---

Two threads today, one on the ledger, one at the gateway. Bloomberg reports that legal AI company Harvey's gross margin fell from 50% to -50% in half a year, and went back to positive after switching to Chinese open-source models. Meta's Muse has been live for 12 days, and downloads have surpassed ChatGPT. The Hangzhou Yunqi Conference ran from September 22 to 24, and Muse was the most-mentioned word during those days. Domestic moves are more concrete: Huawei Kunpeng scaled the machines used for AI assistants to 4096 nodes. OPPO says AI phones are judged by whether they remember you. A-shares are closed today for Mid-Autumn Festival, so the board still shows the September 24 closing prices.

---

📈 Six, Macro and Market Data

OPPO says AI phones are judged by whether they remember you

GeekPark interviewed Jiang Yuchen, OPPO ColorOS Smart Product R&D Director. She gave a definition of the AI phone: the user says one sentence, and the phone gets the job done, without opening apps one by one. She measured products on the market with this ruler, including OPPO's own. The conclusion is that no phone currently qualifies. Her judgment is that there's no gap left at the model layer, and the gap between phones has only just started to open. OPPO is betting its effort on on-device models (models that run locally on the phone and aren't uploaded to the cloud) and memory. The "one-tap quick note" in the system is a cross-app collection box; for meal pickup codes alone, cumulative calls have exceeded 1 billion. This year at ODC they also launched a new chest-worn hardware called the "Heart Power Ball." When switching phones, you can first check whether it remembers you. The pre-purchase test is simple: ask the phone to find last year's air conditioner installation order and see whether it recognizes it. If it can do that, the AI features have actually landed in daily life.

The new IP business in the AI era: one Backrooms movie, $10 million cost for $400 million box office

This article reposted by TMTPost sorts through two kinds of content: America's "Backrooms" and East Asia's "dreamcore." The Backrooms started with a corridor photo posted on a forum in 2019, and over several years grew into films, games, offline immersive experiences, and merch. This year's Backrooms movie had a production cost of about $10 million and global box office of over $400 million. Its setting was filled in by netizens together, so the filmmakers didn't have to invent the story from scratch. Japan's small game "Exit 8" sold 3 million copies, following the same path. Domestically, the two terms "dreamcore" and "Chinese dreamcore" have a combined 20 billion plays on Douyin, and right now they're mostly buy-to-play games and user short films made by small teams. People making content can see the order of this path clearly: first gather people through emotional resonance, then turn the setting into a framework that can be reused. What Chinese dreamcore lacks is the second step, namely distribution and cross-media development. Teams sitting on nostalgic subject matter should first write a solid setting document; that's more practical.

The more AI is used, the more money is lost: this legal AI company's gross margin fell to -50%, and turned positive after switching to Chinese models

Bloomberg reports that legal AI company Harvey is valued at $15.6 billion. In March this year it updated its AI assistant, customer usage surged, but gross margin fell from about 50% at the start of the year to -50% in June. The loss is on the bill. What these companies do is call large models (programs that write text like GPT) to process legal documents; the more usage, the more they pay upstream. In August, Harvey released its own model, built on China's Kimi K3, at much lower cost. People familiar with the matter say its gross margin has already returned to positive. Teams building AI applications can copy this one move: replace the most expensive call. The method is to first find which step costs the most, then put a cheaper open-source model on top, and tune it with your own data to fit the business better. Sequoia and General Catalyst are both betting on this path. Once costs come down, there's room to negotiate user renewal prices.

Musk wants to double the Nvidia chips in Colossus 2 before year-end

Bloomberg reports that Musk plans to double the number of Nvidia chips in Colossus 2 before the end of the year. This machine is what he uses to train AI. Doubling isn't just placing orders. Packing more chips into one machine means power, cooling, and data centers all have to expand along with it, and projects like this most easily get stuck on power. For Nvidia, this is an order that can be counted before year-end. People watching this thread can note one method: chip counts and the progress of power expansion say more about who's still running forward than the model scores at a launch event. Companies doing data centers, power, and cooling support usually see orders appear along this chain.

• A piece from Hupu's tech desk recorded a scene: at the Hangzhou Yunqi Conference from September 22 to 24, Meta's Muse was the most-mentioned word. Its downloads in 12 days surpassed ChatGPT. Discussion among domestic vendors centered on whether ByteDance or Alibaba would be first to catch up with this form factor. Fast-fading hype is normal for this kind of product; the first week's numbers look good, but the second week is the real test. Readers building products can watch one metric: whether users still open it on the second day. There's no domestic counterpart to this kind of product for now, so people who want to try it can only wait for ByteDance and Alibaba to make their move.
• A TMTPost piece went to Shanghai's Yu Garden: tourists wearing Qwen AI glasses hear the story of this teahouse when they reach the Mid-Lake Pavilion, and say one sentence at the door of Nanxiang Steamed Bun Restaurant to pick up the xiaolongbao they just ordered. On the outbound travel side, iFlytek's glasses became popular thanks to real-time translation. Industry numbers are rising fast: in the first half of 2026, mainland China AI glasses shipments grew 306% year over year. The bottlenecks are also out in the open: weight, battery life, and display are hard to balance, and mainstream screen-equipped models cost 2,000 yuan to over 10,000 yuan. For just a few trips a year, most people can't bring themselves to buy one.
• ITHome news on September 25: there's a rumor that Transsion Holdings plans to start Hong Kong listing promotion next week, and the company's relevant personnel responded that official announcements should prevail. This phone company already filed with the Hong Kong Stock Exchange in June this year, with CITIC Securities as sole sponsor, and first-half revenue of 35.431 billion yuan. Its home turf is Africa, where it mostly sells low-priced phones, and in recent years it has been leaning toward AI features. People playing the Hong Kong IPO subscription game can first check revenue and gross margin in the prospectus; prices during the rumor stage don't count. If it follows the rumored pace, the prospectus will be made public around the promotion period, and inside you can see its share of investment in AI features.
• A Hupu repost of an APPSO piece puts this side by side with earlier news of AI causing trouble. The last time a large batch of AI assistants acted together, they were attacking various systems. This time, Anthropic put 950 AI assistants into a massive DNA database and found a new class of enzymes. The number in the article's title is even bigger: 37,000 AI employees teaming up to do pharma companies' work. Readers in biomedicine can note one change: in this field, AI has gone from helping screen molecules to finding answers on its own at scale. If you really want to use it, the experimental validation step can't be skipped. The barrier in this field is validation; finding a candidate doesn't equal making a drug.
• ITHome news on September 25, citing tech media Wccftech's September 24 report. New York University Stern School of Business professor Joost van Dreunen says: when Microsoft bought Activision back then, the result reversed, and XBOX's own studios were merged into Activision's system. He has studied the gaming industry for years, and his judgment for players is that future XBOX flagship games may be scheduled first according to Activision's rhythm. Readers in this field can watch two things: how studio reporting lines are drawn, and who decides subscription service pricing. The change players will feel first is in scheduling; game launch times and update rhythms are both clues.
• TechCrunch reports that Lightspeed has narrowed its India strategy to AI, setting a $250 million target for a new fund, investing in companies that are just starting out and still have little revenue. India's advantages are many engineers and low labor costs, and teams building AI products and outsourcing are all trying there. Domestic entrepreneurs can note one thing: once a direction has proven out a low-price route in India, your quote will be compared against it. People looking for opportunities on the going-global thread can watch the list of companies this fund invests in later. Early-stage fund money is usually deployed over several years, and who they invested in will only be disclosed gradually after a while.
• tech.eu reports that Nordic energy company Fortum, through its innovation and investment arm, became a backer of Norrsken Evolve. This fund completed fundraising this summer, at 62 million euros, and has already invested in more than 80 companies, 75% of which later raised another round. The theme around the money is electrification: the only energy Europe can produce at scale locally is electricity. Readers watching new energy and power support can keep an eye on this thread; the power demand of AI data centers is pushing this kind of investment even faster. The fund isn't large, but energy companies stepping in as backers has been increasing for years.
• A product called Reteach appeared on Hacker News today, positioned as a tutor for university courses. The selling point is that it draws and writes on the screen while explaining, rather than just replying with a block of text. It's still in private testing, and you have to queue for a spot. Readers with college students at home, or who are studying for certifications themselves, can note this kind of tool; their difference from question-bank-style products is that they explain the process. There's a simple test for picking an AI tutor: take a problem you already know how to solve and ask it, then see whether it explains it correctly. Spots are limited during private testing, so people who want to try it have to queue and wait.
• Bloomberg reports that Hong Kong-based cross-border shopping platform Buy&Ship is pushing Japan expansion, with Tokyo as its preferred listing location. The company was founded 11 years ago and does consolidation shipping: it helps people order on overseas websites, combines goods into one package, and ships them back to Hong Kong, earning on shipping fees and warehousing. Its business has already broken even, and this funding round brought in an investment institution under Mitsubishi Warehouse and a Japanese government fund. The company says the next step is to apply AI to cross-border shopping, such as package consolidation and customs clearance. Readers who often buy overseas goods online can note: after this kind of platform lists, shipping fee rules usually get adjusted all over again. The consolidation shipping business lives on volume, and the Japan route's volume has grown fast in recent years.
• AITO pushes National Day OTA, adding time-of-use electricity price display for charging stations — AITO pushes National Day OTA, adding time-of-use electricity price display for charging stations. ITHome news on September 25: AITO Auto announced the upgrade contents for the AITO National Day version, and various models already began upgrading successively on the evening of September 24. The screen directly shows which time period is cheaper for charging, so owners no longer have to check app by app. Readers who drive EVs can watch for the car system's upgrade prompt; before long holiday drives, update the system to the latest version, and what you save is charging money on the road.
• Singapore Deputy Prime Minister Gan Kim Yong says the next stage of China-Singapore cooperation is about AI and the green economy — Singapore Deputy Prime Minister Gan Kim Yong says the next stage of China-Singapore cooperation is about AI and the green economy. He said this at Singapore's FutureChina Global Forum on Friday this week. Decades of cooperation before this were underpinned by projects like the Suzhou Industrial Park and the Tianjin Eco-City. For people in foreign trade and going global, there's opportunity here: when governments push AI and green energy cooperation, a batch of pilot projects and supporting funds usually follows. Companies that want to take on this kind of work can first watch the official lists of parks and enterprises released by both sides.
• A commentary reposted by TMTPost talks about AI music — A commentary reposted by TMTPost talks about AI music: platforms can already produce finished tracks with fine arrangement and good timbre. The author's view is that people don't need to worry about being replaced; everyone's experience is one of a kind. Readers doing music and voiceover can compare it against the work in their hands: highly standardized steps like arrangement and orchestration are the first to be taken over by tools. To protect your quote, the custom and live route is steadier. People taking orders can shift their focus there a bit earlier.
• Tech media TechRadar published an opinion piece — Tech media TechRadar published an opinion piece: making AI too convenient also has a cost. The piece says that when employees hand over judgment along with the work, the team's own innovation slows down. The reminder is very concrete: don't submit the first version of an answer the tool gives you directly. Readers equipping teams with AI can set a rule themselves: which tasks must clearly explain why they were done this way, and which can just copy the output.
• Bloomberg reports that Goldman Sachs's Moe says on the AI matter he's in the stronger and longer camp — Bloomberg reports that Goldman Sachs's Moe says on the AI matter he's in the stronger and longer camp. What this means is that he sees the market as a long cycle and doesn't think the hype will fade soon. Readers investing in AI-related names can use it to check their positions: long and short approaches differ, and if you're long, you have to withstand volatility along the way.
• A GitHub repo has organized methods for using AI in search and recommendation into a list — A GitHub repo has organized methods for using AI in search and recommendation into a list. It's called LLMSearchRecommender, and the content is mainly surveys and paper indexes, aimed at engineers who need to build this kind of system. Backend readers doing search and recommendation can use it as a starting directory and save some time hunting for materials.

---

Today's Market Quick Look (in-house quotes, global 24 hours (US stocks 9/24 (Thu) close + A-shares 9/24 (Thu) close (closed today for Mid-Autumn Festival)))

Six names had both gains and losses, all modest. Meta rose 4.5%, the only one up more than one point. Alphabet rose 1.34%. Nvidia, Microsoft, and Tesla edged down. The gainers had product news that day; the decliners were giving back the previous day's gains.

All six were green. The three doing optical modules and server assembly fell hardest, all around 3%. Domestic chips and office software fell one to two points. A-shares are still pricing AI rather cautiously, and volume didn't expand. Today is Mid-Autumn Festival, the market is closed, and these quotes use the September 24 closing prices.

---