---

Two threads today — one on the shelf, one on the books. Chinese home appliance makers brought AI into fridges, washing machines, and TVs at the IFA show in Berlin, talking about a multi-hundred-billion-dollar pie; that same afternoon in the same Berlin, a Chinese smart metro train also made its global debut. Over in A-shares, toy makers, sanitation companies, and film and TV companies are also announcing data center contracts, but by today's close all six AI hardware leaders were in the red. Two overseas items: Anthropic says its own model helped find a possible new gene-editing tool; US regulators have started investigating aftermarket assisted-driving company comma.ai, triggered by multiple fatal accidents.

---

📈 Six, Macro & Market Data

Chinese home appliance makers are putting AI into fridges and washing machines, chasing a multi-hundred-billion-dollar pie

At the IFA show in Berlin, Chinese vendors' booths are basically all saying the same thing: appliances hooked up to AI. Fridges recognize ingredients, washing machines pick their own cycles, TVs act as the home's control console. The piece reposted by Huxiu's take is that Chinese appliance makers are running ahead in this multi-hundred-billion-dollar industry. This path is different from AI on phones and cars. The chips in appliances are tiny, can't run large models, and most features rely on connecting to the vendor's servers to compute. The cost is direct too: cut the internet and it turns back into an ordinary appliance, and part of your household habits get sent over to the vendor. Buyers can pick by asking three questions: does it still work offline, what data gets uploaded, and do features require a separate subscription. The showroom demos are all under connected conditions — asking those first two questions before ordering is more useful than staring at the spec sheet.

Toy makers and sanitation companies are also signing data center contracts — a cross-industry AI infrastructure rush is rising in A-shares

This piece reposted by Huxiu tallies one thing: in 2026, more and more listed companies in A-shares are announcing moves into the data center business. The list includes toy manufacturing, sanitation operations, film and media, new materials, traditional processing, and new energy operations — many with no prior connection to AI. Contracts have gone from the earliest hoarding-cards-and-subletting to deals worth tens of billions. When reading these announcements, watch three things: where the money comes from (own funds or loans), who the client is (is a name written down), and how gross margin is calculated. Data centers are a heavy-asset business — electricity prices, rack utilization, and depreciation decide whether you make money, and companies used to making fast money in their main business may not be able to carry it. Readers holding these stocks can follow the announcement to find the client list and payment terms. Announcements with only a framework agreement and no client name are still far from real revenue.

After multiple fatal accidents, US regulators start investigating aftermarket assisted-driving company comma.ai

Ars Technica reports that the US National Highway Traffic Safety Administration (NHTSA) has opened an investigation into comma.ai, triggered by multiple deaths and injuries. This company makes aftermarket assisted-driving kits: no need to swap cars, just add a set of adaptive cruise and lane-keeping to a car you already own. The difference between aftermarket and OEM is in the chain of responsibility. OEM assisted driving is validated by the automaker, and accidents go through recall and compensation processes; aftermarket kits are installed and driven by the owner themselves, and software version-to-vehicle matching relies on remote updates from the vendor. Who's responsible when an accident happens is hard to pin down in one sentence. There are domestic brands doing aftermarket assisted driving too. If you're a car owner thinking of installing one, confirm two things first: whether the OEM warranty still holds after installation, and how insurance will treat an accident.

Anthropic says its own model helped it find a possible new gene-editing tool

The Information reports that Anthropic claims its model was involved in discovering a possible gene-editing tool. Gene-editing tools are molecular tools that can precisely alter DNA, and they're the most expensive link in pharma R&D — finding a candidate takes a long time. Right now this news is just a single conclusion; neither the paper nor validation results have been made public. The model's role in this kind of R&D is mostly reading literature, proposing hypotheses, and screening candidates — the ones who actually do the experiments are still the lab. What's more worth watching is reuse: swap in a different target with the same method, and can it save time the same way. Ordinary readers won't be buying a new drug because of this anytime soon. Readers in pharma and bioinformatics can note one time point: when the paper goes up, flip to the experimental section and see which lab it was in and how many rounds of validation were done.

• Geely this time is talking about charging speed, directly targeting BYD. The real bottleneck for fast charging isn't the car, it's the charger: how much power the car can accept and what chargers you can find on the roadside are two different things. Domestic fast-charging stations are concentrated at highway service areas and tier-1 and tier-2 cities; old neighborhoods and tier-3 and tier-4 areas are still mostly slow charging. If you're thinking of buying a fast-charging car, first open a map and check whether the route you drive often has chargers at the matching power level — don't just look at the numbers from the launch event. If you've already bought a fast-charging car, keep an eye on the automaker's charging station buildout plans; how fast chargers get deployed directly affects resale value.
• Chinese smart metro train makes global debut in Berlin — The Berlin International Rail Technology Expo is underway, with over 3,000 companies from 59 countries and regions exhibiting, and among 180 global debut results, over 200 Chinese companies are participating. This metro train, described as able to sense and judge, uses sensors and AI models to replace human eyes watching the track and judging equipment status — a global debut. Procurement cycles for rail transit equipment are long; a single debut doesn't mean orders right away, but Chinese rail transit companies going overseas are shifting from selling train cars to selling AI-equipped systems, and both unit price and margin are different. Readers watching high-end manufacturing going overseas can keep an eye on which city's tender list it shows up on next.
• What "Miyun No. 1" wants to solve is the concern over sharing data between enterprises: banks' and hospitals' data can't just be handed out, yet joint analysis requires computing it together. Its approach is to let data participate in computation in encrypted form, handing over only the result, not the raw data. Several players domestically are pushing this, and the real difficulty is whether each party's encryption methods can interoperate — if they can't, everyone's still computing on their own. Companies with cross-institution data cooperation can compare their own integration cost against existing solutions, and the key question to ask is whether it can connect to the other party's system.
• Global payments company Stripe is bringing its flagship event Stripe Tour to China for the first time, and simultaneously announcing the full launch of Managed Payments — put simply, it collects and pays on sellers' behalf, handling tax rebates and compliance. For people doing cross-border e-commerce and overseas subscription services, the biggest headache is tax and getting paid across countries, and that's exactly what Stripe is aiming at. It doesn't directly matter much to ordinary readers, but if you're building AI tools for overseas clients and charging monthly, every step saved in the payment channel is a bit more profit. How far it actually lands on the Chinese seller side still depends on the scope of subsequent rollout.
• Anhui has concentrated panel, memory, and a batch of chip manufacturing projects, and this kind of personnel arrangement usually precedes capital landing. Projects that local leadership personally champions find it easier to get favorable terms on land, power, and supporting subsidies. For companies in the supply chain, this means Anhui's project pace may accelerate over the next year or two, and orders for equipment, materials, and facility engineering will follow. Readers in supply chain can watch the construction start lists announced later, focusing on which projects are continuations and which are new starts. For listed suppliers, the backlog in announcements is more solid than verbal plans.
• miHoYo completes equity custody registration, company responds — The Shanghai Equity Custody Registration Center publicly announced that miHoYo completed equity custody registration on September 23 — put simply, this means handing the shareholder register to an official institution for safekeeping, usually a preliminary step for corporatization reform and IPO preparation, so the market immediately jumped to IPO speculation. miHoYo responded the same day: there are currently no plans to list or raise funds. Game companies have plenty of cash and don't lack money; what they lack is a venue for pricing tradable shares. Where this step leads and when it lists remains to be seen. Investors watching the gaming sector can treat this as a cooling signal for shell-resource rumors.
• OpenAI's customer case — Customer service is the most straightforward scenario for AI deployment: high question repetition, standardized answers, and call recordings you can check line by line. Voice customer service company Ringg uses OpenAI's models to handle up to 65% of incoming calls, with the rest transferred to humans. That ratio is high for the industry — previously public cases are usually around forty percent. For small and medium businesses, the annual cost of hiring one customer service rep is enough to buy a lot of these tools, and the barrier has dropped to the point where it's worth a try. What to watch is the human handoff step: if a customer asks for a human and can't get one, reputation drops faster than the money saved.
• A seller's daily work is writing titles, making images, figuring out ad copy, and watching inventory — exactly the parts AI tools can most easily replace. Amazon is giving platform sellers free use of AI software Quick for 12 months. Free for a year is the standard play for acquiring users; the real cost is in the renewal price. Readers selling on Amazon can do the math: is the time the tool saves you, converted into money, enough to pay next year's subscription. Another use is to try out your workflow with it, and once it runs smoothly, consider whether to switch to a pricier version.
• YouTube added AI to creator tools at its Made on YouTube event. One assistant goes through videos a channel has posted before, finds the ones that are getting hot again or that happen to line up with the news, and pushes them again. For people running channels, this is like handing the job of picking old videos to a tool for batch screening. The most practical step when using it is to manually double-check the timing of the videos it picks, so you don't collide with a commercial partnership slot. Whether Chinese creators can use it depends on which regions the feature rolls out to later — you can check the creator dashboard first to see if there's an entry point.
• Messaging company Bird borrowed $450 million in new debt and cut staff from a peak of 1,000 down to 120 — nearly ninety percent gone. Doing a bond issue and layoffs at the same time usually corresponds to tight cash flow: the money is used to plug holes, not to expand. This company's clients are enterprises that have done marketing texts and customer service systems, a middle layer that's easy to consolidate. Readers in enterprise services can see a pattern: revenue hasn't collapsed but headcount is cut to a fraction, which usually means the old growth model has stopped. For those following this thread, focus on its debt covenants and repayment schedule.
• Palantir CEO Alex Karp believes OpenAI won't go public, on the grounds that he thinks the company carries too much liability. Going public means spreading risk onto the public market — money-losing lawsuits and model failures all have to be written into financial reports and questioned once every quarter. This judgment has his own stance behind it: Palantir is a public company, and defense and government clients prop up its valuation. Readers watching the AI primary market can flip the question around: which company's risks can be written clearly, and which can't. If it can't go public, how employees' options get cashed out is also worth a quick calculation.
• Model companies renting data centers is like handing off electricity, cooling, and operations as a package, buying only the cards and writing the code themselves. The Information reports that data center service provider Crusoe signed a lease contract with Thinking Machines Labs. The longer these contracts are signed, the lower the rent — it locks in the price for both sides: the buyer locks in costs for the next few years, and the seller uses the contract to raise financing. Readers watching AI infrastructure can check against a list: whoever signs longer contracts with bigger single amounts has more stable cash flow expectations. This kind of news is also a reference for domestic data center operators — long-term contracts are now hard material for getting bank credit.
• The Washington Post reports that a professor produced 200 papers and 14 books in less than nine months this year, and publicly said he used AI assistance. The interesting part here is the research evaluation system: paper count has long been a hard metric for assessment, and once generation speed goes up, count no longer indicates quality. Readers in academia or preparing for grad school can watch their own institution's new rules on AI authorship and submissions — some journals already require declaring which tools were used. What will really change is the review method, shifting from counting papers to seeing whether results can be reproduced, and this shift is already happening in some disciplines.
• Qualcomm announced on the 23rd, US time, that it reached an agreement to acquire robotics software company PickNik Robotics. PickNik's strength is robot motion planning software, and Qualcomm sells the chips that go inside a robot's head. A chip company buying a software team downstream is to give buyers a complete solution that runs out of the box — otherwise selling chips alone can't compete with Nvidia. The implication for practitioners is blunt: consolidation in the robotics industry has burned from the hardware layer into the software layer. The deal amount wasn't disclosed, and closing depends on regulators.
• Data labeling platform offers $60 an hour, but the money isn't easy to get — This article describes a US university teacher's summer job hunt: some tasks on data labeling platforms offer $60 an hour, which looks like more than teaching pays. It's not that easy once you start — the high-priced ones are all specialized questions, and content in medicine, law, and biology can only be done right by people who know the field; fail the assessment and you don't get paid. Data labeling is the work of feeding standard answers to models, and the platform is paying for your professional knowledge, not your typing speed. Readers wanting to earn extra money this way should first ask three things: what's the settlement basis, what happens if you fail the assessment, and how long until the money arrives.
• Tsinghua University and Infinigence AI open-source embodied intelligence platform RLark — Tsinghua and Infinigence AI have open-sourced a cloud-native platform for embodied intelligence (AI that teaches robots to understand the physical world). Robot training requires using multiple different models of robots and scattered compute clusters at the same time; previously each party wrote its own scheduling, and RLark pulls these into one control tower: onboard a new robot in 5 minutes, launch cross-cluster tasks in 10 seconds. Open source means small and medium robotics teams don't have to reinvent the wheel.
• Supermicro starts shipping Nvidia Vera Rubin NVL72 racks — This generation of racks turns chips, power, and cooling into a complete turnkey system; buyers order by the rack and no longer have to assemble servers themselves. The upside is fast delivery and short installation time; the downside is narrower flexibility for returns and swaps — if one link in the whole solution needs a model change, you wait for the whole rack's lead time. Readers doing data center procurement can watch the delivery pace, and can also compare cost per watt between turnkey and self-assembled solutions. The chip supply situation behind these racks will directly affect next year's data center rack-up plans.
• Airbus uses a Steam Deck handheld to control an ExoMars rover prototype for ground testing — Before a rover actually goes to Mars, it has to be repeatedly tested with a ground prototype, including how it drives, how it digs, and how it sends data back. Airbus uses a consumer-grade handheld as the control end, aiming for portability and cheapness — one person standing on the test field can operate it. Readers doing hardware testing can see a line of thinking: during the validation phase you don't need to pile on specialized equipment; using off-the-shelf mass-produced hardware is often faster, and if it breaks you just swap in another one. The handheld's interfaces and stability have limits, and the version that actually goes to space still has to go back to industrial-grade equipment.
• A TMTPost article put Marvell on the beneficiary list for this round of data center expansion — A TMTPost article put Marvell on the beneficiary list for this round of data center expansion. It makes custom chips and networking chips: whatever features a big client wants, it builds the chip, which is a different path from general-purpose chips. Jensen Huang previously said it would be the next chip company to cross a trillion-dollar market cap, which later proved premature. Readers following this thread should focus on whether its orders with hyperscale clients can be renewed, and on the gross margin of its custom business. Domestic companies doing similar custom chips can also use it to compare quotes and delivery cycles.
• At a Nikon imaging competition, an AI-generated video won an award, and microscope users are demanding an explanation from the organizers online — At a Nikon imaging competition, an AI-generated video won an award, and microscope users are demanding an explanation from the organizers online. Competition rules usually state that entries must be shot by the entrant themselves, but whether AI-generated content counts isn't spelled out in the rules. This kind of controversy has appeared more than once this year, and photography, illustration, and short film competitions are all patching their rules. Entrants can go through the competitions they're entering and see whether there's a requirement to declare the source of materials. For organizers, adding a declaration requirement is less hassle than stripping an award after the fact.
• Amazon starts rehiring previously laid-off employees for AI-related roles — Amazon starts rehiring previously laid-off employees for AI-related roles. Layoffs and rehiring so close together shows what's missing is people who can do this kind of project — the money saved isn't the point. This approach at big companies has a side effect: whether laid-off people come back depends first on pay and second on trust. Job seekers can compare: what's the skill gap between the laid-off roles and the rehired roles — the gap is usually whether you can actually put AI tools to work in the business. Readers preparing to switch roles should fill that gap first; it's faster than getting another degree.
• Tech.eu reports DiffuseDrive is filling the data gap in the physical world — Teaching machines to work in real environments, the hardest part is the rare and dangerous scenarios: low-probability collisions, extreme weather, temporary construction. Sending real cars to capture these is costly and risky, so when you can't capture them you fill in with synthetic data — that's what DiffuseDrive does. Synthetic data's problems are also very real: a realistic-looking image doesn't mean the machine can learn from it, and if lighting and physics don't match up, the model still errs on a real car. Teams doing autonomous driving and robotics can first figure out what proportion of synthetic versus real-shot data they use, then look at results.
• An award-winning French author is accused of using AI to write, sparking controversy locally — An award-winning French author is accused of using AI to write, sparking controversy locally. The focus of the debate isn't the tool itself, but whether what the award wants is something "written by a human." Publishing and award rules usually lag behind the tools, and this controversy will most likely push organizers to add a declaration requirement. Writers can check the rules of the contests or awards they submit to, and whether there's a requirement to disclose the scope of AI use. For people who live off manuscript fees, there's another practical issue: how publishers decide what price to pay for a manuscript that involved AI.
• Microsoft announces an additional $2 billion investment in the Gulf region — Microsoft announces an additional $2 billion investment in the Gulf region. This money usually goes into local data centers, cloud services, and localized products; the Middle East has been grabbing AI infrastructure orders these past few years, relying on cheap power and land. For cloud vendors, the Middle East is a node that can serve both Europe and Asia, with good latency and compliance arrangements. Readers watching cloud vendors going overseas can compare the power and land terms each has signed locally — these two determine a data center's long-term costs. For people looking for overseas jobs, this kind of regional headquarters is where hiring has been heavier these past two years.
• ChinaAMC launches 3 new ETFs on HKEX, targeting niche strategies — ChinaAMC launched 3 niche-strategy ETFs pointing at Hong Kong stocks all at once, which may later be included in the Stock Connect so mainland investors can buy directly. Readers who do ETFs know well that the fee war on broad-market products has been fought to the bone, and all the growth is in thematic and strategy products — the license position in Hong Kong stocks is still being fought over.

---

Today's Market Quick Look (in-house data, global 24 hours (US stocks 9/23 (Wed) close + A-shares 9/24 (Thu) close))

Among the six, Palantir gained the most, while Microsoft and Tesla closed slightly green. Nvidia, Alphabet, and Amazon closed red, with Alphabet down 3.8%, the worst of the bunch. The day's two items were Google releasing the Gemini 3.8 voice model and Microsoft announcing another $2 billion investment in the Gulf region.

All six were red. Optical modules, domestic chips, and office software pulled back together, with Foxconn Industrial Internet down 3.14%. The news flow that came out this afternoon wasn't bad: a Chinese smart metro train made its global debut in Berlin, and Stripe brought its flagship event to Shanghai for the first time, but A-shares still gave back expectations on AI pricing first. What A-shares want to earn is the money from orders actually landing, and the pace runs half a beat behind US stocks.

---