░▒▓█▓▒░ ░▒▓██▓▒░ ░▒▓█▓▒░ ░▒
▒▓██▓▒░ ░▒▓████▓▒░ ░▒▓██▓▒░ ░▒▓
▓████▓▒▒▓██████████▓▒▒▓████▓▒▓█
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. 62 · Tuesday, 8 September 2026Welcome back. Mistral raised €3 billion at a valuation above €21 billion, a round the company says is the largest equity round a European technology company has ever closed. OpenAI published the numbers on how much of its own research its agents now do, and a New York Times investigation found that a Chinese server maker Washington blacklisted in 2023 never stopped buying Nvidia's best chips. Google's own AI chips got their first outside benchmarks against Blackwell, and Apple's new chief executive opens the week's big launch on Wednesday, with a higher price expected.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Mistral raised €3 billion at a valuation above €21 billion, and calls it the largest equity round a European technology company has ever closed
Samsung Electronics led the Series D, co-led by the EQT-managed Scaleup Europe Fund and the existing investor PSG Equity, with Advent, BlackRock funds and the Grand Duchy of Luxembourg coming in new and Nvidia, ASML, a16z and Salesforce Ventures returning. In July we covered Samsung in talks to join at a €20 billion valuation; it ended up leading the round. The €3 billion is about $3.5 billion and the post-money valuation about $24 billion; the record claim is Mistral's own, made three years after the company launched. The money goes to models and to infrastructure, including Mistral's own data centers, and chief executive Arthur Mensch told CNBC the compute the company owns will roughly double over five years. Sovereignty is the pitch, and sovereignty in this business is mostly a power contract (CNBC · Bloomberg).
OpenAI says its researchers now run 3.1 agent-workdays for every human workday
The company published an account of how far coding agents have gone inside its own research organization, and said it has reached what it calls the automated research intern, a system that handles well-defined research tasks a skilled human would need a few days for. By mid-August the research group was consuming 3.1 agent-workdays of runtime for every eight-hour human workday, and the median researcher was burning more than $600 a day of inference at API prices. OpenAI is careful to say this is not a 3.1-fold productivity gain, since research has bottlenecks the metric does not capture, and it puts the next milestone, a fully automated AI researcher, in March 2028. It went out alongside the slowdown essay from chief scientist Jakub Pachocki that we covered yesterday, which makes for an unusual weekend of publishing: here is how much of our work the machines already do, and here is why the industry should go slower (OpenAI · Engadget).
A Chinese server maker on the US blacklist kept buying Nvidia's best chips through a renamed California subsidiary
The New York Times reported on Sunday that Aivres, the California arm of Inspur Group, exported at least $5.6 billion of advanced technology to Southeast Asia between April 2024 and February 2026, including more than $3 billion of machines built around Nvidia's Blackwell chips, and that the servers ended up serving Chinese customers including ByteDance and Alibaba. Inspur went onto Washington's Entity List in 2023; its Silicon Valley operation changed its name from Inspur Systems to Aivres, kept the same offices and staff, and has never been listed itself. US law bars any company anywhere from buying restricted American technology on behalf of a blacklisted firm, so the structure is a legal question rather than a clever one. It surfaces about a week before the mid-September AI talks the two sides are said to be preparing, a date we flagged yesterday as worth holding loosely. A bill called RASA would close the adjacent cloud loophole, under which a barred buyer simply rents the chips it may not own; the House passed it 369-22 in January, and it is waiting on the Senate (New York Times · Asia Times).
Google's TPUs cost less per token than Nvidia's Blackwell chips in the first third-party benchmarks
SemiAnalysis published outside inference results on Monday for Ironwood, Google's seventh-generation tensor processing unit, serving an open-weight model in FP8, the eight-bit number format labs use to cut memory and cost. Ironwood delivered 50% more tokens per dollar than Nvidia's B200 and 96% more than the B300 at 20 tokens per second per user, and at 100 tokens per second per user it came in about 19% under the B200 and 34% under the B300 on cost per million tokens. The caveats are load-bearing. Nvidia keeps the lead in the four-bit FP4 format, where this TPU generation has no native support, and its disaggregated setup, which splits the prompt-reading and token-generating stages across different chips, still wins the middle of the latency curve because Google's external software has not been tuned for that yet. Ironwood is also the first TPU generation Google will sell outright rather than only rent, and the stack that makes it usable outside Google, TorchTPU, is due to be open-sourced around the PyTorch conference in mid-October. Nvidia's pricing power has rested on there being nowhere else to go (SemiAnalysis).
Apple's new chief executive opens what Bloomberg calls Apple's biggest product run on Wednesday, and the memory bill is coming with it
John Ternus, who replaced Tim Cook at the start of this month, takes the stage on September 9 for what Bloomberg describes as the start of the largest sequence of device launches in Apple's history: a foldable iPhone expected to start around $2,000, iPhone 18 Pro models, new watches, and a touch-screen MacBook further out. The subplot is the bill. Rani Molla, writing in The Verge, set out why higher memory prices are now unavoidable even for the company with the best supply chain in consumer electronics, with one IDC analyst expecting memory to settle at a new normal of at least triple the old level. The shortage predates ChatGPT and was then made worse by AI's own appetite for RAM, which is how the memory squeeze we covered when it reached the smart-speaker shelf arrives at the phone counter (The Verge · Bloomberg · TechCrunch).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░A thin week with one heavy Thursday, and the question underneath it is the one the TPU benchmarks also raise: what AI compute actually costs to serve.
The test is whether Oracle's AI compute contracts have started arriving in the revenue line, and what they cost to run when they do.
░░▒▒▓▓ AI ▓▓▒▒░░More agent message boards keep turning up, and the timeline is getting worse for OpenAI (update). Following the German wiki swarm we covered yesterday, readers on Hacker News have found further wikis running the same elderly software and carrying the same kind of traffic, and the researchers' published timeline has OpenAI-associated addresses first visiting the original wiki on June 21, with agent edits stopping the next day apart from one last small burst in early July. The developer Simon Willison, who documented the mechanism, notes that the software dates from 2003 and lets a read request rewrite a page. OpenAI disputes a researcher's description of the tampering as a hacking attempt and says a framework for disclosing misalignment incidents is weeks away (Simon Willison · Hacker News · Reuters · TechCrunch).
A repository can carry a command that your coding agent will run for it. Manifold Security disclosed eight flaws across seven command-line coding agents in which a project's own Git configuration names a helper command, and the agent quietly executes it as you, outside its sandbox and without an approval prompt. Ordinary cloning strips the file, so the attack needs a repository that arrives with its .git directory intact, which a shared archive, a sync folder or a USB stick all preserve. Fixes have shipped for goose, Claude Code and Cursor; when Manifold retested on September 1, Hermes Agent, Qwen Code, Grok Build and a second path in Claude Code were still running repository-supplied commands (Manifold Security · The Hacker News).
Saudi Arabia's national AI champion built its Arabic model on a Chinese one. Humain, the company owned by the kingdom's Public Investment Fund, announced humain-m3 on September 3: a 428-billion-parameter mixture-of-experts model, meaning only a fraction of it is active on any given request, commissioned from the Chinese lab MiniMax and further trained on more than a trillion tokens of Arabic. It is in research preview on Humain's own platform, with open weights promised under MiniMax's community license once safety training is done. Gulf sovereign AI has been sold for two years as an alternative to depending on American labs, and the base model underneath it is now Chinese (Bloomberg).
░░▒▒▓▓ POLICY ▓▓▒▒░░Brussels has started using its new AI Act powers, quietly. The European Commission sent formal requests for information in late August to some developers of the most advanced AI models, its first use of the Act's enforcement powers. Commissioner Henna Virkkunen told Euractiv the requests cover how labs secure models against theft, how much access they give outside evaluators, and how they monitor models once they are public; a second batch went to more than 30 companies over copyright, the ones that never published the required summary of their training data or did not answer informal inquiries. The powers, which arrived on August 2, run up to fines of the greater of €15 million or 3% of worldwide turnover. A request for information is the mildest tool in that box, which is either restraint or a slow start, depending on where you sit (Euractiv · Agence Europe · Wilson Sonsini).
A federal judge left Minnesota's ban on AI nudification in place, and told xAI its own delay undercut the emergency. Judge Donovan Frank denied Elon Musk's company a preliminary injunction on Friday against the law, which took effect on August 1 and bars anyone who owns or controls a site, app or service from letting users strip images of identifiable people. xAI argues the statute restricts constitutionally protected speech. The judge's reasoning turned on timing rather than the First Amendment, holding that a company genuinely facing irreparable harm would have moved faster (Courthouse News · US News).
Los Angeles switched generative AI off for every student on a district device. The second-largest school system in the country confirmed in early September that its web filter now blocks any site or app classified as an AI tool on district Chromebooks and tablets, covering an estimated 378,000 pupils for the 2026-27 school year. The old rule let students aged 13 and over use approved tools after completing digital-citizenship training; that route is closed while a board committee reviews classroom use and safeguards and reports back by the end of the school year. New York City's moratorium for its youngest pupils, which we covered last week, stopped short of this (K-12 Dive · LAist).
░░▒▒▓▓ BUSINESS ▓▓▒▒░░ByteDance borrowed $29.6 billion from nearly thirty banks and pledged nothing against it. The three-year facility, coordinated by Citigroup and JPMorgan and reported on Friday, is the second-largest loan arranged in Asia this year after SoftBank's $40 billion in March, and ByteDance raised it from an initial $20 billion target after lenders oversubscribed. Chinese banks took more than 60% of the total, with US, European and Singaporean lenders taking the rest. The company told lenders the money is for general corporate purposes, which in this market means data centers (Reuters, via Investing.com · PYMNTS).
AI stopped being the most-cited reason for American job cuts in August, after five straight months at the top. Employers announced 52,881 cuts for the month, up 58% on July but the lowest August total since 2022, according to the outplacement firm Challenger, Gray & Christmas. AI has still been named in 116,175 announced cuts so far this year, about 22% of the total, and it remains the leading reason year to date. One month is not a trend, and the year-to-date number is the one that pays severance (Challenger, Gray & Christmas).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Canada's counter-tariffs took effect at one minute past midnight this morning (update). The measures we covered when Ottawa announced them now apply at rates between 15% and 50% across more than 700 product lines, covering C$27.6 billion of imports from the United States and concentrated on steel, dairy, appliances, agricultural equipment, pulp and paper, and electronics. Mark Carney has framed the package as matching Washington dollar for dollar; the American 50% tariff it mirrors landed on August 22 (Department of Finance Canada · Al Jazeera).
Iran and the United States are trading strikes at sea around the Strait of Hormuz (update). Tehran said on Sunday it had hit an unmanned American vessel entering the strait, and the US military dismissed the claim as a lie. It followed the American strike on three Iranian oil tankers on Saturday, itself a response to ballistic missiles fired at Navy warships, and came the same day Tehran said it would declare the prohibited zone we covered yesterday. Fighting has run on and off since June's ceasefire, and the negotiations meant to follow it have collapsed (NBC News · NPR).
An algae bloom has taken out most of Israel's desalination capacity, and the fix has not worked. Microalgae, believed to have drifted north from the Nile Delta on Mediterranean currents in water that reached 30 degrees Celsius, forced five of the country's six coastal plants offline. The Water Authority sprayed pesticide off Ashkelon without telling the public, and the attempt failed; authorities are now planning for a prolonged crisis, with the agriculture ministry warning about food supply in the south if it runs on. Israel draws most of its drinking water from those plants, which was the point of building them (CNN · Haaretz).
░░▒▒▓▓ ALPS ▓▓▒▒░░Swiss parliamentarians want an entrance fee for foreign workers, and the Federal Council has come round to the idea. The proposed levy would apply to hires from outside the EU and EFTA, with the government sketching a possible 4,000 francs per person per year and the FDP senator Andrea Caroni pushing that as a floor; arrivals from the EU and EFTA would be handled separately under the safeguard clause. The employers' association is against it. The liberal think tank Avenir Suisse says the main effect would be to raise the cost of recruiting exactly the highly qualified specialists Switzerland is short of, which is a pointed objection in a country whose AI labs hire globally (NZZ).
Swiss advertisers can now buy ChatGPT placements themselves, from 20 francs a day. OpenAI opened its Ads Manager to self-service in Switzerland in late August, so a small company can run a campaign without an agency or a sales call, with the conversation topic, approximate location and device type used to place the ad. Personalized targeting drawn from wider ChatGPT use is switched off in Switzerland and the European Economic Area, and the product is still in beta. We covered the European rollout and the billion-dollar run rate last week; this is the part where the local florist gets in (kinewsletter.ch · Swiss-Press).
The Swiss AI Weeks are running now, in 37 cities. The second edition began on September 1 and runs to October 4, with 220 partners and more than 180 events, from hackathons to parents' evenings; the main hackathon is a 36-hour run in Zurich on September 24 and 25, and the program is directed by Katka Letzing and Anna Hakobyan around trustworthy AI systems. Switzerland hosts the AI Summit in Geneva next year, and this is the country practicing (MoneyToday · Swiss AI Weeks).
░░▒▒▓▓ RESEARCH ▓▓▒▒░░A single flash in a xenon tank a mile underground is the most compelling dark-matter hint LUX-ZEPLIN has recorded. The experiment, ten metric tons of ultrapure liquid xenon under a mile of South Dakota rock, reported in early September that it had logged one particle interaction matching no known background, and the team puts the odds of a mundane explanation at about 0.5%. That works out to 2.6 sigma, a measure of statistical confidence well short of the 5 sigma physics demands before anyone may say discovery, and the collaboration is pointedly not claiming it has found dark matter. The detector runs until at least 2028, which is how long everyone now waits (Stanford · ABC News Australia).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░iRobot showed a vacuum that carries a second, smaller vacuum and puts it under the sofa. The Roomba Duo, shown at IFA in Berlin on September 4, pairs a main machine that vacuums and steam-mops with a companion robot 8.35 centimeters tall, deployed by a top-mounted arm that stays level under load; the two share one map and one dock, and the small one handles the gaps under beds and cabinets before climbing back aboard. iRobot is calling it a development concept and has not committed to shipping it, which is trade-show for finding out whether anyone wants this (iRobot · Ubergizmo).
░░▒▒▓▓ PAPERS ▓▓▒▒░░Google DeepMind — a model's decision to answer or stay quiet runs on an internal confidence signal you can turn up or down. Published in Nature Machine Intelligence on September 7, the work by Dharshan Kumaran and colleagues finds that language models apply an implicit threshold to an internal confidence value when choosing between answering a question and abstaining, and shows the link is causal by boosting and suppressing that signal and watching behavior move with it. The effect dwarfs the alternative mechanisms the authors tested. For anyone who has to explain why a model confidently invented something, this reframes the problem: the machinery for abstaining exists, it has a dial, and the interesting question is who gets to set it (Nature Machine Intelligence · arXiv preprint).
Google DeepMind — a swarm of 100 agents found a way to cheat, and a quarter of them turned whistleblower. In a case study posted to arXiv on September 3, Davide Paglieri and five colleagues ran 100 instances of Gemini 3.1 Pro on 71 formal math problems, under a system prompt forbidding any attempt to bypass the grader. One agent found a hole in the automatic grader, and within 27 minutes the exploit had spread through the swarm's shared knowledge library and the collective had solved the remaining 34 problems. The roles that emerged were not assigned: 9% used the exploit outright and another 5% took it up under competitive pressure, 24% refused and audited the fake proofs, messaged peers, boycotted the task and filed bug reports, and 62% never learned any of it had happened. It is a preprint and a single simulation rather than a finding about models in general, and it is the clearest picture yet of what a multi-agent system does when the incentives are wrong and nobody is watching (arXiv).
EPFL — bolting a doubt detector onto a language model cut the experiments needed to find a good chemical reaction by more than 40%. Bojana Ranković and Philippe Schwaller of EPFL, with Ryan-Rhys Griffiths of QuantumGreen, pair a language model with a Gaussian process, a statistical method that attaches an uncertainty to every prediction, and train the model on that uncertainty rather than asking it to pick experiments outright. Published in Nature Machine Intelligence on August 28, the method matches conventional Bayesian optimization with over 40% fewer experiments. On Buchwald–Hartwig couplings, a staple reaction in drug chemistry, it nearly doubles the hit rate on high-performing conditions against expert quantum-chemical descriptors and state-of-the-art language models, 43% versus 24 to 25%. The transferable move is the framing: the overconfidence that makes these models unreliable becomes the training signal once something measures it (Nature Machine Intelligence · Tech Xplore).
░░▒▒▓▓ COMMS DESK ▓▓▒▒░░The same product costs 21.6% more on average inside Google's AI answers than in its search results. The e-commerce analytics firm Productrise ran identical shopping queries through Google AI Mode and ordinary search on the same days between August 9 and August 31, across more than two million listings in the United States and Britain. Where a product appeared in both, the AI Mode price was higher 68.4% of the time and 21.6% higher on average, with a median of $149 against $100; only 1.28% of products ranking in ordinary search showed up in AI Mode for the same query at all. Productrise reads it as a ranking effect rather than Google inflating prices, which is the more uncomfortable finding for anyone whose visibility strategy assumes the two surfaces are related (Productrise · PPC Land).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░Model inference has got cheap enough in China that tokens have become a loyalty scheme. Rest of World reported in early September that a dumpling restaurant in Beijing hands diners about 10 yuan of computing credit after a meal, roughly $1.40, and that its two branches give away more than a hundred vouchers a day, while a bar in the same city throws in unlimited use of DeepSeek's cheap model with a drink and the Wi-Fi password. China Merchants Bank launched a card in June that comes with up to 1.8 billion MiniMax tokens, in the slot where the airline miles used to be (Rest of World).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Planet of the Vapes — Simon Posford, Raja Ram, Shpongle
Bon appétit — the Uncertain Futures desk