---

Today, money in the global AI capital market is moving in two directions.

One direction is "betting early." Memory manufacturers are daring to raise their 2027 capex to the $6 billion level; data assets are seeing preemptive bidding wars; sovereign capital continues to pile into data centers despite the fires of war. The common implication of these moves is that players judge this cycle hasn't peaked yet—they'd rather buy too early than too late. Order visibility for upstream equipment and materials is being extended accordingly.

The other direction is "cost discipline." Meta is reportedly cutting up to 60% of teams due to AI; data centers in Canada and the Middle East are hitting backlash over site selection and geopolitics; Abu Dhabi's G42 is considering selling its US equity stake to secure chip access. The hidden costs of expansion are being priced item by item: labor, electricity, land, equity—each has entered the cost function.

The story on the China side is more specific. The autonomous driving chain is standing at the negotiation table as a "solution provider" for the first time. Maserati is seeking joint development with Huawei and JAC Motors; the Mercedes-Benz GLE is equipped with Momenta's world model; Qijing is pushing ADS 5 into the 250k RMB segment; Li Auto is extending strategic cooperation to power semiconductors, binding an 8-inch silicon carbide production line with Xinglian Integrated. Supplier relationships are turning into shareholder relationships—this is the most noteworthy structural change recorded this period. Money landing and money tightening are fighting for water levels in the same pool; whoever lets go first will see their valuation give way.

This source draft contains 19 items, expanded across six major sections, with market data as of the 9/4 close. Three things worth watching next: whether the Maserati partnership is officially announced and implemented, whether skeptics like Roubini shift further, and what Meta says about layoff rumors during the quarterly earnings call. Until clear answers emerge for these three questions, the two main directions in this issue must continue to be viewed as hedging against each other.

---

🧠 I. Large AI Models

AI Startup Micro1 Preempts Google; Spirit Data Acquisition Faces Late Bidder

Bloomberg reported on September 4 that AI startup Micro1 is attempting to replace Google in acquiring training data company Spirit Data, inserting a direct bidder into Google's data asset transaction. High-quality labeled and evaluation data is already a scarce commodity for frontier labs; the appearance of bidders indicates that pricing power for data assets is beginning to disperse toward sellers and disruptors. For startups, "buying data pipelines" is becoming an arms race project alongside "hoarding compute." If Google is forced to raise its bid, the rising cost curve for training data serves as a correction signal for gross margin expectations across all industry model vendors. Valuation anchors for similar targets in the primary market, such as Spirit Data-type sellers, are also directly pushed upward by this deal.

Kai-Fu Lee: China Will Win the AI Adoption Race

In a Bloomberg interview, Kai-Fu Lee provided very specific criteria for judgment: it's not about the upper limit of models, but about penetration rate in implementation. His logic is that the capability gap between frontier models will converge with open source; whoever integrates AI into more terminals, more industries, and more emerging markets will capture this round of dividends. This is a key endorsement received by the narrative of Chinese AI applications going global in the primary market, and it explains why Sinovation Ventures' investment focus remains heavily on the application layer rather than the foundation layer.

• The Economist: The End of AI Jobs Has Been Postponed; The AI Job Boom Is Already Here. An analysis published on September 4 uses hiring data to show that growth in AI-related positions has offset substitution effects, creating a directional divergence between panic narratives and employment reports. For employers, the salary anchor for AI jobs is still in an upward phase; the cost of grabbing talent is lower than the cost of waiting to transform processes. For the secondary market, the weight of the macro narrative "AI destroys jobs" can yield to empirical data showing "AI creates jobs." This does not contradict Meta's layoff rumors: substitution happens in old roles, growth happens in new roles; friction costs are borne by individuals, while total dividends are absorbed by corporate statistical metrics.

---

💻 II. AI Software

• Google's Personal Agent Gemini Spark Begins Taking Over Google Photos Album Management. After the official announcement of this integration on September 4, personal assistants can organize, retrieve, and clean up albums based on natural language instructions, taking over high-frequency, low-gratification tasks like "managing photos." Entry-level AI assistants have won another battle in penetrating user data depth. Once agents take over highly emotionally sticky assets like photo libraries, switching costs for cloud storage and subscriptions will only increase; the stickiness premium for Google services needs to be re-evaluated.
• AI Security Tool Aisle Finds 6 Vulnerabilities in Curl That Both Mythos and Codex Missed. This case reported by ZDNET is worth noting: the audited object is top-tier open-source infrastructure maintained for thirty years; even after human top-tier security models scanned the codebase, vertical AI audit tools could still dig out new holes. The AI premium for security software now has its first verifiable empirical sample. Similar vulnerability management startups will receive harder comparative data in due diligence. Conversely, the miss rate of general-purpose large models in professional audits should enter enterprise security procurement evaluation tables; the cost advantage narrative of "general models casually doing security" is weakened by one notch.

---

🤖 III. Humanoid Robots

• NRTA Held a Symposium on AI Micro-Drama "Image Fusion" and "Voice Fusion" on September 2, attended by platforms like iQIYI and Hongguo. The agenda revolved around the use of rights involving portraits, voices, etc., in generative AI micro-drama creation. The meeting's stance was "fully respect rights" plus "build industry consensus," which translates to: the compliance framework for AI-generated content is shifting from interview-style management to rule-based management. For platforms scaling through mass-produced AI short dramas, content budgets must start allocating funds for authorization and compliance. For managers and agencies holding portrait and voice assets, this is an asset repricing event. The same rights framework will eventually apply to anthropomorphic interactions of humanoid robots; embodied intelligence companies that negotiate digital human authorizations now will have first-mover clauses in the future.

---

🚗 IV. Autonomous Driving

Rumor: Maserati Teams Up with Huawei and JAC for Joint EV Development

Italian media Milano Finanza reported on September 4 that Stellantis-owned Maserati is accelerating its joint development project with Huawei and JAC Motors in China. Sources say substantial progress has been made, marking a core step in Maserati's brand revitalization plan. It is unprecedented for an ultra-luxury brand to bet its electrification turnaround on the Chinese supply chain plus Huawei's smart driving stack. If realized, Huawei's Automotive BU secures its first European luxury brand OEM-level client, and JAC completes the leap from contract manufacturer to joint developer. The licensing and revenue-sharing standards for "Chinese solutions going global" will be priced by this case first; second-tier luxury brands still observing the joint venture model will have fewer negotiating chips.

• Qijing GX7 Opens Pre-sales at 249,900–311,900 RMB. This new brand jointly built by GAC and Huawei features a mid-large five-seat SUV with range-extender powertrain. Four versions have pre-sale prices ranging from 249,900 to 311,900 RMB, with the entire lineup equipped with Qiankun Smart Driving ADS 5. For the first time, the Huawei ecosystem writes ADS 5 into the standard configuration list for a 250k RMB segment SUV. The price drop for flagship smart driving capabilities completes another notch; configuration sheets for fuel and pure electric competitors in the same price range will need to be rearranged.
• New Long-Wheelbase Mercedes-Benz GLE Set for Launch on September 16. The wheelbase is extended to 3115mm, featuring a triple-screen interior. It is among the first batch to carry the Momenta R7 world model and includes a Doubao LLM virtual assistant. Luxury brands writing Chinese smart driving solutions into global vehicle launch configurations raises the gold content of Momenta's client list another notch. After Mercedes, the selection pressure for other multinational OEMs will only increase.
• Chery's New Jetour X90 Set for Launch on September 9. Positioned as a mid-size SUV, it comes equipped with Falcon 500 assisted driving. Chery's self-developed smart driving stack penetrates down to the 150k RMB price segment, signaling that Chery's technology democratization is starting to deliver from press conferences to product lines.

---

🌐 V. Physical AI

Li Auto Signs New Strategic Cooperation Agreement with Xinglian Integrated; 8-Inch Silicon Carbide Wafers Roll Off Line Simultaneously

Signed on September 3, both parties witnessed the successful roll-off of 8-inch silicon carbide wafers. Previously, Li Auto just announced a 2.65 billion RMB capital increase in Sunwoda Power, subscribing to 8.79% of its post-increase shares. Upon completion, Li Auto-related entities will indirectly hold a combined 11.17% stake in Sunwoda Power, betting simultaneously on power batteries and power semiconductors. The way automakers intervene in core components is upgrading from purchase agreements to deeper equity ties plus joint production lines. 8-inch silicon carbide is the next stop on the electric drive cost curve; Li Auto has pinned itself next to the fastest ramping-up production line. For second-tier new forces who haven't placed bets yet, this marks the beginning of a widening cost gap. Xinglian Integrated's valuation anchor will also switch from "foundry capacity utilization" to "depth of head-client binding."

• "Dr. Doom" Roubini: Not Worried About an AI Bubble. Economist Nouriel Roubini, known for his bearish views, stated in a Bloomberg interview on September 4 that he is not worried about the so-called AI bubble—a contrast to his previous scathing criticism of tech capex. When "Dr. Doom" softens his stance, it carries more information than any bullish report: market divergence is shifting from "is AI a bubble?" to "which segment is the bubble in?" When the most steadfast skeptics begin accepting the rationality of infrastructure investment, it is often a signal that cycle expectations have completed repricing. But note his original words only discuss bubbles, not stating that valuations are already cheap.
• Krafton (PUBG Parent) Adds $250 Million Investment in India, Betting Beyond Gaming. This is another injection following multiple previous rounds. Korean content capital continues to bet on the Indian market; the track merging AI and interactive entertainment is gaining stronger cash-absorbing ability in Asian capital pools. The Indian user base remains hard currency for content globalization valuations. For domestic gaming and social platforms, the same window is closing; the moat of first-movers' user assets accrues interest annually.

---

📈 VI. Macro & Market Data

• Baidu A-Shares Officially Included in Stock Connect Targets under Shanghai-Hong Kong Connect, Effective September 7. Mainland funds can buy directly starting next week. The announcement states this marks an important step in expanding the company's reach to mainland Chinese investors; inclusion is expected to further diversify the shareholder structure. Another internet giant named for southbound fund pool expansion; the convergence of liquidity discounts in HK stocks continues. For targets like Baidu whose valuations have long been suppressed, the elasticity of marginal improvement in funding conditions usually exceeds that of fundamental improvements.
• Meta Internally Rumored to Cut Up to 60% of Teams Due to AI; Previously ~30% of Engineers Transferred to Data Labeling. Sources obtained by Pragmatic Engineer indicate management cited AI efficiency gains as the reason. The first bill for AI efficiency lands on the labor side. Shares closed up 1.00% that day, as the market read cost-cutting as positive news. This pricing direction warrants caution: if layoff benefits are counted into valuation, incentives for large model vendors to compete in cutting R&D outsourcing costs will be strengthened, and those "optimized" people happen to be executors in data labeling and content moderation links.
• Political Costs of Data Center Site Selection Begin to Become Explicit. Bloomberg reports that Carney's push for Canadian data center construction is hitting local backlash; grid, water, and land have become new bottlenecks for North American compute expansion. Middle East data center expansion proceeds under the shadow of war; capital stickiness driven by sovereign funds exceeds expectations; geopolitical risk discounts have not deterred long-term funds. Abu Dhabi sovereign AI company G42 is considering selling US equity stakes to preserve access to Nvidia chips; export controls are rewriting the equity structure of global AI capital. Three reports point to the same thing: in the cost function of compute expansion, the weight of policy and site selection terms is rising; the era of "build wherever there is power" is passing.
• JD Logistics CFO Change: Wu Hao resigned as Chief Financial Officer to take a new position at JD Group; Mao Jun takes over, effective September 4. The announcement notes Wu Hao's new role is at the group and related consolidated entity level, with specific division of labor undisclosed. The replacement of the financial helmsman at the listed entity, overlaid with a new group-level role, points to a restructuring of JD system capital operations. Subsequent financing or spin-off actions in the logistics sector are worth tracking.

Market Data, as of 9/4 Close. A-share AI leaders split evenly between gains and losses: Foxconn Industrial Internet +0.78%, Kingsoft Office +1.65%, Zhongji Innolight +0.12%, Eoptolink +0.32% closed red; Cambricon -2.54%, Hygon Information -1.03% pulled back. Thousand-RMB stocks continue to digest valuations; order realization and application-side funds are flowing back, while training chips consolidate at highs. Within sub-chains, prosperity signal density in device and material segments (memory, silicon carbide) is significantly higher than in complete machine computing, mutually confirming with this issue's news landscape.

US stocks showed violent divergence: Tesla -5.92% led the decline among the Magnificent Seven, hit by a double blow of disappointing Cybercab updates and federal safety investigations; Palantir -4.49%, Microsoft -2.04%, Alphabet -1.11%, Amazon -0.15% closed lower. Nvidia +0.84% and Meta +1.00% bucked the trend to close red. Funds are exiting high-valuation stories and shifting toward AI with cash flow and order realization. If this style switch continues, the upcoming earnings season starting next week will be the first testing window; the scissors difference between capex and labor costs in guidance will be scrutinized company by company.

---