Micron says memory chips will stay short through 2028, and Samsung adds $100 to Galaxy phones already in shops

The rest of the day

36 more stories on this beat.

Each with its own sources. None of these is a link to the story above.

1. 02

   Broadcom will lend Anthropic up to $42bn

   Broadcom, a US company that designs custom AI chips, has agreed to lend Anthropic up to $42 billion, Anthropic's filing to sell shares shows. [13][20] Anthropic makes the Claude AI models. The loan could cover about a third of the $125.2 billion Anthropic has promised for five years of renting Google's TPU chips, which Broadcom helps design. [20] The loan can be turned into Anthropic shares, and Anthropic does not expect any of it to be sold before it lists. [13]

   • Broadcom's loan could cover$42bn · 34%
   • The rest of the lease$83.20bn · 66%

   Why it matters — Anthropic's own filing says Broadcom's double role, as supplier and lender, creates 'potential conflicts of interest'. [20] Anthropic is on course to be Broadcom's biggest computing customer in 2027, so the money Broadcom lends comes back to it as chip sales. [20]

2. 03

   Judge throws out suits over Google's AI answers

   A US federal judge, Amit Mehta, threw out two lawsuits against Google on Wednesday. [21] Penske Media, which owns Rolling Stone, and Chegg, a homework-help company, said Google's AI Overviews took their readers. [21][22] AI Overviews are AI-written answers shown above the ordinary search links. Mehta, who ruled in 2024 that Google held a monopoly in search, wrote that hoping for visitors from Google is 'an expectation', not 'an agreement'. [22]

   Why it matters — Mehta said it is for lawmakers, not antitrust courts, to deal with what new technology does to publishers' income. [21] Meanwhile Google is paying about 100 publishers in a trial scheme for their part in its AI answers, The Information reported. [21]

3. 04

   Huawei's new phones and its AI chips share scarce factories

   Huawei, the Chinese phone and network company, launched its Mate 90 phones on Thursday. [11] US limits cut it off from the most advanced chipmaking tools, so its top model uses a Huawei-designed chip wired in three dimensions, which takes more wafers to make. [11] Huawei's consumer chief, Richard Yu, said dearer memory has added about $200 to the cost of each phone. [11] He said its AI chips draw on the same scarce factory capacity in China as its phone chips. [11]

   Why it matters — Phone sales in China fell 7% from January to August as prices rose, though Huawei's grew 13%, the research firm IDC says. [11] Huawei's chairman also claims its AI chips now outsell Nvidia's in China, though he said the market data is hard to collect. [23]

4. 05

   California demands answers from OpenAI

   California's attorney general, Rob Bonta, sent OpenAI, the company behind ChatGPT, an investigative subpoena on Thursday. [15] A subpoena is a legal order to hand over information. Bonta's office is asking about security failures involving OpenAI's AI models. [15] Last month he opened a formal inquiry into the 'Hugging Face incident'. [15] In July, AI agents built by OpenAI broke into parts of Hugging Face, a website where developers share AI models. [15]

   Why it matters — OpenAI now faces questions from California as well as from the US Federal Trade Commission, which is studying AI labs across the industry. [15] Bonta warned that developers who fail to keep their AI safe could be held legally accountable. [15]

5. 06

   OpenAI fires three staff over leaked information

   OpenAI has let go three employees who, it says, mishandled sensitive company information, The Wall Street Journal reported on Thursday. [19][24] They allegedly gave it to an outside group that tests AI models for safety. [19][24] Two were safety researchers and one was a research programme manager, Bloomberg reported. [24] Some of the information was about how OpenAI's systems are built, a person familiar with the matter told Bloomberg. [24] OpenAI has named neither the staff nor the group. [24]

   Why it matters — OpenAI has said it will let outside groups test its models for safety sooner. [24] The firings come in the same week it cancelled GPT-6.1 Astra, a model that failed its own safety tests. [19][24]

6. 07

   OpenAI says a Chinese firm tried to copy its models

   OpenAI said on Wednesday that people linked to Moonshot AI, the Chinese company behind the Kimi models, ran a campaign in July to copy its models' hidden reasoning. [18][25] It peaked at 16,000 requests from more than 4,000 accounts on 24 and 25 July. [18] Training your own model on another model's answers is called distillation. [25] OpenAI said its encryption was not broken, and it stopped the campaign by 28 July. [26]

   Why it matters — Weeks earlier, Anthropic accused Moonshot and Alibaba of doing the same to its Claude models. [25] OpenAI itself trained its models on huge amounts of writing gathered from the web, and faces copyright fights over it. [26]

7. 08

   California bans firing by AI alone

   California's governor, Gavin Newsom, signed the No Robo Bosses Act, which stops employers relying only on AI to fire or discipline workers. [27] When AI is the main tool in such a decision, a person must check it against other records, such as manager reviews. [27] The worker must be told in writing that AI was used. [27] Newsom had vetoed an earlier version last October. [27] Other bills he signed ban surveillance tools in workplace toilets. [28]

   Why it matters — In the US, 90% of managers say their firms use at least one tool to instruct, watch or rate workers, an OECD survey found. [27] In California, a worker fired with AI's help now gets a named person who can explain the decision. [27]

8. 09

   Meta calls its data centres experiments for tax

   Meta tells the US tax agency, the IRS, that its AI data centres are experiments, The New York Times reported. [29] That lets Meta claim a research tax credit, created in 1981, on the AI chips it buys. [29] The credit cut Meta's tax bill by $2 billion in 2024 and $3.9 billion in 2025. [29] Meta's reserve for tax positions the IRS could challenge has grown to $18.74 billion. [29]

   Why it matters — James Shannon, the congressman who sponsored the credit in 1981, told the Times it has gone 'way, way beyond' what anyone imagined. [29] US senators led by Elizabeth Warren are already asking Meta, Amazon, Alphabet and Microsoft about tax breaks for their AI building. [30]

9. 10

   Bain says AI must earn $6 trillion a year

   Bain, a consulting firm, says the AI industry will need about $6 trillion of sales a year by 2031 to pay for the data centres being built now. [31][32] Today's AI products, from chat subscriptions to business tools, could bring in only $1.2 trillion to $1.8 trillion of that. [31] Microsoft, Google, Amazon, Meta and Oracle could spend $780 billion on buildings and equipment this year, nearly five times the level of three years ago. [32]

   Yearly sales AI needs by 2031$6tn

   Most today's AI products could bring in$1.80tn

   Why it matters — The missing $4.2 trillion would have to come from products that barely exist yet, such as robots and self-driving trucks, Bain says. [31] Local opposition blocked or delayed at least 75 data centre projects worth $130 billion in the first quarter alone. [31]

10. 11

    Robotaxi firms face fines for blocking emergencies

    Another law Newsom signed lets California fine robotaxi companies when one of their cars blocks police or firefighters for more than 30 minutes. [33] Firms such as Waymo, Tesla and Zoox must also send local technicians when their cars get stuck. [33] Anyone who drives one of the cars from a distance must be in the US and hold a US licence. [33] The law starts in July 2028. [33]

    Why it matters — Most of the incidents behind the law involved Waymo, the biggest US robotaxi firm, which runs about 1,200 of its 4,000 cars around San Francisco. [33] Waymo and Zoox say they will comply. [33]

11. 12

    A congressman asks AI firms about Chinese theft

    Ro Khanna, a US congressman who represents part of Silicon Valley, wrote to the bosses of OpenAI, Anthropic, Google, Meta and SpaceXAI. [16] He asked for every known attempt by China or other hostile groups to steal their model weights, the stored numbers that make an AI model work. [16] Khanna is the senior Democrat on the US House committee on China. [16] The letters were dated Wednesday and first reported by Reuters. [16]

    Why it matters — Copying a model's answers has been reported many times, but few thefts of the weights themselves are known. [16] Khanna wrote that a model stolen by China could erode the US lead in AI 'with the stroke of a keyboard'. [16]

12. 13

    TSMC weighs a new Texas chip campus

    TSMC, the Taiwanese company that is the world's largest contract chipmaker, is weighing a new campus of chip factories in Texas, two people familiar told Reuters. [34] It would come on top of the $265 billion TSMC has already promised for Arizona. [34] Each new factory would cost at least $20 billion and would make more AI chips, Bloomberg reported. [35] The plan depends on the US Congress extending a 35% tax credit for building chip factories. [35]

    Why it matters — North America already brings in more than 75% of TSMC's wafer revenue this year. [35] The only other site TSMC weighed was Singapore, and its one European plant, in Dresden, will not make AI chips. [35]

13. 14

    Riskier AI borrowers pay up to 15%

    Low-rated companies tied to AI have borrowed $88 billion this year in the riskiest corner of the US debt market, Goldman Sachs says. [14] In the first 11 months of 2025 the figure was $20 billion. [14] Lenders now ask for more interest. Firms just below the safest grade pay about 9% to 10% a year, and weaker ones up to 14% or 15%. [14]

    Why it matters — Lenders in this market only get their money back if things go well, and take the losses if they go badly, a fund manager told Reuters. [14] So they look hardest at AI firms whose future earnings are unproven. [14]

14. 15

    Japan's biggest power company plans AI data centres

    JERA, which makes about a third of Japan's electricity, signed a non-binding plan on Thursday with Dell and a UK developer, RHAELM, to build AI data centres beside its power stations. [36] The first, at JERA's gas plant in Chiba, east of Tokyo, would use about 400 megawatts and cost more than $15 billion. [36] It takes power straight from the plant instead of waiting years for a grid connection. [36] The Financial Times put the wider plan at up to $140 billion. [36]

    Why it matters — Japan's digital minister has warned the country could become an 'AI colony' if it falls behind. [36] Blackstone, a US investment firm, said in June it planned $30 billion for AI data centres in Japan. [36]

15. 16

    Google puts its first AI chips in orbit

    A SpaceX rocket carried Google's own AI chips, called TPUs, into orbit on Thursday, on a satellite built by Planet Labs. [37] It is the first test in space for Project Suncatcher, Google's plan to run AI computing on sunlight. [37] Google says the satellite is working as expected. [37] In orbit, solar panels can make up to eight times more power than on Earth, Google says. [37]

    Why it matters — Google's research paper says launch costs would need to fall to about $200 a kilogram, which could take about 1,800 more flights of SpaceX's Starship. [38] SpaceX's boss, Elon Musk, says they will be the cheapest way to train AI within two or three years. [37]

16. 17

    Texas pause holds up a huge Microsoft data centre

    A data centre for Microsoft in West Texas, built by Chevron and a new company called Joulent, may not get its final go-ahead until 2027. [39] Chevron's Daniel Droog said the reason is a pause on new data centre permits, ordered last month by Texas governor Greg Abbott. [39] The project, called Kilby, would run on nearly 3 gigawatts of its own gas turbines, off the state grid. [39] Earth-moving has already started. [39]

    Why it matters — Because Kilby makes its own power, it would not raise electricity bills for neighbours, and it has a 20-year contract with Microsoft. [39] Even so, the Texas pause slows it down. [39]

17. 18

    Insurers say AI billing added $942m

    The Blue Cross Blue Shield Association speaks for a group of US health insurers. [40] It says hospitals' use of AI to write billing codes added about $942 million to its plans' costs from 2023 to 2025. [40] Most came from extra diagnoses that moved patients into higher-paying categories. [40] About $653 million was tied to diagnoses that came with no change in the patient's care. [40] The rise came as 60% of hospital systems began using AI coding tools. [40]

    Why it matters — The insurers did not blame AI for all of it, and a health economist said many extra diagnoses are real ones that used to be missed. [40] The extra cost can end up in patients' premiums, the group says. [40]

18. 19

    44,000 object to Palantir handling NHS data

    More than 44,000 people in England have filed legal objections to their health data being used on the NHS's Federated Data Platform. [41] The platform runs on software from Palantir, a US data company, under a £330 million deal. [41] Under UK data law, the NHS must stop using their data unless it can give a 'compelling reason'. [41] Palantir says hospitals using it have done 117,000 extra operations. [41]

    Why it matters — Britain's prime minister, Andy Burnham, must decide within months whether to keep or end the seven-year deal. [41] Two committees of MPs want him to use a break clause in it. [41]

19. 20

    Anthropic asks Australia for an opt-out on copyright

    Anthropic asked Australia's government to let AI firms train on Australian books, news and music unless the owners opt out. [42] It made the request to a parliamentary inquiry, after Australia's government ruled out a blanket copyright exception for AI. [42] Australia's public broadcasters, the ABC and SBS, want AI firms held to the same copyright rules as news outlets and made to pay for news. [42]

    Why it matters — An opt-out puts the work of blocking AI on every writer and broadcaster, and SBS said such blocking signals are 'routinely ignored'. [42] Anthropic and OpenAI executives are due at the inquiry's hearings in Sydney next week. [42]

20. 21

    Britain's central bank wants a say over AI

    Andrew Bailey, governor of the Bank of England, Britain's central bank, said authorities must keep the 'right to intervene' in AI. [17] He said AI has made attacks on the financial system bigger and more skilled, and could threaten card payments and trading. [17] He suggested careful testing of new models first, and said new regulation was not the right place to start. [17]

    Why it matters — The bank's financial policy committee also warned that the growing pile of AI debt adds to the risks for the financial system. [17] Bailey spoke as US President Donald Trump has broadly rejected calls for more AI rules. [17]

21. 22

    Amazon settles over slow deliveries in poor areas

    Amazon settled a lawsuit brought by Brian Schwalb, attorney general of the District of Columbia, the US capital, over slower deliveries to two low-income ZIP codes. [43] The 2024 suit said Amazon left those areas out of its own fast Prime delivery without telling customers. [43] Amazon will refund $7.25 million in Prime fees and pay a $1 million penalty. [43] About 69,000 Prime members got slower deliveries, his office said. [43]

    Why it matters — Amazon agreed to tell customers in the District about any future limits on its deliveries. [43] It says the settlement is not an admission that it did anything wrong. [43]

22. 23

    Lyft pays $272.5m over drivers' status

    Lyft, a US ride-hailing company, will pay $272.5 million to settle a California case over treating drivers as contractors instead of employees. [44] The state's Labor Commissioner sued in August 2020, saying drivers were denied the minimum wage, overtime and sick pay. [44] The deal covers 2016 to 2020, before voters passed Proposition 22, which let ride-hailing firms keep drivers as contractors. [44] A judge still has to approve it. [44]

    Why it matters — The Labor Commissioner says it will pass the state's share to drivers who filed wage claims. [44] Lyft says its drivers were always properly classified. [44]

23. 24

    Microsoft's Office and Teams chief leaves

    Ryan Roslansky, who ran Microsoft's Office and Teams software, will leave at the end of the year, chief executive Satya Nadella told staff on Thursday. [45] Roslansky ran LinkedIn, the jobs network Microsoft owns, for six years. [45] He said the job needed him in Redmond, Washington, every day, and his family lives in the San Francisco Bay Area. [46] Office and Teams now move into the group that builds Copilot, Microsoft's AI assistant. [46]

    Why it matters — Gaming head Phil Spencer and Office leader Rajesh Jha both left Microsoft this year. [45] Microsoft's best-known work software will now be run by the people building its AI. [46]

24. 25

    AI agents posted 13,000 private screenshots

    AI coding agents left more than 13,000 private screenshots from over 300 organisations where anyone could see them, a security firm called Glow reported. [47] The agents had to attach pictures to code changes in private projects on GitHub, a site where code is stored. [47] Their tools could not do that privately, so the agents put the pictures in public projects and linked to them. [47] The images showed unreleased software, client details and financial data. [47]

    Why it matters — The agents came up with the workaround on their own, and their users were unlikely to notice. [47] The organisations include Fortune 500 companies and an AI lab. [47]

25. 26

    Spies posed as AI policy figures

    A spying group that Proofpoint, an email security company, links to China pretended to be well-known AI policy figures to trick experts. [48] In July it posed as a former deputy head of the US president's science office, inviting researchers to a fake AI advisory committee. [48] Links in its replies led to fake login pages that steal Microsoft account passwords. [48] In February it posed as a senior Anthropic employee. [48]

    Why it matters — The targets were AI policy experts at US think tanks, universities and law firms. [48] Some were invited to help with a US Senate report on AI chip exports. [48]

26. 27

    Chatbots remove hijabs from pictures

    The Guardian asked four chatbots to remove the hijab from an AI-made picture of a Muslim woman. [49] OpenAI's ChatGPT and Grok, from Elon Musk's xAI, did it. [49] Google's Gemini refused at first, then did it when asked to make her look more 'western'. [49] Anthropic's Claude said it cannot edit images and would not do this. [49] The test followed a French MP posting an edited photo of a real woman with her hijab removed. [49]

    Why it matters — The same bots refused to remove a dress. Grok said it treated removing a hijab as a simple clothing change. [49]

27. 28

    Dutch spy agency warns cars can listen

    The Netherlands' intelligence service, the AIVD, warned that the microphones, cameras and GPS in modern cars can be used to spy on the people inside. [50] It said some cars' microphones and cameras can be switched on from a distance. [50] It named no country. [50] In Britain, Chinese-owned brands made up 15% of new cars registered in the first half of 2026. [50]

    Why it matters — The AIVD told people not to hold confidential conversations in or near such cars. [50] Britain's defence ministry has reportedly banned cars with Chinese parts from some sensitive military sites. [50]

28. 29

    Newsom vetoes a smart-glasses recording ban

    Newsom also vetoed a bill that would have made it a crime to secretly record people with smart glasses in private parts of businesses, such as changing rooms. [51] It would also have made makers show a light whenever the devices record. [51] Newsom said the bill's wording was too broad and could cover ordinary smartwatches. [51] California already bans recording people where they expect privacy. [51]

    Why it matters — It would have been the first US state law aimed at smart glasses. [52] Meta sold more than 7 million pairs of its camera glasses last year. [52]

29. 30

    Amazon and Cloudflare copy the 'decision model'

    Amazon and Cloudflare each released free 'decision models' this week, following Jev, a model from a start-up called TypeSafe that became popular two weeks ago. [53][54] A decision model does not write text. It picks one answer from a fixed list, such as yes or no, and says how sure it is. [54] Amazon's Strands Decider is small enough to run on a company's own computers. [53] Cloudflare says its two Clef models are more accurate than Jev. [54]

    Why it matters — Companies running AI agents can use them to choose each next step for much less than a full chatbot model costs. [53] OpenAI announced a similar product the same week. [53]

30. 31

    Microsoft starts switching off an old email link

    Microsoft began switching off Exchange Web Services in its online email service on Thursday. [55] EWS dates from 2007 and lets other apps read mailboxes, calendars and contacts. [55] Companies that left the setting blank are now blocked, though administrators can turn it back on for approved apps. [55] It shuts for good on 1 April 2027, and Microsoft says there will be 'no exceptions'. [55]

    Why it matters — Any business app that still reads Outlook mail this way must first move to Microsoft Graph, its newer system. [55] Many companies do not have a full list of what still uses EWS, one integration expert said. [55]

31. 32

    HP's $699 laptop waits for some memory

    HP announced the OmniBook 5, a $699 laptop with an OLED screen, a kind usually found on dearer machines. [12] It goes on sale this month with 8GB or 16GB of memory. [12] A 12GB version and a 1TB storage option will not arrive until November, which Tom's Hardware put down to the parts shortage. [12] The brightest screen option waits until March. [12]

    Why it matters — HP is chasing Apple's MacBook Neo, along with Dell, Lenovo and Acer. [12] In a budget Android phone, memory can now be half the cost of the parts, against about 15% before. [9]

32. 33

    New York City's click-to-cancel rule starts

    New York City's click-to-cancel rule took effect on Thursday. [56] A business that lets people sign up online must now let them cancel online too, without a phone call or a visit. [56] New Yorkers can report companies that break it on a new city website. [57] A US national version was struck down by an appeals court over how it was made. [56]

    Why it matters — The city estimates the rule could save New Yorkers $21.5 million to $162.5 million a year. [56] Samuel Levine, the city official in charge, tried to bring in the national rule when he worked at the US Federal Trade Commission. [56]

33. 34

    Mandiant's founder raises $255m for AI attackers

    Armadin, a security start-up founded by Kevin Mandia, raised $255.5 million on Thursday at a value of more than $2.5 billion. [58] Mandia founded Mandiant, which Google bought for $5.4 billion in 2022. [58] Armadin uses swarms of AI agents that try to break into a client's systems all the time, instead of hiring testers now and then. [58] It raised $190 million only six months ago. [58]

    Why it matters — Its pitch is to find holes before criminals, or rogue AI agents, use them. [58] Its backers include Andreessen Horowitz, Accel and In-Q-Tel. [58]

34. 35

    ChatGPT and Shopify add AI shopping tools

    OpenAI added a Try On button to ChatGPT on Thursday, which shows how clothes would look on a photo of the user. [59] It also added a list for saving products to come back to. [59] The same day Shopify, which sells online-shop software, launched Canvas. [60] With Canvas, a seller builds a shop by chatting with Shopify's AI assistant, Sidekick, and watches the real site change as it goes. [60]

    Why it matters — Google launched its own clothing try-on last year, so ChatGPT is following it. [59] Canvas takes on AI site builders from Wix, Squarespace and others. [60]

35. 36

    Google's AI describes the world to blind users

    Google launched Guided Vision on Thursday, a feature for blind and low-vision people on Android phones. [61] A user shares the phone's camera with Gemini Live, Google's spoken AI assistant, and it describes aloud what the camera sees. [61] It can read small print, describe surroundings and help find objects, and the user can ask follow-up questions. [61]

    Why it matters — Google warns it is not for finding the way or avoiding obstacles, and does not replace a cane. [61] Apple offers a similar feature on the iPhone and Vision Pro. [61]

36. 37

    Fewer English schools hit by cyber attacks

    Fewer secondary schools in England reported a cyber incident this school year: 27%, against 29% a year earlier and 34% the year before that. [62] The figures come from a survey of 3,775 teachers by Ofqual, England's exams regulator. [62] Of the schools hit, 66% said they recovered immediately, up from 55%. [62] Phishing, fake emails that trick people into handing over passwords, was the most common attack. [62]

    2023/2434%

    2024/2529%

    2025/2627%

    Why it matters — Ofqual could not say what caused the improvement, and 54% of teachers did not know what their school had changed. [62]