---

No single giant grabbed the headlines today — where the money and the people are flowing is more worth watching. DeepSeek pushed its new model to 2 trillion parameters, and Liang Wenfeng says the next stop is 8 trillion; Musk's AI employee has been live for a full month, with users topping 400,000, and OpenAI is reportedly building the same thing. Enterprise software has started a price war, with customers' budgets being eaten directly by model companies. Chips, data centers, and power — three sets of books, all laid out over the past few days.

---

📈 Six, Macro and Market Data

DeepSeek is training a 2-trillion-parameter new model, and Liang Wenfeng says the next step is 8 trillion

The Information reports that DeepSeek is currently training a new model of roughly 2 trillion parameters. At a recent investor meeting, Liang Wenfeng said the company's next step is to push parameters to 8 trillion. A year ago, that number only appeared in discussions. Parameters (the number of adjustable knobs inside a model) getting bigger directly maps to two things: training needs more GPUs and more power; and answering a single question costs more. DeepSeek previously built its developer reputation on low prices and open source — going ultra-large means reordering the priorities between cheap and strongest. Ordinary users won't feel the change short-term; phones still run the current version. Companies building apps should watch two things: whether the next-gen model stays open source, and how the price sheet gets adjusted. These two decide next year's API costs.

Alibaba says an omni-modal model will arrive within three years, and Lu Chuan has already used it to shoot a Ming-dynasty disaster scene

At the Yunqi Conference, Alibaba's judgment was: within three years, a native omni-modal unified generation model will emerge, handling text, images, video, and sound in the same model. In the same showcase, director Lu Chuan and his team used AI to recreate a Ming-dynasty disaster scene from 400 years ago — five days start to finish, no sets built, no waiting on actors' schedules. What's saved is the front-end work: scouting locations, building sets, booking people, doing makeup tests. That part used to be billed by the month; now it's compressed to billing by the day. Small content teams can start with short films and ads. Those have controllable budgets and short pipelines — run one through and you'll know if it's worth it.

Musk's AI employee has been live for a month, users top 400,000

Bloomberg reports that SpaceXAI's Grok Bot surpassed 400,000 users in its first month. What sets it apart from an ordinary chatbot is that it can act on its own: book trips, send emails, fill out forms — users give it account permissions and it runs the process to completion. 400,000 isn't big for a consumer product, but among "does work for you" tools it's already leading. What really decides retention is another thing: the tasks it finishes for you — do you dare not double-check them? One mistake and users will pull the permissions back. If you want to try it, start with small things: subscription renewal reminders, price comparisons, tidying up email. Don't hand over payment permissions right away — use it for two weeks, see how it behaves when it errs, then decide how much access to give.

Fei-Fei Li calls for an independent oversight body for AI

Bloomberg reports that World Labs founder Fei-Fei Li publicly called for independent oversight of AI. Her position is blunt: the body regulating this can't both develop and referee — evaluation should go to outside parties. Statements like this have been coming thick and fast lately. Government, academia, and companies are all jockeying to set the rules; the difference is who gets final say. For teams doing overseas markets, these words will turn into concrete actions: whether model evaluations go to third parties, how long logs are kept. Standardize your evaluation process and record-keeping now — it's cheaper than patching it up later.

• Huxiu laid out the hottest job categories in AI right now — the common thread is assisting AI — The work in these roles is very concrete: checking model outputs, turning scattered data into usable tables, watching the pipeline so it doesn't break in the middle. Pay isn't low, but the experience on your résumé skews toward operations. If you want in, run two separate calculations: how fast the salary is rising now, and whether this skill set will still be recognized by another company three years from now. Another thread is hiring location: these roles cluster at big tech and outsourcing firms, and the pay gap between the two is not small. Another thread is hiring location: these roles cluster at big tech and outsourcing firms, and the pay gap between the two is not small.
• OpenAI's version of the "AI coworker that never clocks out" code has already leaked. What these products compete on isn't model strength but the completeness of task execution: can it log into websites, fill out forms, and ask you before paying. Domestic office-software companies are watching this closely because once users get used to letting AI handle workflows, the original software interface recedes into the backend. Those who lag will get squeezed down to the tool layer. Once a tool product can run an entire workflow for someone, users' willingness to switch back to manual operation is very low.
• Enterprise software contracts are usually signed annually, and the scene most likely to play out at renewal this year is the customer saying: the model company already covers this feature, drop your price first. For buyers, this is a rare bargaining window in years — put the two quotes side by side and you can often get a discount, or a few extra seats thrown in. This kind of price cut usually doesn't make it into a press release; it only shows up in the renewal contract. For software companies, cutting price preserves the renewal rate, at the cost of a chunk of short-term margin.
• If a school doesn't assign homework, it's effectively outsourcing the screening to partner companies: whoever gets picked gets hired. This model can be referenced domestically too — big tech running its own bootcamps and hiring directly, skipping mass applications and written tests. For job seekers, what's worth noting is that people who get in through these channels mostly sign project contracts, not formal headcount. The curriculum is mainly product practice; academic training isn't the focus.
• Base44 launched an AI that can answer calls — The most error-prone part of answering calls for you is commitments: the other side asks if you can go cheaper, the AI casually agrees, and you're the one who has to clean it up. Before using it, write the rules clearly — for example, only record, never commit, and route anything involving money to your phone. Set the boundaries first, and it saves hassle without adding chaos. Tools like this have been appearing more frequently this year; answering calls is just one of them.
• French speech recognition company Gladia was acquired by domestic cloud provider OVH and will continue to operate independently — European customers' concerns are very practical: data like medical recordings and government calls isn't allowed to leave the country freely under regulation. If speech recognition can run in local data centers, the compliance hurdle is much easier to clear. If your team builds voice products and is looking at the European market, keep an eye on the APIs and pricing of local clouds like OVH — comparing prices early never hurts. Medical and customer-service scenarios demand higher recognition accuracy than everyday transcription.
• NATO's innovation fund doesn't invest much money; what's valuable is the endorsement. Companies that get it are usually pulled into member states' small-scale trial lists, and the orders that follow are far bigger than the funding amount. Teams in defense and unmanned systems can see the overseas threshold: beyond technology, you're also blocked by supply chain vetting and local partners. European defense procurement has long approval cycles — getting pilot status often takes one to two years. The fund's investment rounds are usually tied to member states' pilot projects: the money arrives slowly, the orders come early.
• To see how hot AI infrastructure is, a prospectus is more reliable than a launch event: it has data center contract volumes, customer concentration, and power and land costs for the coming years. A $720 million raise isn't large, but the fact that these companies are lining up to go public itself shows capital is willing to pay for data center assets. These companies have high customer concentration — the share of the top few customers is the page in the prospectus most worth flipping to.
• The purpose of futures is to let buyers and sellers lock in prices in advance. If GPUs had a standard contract, companies renting machines and parties renting out data centers could fix prices ahead of time, making costs easier to calculate. This proposal is currently stalled at the US Commodity Futures Trading Commission; the sticking point is that GPUs aren't a standardized commodity: models change every year, and delivery standards in a contract are hard to write clearly. As long as GPU leasing has public quotes, buyers can lock in their cost budgets a year ahead.
• Former UK Chancellor Osborne fires back — Protests against building data centers have popped up in many parts of the UK, with reasons centering on power and water: data centers consume a lot of electricity, cooling needs water, and nearby residents worry about electricity prices and their lives being affected. Osborne believes these objections are slowing investment. This kind of conflict will come up domestically too — in site negotiations, power and water conditions are often harder to negotiate than land price. For local governments that want projects to land, the tax revenue and jobs a data center brings are the only chips at the table.
• Some research proposes letting hardware make judgments directly — This route saves on data movement: cameras and robotic arms make judgments on the spot, with lower latency and no network dependency. The difficulty is flexibility — once the hardware is built, changing the algorithm means going through production all over again. Teams building robots and in-vehicle devices can watch this direction, but first calculate the cost of one iteration.
• The Information digs deeper — The article attributes the cause to the deployment stage: model capabilities have risen, but internal processes, data, and permissions at enterprises haven't kept up, so projects stall at the pilot stage. This judgment holds domestically too — money for buying systems is easy to approve, money for changing processes is hard. For teams doing enterprise services, first check whether the client has a dedicated person responsible for process change, then talk delivery timelines.
• Investor Gulati says — His reasoning is about spending structure: the vast majority of money goes to models and GPUs, and security teams' budgets rank behind. Patching after an incident costs far more. For buyers, you can specify vulnerability response times in the contract; for product teams, putting security testing and feature development on the same schedule is more practical.
• Dongfeng Nissan's new N7 all-electric sedan launches, limited-time price from 109,900 yuan, top trim can option lidar — A car in the low 100k range offering lidar as an option shows this perception hardware has entered the mass-market price band. Buyers need to distinguish two things: having this sensor on the car doesn't mean the assisted-driving features are fully included — software features are often tied to higher trims. Before signing, go through the config sheet item by item; it's more useful than listening to the salesperson. At this price point, the sales chart says more about whether it's worth it than the spec sheet.
• Space company AstroForge hands control of its next spacecraft to AI — Handing over control to AI presupposes that communication latency is so large that human intervention is meaningless. What really needs to be written clearly is what to do in abnormal situations: when facing a fault it can't diagnose, does it hold position or return on its own. Projects like this won't produce civilian value short-term, but they push the boundary of autonomous decision-making one step forward. Deep-space missions have limited communication windows — whether the spacecraft judges correctly directly determines whether the mission can be saved.
• The event ticketing business feeds on information asymmetry and inventory turnover. Venues' idle time slots used to be handled by salespeople calling one by one; the platform wants to turn it into a pool that matches automatically. $52 million is enough to roll out in a few cities; next, watch whether it can really pull the venue side on board. Ticketing margins come from service and markup capability — that only becomes a conversation once scale is up.
• A Wired author made AI clones of several colleagues, feeding their writing habits to a model — Making AI clones of your own colleagues sounds like a joke, but it actually exposes a real office problem: a lot of communication is just passing information along, and doesn't require a person to actually show up. The failure point in this article is that the clone spoke on his behalf while the person himself didn't know. If you want to save effort, first draw the authorization boundaries clearly. The trouble clones bring is mostly not a technical problem, but a question of who speaks for whom.
• Data center assets going public one after another shows this round of AI buildout has reached the stage of needing to raise money from the stock market. For secondary-market investors, the point isn't how much the stock rises but the data center's rack utilization rate and customer list: companies that can clearly account for electricity and depreciation are the ones that can sustain their valuations. Data center valuations look at contract duration and renewal prices, not rack count.
• An article about the ride experience in driverless taxis lists all practical details: confirm the pickup point before boarding — driverless cars often stop in places where stopping isn't allowed; if you need a detour or cancellation mid-trip, go through the app process, don't open the door yourself; if there's a scrape, take photos first, then contact customer service. How liability is divided when something really happens varies by platform's terms — worth flipping through before your first ride. Different regions open different road segments; checking the route coverage before you set out saves a lot of trouble.
• On letting AI give financial advice, the article points to a responsibility gap: the advice it gives doesn't count as investment advisory business, it has no license, it isn't under corresponding regulation, and if it's wrong there's no path to compensation. Using it as a reference is fine; anyone treating its conclusions as decisions is taking on all the consequences themselves. For anything involving money, still run it past someone licensed.
• The free version can edit and export; AI-related features are paid separately. The advantage of editing on your phone is that footage doesn't have to be shuffled between devices — shoot and edit right away. Watch out for two limits: export resolution and watermark. Before starting, run one clip through to confirm the free version is enough — don't get halfway through editing and find you can't export.
• Microsoft's latest restructuring consolidates its gaming business organization, placing multiple studios under Activision management, with future Halo titles handed to Activision for development. For players, the short-term question is whether new titles get delayed; for industry people, once studios are reassigned, budgets and personnel usually follow.
• The recycling startup signed by Meta breaks mixed plastics directly down into chemical feedstock, without needing to sort by type first. Packaging waste from industrial parks and data centers is one of its feedstock sources. The signing party usually participates through long-term procurement or investment; the results depend on the actual capacity of the first plant.
• Armenia wants to take orders for US AI infrastructure. The report mentions it's attracting data center projects with power and land costs, and its location makes it convenient to connect routes to Europe and the Middle East. The opportunity for countries like this is picking up the edge-case demand for training and data storage; the difficulty is talent and network bandwidth. If you're looking at data centers going overseas, put it on your site-selection shortlist.

---

Today's Market Quick Look (in-house market data, global 24 hours (US stocks 9/22 (Tuesday) close + A-shares 9/22 (Tuesday) close))

Among the Magnificent Seven, only Nvidia and Tesla closed green; Meta, Microsoft, and Amazon all closed slightly red. What rose steadily was Palantir — the defense and government orders story keeps adding points for it. That night Tesla and Nvidia closed green in sync; the hardware and vehicle lines keep getting money.

The divergence is the same as ever: the three doing optical modules and server assembly closed red, while domestic chips and office software closed green. The Yunqi Conference talked omni-modal and AI assistants all day, but what A-shares are willing to pay for is chips and software, not racks.

---