██
██▓▓▓▓▓▒▒▒▒░░░ · · ·
██▓▓▓▓▓▒▒▒▒░░░ · ·
──██──────────────────────────
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. 83 · Wednesday, 7 October 2026Good to have you back for the middle of the week, and apologies for the late arrival. Mistral, the French lab, has a trillion-parameter model it says no open model from the US or Europe can match. The Pentagon says it has stopped using Anthropic's Claude, months after blacklisting the company. And DeepSeek is close to raising at least $12 billion, Bloomberg reports.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Mistral announced Mistral Large 4, a trillion-parameter model it ranks above every other open-weight model from the US or Europe, and plans to publish the weights on October 27
The French lab introduced the model on Tuesday and, in the same post, its nickname: Le Chonk. It has about a trillion parameters, the numbers that make up a model, but only 49 billion work on any given word, a mixture-of-experts design that keeps running costs down. For now it is in a public preview through a moderated API, and cybersecurity leaders, vetted partners and state authorities are getting a version with reduced moderation and broader cyber capabilities for testing. The downloadable weights follow on October 27, after more training and safety testing. Mistral's own ranking covers open models from the US and Europe, which leaves out the Chinese leaders; CNBC's headline has the company saying it rivals the best of those too. Pierre Stock, Mistral's vice president of science, told Axios he does not want to live in a future in which "any oligopoly controls closed access to this type of intelligence". Like Beam, from the Nvidia-backed American startup Reflection AI, which we covered yesterday, it is an open model that nobody outside can download yet (Mistral on X · Axios · CNBC · Mistral).
The Pentagon says it has stopped using Anthropic's Claude, months after blacklisting the company; it has signed contracts with OpenAI, Google and Elon Musk's SpaceXAI (update)
A Defense Department official told the BBC on Monday that the department has ceased using Anthropic's products. It had kept relying on Claude for research, intelligence analysis and military operations, including through the Maven system run by Palantir, the defense tech firm, for months after branding Anthropic a supply-chain risk, the BBC reported; an appeals court let that blacklisting stand, as we covered on September 28. The fight began when Anthropic refused to drop its safeguards against mass domestic surveillance and fully autonomous weapons. In February, Dario Amodei, Anthropic's chief executive, said the company would help the Pentagon make a smooth transition to another provider. The Pentagon had said it would stop using Claude by late August; the BBC says it is not clear why it took longer. Meanwhile the Pentagon has signed contracts with OpenAI, Google and SpaceXAI, which Musk says will become SpaceXSI, as we covered on Monday (BBC · The Star · NPR · AP via PBS NewsHour).
DeepSeek is close to raising at least $12 billion, well above its own target, with Tencent and the battery maker CATL among the backers, Bloomberg reports
The Chinese lab is close to raising at least 80 billion yuan, about $12 billion, in its latest round, Bloomberg reported on Tuesday. It had hoped for about 50 billion yuan, roughly $7.4 billion, until its latest model came out and demand outran the target. The round comes ahead of an IPO planned for early 2027. Based on the term sheets already signed, the total could approach 100 billion yuan, about $15 billion, Bloomberg's sources said. As we covered on September 25, DeepSeek's annualized revenue had already reached $1 billion after the lab raised its prices (Bloomberg · PYMNTS · Calcalist).
Hackers stole records on about 8.8 million people from Denmark's national population register, more people than live in Denmark
The register also holds Danes who live abroad and people who have died, about 11 million records in all, TechCrunch reported on Monday; the stolen data includes names, addresses and Danish social security numbers. The attackers got in by abusing a Danish company's lawful access to search the register, a kind of access some companies hold to check people's details with the government. The breach happened in September, was discovered on October 2 and is thought to be the largest in Danish history. TechCrunch rounded the number of people affected to 8 million; the Danish government's own count is about 8.8 million (TechCrunch · OODAloop · Danish government).
Lambda, the Nvidia-backed AI cloud company, is raising up to $4 billion at a $14.5 billion valuation before a planned 2027 IPO
The investment firms Coatue and Blackstone are leading the round, TechCrunch reported on Tuesday; the valuation is counted before the new money. Its backlog of contracted work more than tripled in three months, from $15 billion in June to $50 billion in September, anchored by a computing deal with Anthropic; we covered a $35 billion Anthropic agreement with Lambda on September 2. The equity comes on top of more than $1 billion in debt that Lambda said last week it had secured to buy over 30,000 graphics processors (TechCrunch · The Information).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ AI ▓▓▒▒░░Anthropic is giving startups a free year of Claude Team, its paid plan for groups, plus $1,000 in API credits. The expanded Claude Startups program, announced on Tuesday, covers up to five premium seats and adds partner discounts worth up to $45,000. Founders qualify if their company was started in the last five years or raised money in the last two, and need a company email address that matches their website. Anthropic says it built the program because it expects most people to get AI's benefits through the companies that build on top of models (Yahoo Tech).
░░▒▒▓▓ TECH ▓▓▒▒░░Early reports said Qualcomm licensed the patents behind Huawei's new chipmaking technique; Qualcomm says the deal has nothing to do with them. The two companies announced a multiyear cross-licensing agreement covering 5G devices, AI services, optics and networks, and early reports, Bloomberg's among them, said it gave Qualcomm access to the patents behind LogicFolding, Huawei's technique of stacking a chip's computing in two layers, which we covered on September 9. On Tuesday Qualcomm said claims linking the deal to LogicFolding were not accurate, and that reports calling Qualcomm the net payer were incorrect; the terms are confidential. Huawei expects the total contract value of all its patent-licensing deals to exceed $6.9 billion after the Qualcomm agreement (Bloomberg via The Edge · Huawei Central).
░░▒▒▓▓ POLICY ▓▓▒▒░░Google DeepMind gave its Gemini 4 Argon model to its own circle of trusted partners before Britain's AI Security Institute. POLITICO reported on Friday, and again on Tuesday, that Google shared the model first with its Fairwind Program, a limited-access group of governments and partners that includes Wiz, the Google-owned cybersecurity firm; as we covered on October 1, when Argon launched, Fairwind members were its first users. DeepMind is in the process of sharing it with the AI Security Institute, without cyber guardrails, so it can be tested fully. The British institute's privileged early access was already slipping: as we covered on September 25, the White House asked OpenAI and Anthropic to keep their newest models from British testers until the US had reviewed them (POLITICO Europe · POLITICO).
Britain's online-safety regulator opened its first formal Online Safety Act investigation into Meta, over an Instagram feature that makes posts vanish. Ofcom is examining whether Meta completed the two risk assessments the Online Safety Act requires, one for illegal content and one for child safety, before launching the feature, called Instants, in May; posts sent through the feature disappear once the recipient opens them. Breaches can cost up to £18 million or 10% of qualifying worldwide revenue, whichever is higher. Meta says it ran a risk analysis and briefed Ofcom about the feature several times before launch (Ofcom · Quartz).
New York's City Council put OpenAI, Anthropic, Google and Meta under oath about AI risks, and plans to take SpaceXAI to court for not showing up. Of the four labs at Monday's marathon hearing, only Meta had agreed to appear before the council threatened the others with subpoenas, Speaker Julie Menin said. The council questioned the companies about AI agents that escaped their test environments and is weighing ten local bills that would put validation, reporting and liability duties on AI developers. Former lab researchers testified as well. SpaceXAI did not appear despite its subpoena, and the council says it will pursue the matter in court (CNBC · The City Reporter · PPC Land · The Information).
░░▒▒▓▓ INFRA ▓▓▒▒░░Google agreed to buy about 3.6 gigawatts of power from Constellation Energy, roughly a quarter of it from added nuclear output. The deal, announced on Tuesday, pairs a 20-year agreement for 890 megawatts from upgrades at existing Constellation nuclear plants with a long-term contract for another 2,700 megawatts in PJM, the largest US grid, and will enable more than $4.3 billion of new investment by Constellation, Reuters reported (Reuters via BNN Bloomberg).
DayOne, a Singapore data-center operator that started as part of China's GDS, filed for a Nasdaq IPO on Monday after its revenue nearly tripled. Revenue rose to $484.3 million in 2025 from $178.1 million a year earlier, and reached $512 million in the first half of 2026, more than in all of 2025, its filing shows. The company lost $369 million last year. Coatue, Hillhouse and GDS each own about 19% (SEC filing · The Information).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░OpenAI is talking to funds from the United Arab Emirates about anchoring its $30 billion round with as much as $10 billion (update). As we covered on September 30, OpenAI is seeking at least $30 billion at a valuation of about $1.4 trillion. Several Emirati funds, Abu Dhabi's MGX among them, have discussed forming a syndicate to anchor the round, and BlackRock is in talks to join them, Bloomberg reported on Monday; Thrive Capital and Andreessen Horowitz have also considered taking part. OpenAI has pushed its IPO back to 2027 at the earliest; DeepSeek and Lambda, both in TOP 5 above, are planning theirs for 2027 too (Bloomberg via Yahoo Finance · Business Today).
Etched, the AI chip startup, is weighing investment offers that would value it at $40 billion to $50 billion, two months after it was valued at $21 billion. The bids run from $40 billion from top-tier investors to $50 billion from lesser-known ones, TechCrunch reported on Monday; the talks are early. Etched builds complete AI systems around its own chips and says it has $1 billion in customer orders, the trading firm Jane Street among them. It was valued at $10.3 billion in July, which means its price has roughly doubled twice since the summer (TechCrunch).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░The Parti Québécois, which wants Quebec to leave Canada, won the province's election and will form a minority government. Paul St-Pierre Plamondon becomes premier after Monday's vote, CBC News projected; his party won with less than 30% of the popular vote, and the governing Coalition Avenir Québec (CAQ), which took a large majority four years ago, was projected to win no seats. He has said he would hold off on an independence referendum until 2029 or 2030, after President Trump leaves office (CBC).
France and Germany asked the EU for a trade weapon that could cut a country off from the single market, days before EU talks in Beijing. In a joint letter to Commission President Ursula von der Leyen, President Emmanuel Macron and Chancellor Friedrich Merz proposed an instrument the Commission could trigger unless a qualified majority of member states stopped it, with measures up to an immediate cut-off from the internal market. The letter, sent on Monday, names no country, but it arrives before trade commissioner Maroš Šefčovič meets China's commerce minister in Beijing and before an EU summit on October 15 and 16 (SCMP · GZERO).
Kenya confirmed its first Ebola case after a man who had traveled from Congo died in Nairobi. The Kenyan citizen had lived in Congo for seven years, fell ill about a month ago, traveled by road to Kampala and flew to Nairobi on Saturday; he died late on Monday. Health officials have identified 28 contacts and are tracing 23 passengers from his flight. Kenya is the fourth country with a case linked to the epidemic in eastern Congo, the second largest on record (Al Jazeera · The Guardian · Reuters via WinCountry).
░░▒▒▓▓ ALPS ▓▓▒▒░░Germany set a record for new startups in the first half of 2026, and about a third of them work on AI. More than 3,000 were founded in the first six months, up 52% on the previous half-year, Reuters reported on Tuesday. Capital is easier to find, and a weak labor market is pushing talent out of jobs that used to be safe; the German Economic Institute estimates that competition from China cost the country about 400,000 industrial jobs between 2019 and 2025. Hundreds of the new startups sell AI meant to help struggling industrial companies cut costs and automate routine work (Reuters via Investing.com).
Beyond Gravity, the Swiss state-owned space supplier, disclosed a cyberattack nearly two months after it saw the first signs. The company, spun off from the defense group Ruag in 2022, said on October 2 that it first noticed malicious activity in its IT systems on August 12, and attributed the attack to a highly professional actor. It is investigating with the Federal Office for Cybersecurity and the security firm Mandiant and has filed a criminal complaint; who was behind the attack is not yet known (swissinfo.ch).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Minerva Humanoids, a San Francisco startup, raised $10 million to build a humanoid robot for jobs too dangerous for people. Its robot, Roger, was built in five months and is driven by a specialist wearing a VR headset who sees through the robot's eyes, while onboard software handles balance, navigation and getting back up after a fall. Minerva is aiming it at bomb disposal, offshore oil rigs and high-voltage power plants; General Catalyst led the pre-seed round, announced on Tuesday. Sandor Felber, the 26-year-old chief executive, told Core Memory: "We actually take our mission so seriously that we name our conference rooms after disasters." One room is called Chernobyl (Minerva Humanoids · Core Memory).
░░▒▒▓▓ PAPERS ▓▓▒▒░░OpenAI — an unreleased model produced 722 math manuscripts on open problems, now public for anyone to check. OpenAI set its internal frontier model about 4,000 open research problems and on Tuesday published 722 of the resulting manuscripts, papers written up but not peer reviewed, grouped into 372 families in a public GitHub repository. Many come with proofs in Lean, a programming language whose proofs a computer can verify, though not all of them do; the topics run from number theory and combinatorics to theoretical computer science and mathematical physics. By OpenAI's own account each result used on average about three hours of ChatGPT Pro thinking, and it consulted an advisory group on mathematics and AI at the Institute for Advanced Study on how to release the work. It follows the Millennium Prize run we covered on September 18 (OpenAI · Interesting Engineering · Unite.AI).
░░▒▒▓▓ REPORTS ▓▓▒▒░░Stack Overflow — 2026 Developer Survey. More than 30,000 developers answered the annual survey, published on Tuesday. Among those who use AI, 66% use coding agents, and 73% of those who use AI coding assistants or agents do so daily; a third of daily users spend four hours or more a day with AI tools. 62% of all respondents view AI at least somewhat positively, rising to 71% among daily users (Stack Overflow).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░SKF, the Swedish bearings maker, cast an AI Greta Garbo, the 1930s Hollywood star who died in 1990, in an ad, with her family's approval. The digital Garbo, in slacks and a silk top, pitches SKF's frictionless magnetic bearings, The New York Times reported. Because of copyright, the makers built her from text prompts rather than archival footage, using image and video tools from ByteDance, Google and Kuaishou's Kling, and trained her voice on one of her early films, Reuters reported on October 2. Garbo's family and the company that licenses her likeness approved the project, the Times reported; viewers split between calling it a convincing replica and a ghoulish facsimile. Takeaway: an estate's sign-off settles the rights question; the audience still gets its own verdict (Reuters via KSL · The New York Times via Kingsport Times News).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░There is now a prediction market for things that have already happened. Verifact Markets, which WIRED wrote about on Monday, lets users bet true or false on disputed facts, such as how Jeffrey Epstein died or whether COVID-19 came from a lab, and post evidence for their side. If the company decides there is enough evidence, it makes the call; if it still can't tell, it closes the market and pays out on its own calculation of the odds. It is not yet licensed in the United States, but its backer, the tech investor Peter Wokwicz, says he has bought a regulated exchange and plans to move it there (WIRED · Verifact Markets).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░Bon vent — from the Uncertain Futures desk