---

Today's main threads are pricing and capital. OpenAI set its Pro plan at $500 per month at Developer Day, also launched dots, an always-on assistant, and casually moved the office trifecta into the chat box. MongoDB released 9.0 and Atlas Infinite the same day, scaling up data capacity for AI applications. On the capital side, one cold and one hot: data center developer 5C is preparing for an IPO, while smart ring company Oura postponed its IPO due to market instability. Domestically, Qianli Technology spent 34.4388 million yuan to buy Zeekr's smart driving R&D assets. On September 29, A-share AI hardware saw five declines and one gain, with Cambricon the only one in the list to close green.

---

Six, Macro and Market Data

Qianli Technology spends 34.4388 million yuan to buy Zeekr's smart driving R&D assets

IT Home reported on September 29 that Qianli Technology announced its controlling subsidiary Qianli Smart Driving signed an asset transfer agreement with Zhejiang Zeekr. What was bought was electronic equipment and installation work for smart driving R&D, priced at 34.4388 million yuan. Zeekr selling R&D equipment to a company in the same system saves it the expense of maintaining its own smart driving line; Qianli Smart Driving taking over the equipment is equivalent to handing the autonomous driving work to a more focused party. The Geely system is doing subtraction on smart driving. For those holding auto parts and smart driving concept stocks, watch whether this kind of asset shuffling becomes more common. Automakers spend money on self-developed teams; selling equipment to specialized companies and then buying services back is the cheaper route. The amount isn't large, but it points to a direction.

OpenAI Developer Day: always-on assistant dots launches, subscription sells for $500 per month

Bloomberg reported on September 30 that OpenAI launched dots at its San Francisco Developer Day. This is an always-on AI assistant, not just one waiting for you to ask questions in a dialog box. At the same event, OpenAI set its Pro plan at $500 per month. The expensive tier bundles Astra Ultrafast, which the company says boosts answer generation speed by up to 8x. OpenAI also released ChatGPT data spaces, dynamic pages, and collaborative slides, moving the office software trifecta into the chat box, directly taking on Microsoft Office. What matters to you is your wallet. Free tier quotas were cut, and using the fast model requires paying more; ordinary people won't spend $500 a month. OpenAI isn't public yet, so what you can touch are its rivals: will Microsoft Office and Google Workspace cut prices in response? Funds holding those two should take a closer look.

300 media outlets join forces to push a bill setting thresholds for AI scraping content

Digiday reported on September 29 that executives from about 300 media outlets including Condé Nast and Hearst jointly pushed a federal bill. The bill targets bots that automatically crawl web content, which the industry calls the bad bots bill. What media fear is training data being taken for free. After news and images are scraped in bulk, readers ask AI directly instead of clicking into websites, and ad revenue drops accordingly. This time they're pushing for unified federal rules, no longer relying on each website to block on its own. For those holding content platform stocks, watch whether copyright negotiations can turn into revenue. For companies doing AI training, data procurement costs will go up. The news ordinary people scroll through will increasingly come from paid licensing rather than casual scraping.

• Hushu published an article by Zhanmo Tianmo on September 29: NIO has built 4,000 battery swap stations and begun opening them to external brands. Swap stations are heavy assets with substantial per-station investment, and serving only your own car owners makes it hard to recoup costs. Opening them to others dilutes costs and raises utilization. What it means for you: in the future, buying an EV from another brand might also let you use NIO's swap network, adding an option for long-distance recharging; for those investing in the new energy vehicle supply chain, watch whether shared recharging becomes standard. Once the swap network opens up, the test is whether NIO can hand its operations standards to others without wrecking its own reputation.
• Tencent Marvis positions itself as a "butler" — Leiphone reported on September 29 that Tencent's Marvis positions itself as a butler inside your computer, handling files, apps, and browser pages, not chatting with you. Overseas, the phenomenon of installing more and more assistants with increasingly scattered permissions is called "assistant sprawl," and what Marvis wants to answer is who manages all these assistants. What it means for you: if your desktop files are a mess and you're slow at finding things, you can wait for this kind of tool to mature and try it, first seeing whether it dares to open up system permissions before deciding whether to hand over your computer.
• IT Home reported on September 29 that Alibaba announced it is extending the limited-time free period for Qwen3.8-Flash on the Qoder platform. Originally set to end at 23:59 on September 30, it will now continue to be free starting October 1. Qoder is Alibaba's AI coding platform, and extending the free period again and again clearly shows it's using low prices to grab developers. What it means for you: if you write code yourself, try it while it's free; pay-as-you-go budgets can be pushed back a bit, and you can pay once the free period really ends, so don't rush to renew.
• South China Morning Post reported on September 29 that an Accenture survey shows only about one-tenth of enterprises in Hong Kong have deployed generative AI, lagging behind the mainland. The report attributes the cause to too many legacy systems: old IT architectures in industries like banking and real estate are slow to change. The gap isn't about willingness to use it, but about old systems dragging things down. What it means for you: for companies doing enterprise software and system integration, this wave of transformation in Hong Kong is a source of future orders; teams that have done similar projects in the mainland can directly transfer their experience over.
• A Hushu report on September 29 covered a new business: helping AI remember the context of users and tasks. The model itself doesn't remember things long-term, and every new conversation requires re-explaining the background. Making memory a separate layer saves repeated input and makes the assistant understand you better. The article says this is already a business worth hundreds of millions. What it means for you: for companies doing customer service and coding assistants, whoever makes this memory layer solid will have better renewal data. Once memory is built into a product, the switching cost for users wanting to change tools goes up, and that's also the confidence behind these companies' renewal pitches.
• An article reposted by Hushu on September 29 described an office: over fifty second-hand phones, hundreds of accounts, and proven short videos rewritten into many versions and posted repeatedly. The author's point is that after AI pushes production costs close to zero, what matters isn't who can make it, but whose accounts and distribution are more stable. The same piece of content gets copied into dozens of versions, and platforms can't tell who's original either. What it means for you: if you want to do content as a side hustle, first think clearly about distribution channels, and don't just focus on whether you can make it. Platforms are also stepping up original-content detection, and accounts doing bulk rewrites are getting banned faster and faster; what survives are those with a real audience.
• The Information reported on September 29 that OpenAI's annualized revenue has already approached $70 billion. This measure multiplies recent monthly revenue by 12, which isn't the same as actual full-year receipts, but it shows the growth rate. In the same week it was still negotiating a new funding round with investors, reportedly around $30 billion. What it means for you: if you really want to value OpenAI, you have to wait for its IPO filing; for now, what you can reference is what comparable businesses at Microsoft and Google sell for, plus the pricing of this funding round.
• MongoDB released MongoDB 9.0 and Atlas Infinite on September 29, the latter being its cloud database service. The new version focuses on withstanding the data volumes brought by AI applications. The data fed into AI applications is large and messy, and old databases easily choke. The company calls it the best version ever, with the focus on stability and scaling. What it means for you: for small teams building their own backends, calculate migration costs before upgrading; for those using cloud services, check the scaling bill in advance so you don't discover overspending only at month-end.
• Moneycontrol reported on September 29 that Meta brought in MongoDB CEO Chirantan Desai to head its enterprise AI business. Meta has spent the past two weeks pushing its enterprise platform and releasing a small-business version of its assistant, and it lacks someone who knows how to sell software to companies. Using the head of a database company, what they value is precisely his batch of enterprise customers. What it means for you: for those doing enterprise software, Meta entering means one more competitor for deals; for local channel players, watch whether it pulls service providers in to sell together.
• Bloomberg reported on September 29 that Brookfield-backed AI data center developer 5C is preparing for an IPO. Most data center money is borrowed, and going public can convert part of the leverage into equity. Among this batch of projects, many are still blueprints and site selections, with few actually powered up. What it means for you: when looking at this kind of company's prospectus, first check the ratio of what's built versus under construction, then look at signed area, and don't be misled by numbers propped up by letters of intent. Several similar companies are queuing up to list this year, and the market's attitude toward this kind of heavy asset will be clear within weeks.
• Bloomberg reported on September 29 that data center construction sites bring about $3 billion in opportunities to catering companies. Sites often have thousands of people and are built over several years, making cafeterias and meal delivery a stable business. This money isn't in chips and electricity bills; it's hidden in the pots and pans at the construction site. What it means for you: for companies doing group meals and food supply chains, these large sites can sign long-term contracts; for investors, this is an inconspicuous side branch with steadier cash flow than equipment makers. This business doesn't rely on price increases, but on long enough construction periods and enough workers. The more data centers are built, the steadier this kind of supporting order becomes.
• Tech.eu reported on September 29 that Berlin-based Apheris and Ginkgo Datapoints launched an antibody developability consortium, with initial members including several pharma companies. All parties contribute about 10,000 antibody data points to train models together, with raw data not leaving their respective systems. Pharma companies fear data leaks most, and this approach of pooling data without handing it over fits the problem exactly. What it means for you: for those doing drug R&D, this approach can save the cost of building your own dataset, and it's worth asking how to join the consortium.
• TechCrunch reported on September 29 that Wabi, which does "type to generate an app," has pivoted to a messaging format. Its old path was letting users generate small apps from a paragraph of text, and now it's putting AI into chat instead. Demand has changed: assistants like Meta's Muse have taken away the business of independently making small apps. What it means for you: for teams making small tools and standalone apps, first see clearly whether users actually want an app or want to get things done in a chat box.
• A Hushu article on September 29 tallied the "15th Five-Year Plan" of 291 cities, of which 198 listed new energy in their industry lists. The author reminds that being written into a plan doesn't equal projects landing, and part of it is empty promises. The overcapacity in solar and energy storage over the past few years has already taught a lesson. What it means for you: when looking at city industry funds and investment promotion projects, first distinguish which are slogans and which already have real capacity before deciding whether to follow, and don't be intimidated by the word count in the plans. This round focuses more on consumption and the grid: the more capacity piles up, the easier it is for curtailment to be brought up and accounted for.
• Hushu reported on September 29 that a 712 million yuan energy storage contract was terminated 48 days after signing, with risk exposure of about 800 million yuan. The article says the funders are shifting the risk of independent energy storage projects to shareholders and construction parties. Signed fast and withdrawn fast shows the accounts weren't squared. What it means for you: for companies in the energy storage chain doing construction and equipment supply, payment collection risk is rising; when looking at orders, pay more attention to who is paying and whether there's a guarantee, and don't just look at the contract amount. Similar terminations this year in independent energy storage projects are no longer the first, and the gap between signed amounts and final implementation is getting wider and wider.
• Bloomberg reported on September 29 that some UK data centers have switched to natural gas to generate their own power because grid connection schedules are too long. Local officials are pushing to shorten grid connection waiting times. Self-generation is more expensive and dirtier, but projects can't wait. What it means for you: where data centers cluster, power supply will get stuck first; when choosing a site, first look at how much margin the local grid still has, and don't just look at land prices and government subsidies, because the grid connection schedule may be longer than the construction period. Building your own generator sets can quench immediate thirst, at the cost of higher electricity bills and carbon emissions, so the accounts need to be clear.
• The Information reported on September 29 that General Catalyst invested in Proximal, which makes coding datasets, and this company's revenue has also reached a new milestone. This kind of company supplies code data for training and evaluation to AI coding tools. Whether the model writes correctly depends on this batch of data for scoring. What it means for you: for those watching the AI coding track, there are already specialized companies in the data supply link, so don't just watch the models and tools themselves; the money will flow first to the data sellers.
• The Information reported on September 29 that Amazon is adding new automation tools to its advertising business to automatically allocate placements for merchants. Advertising is one of Amazon's fastest-growing profit sources. Handing bidding and price adjustment to the system makes the platform's cut more stable. What it means for you: for sellers opening stores on Amazon, placements are increasingly decided automatically by the platform, leaving less room for manual price adjustment; in your budget, keep a portion to test the suggestions it gives, and don't hand everything over at once. With the platform managing placements itself, seller data also flows back to the platform, leaving even less room for bargaining later.
• An analysis on September 29 asked when the AI price war will begin. The author says that right now everyone sells their most expensive model as the flagship, and a real price cut must wait for hardware and data costs to come down first. At present, the cheaper tiers are quietly shrinking their quotas. What it means for you: for small and medium enterprises procuring AI services, don't sign long-term contracts first; models change every quarter, and pay-as-you-go is more cost-effective and makes it easier to switch suppliers at any time. The most direct change is already happening: free quotas are shrinking, while high-priced tiers are actually getting more.
• An article on September 29 said that both hospitals and insurance companies in the US are using AI to handle claims and appeals, with both sides automating the same dispute. The author worries that escalating on both sides will only make the process more convoluted. What it means for you: for companies doing healthcare IT, products that can serve both payers and hospitals are easier to roll out than those standing on only one side. Once both sides use it, whoever gets it working first gains an advantage in the process, and the slower side can only add people to catch up.

---

Today's Market Quick Look (in-house market data, global 24 hours (US stocks 9/29 (Tuesday) close + A-shares 9/29 (Tuesday) close))

Of the six, four fell and two rose. Meta rose 3.24% as the strongest, Amazon edged up 0.21%. Tesla fell 1.29%, Nvidia fell 0.72%, Alphabet fell 0.53%. Microsoft barely moved, down 0.05%. This is the close of US stocks on Tuesday, September 29.

Of the six, five fell and one rose. Cambricon rose 2.55%, the only one in the list to close green. Kingsoft Office fell 2.30% the most, Hygon Information fell 1.92%, Eoptolink fell 1.70%. Zhongji Innolight, which makes optical modules, fell only 0.24%, basically flat. This is the close of A-shares on Tuesday, September 29.

---