Beijing collected $32.3 billion in stamp duty on stock sales from January through August, up more than 80% year over year, with average daily turnover running 72% higher. The tax office is, in effect, the cleanest read on the AI trading frenzy that’s gripped mainland retail investors this year; you can argue about whether the rally is justified, but the receipts are real money, and they’re landing in the same state treasury that’s been underwriting the buildout everyone is betting on. The frenzy has a strange mirror image further down the income ladder. More than 7 million one-person startups were registered in China last year, roughly 42% more than in 2024, most of them launched by young people using AI tools to substitute for staff they can’t afford and to escape a job market that isn’t hiring them. Competition among them is brutal. Two stories, one economy: capital piling into the AI trade at the top, and at the bottom, people using the same technology as a last resort.
That tension is now an official concern. China has made AI a national mission, but the anxiety about displacement among Chinese workers looks a lot like the anxiety in Detroit or Leeds, and the possibility of unrest is the one variable that could push the Party to slow its own program down. Nobody in Beijing wants to be the government that automated its way into a social crisis while celebrating the milestone.
On the diplomatic side, Scott Bessent says Washington has proposed an AI incident notification mechanism to China, and that both governments have agreed to stand up a formal AI dialogue ahead of the Trump-Xi summit. Incident notification is a nuclear-age idea being retrofitted for models: you don’t have to trust the other side, you just have to agree to pick up the phone when something goes wrong. It’s modest, and modest is roughly the ceiling for what these two can agree on right now.
Domestically the mood is less collegial. Politico has a long reconstruction of the 19-day standoff between the White House and Anthropic, which began with a jailbreak dispute and ended with officials bluntly telling Dario Amodei to take Fable down. Whatever the merits of the underlying safety argument, the precedent being set is about who gets to decide a model comes off the market, and the answer in this telling wasn’t a court or a regulator with a statute behind it. Jensen Huang, meanwhile, offered the least charitable possible reading of his peers’ regulatory enthusiasm: the AI leaders asking for rules don’t want new legislation, he says, they want to be “relieved of the laws we do have,” and he thinks the reasons are ulterior. It’s a cheap shot that happens to describe several real lobbying positions. At the state level, analysts and lawmakers are warning that several AI chatbot safety bills contain language broad enough to function as loopholes, which is what tends to happen when the industry helps draft the text.
Capital kept moving regardless. SoftBank is reportedly looking to raise $10 billion and €1 billion in debt to fund its OpenAI position, which would rank among the largest junk bond sales ever attempted. Masayoshi Son borrowing at high yield to buy into the most expensive private company on earth is either the defining trade of the cycle or its epitaph, and there’s no way to know which until the paper matures. Elsewhere the rounds were smaller and more legible. Hong Kong’s Qupital, which finances cross-border ecommerce for small merchants, raised a $300 million Series C led by M Capital and is weighing an IPO. Adaptive, selling accounting software to construction contractors, took a $30 million Series B led by Tidemark, bringing it to $57 million total. Altis Labs raised $25 million for a model that reads CT scans collected during oncology trials and produces survival-linked predictions. Mithrl raised $20 million from Obvious Ventures to wire its biomedical world model into pharma R&D systems. And in the strangest deal of the day, London’s Unit1 raised nearly £15 million from Balderton and others to build “hyper-realistic” digital avatars of musicians and recreate classic concerts, which is a business whose legal exposure probably scales faster than its revenue.
The Pentagon spread its bet on drones across nine startups, including Eric Schmidt’s Perennial, splitting orders for 60,000 units under a competition that bars China-made parts. Nine winners means the department is buying supplier diversity as much as hardware, and the no-China-components rule is the actual product being procured. On the components side, CXMT says its fifth-generation DRAM platform has entered mass production, with two 24Gb LPDDR5X mobile memory parts built on it. Every one of those announcements is a small subtraction from the pricing power of the incumbent memory makers, and they’re arriving faster than the consensus assumed.
Model releases kept their usual pace. Alibaba shipped Qwen-Image-2.1, a 7-billion-parameter open-weight model it claims beats most closed competitors, with native transparency support and up to ten reference images. Damo Academy separately open-sourced RADAR, a medical vision-language model that reads CT scans and flags around 150 abdominal conditions, cancers among them. Clinicians in the UK are sounding notably less enthusiastic about medical AI once it moves past diagnostics and imaging, where the performance data thins out quickly. Both things can be true: the models are getting good at the narrow tasks, and the broad deployment case still rests on evidence nobody has produced. In tooling, Vercel and Cloudflare moved fast to adopt Jev, which speeds up and cheapens AI tool selection, and TypeSafe’s evals put it level with GPT-5.6 and Sonnet 5 on workflow tasks.
A researcher used GPT-6 Astra to crack a German WWI radio transmission from 1918, one of 50 ciphers a German science blog lists as unsolved. It’s a parlor trick with real archival value, and a reminder that the most convincing demos are still the ones where the answer can be checked.
The physical economy around all this keeps showing up in rents and wages. AI companies, OpenAI and Anthropic among them, are pushing up Singapore office rents as they expand into a market the government spent years courting. In the US, Indeed data puts a roughly 42% pay premium on data center maintenance and installation roles compared with equivalent hourly work elsewhere. The compute buildout is a construction and facilities story before it’s a software story, and it’s the tradespeople who are getting paid first.
The rest of the day’s news was regulatory housekeeping with teeth. Meta is challenging Ofcom’s Online Safety Act categorization of WhatsApp and Instagram, which would impose extra obligations; Roblox and Quora are contesting theirs too. Polymarket’s Shayne Coplan reportedly waved off internal concerns after fraudsters tried to move more than $10 million from stolen cards through clean accounts using the platform’s wagers back in February. And Raspberry Pi’s Eben Upton, asked about AI, came out skeptical of the evangelists while noting that the same wave is driving real demand for edge compute hardware. Both of those positions are correct, which is more than most people in this business manage.