---

Today's main thread is AI tightening at both ends at once: money on one side, rules on the other. The US Federal Trade Commission has opened an investigation into OpenAI, Anthropic and others, looking at what risks the products themselves pose. On the money side, banks like Societe Generale and Mitsubishi UFJ have pulled out of data center loans, and Wall Street and Silicon Valley are arguing over AI's price tag. Amazon, meanwhile, has stuffed Alexa+ into Fire TV, and AI assistants are starting to move into televisions. On the China side, who gets the ticket when a self-driving car runs a red light, and who hands over the driving data, are still at the draft stage. On Wednesday, September 30, five of six A-share AI hardware names fell and one rose, with Kingsoft Office the only one closing green.

---

Six, Macro and Market Data

A self-driving car runs a red light, the ticket goes to the carmaker, and the carmaker has to hand over the data for evidence

An article reposted by Huxiu on September 30 raised a very concrete question: when a self-driving car runs a red light, who does the traffic police ticket go to. The current practice is to ticket the carmaker, because the car is being driven by the carmaker's system. The trouble is in gathering evidence. Determining a violation requires driving data, and the data sits on the carmaker's servers, so the carmaker has to hand it over itself. The author's analogy is blunt: this is no different from letting students grade their own exams. The article says the relevant draft now out for comment leaves the evidence-gathering step to the carmaker. For drivers, this determines who has to prove what after an accident. Before buying a car, ask one question first: how long does this company keep driving data, and does the traffic police need to go through procedures to retrieve it. For companies doing autonomous driving, whether the data interface is opened and to whom will soon be written into the rules. The draft puts the responsibility on carmakers, so carmakers have to find a balance between keeping records and privacy: keep too little and you can't explain what happened; keep too much and users' trips end up on someone else's servers.

US FTC opens investigation into OpenAI and Anthropic, targeting the risks of the products themselves

CNBC reported on September 30 that the US Federal Trade Commission (FTC) has launched an investigation into AI companies including OpenAI and Anthropic, on the grounds that their products may pose risks. An FTC spokesperson confirmed the investigation to the media. The focus of the probe is how the products will be used and who they might harm, which is broader than a typical merger review. In the same week, another regulatory push is also advancing: tech giants are being pressed on why data center deals are being signed in secret. For people who use ChatGPT and Claude every day, there's no visible change in the short term. For companies building AI products, the compliance checklist gains one more item: lay out the risk points before launch, don't wait for regulators to come asking. Companies aren't entirely passive here either: whoever puts out safety test reports and incident-handling procedures first has one less hurdle when bidding for government and bank clients.

Tech giants pressed: why are the big data center deals being signed in secret

The Wall Street Journal reported on September 30 that several tech giants are being pressed over their AI data center deal arrangements, with the question being why those multi-billion-dollar leases and power supply contracts weren't fully disclosed. Data centers are the heaviest asset in this round of AI expansion — buying chips, building facilities, pulling power, the money is all piled up here. The report says the more complex and opaque the deals, the harder it is for outsiders to judge how much long-term obligation each company is carrying. What ordinary investors can do is very concrete: dig through a company's financials, find the data center-related long-term commitments, and compare them against its cash flow. For suppliers of servers, power and racks, whether the customer list is public directly determines whether orders can be counted in their own earnings guidance. Once power contracts and long-term leases are public, peers can work out a rival's capacity rhythm, which is also why many companies don't want to disclose.

• IT Home, citing Korean media Sammobile on September 30, said AMD CEO Lisa Su is expected to visit Korea next month for talks with Samsung Electronics executives. Samsung supplies memory for AI chips, AMD makes server chips, and the two companies' products often end up in the same machine. In recent years the bottleneck for AI machines has often not been the chip itself but the supporting memory and storage, and whoever can guarantee supply has the say. What Su is going to discuss on this trip is currently only per Korean media, with no formal announcement from either side yet, so treat it as a rumor for now. Korea's memory makers have all been scrambling for AI memory capacity lately, and whoever talks deeper with AMD gains one more long-term customer.
• The Information analyzed on September 30 that terms for data center financing are loosening. Lenders are starting to demand longer repayment buffers, higher interest, and even require project owners to put up some of their own capital first. These concessions look like protecting creditors, but in practice they push construction costs up. Data centers mainly recoup costs through long-term leases, so when costs rise, the payback period stretches out. The article's judgment is that once financing tightens, the start dates for the next batch of data centers will be pushed back, and the companies renting out machines will scale up more slowly too. For companies renting machines, this is bad news: new supply slowing down doesn't necessarily mean rents come down.
• The Information reported on September 30 that several banks including Societe Generale of France and Mitsubishi UFJ of Japan have pulled out of several data center loans. Data centers are where AI companies put their servers, and one costs anywhere from over a billion to several billion dollars, with most of the money borrowed from banks. When banks pull out midway, project owners have to find someone else to take over, or put up more of their own money. The article also notes that this kind of lending grew very fast over the past year, and now underwriting is slowing: more materials to review before signing, fewer banks willing to take it on. For projects already under construction, the most direct impact is higher borrowing costs.
• Bloomberg reported on September 30 that cloud services company Vultr placed a $1.2 billion order with HPE for servers equipped with AMD chips, to expand the scale of machines it rents out. This is HPE's latest AI hardware order. The order is large and will be executed over several quarters, so it spreads out over HPE's results going forward. What to watch is whether this kind of big order keeps coming; a single order's size doesn't say much. For companies doing server contract manufacturing and components, the rhythm of these orders is their production schedule. AMD winning this order also shows server makers are giving customers an option beyond Nvidia.
• The Information reported on September 30 that sports prediction trading company Novig is valued at $2 billion in a new funding round, quadruple its previous round. Among its investors is actress Sydney Sweeney, and the star power has drawn it plenty of attention. In the prediction market space, Kalshi and Polymarket are fighting over the same pool of users, and Novig wants to cut into the more mainstream sports angle. Sports events are frequent and results come fast, making them better suited to a daily-use product than election-type contracts, and also harder to regulate. It will have to clear the regulatory hurdle sooner or later, and several US states are already tightening the rules on sports betting.
• Bloomberg reported on September 30 that prediction market platform Kalshi is finalizing a new funding round at a $40 billion valuation, timed just before its IPO. Kalshi's business is turning people's judgments about elections, games and economic data into tradable contracts. A $40 billion valuation would have been hard to imagine a year ago. Whether the round closes at that price depends on whether investors sign in the coming weeks. Raising again before listing is usually about padding the cash on the books and being less hostage to roadshow conditions. Prediction markets have walked a years-long compliance path in the US, and listing is its exit for investors.
• TechCrunch reported on September 30 that Flow Engineering closed a $50 million Series B at a $750 million valuation, with investors including Sequoia, Valor and Atreides. The company uses AI to assist hardware design, helping teams cut a few rounds of prototyping and validation. Hardware design cycles are long and trial-and-error is expensive — a chip goes through several validation stages from idea to mass production, and shortening one of those stages saves money for companies making chips and electronics. Its customer list and renewal rate are the two sets of numbers to watch next. Bloomberg also reported the funding the same day, and the two outlets' accounts match.
• JPMorgan executive — Bloomberg reported on September 30 that JPMorgan executive Sundar described the current AI market as "healthy turbulence." He spoke of two cycles: one is capital expenditure, meaning the money to build data centers and buy machines; the other is application deployment, meaning companies actually putting AI into their business. His point is that the spending cycle and the adoption cycle aren't in sync, so the market will swing back and forth. For investors, the use of this is not to judge by a single rhythm — when you see a pullback, first figure out which cycle is retreating. This line is also useful for people in foreign trade and manufacturing: the hot and cold of orders will come in two waves.
• A September 30 article from The Information says Wall Street and Silicon Valley are diverging on what AI is worth. One side is calculating whether the investment can still turn into profit, the other is looking at a technology curve that hasn't finished running, and the two sides' books use different assumptions. The disagreement shows up in concrete places: for the same company's valuation, the gap between the seller's price and the buyer's price is growing. For ordinary investors, buying a broad index at times like this is less stressful than betting on a single company, and less volatile too. If the disagreement drags on, AI-related stocks will swing more than the broader market.
• Cerebras chief speaks — TechCrunch interviewed Cerebras CEO Andrew Feldman at the Disrupt conference on September 30, asking whether AI models can keep scaling. His answer was roughly that bigger models need more machines and power, and this path will hit a wall eventually, but he thinks it can go a while further. Cerebras builds chips that turn an entire wafer into one piece, betting on "big." His company's orders and customer renewals say more than the founder's statements. He gave no specific timeline, only saying that wall hasn't arrived yet.
• Engadget reported on September 30 that Amazon gave Fire TV a new interface, revamped the mobile app too, and built in the upgraded Alexa+. The change is that operations on the TV box and phone now use conversation to find content, instead of clicking through menus layer by layer. Not many people use Fire TV in China, but putting AI assistants into TVs, speakers and cars is a direction everyone is moving in, and Xiaomi and Huawei's TV boxes are also adding voice entry points. If you're buying a new box, first check whether older models can be upgraded, and don't switch devices just for one redesign. The standards for voice entry points aren't unified yet, so before buying, confirm which company's large model it connects to.
• Bloomberg released an interview video with CoreWeave's CEO on September 30, discussing new products and demand for renting machines. CoreWeave rents out rows of GPU machines to AI companies, and its customers include a batch of companies building large models. Its revenue is tied to how fast its customers burn cash, so its statements are often treated as an industry thermometer. To read the AI infrastructure market, this company's orders and renewal prices are more useful than its statements — a lease is signed for years and can't be walked away from. Its own machines are rented too, so when rents and power prices rise, its profit gets squeezed in the middle.
• Tech.eu reported on September 30 that travel booking company Tryp.com raised 1.9 million euros, with a team that started as six engineering students. The platform combines flights, hotels and itineraries from various countries, and its user base has reached 10 million. The funding amount isn't large; it's betting on the speed of user growth, and small companies like this compete on making price comparison more granular than big platforms. For ordinary users, one more price comparison entry point doesn't hurt, and flight and hotel price gaps are worth opening one more page to compare. Whether a small platform survives depends on not burning through its subsidies.
• Bloomberg reported on September 30 that AI infrastructure company Accelevation closed down 2.5% on its first day of trading. Companies with AI in their names are starting to show mixed listing performances lately: some pop on day one, others open lower. For IPO subscribers, whether the issue price is set high matters more than what the company is called. Breaking below the issue price on day one is mostly a case of expectations being maxed out during pricing, and the trend afterward depends on the orders it actually lands. The AI infrastructure line has slow revenue recognition and heavy capital expenditure, and the market won't be patient for long.
• TechCrunch reported on September 30 that Restate raised $20 million for low-level tooling that keeps programs from losing progress midway when running long tasks. AI programs often have to run several steps in a row and wait for external systems to respond, and if one step breaks, the whole thing has to start over. Tools like this don't show up in front of users, but once a company scales up it can't do without them — things like bank reconciliation and inventory sync rely on them as a backstop. The founder started on this in 2022, betting that programs would get longer and longer. Once customers pile up, tools like this are like water and electricity: when they break, everything stops.
• CNBC reported on September 30 that dating app Grindr acquired telehealth company PurposeMed for $250 million, its first step outside the dating business. Its user base is concentrated in a specific group, and buying a telehealth service is meant to keep health consultations inside its own app too. Beyond ads and subscriptions, one more revenue line. The point to watch with this deal is whether longtime users are willing to see a doctor inside a dating app, and paid conversion will need a few quarters of data to verify. The compliance cost of a medical business isn't low, and it'll take years to see whether this money was well spent.

---

Today's Market Quick Look (in-house data, global 24 hours (US stocks 9/30 (Wed) close + A-shares 9/30 (Wed) close))

Five of six rose and one fell. Amazon rose 1.01%, the strongest, Alphabet rose 0.93%, Microsoft rose 0.77%, Tesla rose 0.56%, Nvidia rose 0.51%. Meta fell 1.84%, the only one in the list to close down. This is the September 30 Wednesday US stock close.

Five of six fell and one rose. Kingsoft Office rose 0.88%, the only one in the list to close green. Hygon Information fell 3.85%, the most, Cambricon fell 3.54%, Eoptolink fell 1.00%. Optical module maker Zhongji Innolight fell 0.56%, and server contract manufacturer Foxconn Industrial Internet fell 0.38%. These are the September 30 Wednesday A-share closing prices.

---