---

Today's issue watches two threads. One is money. The Information says the debt behind AI data centers is starting to surface in all sorts of places; Amazon is putting up $1 billion to respond to the backlash caused by its server farms; Lambda borrowed $1 billion specifically to buy the GPUs that run AI. The denser the server farms get, the more the bill will eventually land on someone's electricity and rent — that's an unavoidable question. The other thread is people. OpenAI brought in an official who once handled US AI policy, to lead on national security; data center company Nscale brought in a Meta executive as its number two, paving the way for an IPO; in a UK poll, three out of five adults said they don't have much goodwill toward AI giants. On the chip side, onsemi switched to an all-cash acquisition of Synaptics at $123 per share, and both stocks rose together in pre-market. On the hardware side, Nvidia rolled out a 64GB desktop dev machine, so developers can run open models locally. On Wednesday, September 30, five of six A-share AI hardware names fell and one rose, with Kingsoft Office the only one closing green, after which A-shares entered the National Day holiday closure.

---

Six, Macro and Market Data

onsemi switches to all-cash acquisition of Synaptics, both stocks rise together in pre-market

TMTPost reported on October 2 that onsemi revised its acquisition plan for Synaptics. The new plan is $123 per share, paid entirely in cash. Once the news broke, both companies' stocks moved up in pre-market. Synaptics makes chips for touch and IoT. onsemi's original plan included stock; this time it's all cash, the terms are more direct, and shareholders went along with it. Synaptics' touch chips are used in phones and cars, and together the two want to grab a bigger piece of consumer electronics chips. For those watching semiconductors, this is a sizable industry consolidation, with chipmakers filling in the links they lack. For ordinary investors, before the acquisition actually closes, the stock price tends to swing with the news — don't make decisions chasing a single day's moves.

OpenAI brings in a former US government AI official to handle national security

The Information reported on October 3 that OpenAI brought in a senior official from the Trump administration who handled AI affairs, to do national security-related work at the company. Bringing in people who understand policy is a common practice for model companies these past two years. The government sets the rules, companies get an early read on the direction, and people flow back and forth between the two sides. This kind of personnel change often happens right around when regulations come out, as companies want an early read on which way policy is heading. For AI companies, the line between government and industry is getting blurrier, and people on the team who understand policy will be worth more. For ordinary readers, the closer model companies get to the military and government, the more we should ask where and how the technology is used.

Nscale brings in Meta executive as COO, preparing for an IPO

Bloomberg reported on October 3 that AI data center company Nscale brought in Meta executive Justin Osofsky as chief operating officer. The company is preparing for an IPO. Nscale builds server rooms and servers for AI — the kind of machines needed to run models. Filling out the management and operations ranks before an IPO is a common move. The company has been expanding its server farms for the past year, adding headcount along the way, and is getting its management layer in place ahead of the listing. For those watching this sector, the IPO pace of such companies shows how strong data center demand still is. For job seekers, AI infrastructure companies are competing for people with operations experience — the opportunities are on this line.

• ITHome news on October 2: Rivian is pushing an October infotainment system update to its R2 model, with the version number expected to be 2026.35. What's added this time is pretty life-oriented — a pet camera, and the power tailgate can adjust its opening height on its own. For owners, the car can still gain features through software updates after purchase, so you get something new without changing cars. For companies making in-car software, this kind of regular update is a way to retain users, and how features iterate directly affects reputation and repeat purchases. This update will roll out in batches; owners can install it over the internet after getting the notification, without a special trip to a store, and old features will get a touch-up along the way.
• The Information reported on October 3 that compute company Lambda secured a $1 billion loan specifically to buy GPUs — the hardware that's most in demand for running AI models. The company rents machines to people training and deploying models. For model companies, renting is more flexible than buying, but the unit price can change at any time, so budgets need some margin. For investors, borrowing to expand capacity shows everyone is still betting demand will keep rising, and payback depends on whether the rent keeps coming in. This kind of loan is mostly secured by the machines and equipment, with rates higher than ordinary borrowing; machines depreciate and rent fluctuates, and these risks all fall on the borrower.
• The Information — A October 3 article from The Information says the debt behind AI data centers is starting to show up in all sorts of places, from bank loans to the bond market. More and more money is being borrowed to build server farms, but repayment depends on slowly collecting rent by leasing out the machines later. For investors, the risk in this kind of debt is hidden in interest rates and lease terms — don't just look at how big the server farm is. For ordinary readers, the faster server farms expand, the more easily the electricity and rent bills eventually land on the people using AI. The report warns that if rental income falls short of expectations, the repayment pressure will first hit the few companies that borrowed the most aggressively without matching scale — no one in the chain escapes.
• Bloomberg reported on October 2 that, facing the backlash caused by data centers, Amazon put up $1 billion in response. Nearby residents complain about rising electricity bills, noise, and heavy water use, and the company has to put up money to smooth things over. For people living near server farms, this kind of commitment depends on where it lands and how it's actually spent — don't just listen to the number. For infrastructure companies, site selection and how well they get along with the community are becoming key to whether they can break ground smoothly. This money will mostly go to local grid upgrades and community projects; how it's divided and how much, we'll have to see from the details the company releases later.
• Bloomberg reported on October 2 that Microsoft's and Amazon's cloud businesses will face closer scrutiny under the EU's new tech rules. The EU is looking at how cloud services are priced, where data is stored, and whether customers can easily switch providers. For companies doing business in Europe, choosing a cloud means keeping a backup path — don't bet critical operations on just one. For these two cloud vendors, compliance costs will rise, and services and prices may change along with them. Once found non-compliant, fines could be calculated as a percentage of the company's global revenue — not a small sum — and remediation takes time.
• TechCrunch reported on October 3 that a company called Circuit Breaker Labs is building tools to make AI safer for kids. It focuses on dangerous topics in chats like self-harm and bullying, aiming to help parents and platforms spot warning signs early. For parents, relying on platform reminders isn't enough — best to know which app your kid uses and whether there are parental controls. For product teams, content filtering for minors is turning from a bonus into a hard requirement, and doing it poorly means penalties. The company says it will first define the categories of dangerous content clearly, then hand it to models to identify, with humans doing a final confirmation to minimize misjudgments.
• Wired reported on October 3 that Trillium Labs wants to do high-risk AI research openly, so the process can be seen from the outside. Big companies often tune models locked in labs, with no outside view of what happens inside. For people outside the company, open research gives the discussion more evidence and less guesswork. For researchers, transparency also means facing outside scrutiny earlier, and the safety and data lines have to be held first — otherwise openness just adds chaos. The organization says it will publish experimental methods, data, and conclusions together, making it easy for peers to review and for outsiders to find faults; how far to open up will be tested gradually.
• Business Insider reported on October 2 that most VCs are still pouring money into AI startups, but a few have turned to betting on the old concert business. Their judgment is that AI makes it easier to kill time at home, so people actually want to go out and see live shows more. For people in offline entertainment, this wave of money can help venues and tours expand. For investors, this bet is on people being willing to pay for live experiences, and whether it works ultimately depends on the box office — don't mistake the buzz for returns. Over the past year, live performance box office has been steady, the money isn't so burn-heavy, and returns are visible — that's also why VCs took a liking to this old business.
• TechRadar reported on October 2 that a new YouGov poll shows three out of five UK adults say they don't have much goodwill toward AI giants and don't trust them much either. The controversy centers on data, copyright, and jobs. For companies making AI products, user distrust directly lowers willingness to pay — no matter how good the product, you have to explain the whole story clearly first. For ordinary people, these feelings will push regulation to tighten, and new features you could use later may have to wait. The poll also found that younger people are actually more cautious about AI, which is different from the "young people love trying new things" assumption of the past few years.
• Bloomberg reported on October 2, citing a report, that New York's AI office buildings are going up hot, but fresh graduates are finding it harder to get jobs. Server farms and offices take up large swaths of space, but entry-level positions haven't grown along with them. For CS graduates, the bar for getting into big companies is rising, so practice early in directions where you can get hands-on and build a portfolio. For those concerned about employment, just looking at office buildings and output isn't enough — you also have to see who the jobs land on. The report says the ones actually hiring are mostly established financial institutions, not the newly built AI companies; the two sides don't line up, making it even harder for fresh graduates to land.
• AP reported on October 2 that in Arizona, an AI-generated victim statement was used in a case, and the verdict was subsequently overturned and redone. AI-written text looks real, but the content may not match the facts. For people in litigation, don't use AI-fabricated material as evidence — once it's found out, all the prior effort may be wasted. For courts, how to identify AI-generated content is becoming an unavoidable new problem. Defense lawyers found the details in the statement didn't add up, which is how this came to light; the court then overturned the original verdict and the case started over.
• A October 2 article from TechRadar says the next risk for financial institutions using AI lies in processes no one can fully explain. Models make judgments for people, but the steps they rely on are scattered across systems, making it hard to assign blame when something goes wrong. For risk control people, you have to sort out the process first, then plug AI in — otherwise you can't say which step went wrong. For depositors and investors, when banks use AI for lending and risk control, the rules should be more transparent — don't leave people in the dark. The article suggests leaving a record at each step first, then letting AI take over the judgment, so when something goes wrong you can trace the source and say clearly who's responsible.
• Nvidia introduced the 64GB version of DGX Spark on its official blog on October 2, giving developers a local machine that can run beside their desk. Running open models doesn't require connecting to the cloud — you can get to work locally. For people doing development and testing, running locally saves money, and data doesn't have to leave the machine. For small teams, this kind of device means renting fewer cloud machines, at the cost of buying the machine itself and bearing the upgrades yourself. This machine uses Nvidia's own chips, comes with supporting software tools, and works right out of the box, saving a lot of the hassle of setting up your own environment.
• TechCrunch reported on October 3 that Napster founder Sean Parker wants to rebuild Stability AI around music. The company was originally known for image models, and now it's shifting its focus to generating music, and says it will play by the rules. For people making music and sound effects, more tools and a lower barrier, but how copyright is counted is still undecided. For this company, switching tracks is survival — whether it can stand depends on whether the work sells and whether licensing can be negotiated. He stressed that this time it will settle music licensing first, no longer going down Napster's old lawsuit-prone path — only with the right direction can it talk about growing big.
• TechCrunch reported on October 2 that Rivian sold a record high number of EVs in the third quarter, mainly on the back of its new cheaper model, the R2. Push the price down and orders follow. For people wanting to buy an electric SUV, more models entering the market means more choice and bargaining room — you can compare a few more. For carmakers, the entry price point is the key to volume; high-priced cars alone can't support factory capacity. The company also said R2 production will keep rising over the next few quarters, delivery times are expected to shorten, and those waiting can suffer a bit less.
• The MariaDB team said on its official site on October 3 that, with the help of AI tools, they found many program vulnerabilities this year that went undiscovered before, and security fixes have been much busier than in past years. Databases store core data of enterprises and individuals — one hole is enough to cause trouble. For companies using open-source databases, patching promptly is more practical than anything — don't drag it out until something happens. For developers, plugging AI into code review can cover spots humans easily miss. The team compiled these findings into a list, reminding organizations still on old versions to upgrade as soon as possible — don't wait until something actually happens to go back and patch, the cost will be much higher.

---

Today's Market Quick Look (in-house quotes, global 24 hours (latest US market close (as of Asia-Pacific session) + A-shares 9/30 (Wednesday) close))

All six rose. Tesla rose 4.65%, the strongest, Alphabet rose 1.56%, Nvidia rose 1.34%, Amazon rose 1.33%, Microsoft rose 0.92%, Meta rose 0.30%. Tesla's gain is tied to third-quarter deliveries beating expectations. The above are the latest US market closing prices.

Five of six fell and one rose. Hygon Information fell 3.85%, the most, Cambricon fell 3.54%, Eoptolink fell 1.00% after that. Zhongji Innolight fell 0.56%, Foxconn Industrial Internet fell 0.38%. Kingsoft Office rose 0.88%, the only one closing green on the list. The data is from the Wednesday, September 30 A-share close, after which A-shares entered the National Day holiday closure.

---