---

Today's main thread is money tightening and goods getting more expensive, both moving at once. Storage tightens first: Micron's CEO says the memory shortage will drag on until 2028, and 75% of 2027 output is already booked. On the money side, the Bank for International Settlements called out circular investment in the AI world — from 2021 to 2025, these kinds of mutual-investment deals accounted for 28.7% of AI investment dollars. The same day, Carlyle warned that credit risk is mixed into the AI hype. On the China side, Huawei and Seres signed a new five-year partnership, and AITO users have passed 1.2 million. On Wednesday, September 30, five of six A-share AI hardware names fell and one rose, with Kingsoft Office the only one closing green.

---

Six, Macro and Market Data

Huawei and Seres sign another five years, AITO users pass 1.2 million

ITHome reported on October 1 that HarmonyOS Smart Mobility announced Huawei and Seres reached a new five-year partnership to continue upgrading the AITO business. AITO users have now surpassed 1.2 million. The signing was held in Shenzhen on September 30, with Huawei's Yu Chengdong and Seres Group Chairman Zhang Xinghai both present. Extending the partnership another five years shows the AITO line will keep getting money. For car buyers, AITO's models and service network will follow Huawei's rhythm, so before committing, check the new model schedule for the coming year. For those watching auto stocks, who funds R&D and who carries production capacity in the deal terms determines how profits get split.

Tencent-backed AI cloud company files for Hong Kong IPO

Bloomberg reported on October 1 that a Tencent-backed AI cloud unicorn has submitted a Hong Kong listing application. The company rents out compute (the machine resources needed to run AI) to enterprises. Over the past year, Hong Kong has become the main destination for domestic AI companies to list, with a smoother process than going to US markets. Tencent is both a shareholder and potentially its customer and channel. For IPO subscribers, whether AI cloud revenue can persist depends on whether customers sign long-term — start by flipping through its prospectus to check customer concentration. For investors, valuation swings in the early listing period for these companies are usually bigger than on the main board.

Memory shortage to drag on two more years, Micron has booked 75% of 2027 output

Ars Technica reported on October 1 that Micron CEO Mehrotra told investors the memory crunch will last at least two more years, with demand exceeding available supply. Micron mainly sells high-bandwidth memory and server memory for AI servers. He gave two numbers: 75% of Micron's 2027 memory output is already booked, and current talks are mostly about 2028 orders. Rival Samsung's executives also said high-bandwidth memory will take nearly 30% of memory makers' capacity in 2027, up from 20% this year. New fabs won't come online until 2028, and shipments will have to ramp up slowly. Memory for phones and PCs will give way to servers. For those building PCs, next year's new products may be more expensive; for those watching storage stocks, keep an eye on long-term contract prices and new fab progress — these two reflect the market earlier than quarterly revenue.

BIS calls it out: AI world investment, money flowing back to its own people

The Bank for International Settlements released a report on October 1 tallying circular investment among AI companies: one AI company gets money, then goes and invests in another AI company. The report says that from 2021 to 2025, by deal value, 28.7% of AI investment deals had the invested party also being an investor. Mutual investment pushes valuations up faster than actual business. For those buying AI concept stocks, this structure means faster losses when the market drops — one company's cash crunch connects to another's. To gauge risk, first look at how much of a company's revenue comes from peers rather than end customers.

Mandiant founder starts again, Armadin raises $255 million at $2.5 billion valuation

TechCrunch reported on October 2 that cybersecurity entrepreneur Kevin Mandia's new company Armadin raised $255.5 million at a $2.5 billion valuation. Mandia previously founded Mandiant, which Google acquired for $5.4 billion in 2022. The new company has a group of AI programs divide up the work of watching for vulnerabilities, replacing part of manual security defense. The security industry has always been short-staffed, and using automated programs to fill the gap is this round's selling point. Companies doing security procurement can put it on their candidate list; for investors, the valuation is already at $2.5 billion, so next watch whether it can sign big customers.

• An article by Yuyuantan reprinted by Huxiu says China and the US still have plenty of ready-made AI cooperation opportunities left unpicked. The author visited San Francisco and saw dense AI activity and active communities there. The point is that with regulation tightening on both sides, the parts that can be done together should be done quickly. For companies doing China-US business, this kind of judgment helps decide where to put teams; for investors, compliance costs for cross-border AI projects are the first thing to calculate.
• Nvidia pressed for answers — Bloomberg reported on October 1 that Nvidia is being pressed over multiple smuggling cases involving AI chips to China. The US restricts exports of advanced chips, and once chips change hands via a third location, tracking becomes very hard. For Nvidia, these cases affect whether it can keep selling China-specific models; for those doing server procurement, chips of unclear origin have no warranty or compliance protection, so don't touch them even if cheap.
• Bloomberg reported on October 1 that a Chinese company with state backing disclosed transactions involving Nvidia Blackwell chips. Blackwell is Nvidia's new-generation chip for AI servers. Disclosing these transactions shows the procurement has been put on the table. For those in servers and data centers, how many Blackwells arrive directly determines next year's machine rental prices; for investors, seeing who gets the goods is more useful than seeing who talks the loudest.
• Ars Technica reported on October 1 that a US federal judge dismissed the antitrust lawsuit by Chegg and Penske against Google's AI search. The two companies had claimed Google used AI summaries to divert their traffic. The ruling is good for Google, and the search ad model is stable for now. For companies in content and education platforms, a business relying on search referrals still needs to find a second source — don't bet traffic on one player. This case ran for over a year, and after the ruling, similar suits will drop off in the short term.
• The Information reported on October 2 that Sharon AI got a $356 million loan collateralized by GPUs (graphics processors, the chips needed to run AI). Pledging machines for cash is a common move in this round of AI expansion. For investors, note that chips depreciate fast, and collateral value shrinks as new-generation chips hit the market — the more borrowed, the bigger the risk. Interest on these loans is usually higher than ordinary corporate bonds, because chips lose value fast as collateral.
• The Information reported on October 2 that data center support supplier Accelevation's stock fell after listing. This company does supporting work for data center construction, the tool-selling type in the AI expansion. Among the batch of AI infrastructure companies that listed together, several have already broken below their issue price. For IPO subscribers, order visibility matters more than the concept for these companies — data centers get built fast, and support vendors' revenue swings widely too. Support vendors feed on the pace of data center construction; whether orders are full says more than the stock price on listing day.
• Bloomberg reported on October 1 that private equity giant Carlyle warned credit risk is mixed into the AI boom. Data centers and chip expansion rely heavily on borrowing, and if rates don't fall, repayment pressure will hit thinly capitalized companies first. For ordinary investors, to judge whether an AI company can hold up, first look at how many months its cash can last, not how big the story it tells.
• Bloomberg reported on October 1 that ECB President Lagarde said Europe must develop its own AI capabilities and cannot rely entirely on external supply. The takeaway is investment: Europe will put up more public money and subsidies to support local chips, models, and data centers. For hardware and cloud vendors doing business in Europe, local orders will increase; for investors, EU policy documents reveal where money is going earlier than company press conferences.
• An article reprinted by Huxiu says that as AI moves from answering questions to acting on people's behalf, the shape of software is changing. Before, people found buttons and filled in parameters; now you give an instruction and AI calls tools and runs the process itself. For product teams, the focus shifts from building interfaces to building APIs and permissions; for users, while it saves effort, tighten account permissions — don't let one assistant touch everything. Assistants available today still make frequent mistakes, so don't fully let go on important matters.
• A The Information article on October 1 compared two robotics paths: Google builds an open robot system, like Android, for various vendors to use; Tesla builds its own complete machine and software, like Apple. The open path spreads fast, the closed-loop path gives a more consistent experience. For those watching robotics stocks, the winners differ on the two paths — one depends on who can gather the most hardware partners, the other on who can push costs down first.
• The Guardian reported on October 1 that Anthropic suggested the Australian government give conditional approval for big companies to train models on Australian copyrighted content, while the Australian Broadcasting Corporation warned local news traffic would be siphoned off. For content people, how licensing fees are set directly determines next year's revenue; for model companies, without local-language corpora, model performance in local scenarios will be discounted.
• Tech.eu reported on October 1 that French fleet management software company Voltaback raised €2.8 million from Serena and several angel investors. It automatically records and reimburses company vehicle charging costs. After European fleets electrified, charging bills became scattered and hard to reconcile. For teams doing overseas software services, this kind of niche tool has a low barrier — the competition is whether it can plug into a company's existing financial system. When electricity prices and tax rules change, these tools have to be rebuilt.
• TechCrunch reported on October 1 that AI startup Photon raised $4.5 million and even held a real funeral in New York for mobile apps. It's betting that in the future, ordering takeout and booking tickets will be handled directly by AI assistants, and users won't need to open app after app. For app teams, the entry point being taken over by assistants is something to think about in advance; for users, the price of convenience is handing every step to the same assistant.
• Positron published on its official site hardware designed specifically for generative AI, touting high performance and low power. Generative AI inference (the step where the model gives an answer) is very hungry for machine power and memory, and general-purpose servers aren't necessarily cost-effective. For enterprises building their own AI services, these specialized machines can lower unit costs — to judge if it's worth it, first calculate how many requests per watt, then compare prices. Also check whether the software ecosystem keeps up; no matter how fast the machine, if the model won't run, it's useless.
• Cloudflare announced on October 1 that AI Search is officially commercial. It lets websites turn their own content into a knowledge base that AI can retrieve, and site owners can control which content models read. For content companies, this is a way to hold the traffic entry point back in their own hands; for developers, plugging into a ready-made search API saves effort versus building one, at the cost of factoring data and fees into the budget first.
• Tech.eu reported on October 1 that Danish AI transcription company Good Tape raised new funding from Nordic Makers, and its tool has 3 million users in over 150 countries. Transcription is turning recordings into text. It also launched Explore, used to dig buried stories out of old interviews. For media and interview teams, these tools save a lot of organizing time, but first confirm where the recording data is stored.
• Tech.eu reported on October 1 that Belgian tech companies raised €1.09 billion in the first half of 2026, with money highly concentrated at the top — Kpler alone took a big chunk. Excluding that one deal, the average raise for the rest isn't high. For teams looking for money in Europe, Belgium isn't the hottest market; for investors, when looking at these regional rankings, strip out the top outlier, or you'll overestimate overall heat. When looking at regional funding rankings, first see how much the biggest deal took, then see how much the rest of the small and mid-size companies got.
• Tech.eu reported on October 1 that Ukraine-backed Osavul raised €8.5 million. Its tool helps NATO and governments detect hostile attacks on public opinion in advance. Government orders are stable but decisions are slow. For companies doing public sector business, these contracts are large and long-cycle; for investors, revenue depends on government budgets, and the rhythm is completely different from commercial customers — what to watch is the contract renewal rate. Once a government customer renews, revenue is stable; conversely, losing one deal is hard to make back.

---

Today's Market Snapshot (in-house data, global 24 hours (US stocks 10/1 (Thu) close + A-shares 9/30 (Wed) close))

Three up, three down among the six. Palantir rose 1.60%, the strongest, Nvidia rose 1.09%, Meta rose 0.10%. Alphabet fell 1.70%, the most, Tesla fell 0.20%, Microsoft fell 0.02%, basically flat. This is the October 1 Thursday US stock close.

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

---