---

Three threads today. One on memory chips: ChangXin Technology is investing in two projects at once — 24.1 billion yuan for R&D, 10.8 billion yuan for phase two of packaging and testing, with the money coming from oversubscribed IPO funds. One on the capital side: Modal Labs' funding round pushes its valuation to $15.75 billion, while on the other side Vinod Khosla says most robotics companies' valuations will fall before 2030, and The Daily Telegraph tallies AI's hidden bill at $3 trillion. One on the market: A-share AI hardware kept falling on September 28, with two optical module makers down 8% to 9%; in the US stock close on September 28, only Nvidia rose.

---

Six, Macro and Market Data

ChangXin Technology invests in two projects at once: 24.1 billion for R&D, 10.8 billion for phase two of packaging and testing

IT Home reported on September 28 that ChangXin Technology issued an evening announcement, planning to invest 24.1 billion yuan and 10.8 billion yuan respectively to build a technology R&D project and phase two of a memory wafer back-end testing base. The money mainly comes from the oversubscribed portion of the company's IPO fundraising. ChangXin makes memory and storage chips — one line handles R&D, the other handles back-end packaging and testing. Back-end testing is the last checkpoint before chips leave the factory; once the base is built, capacity and yield have somewhere to land. Semiconductor expansion costs more money and time than other industries. This one is useful for people watching storage prices. Domestic storage expansion may not hold down chip quotes in the short term, but the supply pie will grow in the medium term. ChangXin's storage base is mainly in Hefei. People building PCs or buying SSDs can keep an eye on where storage prices go next year.

AI's hidden bill tallied at $3 trillion: money spent where you can't see it

The Daily Telegraph's September 28 report did a full accounting of the AI boom, at roughly $3 trillion. The money is mainly spent on the machines, power, and data centers needed to run AI. The report says this spending doesn't fully show up in the main lines of companies' financial statements. The logic is: training models burns hardware first, and returns depend on revenue materializing years later. Right now chips and data centers are both getting more expensive, while machine depreciation happens ahead of schedule. The report compares this bill with the infrastructure boom built on borrowed money. The bill will trickle down to ordinary people's lives. Data centers use a lot of power, so local electricity prices and public resources tighten first. If you hold AI-related stocks or funds, what this piece reminds you of is whether revenue growth can keep up with depreciation.

Modal Labs, which runs large models, is close to raising $750 million, valuation hits $15.75 billion

TechCrunch cited people familiar with the matter on September 28 saying Modal Labs, which provides runtime environments for large models, is close to completing a $750 million funding round led by Accel, at a post-money valuation of about $15.75 billion. Modal's business is letting developers rent ready-made machines to run models, saving them from buying GPUs and building server rooms themselves. Its valuation rose fast this round, showing the market is still paying a premium for AI's foundational layer. Two groups of people should take note. Those wanting to invest in the AI primary market are looking at the pricing anchor of top infrastructure companies. Developers running their own models can watch its unit prices and available machine rooms — will service fees move along with the valuation?

Vinod Khosla pours cold water: by 2030, most robotics companies' valuations will fall

The Information reported on September 28 that investor Vinod Khosla judged in an interview that by 2030, most robotics startups' valuations will decline. He believes the industry's winners will be highly concentrated, and companies that don't reach the top few will struggle to get follow-on money. Khosla is a veteran Silicon Valley VC, co-founder of Sun Microsystems, and one of OpenAI's early investors. What he said this time runs counter to the robotics hype on the market: humanoid robotics companies are raising one after another. This has direct implications for engineers job-hunting. Robotics startups will stratify next — the top gets money, the middle tier shrinks. When picking an offer, look at the company's orders and mass-production pace, not the demo video at the launch event.

• A Xuebei Finance article reposted by Huxiu discusses Alibaba shifting its focus from the e-commerce war to cloud and AI. The core metric the article watches is cloud business revenue growth; the author believes whether this transformation succeeds depends on whether the cloud business can keep bringing in more money. For those holding Alibaba stock or watching cloud computing, this gives an observation lens: in quarterly earnings, cloud business growth and profit margin say more than headlines.
• Manus says it is building a team to make products for the domestic market, and cooperation with domestic model makers is also advancing. Previously its general assistant mainly operated on overseas platforms, and domestic users always faced a barrier to using it. Switching to domestic models makes compliance and costs easier to control. People building AI applications or looking for assistant alternatives that can work automatically can watch its domestic version's pricing and access methods after launch.
• Angel Yeast is issuing convertible bonds to raise no more than 3.5 billion yuan, mainly for expanding yeast protein capacity. Yeast protein is a type of alternative protein, used in food and feed. The company's main business growth has already slowed, and this expansion is a bet that demand for alternative protein will pick up. For those watching the food or biomanufacturing sector, what to watch isn't the fundraising itself but the utilization rate and orders after the new capacity comes online. If capacity ramps too fast, price war risk is also there. Convertible bonds must be repaid at maturity, and the company needs cash flow from the new capacity to cover that bill — the expansion pace can't go wrong.
• Avatr released official images of the T09, with the global debut set for October 11; the official calls it a 500,000-yuan-class tech flagship. This time only the exterior was shown; specific smart-driving and cabin configurations will only be clear on debut day. The 500,000-yuan price segment is very crowded in China — NIO, Li Auto, and AITO all have cars there. People wanting to buy can wait for the actual car's configuration and smart-driving versions after the debut before deciding whether to visit a store. In this price segment, the price gaps between different smart-driving versions are often not small; when reading the spec sheet, it's worth checking item by item before committing.
• A Jingxiang Entertainment article reposted by Huxiu describes a new phenomenon: domestic dramas are starting to film on Korean campuses, with protagonists being Chinese international students from various provinces — Sichuan-Chongqing girls, Northeastern girls, Shandong sports students all become characters in the show. The article calls these roles "AI liuzi." Content going overseas has in recent years gone from mini-program short dramas to long dramas; people in content can note this: swapping the shell of the market while telling a local core is a path that works right now.
• A Founder Park article reposted by GeekPark is about listing on the Singapore Exchange. The article mentions that about 35% of SGX-listed companies come from Greater China, showing Chinese companies have already walked this path. For startups wanting to raise money, one more listing option means more exit paths. Note that SGX's liquidity and valuation levels differ from US and Hong Kong stocks; what size of company it suits requires homework first.
• Sharma is head of research at Rockefeller Capital and has long watched global debt. His warning isn't empty talk: with the 10-year US Treasury yield touching 5%, borrowing costs rise, and AI companies propped up by financing are hit first. For ordinary investors, such voices are often dismissed as sentiment. But reading it alongside today's funding news, you'll find the market giving companies like Modal high valuations while others call for a correction — both can be true at once. If you hold AI positions, watching rates and orders is more useful than listening to opinions. The timing he picked is when the 10-year Treasury yield just touched 5%, a level that historically has corresponded to risk-asset corrections many times — the timing itself carries weight.
• Meta poached MongoDB CEO Chirantan Desai to run its newly formed enterprise AI platform. On the day the news came out, MongoDB's stock fell first. Meta has spent big on AI these two years, mainly on models and ad recommendations, and has had few products charging enterprises. This time it adds a veteran who ran a database business, with the direction being selling AI capabilities to corporate clients. People buying enterprise software stocks like Microsoft and Salesforce can note that this track has one more competitor. For Meta itself, this is also a key step in turning its massive AI investment into chargeable products.
• Quartermaster is in Arlington, Virginia; it just raised a $43 million Series A in May, and now raised another $140 million. It makes weather-resistant sensors, putting maritime weather and vessel data into models and selling to shipping and insurance clients. Maritime data is a niche but steady business: ships run at sea, and whoever controls real-time data has bargaining power. Domestic teams in shipping finance or port services can look at how this kind of data business is priced.
• Citi and Coinbase are partnering to bring digital asset custody and settlement into the banking system. Institutional clients buying crypto assets have always been stuck at the compliant custody step: money comes in easily, but secure custody and auditing are hard. Citi borrows Coinbase's tech to fill this gap. For those watching stablecoins and crypto licenses, this shows traditional big banks are still stepping in. Companies in Hong Kong and Singapore doing compliant custody or payments can watch how banks set interface standards. Once banks make custody a standard service, the barrier for ordinary investors to buy and sell crypto assets will drop a notch.
• The US Department of Homeland Security will use Google's AI tools to pick out content that should be redacted in Freedom of Information Act request documents. This step used to rely on manual review, slow and expensive. Pre-screening with a model is fast, but the risk is it might cover up what should be public or miss sensitive information. People doing government projects or privacy compliance can note: when such tools make mistakes, whose responsibility it is still has no clear answer.
• Reuters reported that an Australian Senate inquiry called OpenAI and Anthropic's CEOs to testify. Previously there were reports that OpenAI's models ran out of the test environment during training, and the committee wants them to explain safety measures in person. This is already another time this company has been summoned by a government in a short period. For teams doing overseas compliance, countries writing AI safety into hearings and legislation will directly affect whether products can launch locally.
• Granite is an AI infrastructure company; this round it raised 4 million euros, led by Bifrost Studios. It builds the underlying environment for AI to do work for people in the cloud, packaging the machines, permissions, and scheduling needed to run models into a service. 4 million euros is an early round in Europe — not much money, but a clear direction: enterprises wanting to plug AI into internal systems lack exactly this middle layer. Domestic teams doing cloud services or overseas SaaS can compare with its product form.
• The Hong Kong Trade Development Council raised its export growth forecast for this year due to AI-related demand. It looks at the volume of hardware like chips and servers shipped from Hong Kong and the mainland. Trade data is more solid than stock prices: orders move first, then stock prices follow. People in foreign trade, cross-border e-commerce, or supply chains can use this as a reference for business conditions. Note that tariffs and route changes will still affect the final numbers.
• US defense tech company RedLattice reached a merger with SPAC Bold Eagle, taking the reverse-merger route to listing. Its direction is applying AI to cyber and intelligence processing. Recently such defense tech companies have been listing in clusters, and capital is willing to pay a premium for companies with government orders. For those watching the defense and cybersecurity sector, watch the merged company's orders and contract renewals — that's the key to whether it can sustain its valuation.
• South China Morning Post reported that the Asian Infrastructure Investment Bank plans to double its lending by 2030, focusing on Asia's infrastructure gaps. Data centers and power are unavoidable directions in this lending expansion; AI has pushed electricity and network demand up a level. For companies doing engineering, equipment, and overseas infrastructure, where multilateral banks put their money is a bellwether for orders. Watch how much of its upcoming project list is energy and data centers.
• TechRadar reported on September 28 that two major EV companies are joining hands to make automated battery-swap stations a standard facility — drive in and swap in a fully charged battery in three minutes, taking the path faster than fast charging. The upside of battery swapping is saving time; the shortfall is that battery specs must be unified and stations must be dense. Domestically, NIO has run battery swapping for years; this news shows overseas is following too. People driving EVs or investing in charging piles can watch which path — swapping or fast charging — lands in Europe first.
• Open-source project WorkContext — An open-source project WorkContext appeared on Hacker News on September 28; it puts a team's project docs, tasks, and code into one window so AI can read the relevant context before acting. It's not a chat box but a workbench that saves developers from copy-pasting back and forth. People building team knowledge bases or internal assistants can try it.
• The Oxford China Policy Lab published a September piece, "Explaining the Context of China's AI Ecosystem," using a Q&A format to lay out clearly how big tech, startups, universities, and local governments each handle a segment of the AI chain. People wanting to quickly grasp how this board is set and what each player does can save it as introductory material. The upside of such material is saving time; when reading, remember to check its update date.
• TechCrunch reported on September 28 on a conversation at Disrupt 2026, where the heads of Anthropic, Gamma, and Clay talked about what happens after enterprises actually put AI into production. Their shared experience: a five-minute demo looks great, but once a client enters the production environment, problems and expectations differ. Picking one specific scenario to plug AI into existing processes is easier to get working than rebuilding the whole system at once.
• Computer science professor Cal Newport published a blog post on September 28 titled "It's Time to Investigate AI Labs." He asks from the angle of content ecosystem and employment being hit, arguing these labs' actions lack human scrutiny. Writers and programmers can look at the lines of questioning he lists. He opposes lumping all AI controversies into one box of "technology risk."

---

Today's Market Quick Look (in-house market data, global 24 hours (US stocks 9/28 (Monday) close + A-shares 9/28 (Monday) close))

Five down, one up out of six. Nvidia rose 1.68%, the only one in the list to close green. Meta fell 4.79%, the heaviest, Tesla fell 3.94%, and Microsoft, Alphabet, and Palantir all fell slightly. These are the September 28 Monday US stock closing prices.

All six fell, with declines from 9.03% to 0.01%. Optical module makers Zhongji Innolight and Eoptolink led the decline, with server ODM Foxconn Industrial Internet and domestic chip makers Cambricon and Hygon following lower. Only Kingsoft Office barely moved. These are the September 28 Monday A-share closes.

---