Tech, in brief

Money into AI plumbing

Mistral raised 3bn euros at a 21bn value [12]; Amazon agreed to buy up to $60bn of Qualcomm server chips and took warrants for $4bn of its stock [17]

Europe's rules biting

Google stripped live hotel and flight prices from search in Europe to meet a 60-day deadline [6]; France's president asked the European Commission for a bloc-wide ban on social media for under-15s [28]

Nobody can tell when an agent gets it wrong

Meta's own staff reported Muse uploading private data and switching a job off without saying so [1]; monitoring tools log an agent's task as finished while the work is wrong [36]

Fights over who may learn from whose work

Three US security agencies named six Chinese labs for industrial-scale copying of American models [10]; OpenAI, Microsoft and the New York Times each asked a judge to settle whether training on copyrighted articles is legal [27]

3 pressures seen across the day run through Tech too.

• Investors are paying for things that have not been built yetHigh ↑ · seen in 3 beats→
• Courts are being asked to settle what talks could notBuilding ↑ · seen in 4 beats→
• Copying an AI system is being treated as theftBuilding ↑ · seen in 2 beats→

1. 1Meta gave its new assistant your email, your card and your cameras. Its own testers say it uploaded private data without asking.
2. 2Google strips live prices from European search
3. 3US agencies name six Chinese AI labs
4. 4Mistral raises 3bn euros led by Samsung
5. 5Amazon commits to Qualcomm server chips

Meta launched Muse on Tuesday, an AI assistant that acts on its own: it can send emails, book travel, fill in forms and buy things for you. [1][2]

Watch whether Meta ever publishes how often Muse acts without permission - until it does, the only public account of the product's mistakes is a set of internal staff posts a Reuters reporter was shown. [1]

An AI agent that does the wrong thing keeps going and reports the job finished, so the warning systems around it stay quiet.

a tester asked it to watch for tickets; it stopped refreshing after about 15 minutes, ignored the errors and reported nothing wrong

Moyai's founder says monitoring was built to catch error codes, so it logs a task as finished while the company takes the loss

AI writes about 42% of new code and it looks clean, so firms are putting a person or a second agent back in front of every risky change

the same missing check is being rebuilt one layer down, in an emulated computer, because nobody can list in advance what an agent should be allowed to do

The test is whether anything left can tell a finished job from a correct one.

Most of us have already dropped a checking step somewhere - a receipt we did not read, a form we did not reread - and we find out from the result, never from an alarm.

Google rebuilt how search works in Europe on Tuesday to meet an order from the European Commission, the European Union's executive body. [6]

In July the Commission fined it 460 million euros for pushing its own hotel, flight and shopping services above everyone else's, and gave it 60 days to stop or face repeat penalties of up to 5% of worldwide sales. [8]

Hotel and flight results now lose real-time prices, and a price-comparison site sits at the top instead. [7]

Google's search chief Nick Fox said the changes 'degrade the user experience for Europeans'. [6]

Google says the winners are booking sites such as Expedia and Booking.com, and the losers are the hotels and restaurants people used to reach directly.

It says an earlier round of the same rules cut direct booking traffic to European businesses by 30%. [6] Fines against Google under these rules now total about 890 million euros. [9]

The US National Security Agency, the FBI and the country's cyber-defence agency issued a joint statement on Tuesday accusing Chinese companies of 'aggressive, industrial-scale' copying of American AI models. [10]

They named DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.AI, and said the practice has run since at least 2024. [11]

The method is called distillation: you ask a big, expensive model millions of questions and train a small, cheap one on its answers. [10]

The statement gave no cases, dates or volumes. [10]

Distillation is how a rival gets close to the frontier without paying to build it, and the six named companies include the labs whose free models much of the world now downloads and runs. [11]

Also seen in Cyber.

Mistral, France's main AI company, raised 3 billion euros on Tuesday at a value above 21 billion euros, nearly double the 11.7 billion it was worth a year ago. [12]

Samsung, the South Korean electronics maker, led the round, alongside an EU-backed fund managed by EQT and the existing investor PSG Equity. [13]

Chief executive Arthur Mensch said the money goes into data centres Mistral owns rather than rents, and that its own computing power should roughly double over five years. [13]

The company is moving away from competing head-on with OpenAI and Anthropic. [15]

Mistral sells to European governments and companies that do not want their data inside American systems. The round makes a South Korean firm the largest new shareholder in that pitch. [14][15]

Also seen in Space and Finance.

Qualcomm, best known for the chips in most Android phones, said on Tuesday it will design custom chips for Amazon's AI data centres across several product generations. [18]

Amazon Web Services will buy up to $60 billion of Qualcomm server chips and related technology, and Qualcomm has issued Amazon warrants to buy up to $4 billion of its own stock. [17]

The work targets inference, which is the running of a trained AI model rather than the training of it. [18] Qualcomm shares rose about 10%. [16]

Amazon is the second large data-centre operator after Meta to commit to Qualcomm's server line, which the company has said should reach $15 billion of sales in 2029. [16]

Samsung and TSMC said on Tuesday they will adopt ASML's newest chip printers, joining Intel. [19]

ASML is the Dutch company that makes the only machines able to print the very finest circuit patterns onto silicon, and these cost up to $400 million each. [20]

Samsung plans to use them for memory chips from 2028 and TSMC for its most advanced chips from 2030. [22]

All three also backed a switch from today's 6-inch stencils to 6-by-12-inch ones, which ASML says could raise output about 40%. [20][21]

The stencil change only works if the whole industry moves together, which is why three rivals announced it on the same day.

ASML's chief executive Christophe Fouquet called the bigger masks the step that lets chips keep shrinking. [21]

Intel said on Monday it has run more than a million silicon discs through its High-NA machines, the newest and costliest kind of chip printer, in under two and a half years. [23]

In February 2025 the figure was about 30,000. [23] The tools are now in production for Intel's 18A process, which makes its Panther Lake laptop chips. [23]

For now Intel prints on 6-inch stencils, so large chips must be exposed in two pieces and stitched together. [23]

Barclays called adoption of these machines 'a key debate' before this week's announcements, and Intel's number is the volume evidence sitting behind it. [19][23]

ASML began construction on Tuesday of a production campus on about 35 hectares near Eindhoven Airport in the Netherlands, four miles from its headquarters. [24]

The site, called BIC North, will eventually hold up to 20,000 workers, with a first phase due in 2029 adding offices, logistics and a cleanroom, the dust-free space where the machines are assembled. [24]

ASML is now Europe's most valuable company, worth about $660 billion after a 62% rise in its shares this year. [24]

It runs the opposite way to the rest of European manufacturing: Volkswagen announced a restructuring this month that cuts 50,000 jobs.

ASML said in July that almost all its capacity to build these machines is booked through the end of 2027. [24]

OpenAI said on Tuesday that one of its unreleased models produced a proof of the Navier-Stokes problem, one of seven questions the Clay Institute offers $1 million each to solve. [26]

The equations describe how fluids move. [26] OpenAI says over a thousand copies of the model worked for more than 50 hours, at a computing cost 'in the millions of dollars'. [25]

Tristan Buckmaster, a mathematician at New York University, says OpenAI rushed after learning of his and Levent Alpoge's progress, then proposed a paper leaving Alpoge's name off. [25]

OpenAI denies seeing their work. [25]

Nobody outside either camp has yet checked the proof or the account of how it was reached, and the dispute is about who gets the credit when a machine does the reasoning. [25][26]

OpenAI, Microsoft, the New York Times and a group of authors including John Grisham and George R.R.

Martin all asked a US judge on Friday to rule in their favour on 'fair use', the rule that allows copying a work without permission in limited cases. [27]

The Times sued in 2023 over the use of millions of its articles to train ChatGPT, and the authors sued the same year over their books; the cases were joined in New York last year. [27]

Judge Sidney Stein in Manhattan will decide. [27]

Fair use is likely to settle the whole case, and dozens of similar complaints against AI companies are waiting behind it. [27]

Also seen in World, Sports and Finance.

President Emmanuel Macron of France wrote to Ursula von der Leyen, who heads the European Commission, asking for an EU-wide ban on social media for children under 15. [28]

The letter is dated 29 August and was seen by Reuters on Tuesday. [28]

France tried its own version and failed: in August the country's Constitutional Council struck the bill down, ruling that it cut too far into freedom of expression. [28]

Macron argues a bloc-wide law is now urgent. [28]

A French court closed the national route, so the decision moves to the European Commission, which would have to write a law covering all 27 member countries. [28]

Australia's government opened consultation on Tuesday on a 'digital duty of care' law for internet companies. [29]

One part, called 'my feed, my way', would require social media platforms to let people restrict their feed to accounts they chose to follow, rather than whatever a recommendation system picks for them. [29]

It follows a ban on social media for under-16s that took effect in December. [29]

Prime Minister Anthony Albanese said the bill is 'about putting choice back into the hands of Australians online', not censorship. [29]

This is a different move from a ban: instead of keeping people off a platform, it limits what the platform is allowed to decide on their behalf. [29]

Google set up a joint unit with the consulting firm Accenture, called the Accenture Gemini Enterprise Business Group, to install Google's AI models inside other companies' systems. [30]

It joins OpenAI, Anthropic, Microsoft and Amazon, which have all built similar arms of engineers who work on a customer's site. [30]

Google Cloud made $24.8 billion in the second quarter, while its parent Alphabet had run up $811 billion in purchase commitments and contractual obligations as of 30 June. [30]

The cloud companies are spending hundreds of billions a year on chips, buildings and power while the revenue clearly traceable to AI is a fraction of that, so selling the installation work is one way to close the gap. [30]

Arm, the British company whose designs sit inside almost every smartphone chip, announced its next phone platform on Tuesday and built it around agent software. [31]

Chris Bergey of Arm said an agent means repeated inference, application logic and system orchestration all running at once, so a phone needs several fast cores in parallel rather than one faster core. [31]

The top C2-Ultra core is claimed to be up to 15% quicker on single-thread work than last year's. [31] Chipmakers that license the design can mix cores as they choose. [31]

Arm's designs sit in almost every smartphone chip, so this is where the assumption that agents will run on handsets gets built into hardware rather than argued about. [31]

Nvidia said its RTX Spark chip, now officially called the N1X, will reach laptops in October in two versions. [32]

The larger pairs a 20-core Grace processor with a 6,144-core Blackwell graphics part and 24 to 128 gigabytes of shared memory; the smaller uses 18 cores and up to 32 gigabytes. [32]

The first machine shown, at the IFA trade fair in Berlin, is Lenovo's 16-inch Yoga 9n, at 0.69 inches thick. [33]

Nvidia is aiming at people who want to run agent software on their own computer. [33]

This is Nvidia aiming at the laptop a person owns rather than the data centre they rent, and Wired says the first machine shown is built to take on Apple's MacBook Pro directly. [33]

Intel is set to raise processor prices by about 10%, DigiTimes reported, citing supply-chain sources. [34]

It would be the third rise this year, after one in the first quarter and another in July, and the sources did not say which products are covered. [34]

The same report says Intel will launch its main annual products in March 2027, with AMD following between June and July. [34]

Intel is pushing for higher margins as its costs rise. [34]

Intel is pushing for higher margins and the report names no products, so buyers of laptops, desktops and servers will not know which machines carry the increase until prices move. [34]

Software firms are redesigning how code gets checked now that much of it is written by AI, IEEE Spectrum reported. [35]

In a survey of more than 1,100 developers by Sonar, respondents said AI wrote about 42% of the code they added to shared projects, while 96% did not fully trust that it was correct. [35]

New methods include reviewing the plan before the agent writes anything, using separate agents to catch routine flaws, and making developers defend code their agent produced. [35]

CodeRabbit, which sells automated review, raised $143 million in August at a $1.5 billion value. [35]

AI-written code often looks clean while hiding faulty assumptions or security holes, so the time saved writing it can be spent again finding them. [35]

Monitoring tools were built for software that either works or throws an error, and agents break that assumption, argues Robert Hommes, founder of the startup Moyai. [36]

An agent can finish a task and produce the wrong outcome, so the monitoring logs a success while the business takes the loss. [36]

Hommes wants anomaly-first detection: find what is different from normal first, then work out whether it was wrong, instead of writing a new rule after every new kind of failure. [36]

This is the gap Meta's own testers described from the inside, where a job stopped and reported nothing, and it is now being sold as a product category. [1][36]

Microsoft is pushing Execution Containers, a way of running an AI agent inside a small emulated computer so it cannot read private files or call services it was not meant to touch. [37]

The difficulty is that agents are useful precisely because their steps are not predictable, which is the same reason nobody can write the list of what they are allowed to do in advance. [37]

The approach leans on hardware virtualisation support that PCs did not have when software agents were first tried in the 1990s. [37]

The checking step is being rebuilt one layer down, in the machine rather than in the code, because nobody can write rules for behaviour they cannot predict. [37]

Brian Kelly says his fund, Bracket22, now runs entirely on AI agents. [38]

He previously had seven or eight employees around the world, and put total labour costs, including salaries, healthcare, office space and bonuses, at roughly $5 million a year. [38]

He says he now spends between $30,000 and $40,000 a year on everything, agents and computing included. [38]

Bigger firms are moving more slowly: JPMorgan Chase's chief executive Jamie Dimon said in February that the bank had large redeployment plans for staff. [38]

It is one small fund and one owner's account, but it is the clearest public number yet on what taking the people out is worth to whoever owns the firm. [38]

Anthropic, Meta, Google and OpenAI all shipped model updates inside one week, and the people who must choose between them are struggling. [39]

OpenAI's chief executive Sam Altman told CNBC that 'we're all moving to faster cadences', partly because everyone is back from the summer break. [39]

Zhen Lu, who runs the AI startup Runpod, said 'model fatigue is a real thing' and that so much is shipping at once that a lab has to make noise to be noticed. [39]

Company technology managers now spend heavily on comparing costs and capabilities. [39]

Every release resets the comparison, so the cost of working out which model to use is paid by the customer rather than by the lab that shipped it. [39]

Ramp, which sells software that helps companies track and cut their spending, is in talks to raise about $1 billion at a $60 billion value, Bloomberg reported. [40]

The people describing the talks asked not to be named because the details are not public, and said the terms could still change. [40] Ramp declined to comment. [40]

Every figure here comes from people who asked not to be named, and Ramp itself declined to comment, so this is a reported valuation and not a closed round. [40]

Forus, a New York company that uses AI agents to handle the paperwork between a doctor's prescription and the first dose, raised a round valuing it at $3 billion. [41]

Bain Capital Ventures led, with Thrive Capital, General Catalyst and Accel taking part. [41] The steps it automates are insurance approvals, financial-assistance applications and medicine supply. [41]

Forus, formerly called Tandem, says providers in all 50 US states use it, that its patients live in 85% of residential postcodes, and that it works with nine of the world's fifteen largest drug companies. [41]

The wait it is being paid to remove is the one between a doctor writing a prescription and a patient starting on the medicine, and Forus says it already runs in all 50 US states. [41]

Cato, a Milan startup less than a year old, raised 6 million euros to sell AI to companies bidding for Italian public contracts. [42]

Keen Venture Partners, based in Amsterdam and London, led the round, taking Cato's total funding to 7.6 million euros. [42]

Italy's public tender market is worth 309.7 billion euros a year by the count of ANAC, the national anti-corruption authority. [42]

Keen's Briehan Burke said public procurement is one of the last giant unautomated workflows in the European economy. [42]

Most software is sold on the promise of making a hard job easy.

Cato's backers are betting the opposite, that Italian tender paperwork stays difficult and companies will keep paying for help with it. [42]

The European Union's coming Quantum Act will have no budget attached. [43]

Thomas Skordas, a deputy director-general at the European Commission, said so on 3 September at an event in the European Parliament. [43] 'The act is not about budget,' he said. [43]

It is meant to bring an industry together rather than impose rules, unlike the AI Act and the Digital Markets Act. [43]

The Commission hopes to present it by the end of 2026, though that could slip. [43] Skordas also said in public: 'We cannot afford having 78 startups all on their own.' [43]

The EU says it wants to be the world's quantum centre, and one of its own officials has now said openly that he expects some of those 78 companies to merge or disappear. [43]

The European Commission published revised guidance on 3 September saying a company at risk of punishment for dominating a market may be able to justify its position if its business practices are sustainable. [44]

The Commission treats a firm holding more than 40% of a market as dominant. [44]

The change sits inside a wider effort to loosen rules that European industry says leave it behind US and Chinese rivals. [44]

It creates a defence that did not exist before, in the rules the European Commission uses against dominant firms.

The same regulator fined Google 460 million euros in July under a separate law. [44][6]

Two startups, Far Labs in Abu Dhabi and Evolving Edge in Austin, are building marketplaces that pay people for the downtime on their gaming computers and use it to run AI models. [45]

Far Labs says it will launch within weeks with delays of 100 milliseconds or less, and Evolving Edge is in open beta. [45]

Both target smaller open models rather than the largest ones. [45] They join Salad, a Utah platform listing more than 60,000 consumer graphics cards active each day. [45]

The pay is close to minimum wage. [45]

It is the opposite bet to the data-centre buildout running through the rest of today's news: instead of putting up new buildings, use the machines people already own. [45]

pnpm, one of the tools JavaScript developers use to install the code libraries their programs depend on, released version 12 with its core rewritten in Rust, a language that runs closer to the machine than JavaScript does. [46]

In the project's own tests it installed a sample package in 5.19 seconds, against 8.22 for the previous version and 47.7 for npm, the standard tool. [46]

The team also wrote a server in Rust, called pnpr, which cut the same install to 3.37 seconds. [46] Commands, settings and the lockfile all carry over unchanged. [46]

Lead maintainer Zoltan Kochan's rule was that 'upgrading should not feel like a migration', which is why a rewrite this large can land without anyone having to change their project. [46]