AI today, 19 October 2026: Anthropic's 5% public fund, Microsoft quits Nscale UK site, neocloud shakeout
8 things that happened in AI, each in under a minute. A quiet weekend for model launches, so the news was about money and infrastructure. Anthropic set aside a large block of its own equity for the public just before its IPO, while Microsoft's exit from Nscale's grid-starved Essex site and reports of a coming neocloud shakeout showed how fragile the data-centre boom has become. Meanwhile, Anthropic is asking lawyers how agents that act for users should bargain.
-
Business · Big
Anthropic commits 5% of its shares to a new Public Benefit Fund, pending stockholder vote
Anthropic said on Sunday it will put 5% of its shares into the Anthropic Public Benefit Fund, subject to stockholder approval, with the board able to add 2% more for every 50% rise in company value above a baseline it has not yet disclosed. The fund is a wholly owned LLC rather than a charity, and Anthropic says none of its equity will be used for lobbying or political giving. Possible uses include large donations to nonprofits or direct payments to specific groups, with AI-driven job disruption given as an example; Anthropic had already set aside $200 million for that issue. Counting the co-founders' pledge to give away 80% of their equity and an employee donation match, the company says roughly 20% of its shares are now earmarked for public benefit. CNBC reports the IPO is slated for November.Why it matters: It is one of the largest corporate equity pledges tied to AI's social effects, made weeks before Anthropic's expected IPO. How much good it does will depend on governance details the company says it will publish 'in the coming months'.
-
Business
Microsoft reportedly quits Nscale's Essex AI data centre over grid power delays
Microsoft has ended its agreement to be the main customer at Nscale's planned data centre in Loughton, Essex, The Telegraph reported on Sunday. The site was meant to house about 23,000 Nvidia GPUs, but UK grid officials said they could not supply enough power until the 2030s (SiliconANGLE). Nscale called its Microsoft relationship 'strong and long-term' and said it is exploring alternative power sources for the site. The loss comes as Nscale prepares a New York IPO; Microsoft and Anthropic make up about 85% of its contract value according to its filing. Bloomberg separately reported, citing people familiar with the matter, that some investors are pushing for a valuation of $16 billion to $19 billion rather than the roughly $30 billion Nscale is seeking (the Telegraph puts the target at up to $35 billion).Why it matters: The Essex site was a flagship UK AI project. Its loss shows that grid connections, not chips, are now the hard limit on AI buildouts, and it puts the business model of 'neoclouds' (GPU rental firms) under more pressure.
-
Business
Executives expect only 7–10 neoclouds to survive a shakeout, The Information reports
Industry executives expect only 7 to 10 of the hundreds of 'neocloud' GPU-rental providers to survive a coming shakeout, The Information reported on Sunday (via Techmeme). According to the report, Nvidia is supporting the sector with easier credit terms, help finding sites, and about $20 billion in data-centre leases (Asianet Newsable).Why it matters: Read it alongside the Nscale story above: the firms renting out GPUs depend heavily on a few customers and on Nvidia's backing, and executives expect most of them to be gone or absorbed.
-
Products
Anthropic is asking law firms how AI agents should negotiate deals, FT reports
Anthropic is consulting law firms on negotiation tactics as it builds AI agents that bargain on people's behalf and deal with other companies' agents, the Financial Times reported on Sunday, citing two people familiar with its thinking. Open questions reportedly include how much an agent should check back with its user during a negotiation, and how wary it should be of rival agents without becoming 'paranoid' (PYMNTS). Anthropic declined to comment to the FT.Why it matters: Agent-to-agent negotiation is moving from lab experiment to product, and the norms for it, such as what an agent may concede or disclose, are being written by the companies themselves.
-
Business
Disney agrees to hire most of bankrupt Character.AI's staff
Disney will offer jobs to at least 60 Character.AI employees, the majority of the chatbot startup's workforce, with roles carrying an estimated $60 million in annual pay, TechCrunch reported. Character.AI filed for bankruptcy earlier in the week. Disney had already agreed to buy some of its assets for $20 million as a 'stalking horse' bid, meaning a rival bidder can still offer more. Character.AI's former CEO Karandeep Anand recently became Disney's first CTO.Why it matters: One of the best-known consumer chatbot startups is ending up as a talent purchase by a media company, a sign of how hard it is to run a standalone AI companion app.
-
Policy
Anthropic's Jack Clark tells IMF meeting AI could lift US GDP to $44 trillion
Speaking at the IMF and World Bank annual meetings in Bangkok on Saturday, Anthropic co-founder Jack Clark said an internal Anthropic model projects that, under certain conditions, US output could grow by nearly half by 2030, to about $44 trillion from just over $30 trillion; he admitted it 'sounds very outlandish' (Reuters via Manila Times). He called for international cooperation on 'unusual and serious rules' for AI, said policymakers have 'a small number of years' to act, and acknowledged AI is already costing jobs in call centres and customer service, while adding them in scientific research.Why it matters: It is a frontier lab pitching a growth forecast and stricter rules to finance ministers at once. The forecast comes from a company model, not an independent one.
-
Research
TreeOPSD trains long-horizon research agents better than GRPO without a critic
A new paper, Tree-Aware Posterior Distillation for Long-Horizon Agentic Reinforcement Learning, branches each multi-step agent run into a tree of alternative continuations and uses the final rewards to work out which individual turns helped. It needs no separate value model or teacher. The authors report it raises Qwen3-30B-A3B-Thinking-2507's average accuracy on deep-research tasks by 4.97 points over GRPO, a widely used training method.Why it matters: Credit assignment, working out which of many steps earned the reward, is a main bottleneck in training agents for long tasks. This is a cheaper approach, though the results are the authors' own.
-
Research
MIT model reads 3D crystal structures and text together to find new materials
MIT researchers described a multimodal model, published in Nature Communications, that links 3D atomic structures of crystals with synthetic text descriptions generated from papers and databases. Unlike a version trained only on text from papers, it learned to match structures to descriptions and predicted material function better; for solid-state battery electrolytes, its top picks included materials other groups had independently flagged as promising. It can also answer plain-language queries for materials with a given property.Why it matters: Pairing structure with language lets scientists search materials by describing what they want, which could speed up early-stage discovery for batteries and other uses.
How this was made: compiled by an AI model (Claude) from the linked sources and checked item by item in a separate AI fact-check pass. Corrections: [email protected].