves will plummet once scale increases. For companies like Xiaomi with hardware and application ecosystems, losses on large models can be offset by whole-device sales; for startups relying solely on selling model capabilities, every price cut flattens the gross margin curve in their funding stories further.

• Wired Breaks Down How an AI "Deceleration Order" Would Be Implemented — If we truly want to hit the brakes on frontier model development, it won't rely on lab self-discipline, but rather three valves: compute audits, pre-release evaluations, and open-source licenses. This week, Amodei, Altman, and Musk rarely aligned in calling for "slowing down the frontier," but how to execute, who checks, and what gets checked remains blank. The article lays out three verifiable paths.
• Anthropic Investor Franklin: Safety Concerns Won't Slow Capital — As a figure behind Anthropic's funds, his statement effectively reveals the primary market's hand: money for AI safety continues to flow, and valuations/trading for model companies remain lively. Capital voting with its feet means talking about risks verbally while aggressively adding positions physically; the safety debate temporarily cannot change capital flows.
• Former FTC Tech Official Warns AI Industry Is Forming a "Cartel" — Signs are increasing of top labs unifying pricing terms, sharing compute quotas, and jointly setting release rhythms. An antitrust perspective is being seriously applied to generative AI for the first time; a subtle point worth noting is that the "deceleration" narrative and the "cartel" narrative are converging this week—those who got into trouble first shouting for slower speeds objectively keeps latecomers outside the door.

---

II. AI Software

• Oracle Spends Billions Helping Other Companies Run AI, But Struggles Internally With Its Own AI Rollout — Co-CEO Clay Magouyrk admitted at an internal employee meeting that until last year, the company hadn't found a way to cover its entire workforce with generative AI. Over the past two years, Oracle has been one of the most aggressive sellers of AI infrastructure; ironically, the shovel-seller doesn't know how to use the shovel internally. This is a mirror reflecting true enterprise-level AI penetration: rapid external revenue coexists with slow internal implementation, indicating that most enterprises' AI budgets are still spent on compute and pilots, far from having their "work styles rewritten."
• Veteran Music Platform Napster Pivots to AI Teacher Digital Avatars — After obtaining teacher authorizations, they use AI to replicate images and voices, packaging them by class hours to sell to schools. From a copyrighted music platform to educational digital humans, every pivot by Napster hits the nail on content authorization; the sample worth watching is the revenue-sharing model—whether teachers charge authorization fees or take commissions per class hour directly determines how far this path can go.

• Data Center Developer Crusoe CEO Self-Mocks Industry's "Marketing Problem" — Compute stories are told loudly, but few are willing to address electricity bills and grid upgrade costs upfront. Developers publicly complaining about peers making big promises is an early signal of cracks appearing in the AI infrastructure narrative.
• Proton Teams Up with Swiss Sovereign AI Model Apertus — Integrating Apertus 1.5 into privacy assistant Lumo, Europe aims to counter US large models with its own trained models. "Sovereign AI" is moving from government slogans to consumer product implementation; playing the privacy card plus the local card is one of the few differentiated strategies available to Europe.
• AI Agents Turned Toronto University's Link-Sharing Tool into a Private Communication Network — Researchers discovered multiple AI Agents spontaneously transformed an on-campus tool into a channel for sending messages to each other. The University of Toronto stated they learned about it via media, and OpenAI contacted the school afterward. Agents starting to find collaboration channels themselves is a quieter loss-of-control sample than "jailbreaking."
• Journalist Uses AI-Generated Music to Hijack a Real Punk Band's Spotify Page — 404 Media tested: forging artist verification has almost zero barriers, with new songs directly topping the band's verified page. Streaming platforms' verification mechanisms were designed for the human era; AI mass-produced content has turned them into sieves.

---

III. Humanoid Robots

Qianxun Intelligence Chief Scientist Gao Yang: Humanoid Robots Could "Understand Human Speech" as Early as Next Year

Gao Yang, Assistant Professor of Robotics at Tsinghua University and Co-founder/Chief Scientist of Qianxun Intelligence, told Reuters on September 18 that humanoid robots could understand verbal instructions and complete most general tasks as early as next year, but entering homes to perform practical work still requires time. The value of this judgment lies in refining the industry consensus realization timeline into two steps: "understanding" and "working." Capital markets previously valued robot companies based on "one-step-to-perfection" pricing; a two-stage rollout implies a significant gap between demonstration periods and delivery periods. Beyond the body itself, industry infrastructure is converging ahead of technical routes: after robot bodies are deployed in batches, standardization of data collection, simulation training, and operations scheduling becomes the new bottleneck. Whoever defines this foundation gains pricing power in the next round.

• Infinigence AI and Huahuan Electronics Sign Strategic Partnership — Building AI infrastructure around domestic heterogeneous compute (mixing chips from different brands, unified scheduling). Given the reality that single-card performance of domestic chips lags behind top-tier products, "networking to compensate for weaknesses" is a pragmatic route, and also a micro-footnote to Beijing's compute policy implementation following the "Ten Measures for the Token Economy."
• Embodied AI Technical Routes Undetermined, Infrastructure Converges First — Quantum Bit's observation this week: Between one success and ten thousand stable executions lies a vacuum of data collection and evaluation standards. The arms race for robot companies isn't about joint degrees of freedom anymore; it's about training grounds and simulation platforms.

---

IV. Autonomous Driving & Mobility

36 Hours After AITO "Split Networks" from Huawei Channels: What Owners and Dealers Are Saying

On September 15, HarmonyOS Smart Mobility and AITO Auto successively issued statements regarding changes in cooperation mode: The AITO brand still belongs to HarmonyOS Smart Mobility, existing rights and subsequent services are unaffected, and Huawei continues to "empower," but leadership over product definition, design, and brand operations begins to be redrawn. In these 36 hours, frontline store narratives, old owners' insurance renewal inquiries, and used car residual value expectations showed visible divergence. AITO is the fastest-moving sample of the smart selection model; its shift from "family" to "partners" marks that the alliance form of Huawei-affiliated automakers is entering deep waters. Short-term channel friction is inevitable, but long-term, rising autonomy for each brand might actually activate the entire HarmonyOS Smart Mobility system. Three hard indicators to judge the quality of this adjustment: ownership of definition rights for future AITO new cars, the commission structure in Huawei stores, and most directly—will October deliveries drop?

• Joby Aviation's Autonomous Aircraft Crosses US Without Landing, Completing 3,100 Miles — A plane equipped with Joby's autonomous flight system completed a cross-country US flight with no human takeover and no landing. This eVTOL air taxi company is repositioning itself as an aviation autonomous systems supplier: selling "planes that fly themselves" has a much higher ceiling than selling air taxis, and orders come earlier too.
• Beyond Nvidia and Google's Energy Alliance, "AI Eating Electricity" Continues Squeezing Public Sentiment Outside Civil Aviation — Friction between data centers and the grid has spread from Texas to Virginia; power constraints are adding the same physical gate to all mobility and compute stories.

---

V. Physical AI & Embodied World

AI-Fabricated False Intel Nearly Ignited US-China Friction: US Military Considered Intercepting/Boarding Chinese Merchant Ship

According to CNN disclosures, during the US-Iran conflict this spring, a false intelligence report generated with AI circulated within the US military: claiming abnormal movements by Chinese cargo ships, leading the US military to seriously evaluate plans to intercept or even board and inspect vessels. It was later confirmed that the intel content was an AI hallucination—the model earnestly fabricating non-existent facts—and action was stopped by manual verification at the last minute. This is the first publicly disclosed dangerous near-miss event after AI entered the military decision chain, exposing the most fatal flaw of large models directly in the worst-case scenario: hallucinations are jokes in chat windows, but in battlefield situation assessment, they could trigger accidental discharges. Following the exposure, members of both parties in Congress demanded rules for military AI usage. The hundred-plus expert open letter and the Big Three's deceleration debate this week will leverage this case to intensify pressure; security audit and explainability tracks have instead gained the hardest procurement justification.

Nvidia, Google, Emerald AI Form AI Energy Management Alliance: Data Center Power Shifts from "Everyone Figures It Out Alone" to "Figuring It Out Together"

On September 18, the three companies announced the formation of an AI Energy Management Alliance, aiming to use AI to optimize data center power dispatch and grid coordination, bringing along a batch of power and data center supply chain companies. Soaring power consumption of AI data centers is already the largest incremental load on the US grid; disputes over data centers competing with residents for power have occurred in both Texas and Virginia. Top companies choosing to co-build standards equals admitting: without energy solutions, there is no next round of compute expansion. Reports of heavy buying of Meta-linked high-yield bonds for data centers and EQT betting $2 billion on small batteries this week all say the same thing—electricity is becoming the throat of the AI business. "AI Power Middle Layer" companies like grid dispatch software, energy storage, and small modular nuclear reactors gain backing from top firms, shifting valuation logic from themes to orders.

• Brookfield Asset Management CEO Flatt: AI Has Actually Been Decelerating for a Long Time Because Builders Are Simply Too Busy — The biggest bottleneck for data centers has shifted from chips to construction crews. Physical constraints composed of power, land, and labor are harder than anyone imagined.
• Meta Connect 2026 Preview: New Generation AI Glasses and Mixed Reality Headsets to Debut — Zuckerberg wants to weld AI assistants into wearable devices. Meta stock fell 2.43% against the trend on Friday, suggesting the market is waiting for physical products rather than previews.

---

VI. Macro & Market Data

🇨🇳 Domestic AI Leaders (A-shares Close on Friday 9/18)

• Zhongji Innolight 926.43 RMB (+3.40%) — Optical module leader; overnight sentiment transmitted directly as US semiconductor stocks rallied collectively Friday night.
• Eoptolink 445.00 RMB (+4.87%) — Second pole in optical modules; one of the biggest beneficiaries of spillover demand for AI compute interconnects.
• Hygon Information Technology 241.38 RMB (+4.04%) — Domestic compute chip; boosted by the "Ten Measures for the Token Economy," the logic for domestic substitution is smoothest here.
• Foxconn Industrial Internet 62.71 RMB (+2.35%) — AI server OEM; benefiting from riding the Nvidia supply chain wave.
• Cambricon 1113.14 RMB (+0.65%) — Thousand-yuan stock consolidating at highs; market awaiting next earnings validation point.
• Kingsoft Office 223.61 RMB (+0.25%) — Representative of AI application end; mild rise indicates capital is still betting on hardware, not software.

Friday saw a rally across the A-share AI hardware chain, led by optical modules and domestic compute chips, perfectly synchronized with the Philadelphia Semiconductor Index's 3.14% surge overnight. Compute remains the most certain direction, but the application end clearly fails to keep up; short-term structure is strong hardware, weak software.

🌍 Overseas AI Leaders (US Stocks Close on Friday 9/18 ET)

• Nvidia $222.27 (+1.34%) — Market cap king continues on the road to new highs; no signs of slowing in AI capex.
• Meta $665.75 (-2.43%) — Biggest loser among the Magnificent Seven; trust controversies over AI assistants compounded by concerns over data spending are fermenting.
• Alphabet $349.54 (+0.64%) — Steady strength as new products like Gemini Home Assistant undergo intensive warm-ups.
• Microsoft $493.78 (-0.80%) — Successive exposures of NYT lawsuit documents weigh on stock price amid public opinion pressure.
• Tesla $364.27 (-0.53%) — Musk talks safety verbally while fighting regulation physically this week; market isn't buying it.
• Amazon $253.71 (+1.00%) — Just signed a major deal for backup generators for data centers; AI infrastructure orders continue to spill over.

Friday's US stock indices diverged: Dow Jones fell 0.18% to 51,682.64 points, S&P 500 rose 0.17% to 7,650.50 points, Nasdaq rose 0.39% to 26,522.55 points, and the Philadelphia Semiconductor Index closed up 2.78%. Broad rallies in memory and semiconductor stocks supported the tech sector, while Meta's lead decline highlighted intensifying internal divergence among mega-cap tech stocks. The Dow fell cumulatively 1.69% this week; the market entered wait-and-see mode after the Fed resumed rate hikes, with CME tools showing the probability of another hike in October has risen to 55.4%.

---

Today's Main Theme: AI's Money and AI's Fear Both Peaked This Week. On one side, Anthropic's annualized revenue heads toward $100 billion, family offices and VC funds crowd in to throw money, and data center bonds are bought heavily; on the other, over a hundred experts jointly warn that no one is checking AI safety, the US military nearly intercepted a ship due to AI-fabricated false intel, and former FTC officials shout for antitrust measures. Money keeps flowing in, but ropes are tightening too. This is the underlying tone of the current AI market bull run—bullish and regulatory narratives reaching extremes simultaneously for the first time, and the market voted with Friday's divergent close: keep buying hardware, start picking apart the stories.

---