░▒▓██████▓▒░
──────────▓██████████████▓──────────
░▒▓██████▓▒░
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 80 · Friday, 2 October 2026Hello from the end of the week. Broadcom has agreed to lend Anthropic up to $42 billion so the lab can rent chips Broadcom helps design, according to the Anthropic IPO filing we covered earlier this week. OpenAI cut ties with three safety researchers it says mishandled sensitive information. Meanwhile, Nvidia wants banks to lend against its chips to fund its customers' data centers, and the banks want more guarantees first.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Broadcom agreed to lend Anthropic up to $42 billion to rent the chips Broadcom helps design
Anthropic's confidential IPO prospectus, which we covered earlier this week, also shows Broadcom agreeing to lend it up to $42 billion through a convertible note, Reuters reported on Thursday. The money could cover about a third of the $125.2 billion Anthropic has committed to a five-year lease of computing capacity on tensor processing units (TPUs), the AI chips Broadcom helps Google design. If the note converts, Broadcom or a financing partner it names could end up holding Anthropic shares, and Anthropic itself flags that a supplier doubling as its lender creates potential conflicts of interest. Bloomberg reported on Thursday that Anthropic has pushed its listing back to mid-November at the earliest, after the midterm elections, a slip we flagged as possible on September 21 (Reuters via Investing.com · Semafor · Bloomberg).
OpenAI cut ties with three safety researchers it says mishandled sensitive company information
The three allegedly shared confidential material with an outside AI-safety organization, The Wall Street Journal reported on Thursday, as relayed by TechCrunch. OpenAI confirmed the departures and said its investigation found they broke its rules on handling sensitive information, but it has not confirmed who the three are, which organization received the material or what it was; the Journal has since named them. The exit comes two days after The New York Times reported that OpenAI executives had brushed aside staff warnings about its safety practices, a report we covered on Wednesday (TechCrunch · Decrypt · RTÉ).
Bankers want more guarantees before lending against Nvidia's chips under its $500 billion financing plan
Nvidia wants its chips to serve as collateral for loans that pay for its customers' data centers, but banks and asset managers told Reuters on Thursday that they doubt the chips hold their value long enough. Banks write the chips off over three to four years, while Nvidia says its top chips can earn money for up to a decade. The $500 billion financing plan rests on memorandums of understanding that Nvidia signed in August with financiers including Blackstone, Apollo and KKR, not on committed funds, as we reported on August 11. Loan deals worth tens of billions of dollars are in the pipeline and likely to carry strong guarantees, one source said, and Nvidia told Reuters that AI compute is a durable asset that can support long-term financing. Separately, The Information reports that Société Générale, Sumitomo Mitsui Banking Corp. and Mitsubishi UFJ Financial Group have become more selective in lending to data-center projects (Reuters via BNN Bloomberg · The Information).
Meta's Muse agent passed 3 million weekly users, The Information reported a day after OpenAI launched its rival Dots
More than 3 million people now give Muse, Meta's always-on personal AI agent, at least one task a week, and more than 1 million do so daily, according to internal data seen by The Information. About a week after its September 8 launch, more than 500,000 people had tried it. Muse has a free tier, while Dots, OpenAI's rival agents we covered on Wednesday, are rolling out first to Pro and Business Premium subscribers ⚠ (The Information · The Information, September 22).
China-aligned hackers posed as a former senior White House science official to phish experts who work on AI regulation
The cybersecurity company Proofpoint said on Thursday that a China-aligned espionage group it tracks as TA419 sent emails purporting to come from AI and statecraft experts, among them Lynne Parker, formerly principal deputy director of the White House Office of Science and Technology Policy. The targets were people at think tanks, universities and law firms working on AI regulation, export controls and national AI strategy. Proofpoint bases the attribution on the malware, the infrastructure and a target list that matches Chinese intelligence priorities (Reuters via ThePrint).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░America.gov, the government chatbot built on Gemini and Grok, gave mixed answers on the 2020 election, sometimes contradicting the president and sometimes refusing to answer (update). At its Tuesday launch, which we covered on Wednesday, it contradicted him on the 2020 election and confirmed that Barack Obama was born in Hawaii, Axios and other outlets found. By Wednesday morning some users were getting a refusal instead, with the bot explaining that it does not provide political commentary; by the time Axios published later that day, its reporter was again getting the launch-day answers (Axios · RTÉ).
░░▒▒▓▓ POLICY ▓▓▒▒░░The Pentagon plans a four-star command for drones and robots, and Elon Musk will help recommend which weapons the military should develop. Defense Secretary Pete Hegseth announced the Autonomous Warfare Command on Wednesday, meant to scale autonomous and robotic systems across the military and to be established by October 2027, pending legislation from Congress. He also named Musk, Anduril co-founder Palmer Luckey and former House Speaker Newt Gingrich to lead Project Meridian, a 120-day study of future warfare that will recommend weapons and capabilities for the military to develop. Musk's SpaceX and Luckey's Anduril both hold multibillion-dollar defense contracts with the US government (DefenseScoop · Stars and Stripes).
░░▒▒▓▓ INFRA ▓▓▒▒░░Google put its first AI chips in orbit, and its own study says making launches cheap enough for space data centers would take about 1,800 Starship flights (update). The prototype satellite, built with Planet Labs, went up on a SpaceX rocket on Thursday to test whether Google's TPUs survive space, as we noted on September 28. Launch prices are a critical part of a space data center's cost, and Google's study says getting them to about $200 a kilogram by the mid-2030s would take roughly 1,800 Starship launches, about 180 a year (TechCrunch · arXiv).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░Meta's research tax credit reached $3.9 billion in 2025 after it began calling its AI data centers experiments, The New York Times reported. For tax purposes Meta classifies the data centers as "pilot models" and claims the credit on the AI chips it buys from Nvidia and others, the Times reported on Wednesday; the same credit was worth $700 million to Meta in 2023, the year before it adopted the approach. Meta's own accountants consider the strategy legally shaky, the paper said (The New York Times · The Next Web · The Decoder).
Peter Lee, who led Microsoft Research for years, and former LinkedIn CEO Ryan Roslansky are leaving Microsoft. Lee, most recently president of Microsoft Science, announced his departure on Wednesday without naming his next job; Roslansky, a Microsoft executive vice president since last year, will leave at the end of the year. They join a run of senior exits this year that includes Windows and Office chief Rajesh Jha, Xbox chief Phil Spencer and communications chief Frank Shaw, whose exit we covered on September 14 (The Information · GeekWire).
░░▒▒▓▓ ALPS ▓▓▒▒░░Switzerland and Germany signed a letter of intent naming AI and robotics as research priorities. State Secretary Martina Hirayama and her German counterpart Marcus Pleyer signed the letter after talks in Bern on Wednesday; the two countries have already run just under 1,300 joint projects under Horizon Europe (swissinfo · SERI).
Allianz Commercial's chief executive says some AI risks cannot be fully insured. Thomas Lillelund, the chief executive, told Handelsblatt, in an interview published on September 29, that damage hitting many customers at once, because they depend on the same few AI models, needs public-private solutions. He called for a state-backed fund (Handelsblatt · Versicherungswirtschaft-heute).
░░▒▒▓▓ PAPERS ▓▓▒▒░░EPFL and Apple — in machine-learning engineering, the framework built around an AI coding agent adds little; the model does the work. In a preprint posted on September 30, researchers tested agents that try to build machine-learning systems on their own and found that, with the same time budget and the same frontier model, leading open-source frameworks gave no advantage over one well-prompted coding session with basic tools. The paper is not yet peer reviewed (arXiv).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░Americans prefer "artificial intelligence" to "super intelligence" by 53% to 9%, a YouGov poll found (update). The rename the president made official for the executive branch, in the executive order we covered on Wednesday, has yet to catch on, Axios reported, and some AI companies plan to use the new name, SI for short, when meeting the president and working with the federal government (Axios).
Seattle Times journalists picketed over AI on Wednesday. Management has agreed that AI will not interview sources or write stories, but has not agreed to protect copy editors, photographers and designers, roughly half the newsroom by the union's account; the paper itself sued OpenAI and Microsoft over its archives, as we reported on September 7 (The Stranger · GeekWire).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Ever since the president told the UN on September 22 that AI would be called super intelligence, or SI, in government documents, people have been buying Slovenian web addresses. Slovenia's .si registry logged 13,603 new domains in the 24 hours to October 1 and had more than 241,000 in all by then. The registry, which charges registrars €10 a name, has rejected reports that the surge is earning Slovenia millions and cautioned against comparisons with Anguilla's .ai windfall, making it the rare beneficiary of the AI boom talking the numbers down (CircleID).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 The Diver — Juan Martin, Mark Isham
Lep vikend — the Uncertain Futures team, still on .org