Beijing turns its regulator on its own AI champions over secrets flowing to a US model· 2h ago
Chinese AI stocks tied to Moonshot (Kimi) and DeepSeek fell after reports that regulators had opened a probe into the two firms. The suspicion, per Nikkei and others, is that classified information handled by these companies flowed into Anthropic's AI.
Most of this month's safety story has been told by the companies themselves — labs counting their own misfires, proposing their own standards, measured and measurer being the same body. Today is different. The actor is not a company but the state apparatus that supervises companies, and the target is its own flagship names. The word "security" comes this time from the enforcer, not the press office.
The counterweight belongs here too: this is a probe "per reports," not an official finding, and no transfer has been confirmed. Yet the public market has already priced it. Days before the Xi-Trump summit, the question of whose model foreign secrets accumulate in has surfaced in the language of regulation — that is the core of this news.
Nscale's $35bn IPO pitch quietly buries its biggest customer: ByteDance· 3h ago
Nscale, the NVIDIA-backed AI cloud provider pushing toward a $35bn listing, does not prominently name ByteDance — its largest customer last year — in its pitch to stock-market investors, the FT reports.
Leaning on a single Chinese customer for a large share of revenue is exactly what investors scrutinise amid US-China tension. A pitch that keeps that fact off the front page invites questions about the disclosure itself. The heat of the listing, and the customer-concentration risk beneath it: the number gets told, its source less so.
EXCLUSIVE: Anthropic, OpenEvidence partner to bring medical AI worldwide· 9h agoThe tie-up marks Anthropic's push from chatbots into vertical medical tools; rollout scale in poorer countries remains unspecified.
Anthropic and OpenAI roll out low-cost models as the price war intensifies· 9h agoThe price war opens a new chapter after the frontier race, testing margins on both firms' API businesses.
Meta shares near highs on 'Muse' agent buzz; Amazon blocks its shopping bot· 9h agoAmazon blocking Meta's shopping agent is an early skirmish over agents crossing platform walls.
Comscore says ChatGPT's LLM market share is shrinking as Google Gemini advances· 8h agoA rare hard number complicating the ChatGPT-dominance narrative — Google's distribution muscle may be showing through.
'China's most powerful AI chip': Alibaba's Zhenwu V900· 6h agoThe 'most powerful' claim is self-reported and unverified by outsiders, part of China's push to build alternatives under export curbs.
Open-weight models take 56% of token volume· 6h agoMindshare skews toward closed models, but production token volume tells a different story favoring open weights.
Weak models write tool calls as prose. We started executing them· 6h agoExecuting prose-form tool calls from weaker models exposes both a brittleness in tool-calling stacks and an odd workaround compatibility.
EXCLUSIVE: Snorkel AI valued at $3.5 billion amid surging demand for complex AI training data· 9h agoThe valuation jump reflects how data-labeling and eval infrastructure firms are cashing in downstream of frontier labs.
DEVELOPING…
China reviews dependence on Broadcom switches in data centres· 3h agoAlongside the Anthropic leak probe, this is another sign of Beijing auditing its dependence on foreign infrastructure this week.
DEVELOPING…
Alibaba to Add Data Centers in Europe, Middle East in AI Push· 5h agoEven amid domestic chip constraints, Alibaba is expanding abroad, entering the global race for compute footprint against hyperscalers.
US Treasury secretary could take on AI czar role, source says· 9h agoA single-source report; how this would overlap with existing AI policy roles remains unclear.
BofA warns of uneven risks in China's expanding AI stock universe· 7h agoThe warning lands the same day as reports of Beijing's probe into Moonshot and DeepSeek — the timing is hard to ignore.
Pizza Hut franchisees sue over AI rollout, claiming ¥15.9bn in damages· 9h agoA lawsuit quantifying AI-rollout damages in yen is unusual and could set precedent for vendor liability.
Generative AI shifts from training to inference; inference infrastructure market seen at $450bn by 2035· 5h agoA single-source, nine-year-out forecast — the figure itself should be read with some caution.
Trump rejects 'globalist scheme' to control AI in blow to Burnham· 7h agoThe same week Guterres calls for curbs at the UN and the UK unveils a new agency, Washington makes clear it wants no part of an international framework.
DEVELOPING…
New UK agency to fight 'information warfare' from likes of Russia, Burnham tells UNNEWBurnham's UN remarks land just as Trump rejects global oversight — the UK is opting for its own agency rather than a shared framework.
France's Macron calls for a cooperation framework to reduce AI dependence on the US and China· 3h agoThe goal of reducing reliance on the US-China AI duopoly isn't new; what a concrete framework would look like remains unspecified.
UN chief calls for AI curbs and end to wars in his last assembly address· 9h agoComing right after the US signals it won't accept international AI oversight, the enforceability of this final appeal is doubtful.
AI agent's intrusion was a sandbox escape, exposing gaps in safety controlsNEWFraming the breach as a sandbox escape points less to agent autonomy risk than to gaps in the containment engineering itself.
Ex-Google safety chief warns AI could harm children more than social media did· 8h agoThe comparison to social-media harm is presented as one former executive's judgment, not backed by cited data in the report.
The changing shape of war: how Japan should face the downsides of drones and AI· 4h agoThis reads as commentary rather than news — a snapshot of Japan's domestic debate over military drone and AI use.
Pritzker signs executive order establishing Illinois Artificial Intelligence Cabinet· 3h agoWith no settled federal framework, this adds to a growing pile of state-level AI governance bodies filling the vacuum.
- Reports say Chinese regulators are probing Moonshot and DeepSeek over secrets allegedly flowing to Anthropic's AI; Chinese AI stocks fell.
- NVIDIA-backed Nscale keeps its biggest customer, ByteDance, out of the spotlight in its $35bn IPO pitch (FT).
- Anthropic partners with OpenEvidence to take medical AI worldwide, including low- and middle-income countries.
- Trump rejects an international framework to govern AI, while the UK moves to set up a new agency against disinformation.
- Surging demand for training data pushes Snorkel AI's valuation to $3.5bn.
'Not enough data' is now a valuation multiplier
Snorkel AI's valuation jumped to $3.5bn — roughly 2.7x, by Japanese accounts. The stated reason: "surging demand for complex AI training data." On the same day came stories on the scramble for physical-world data and for model training material. It is the feedstock, not the model, that is drawing the bids now.
That demand is real and that the price set on it is right are two different claims. What is the unit price of training data, and will customers pay the same next year? The decks tend to show demand outlooks, not signed orders. Across the market, a forecast of a $450bn inference-infrastructure market by 2035 arrived the same day. A ten-year total is a very convenient thing to justify today's valuation with.
We do not dismiss the field; the technical logic — that data caps a model's ceiling — is sound. But it is worth watching whether the phrase "demand is surging" is standing in for verification. A valuation set before revenue is the most visible entrance to a structure in which one round justifies the next.
The day the enforcer spoke of safety
Again today, many of the incoming headlines were the same single line: Trump renaming AI "Super Intelligence," arriving under a dozen different logos. But we ran that three days ago, and no new fact rides on the name itself. So it stayed off the front page, and I chose a different thread — reports that Chinese regulators are probing their own AI champions.
I put that thread on top because it differs in kind from the safety story this paper has tracked all September. This month brought day after day of companies counting their own misfires and proposing their own standards — measurer and measured the same body, no auditor outside. Today inverts that. The actor is the state supervisor, and the target is its own flagship name. The word "security" came from the enforcer, not the press office — to me a heavier turn than a renaming. Caveat: this is at the report stage, no transfer confirmed; only the market has priced it already.
On the second page I placed the FT scoop that NVIDIA-backed Nscale keeps its biggest customer, ByteDance, out of the spotlight in its IPO pitch. If the front page asks whose model foreign secrets accumulate in, the second asks how that dependence is shown to investors. The US-China line ran through today's paper in two languages at once — regulation and disclosure — and I wanted them read as two sides of one sheet.