---

🧠 I. Large AI Models

Musk speaks at G20, predicting AI will increase global economic scale by 20%-30%

Musk delivered a speech at the G20 meeting on the evening of September 1, predicting that AI will increase the global economic scale by 20%-30%, equivalent to an annual output increase of approximately 134 trillion to 202 trillion RMB at current exchange rates. This is the first time a top industry leader has provided a quantified range for AI's incremental impact on the macro economy and placed it on the G20 agenda. Such statements serve as one of the anchors for capital markets to price the AI narrative.

However, the value of the anchor depends on the path to realization. Musk's macro statements have historically been aggressive, and the market needs to distinguish between technological possibility and the recoverability of ten-year capital expenditures. On the same day, investor Kedrosky publicly pointed out the decoupling between AI valuations and revenue growth rates, and Tesla itself closed down 3.22% that day, indicating the market did not buy into this statement.

Baidu CFO Haijian He: AI investment expected to soon generate profits and cash returns comparable to search

According to Bloomberg, Baidu Chief Financial Officer Haijian He stated that large-scale investments in the AI field are expected to soon generate profits and cash returns comparable to the search business, proving that Baidu's transformation from an internet company to an AI company is on the right track. Currently, AI-related revenue covers multiple business areas including cloud computing and applications. Search is Baidu's profit engine, and this is the first time a major Chinese tech company's CFO has described AI payback to a degree that benchmarks against its cash cow business.

This forms mutual verification with the surge in profits among downstream server manufacturers on the same day, as two markets answer the same question in two ways: money invested in compute is starting to turn into profits on financial statements. Baidu has long traded at a discount due to "search decline + unfulfilled AI promises." If profit metrics gradually approach the CFO's commitments, the valuation anchor for Chinese concept AI applications will rise overall, and other Chinese giants still burning cash will also face pressure to reconcile their books.

NIO Q2 Report: Revenue 32.14 billion (+69%), Gross Margin 18.4% hits new high, but per-vehicle cost to rise another 2000-3000 yuan in H2

NIO released its second-quarter financial report on the evening of September 1, with total revenue of 32.14 billion yuan, up 69.1% year-on-year, and comprehensive gross margin of 18.4%, up 8.4 percentage points year-on-year. Third-quarter delivery guidance is 108,000 to 111,000 units, up about a quarter year-on-year. CFO Qu Yu simultaneously forecast that per-vehicle costs will continue to rise by 2000-3000 yuan in the second half.

A realization-style earnings report collides with headwinds on the cost side. As smart driving cars become more popular, material costs for storage and compute hardware rise faster. NIO's mix of cool and hot results is a sample of the collective situation for EV startups: economies of scale have beaten competition but cannot beat the upstream component cycle. For NIO, valuation elasticity depends on whether Q3 deliveries can be converted into operating profit; for peers, if rising per-vehicle costs become an industry norm, the room for price wars in smart driving models in the second half will be quietly confiscated by upstream costs.

• DeepMind releases Gemini agent video understanding, launching across new models like Gemini 3.7 Flash. Models can call tools while watching videos, autonomously completing retrieval, summarization, and operation tasks. Video understanding moves from "seeing" to "acting while seeing," expanding commercial scenarios for multimodal Agents. Surveillance inspection, content moderation, and long-form video education are obvious markets. Google's route of stuffing model capabilities into its own traffic entry points becomes clearer, with Gemini's iteration rhythm always revolving around distribution via Search, Workspace, and YouTube. This contrasts with OpenAI's independent App strategy—two different businesses.
• Google Pics launches in Workspace, a prompt-based creation tool competing with Canva, focusing on professional-grade image generation and editing. Users can create images just by typing, without learning design software. Google is embedding AI image capabilities into the default entry point of its office suite, shifting the battle for creative SaaS from standalone apps to pre-installed suites. The customer acquisition cost logic for companies like Canva must be rewritten, leaving only workflow stickiness as a defensible moat.
• Macquarie University in Australia replaces face-to-face teaching in two psychology courses with AI chatbots, accompanied by online quizzes and optional online tutoring. Students can ask questions repeatedly at their own pace, and the university states that educational effectiveness will only be assessed after one year. In the tug-of-war between higher education cost reduction and teaching quality, this is the deepest implementation of AI replacing human lecturers. Universities are places that preach long-termism, but the cost sheet doesn't wait a year. Before students vote with their feet, the budget table has already been calculated.
• Luo Yonghao attacks NEV makers' AI large models as "garbage" on Weibo, claiming he tested dozens of cars and none integrated the dialogue capabilities of top-tier full-power models like Doubao, Qwen, or Kimi. What lies between cockpit AI demos at car launches and the voice assistants actually used by owners was put on the table by a livestreamer. For automakers, this is a credibility issue regarding self-promotion; for model vendors, it means the vehicle entry point hasn't truly been handed over yet. The attribution of cockpit large model orders should now be spoken for by sales volume.

---

💻 II. AI Software & Security

• Security firm CloudSEK infiltrates misconfigured servers of hacker group Aurora, exposing internal materials. Attack plans were assisted by Cursor AI, and ransomware was rewritten in Zig language to evade detection. Cybercriminals treating AI as part of their standard toolbox means the production speed and mutation speed of attack scripts will step up. In the defensive AI arms race, opponents have already started fighting. As the attack surface expands, security budgets follow suit, making the revenue story for defenders smoother.

---

🤖 III. Humanoid Robots

• Leaks about Unitree Technology's management trend on hot searches, including claims that "reimbursements under 100 yuan require Wang Xingxing's personal approval" and "rewards/punishments consist almost entirely of penalties." Caijing cited employees saying the CEO personally oversees details as small as screw lengths. Unitree responded that much of the content is untrue and should not be taken seriously. On the eve of the IPO, the founder's strong control style has turned from management details into public discourse. Efficiency myths and employer brands begin to consume each other, and primary market valuations for such companies need to add a governance discount.

---

🚗 IV. Autonomous Driving & Smart Cars

• Avatr officially announces August deliveries of 8,082 new vehicles, with the all-new 07L launched and OTA upgrades rolling out for all models. Monthly deliveries below 10,000 units indicate the ramp-up phase is ongoing. Huawei-affiliated smart driving's scale story needs time to prove itself, while competitors' monthly sales have entered the 30,000-unit range. Beyond product definition, the fight is about channels and capacity.
• Leapmotor Q2 revenue 27.29 billion yuan, up 91.78% YoY, net profit 600 million yuan turning losses into gains, filling the 390 million yuan loss hole from Q1 in one quarter. Analysis from Lianxian Mobility suggests thin margins from low-price high-volume sales cannot support a tech valuation; the robot story is key to Leapmotor's next phase narrative. Money made from selling cars isn't enough to tell the story, and the industry understands this logic. But understood or not, profit margins on financial statements must be accumulated word by word. Two EV startups reporting in the same week as NIO are walking completely opposite cost curves.

---

🌐 V. Physical AI & Compute Infrastructure

• Sugon teases the world's first 64-thread mobile workstation, featuring 16GB VRAM and a body thickness of 16.9mm. It can run 35B MoE models locally with 50 Tokens/s output, officially claimed to lead the industry in AI inference speed. The arms race for edge AI compute has rolled from data center racks into portable machines. The compliance selling point of keeping data off-premises is becoming a new lever for domestic workstations to counter the NVIDIA ecosystem.
• Google signs power purchase agreement with geothermal developer Fervo, with the project scheduled to go operational in 2028. Amidst data center power shortages, tech giants are extending energy procurement to baseload sources like geothermal, which are unconstrained by sunlight and wind windows. This marks the start of testing new supply routes, whose success or failure will be copied by the entire industry.
• Pennsylvania Governor Shapiro publicly criticizes data center developers, stating many don't care about the communities where projects are built. The political cost of siting AI infrastructure is rising. Electricity cost sharing, water resource usage, and local employment commitments are becoming implicit financing terms for projects. This applies to both red and blue states. Developers who only calculate PUE earlier without considering public opinion will see subsequent approval cycles forced to lengthen.
• Samsung Display signs MOU with Chungcheongnam-do government regarding its 67 trillion KRW investment plan in Asan. Asan city government, power, and water companies attended, with parties promising administrative and infrastructural support for investment landing. Advanced display capacity expansion is written into regional infrastructure packages. Negotiating electricity and water before breaking ground is a model action of Korean-style industrial mobilization, and also a signal of the restart of the display panel capex cycle.
• Innolux debuts Chip Last FOPLP fan-out panel-level packaging platform, mass production expected in H2 2027. Panel makers are formally entering the AI/HPC advanced packaging supply chain with existing large-size production lines. The logic that compute bottlenecks are shifting from chip design down to packaging capacity has been discussed for a year; now there is another player with its own factory. The scarcity premium for CoWoS begins to face direct bidding. Panel-level packaging uses ready-made large-size glass substrates to dilute unit area costs, betting on the industry inflection point of AI chip packaging migrating from wafer-level to panel-level.

---

📈 VI. Macro & Market Data

• Nestlé sells vitamin supplement business to private equity fund for $1 billion. Consumer giants continue to slim down portfolios, concentrating capital back into high-growth core businesses like coffee and pet food. This deal is a benchmark transaction for PE acquiring health consumer assets. Clearing low-growth assets for cash follows the same financial logic as tech giants cutting non-core AI peripheral projects—both are freeing up ammunition for the next round of capex.
• Indian ed-tech unicorn Unacademy sold to competitor upGrad for $206 million, ~94% down from 2021 peak valuation. This star asset, once ranked among India's most valuable ed-tech startups, transacted at near-liquidation price during the capital ebb tide. The warning to the primary market is blunt: when share stories lack gross margin support, the downside space for valuation can be nearly zero. The buyer is also a competitor lacking cash flow.
• Prominent VC Kedrosky publicly warns of decoupling between AI valuations and revenue growth forecasts. Primary markets price for winner-takes-all, secondary markets trade for slowing growth. Neither side wants to admit error first, leaving mid-tier AI companies that are neither listed nor profitable caught in the middle. Latest earnings from overseas OEMs prove segments with orders can make money, highlighting how names without orders rely on narratives to sustain valuations. The reconciliation season clears this layer first.
• A-share AI chain mostly closed lower on Tuesday. Cambricon dipped slightly 0.29% to 1109 yuan, taking a small rest at historical highs. Foxconn Industrial Internet fell 2.93%, Hygon Information fell 1.37%. Funds chose to sell expectations on the AI server ODM chain ahead of overseas positive news landing in the same sector. Optical modules bucked the trend, closing higher: Zhongji Innolight up 0.87%, Eoptolink up 0.23%. The demand logic for 800G remains unchanged, but crowded directions yielded first. Divergence in A-shares is in sync with US stocks; both are waiting for the next order data reconciliation.
• Meta was the only large-cap AI stock to close higher in the US, up 1.08%, as the market starts looking for evidence of "AI investment paying off." NVIDIA fell 1.51%, Microsoft fell 1.24%, Palantir led high-valuation application targets down 3.47%. On the same night overseas OEM earnings confirmed orders turning into profits, funds took profits on the most expensive names. Tonight's divergence among AI leaders is the live version of the earnings reconciliation season.

---