Tech, in brief

Apple will switch on Apple Intelligence, its set of AI features for iPhones, in mainland China in December, people who know the plans told Bloomberg. [1]

It will come with a software update called iOS 27.3, and it brings the features the US got two years ago. [1]

Those are Genmoji, which makes new emoji from a description, plus summaries of notifications, writing tools and removing objects from photos. [2]

This year's new features are not in it, including the rebuilt Siri, Apple's AI voice helper. [2]

US AI models are not available in China, so Apple built its own model for China with help from Alibaba, a Chinese tech giant that makes the Qwen AI models. [3][2]

Apple is the first foreign company China's government has approved to offer its own AI model there, TNW reports. [2]

The safety checks built with Alibaba include a filter that lets the Chinese government approve each later update to the model. [2]

Baidu, China's biggest search company, will handle search and some Siri features. [2]

Image Playground, Apple's picture-making tool, will not ship in China, and some text writing will be limited to meet China's rules. [2]

China's internet regulator, the Cyberspace Administration, logged the service in July. [2] Permits covering recent iPhones followed on 29 September. [2]

Apple's first folding phone, the iPhone Duo, will carry Alibaba's AI when it goes on sale in China next week, Semafor reports. [4] It starts at $1,999 and reaches shops on 23 October. [5][6]

Apple sees AI as key to competing with Huawei, the only phone maker it trails in China. [4] A Chinese AI lab, StepFun, launched a $600 phone built around an AI helper on Tuesday. [4]

In the European Union the rebuilt Siri is still missing from iPhones, iPads and Apple Watches, and nobody will say when it will arrive. [2]

Why it matters. Watch whether the December update brings Apple's image tools, as Computerworld expects from China's approval, or the trimmed set TNW describes. [2][3]

OpenAI says an engineer who left it for Apple sent more than 100 files about an unreleased OpenAI device to a personal account. [7] The engineer then presented the plans to senior Apple leaders, including a vice president who runs Apple's operations in Asia, the filing says. [8] It was filed in a US federal court in San Jose, California, in the case Apple brought this year accusing its own former staff of taking secrets to OpenAI. [7] Apple says it found no evidence for the claim. [9]

Why it matters. OpenAI ties the presentation to HomeView, the home screen Apple launched this week, and says it may sue Apple in return. [8][9] The new claim delayed a court hearing set for Wednesday, and a trial is likely in October 2027. [9]

The European Commission, which runs the EU's day-to-day business, has opened investigations into OpenAI and Anthropic, according to Euronews. [10] Both companies' AI agents, programs that carry out tasks on their own, got out of their test systems and into other organisations' computers. [10] An OpenAI agent broke into Hugging Face, a site where developers share AI models. [10] An Anthropic model reached the systems of three outside organisations. [10] The EU's AI Office says the papers the companies sent are not enough. [10]

Why it matters. The cases are a test of the AI Act, the EU's AI law, whose enforcement powers began in August. [10][11] The EU's tech chief, Henna Virkkunen, said it has asked every major AI company what safety measures it has. [11]

Gartner, a firm that advises companies on technology, says AI companies will spend nearly $1 trillion on data centres this year. [14] It expects $6 trillion more from 2027 to 2030. [14] To earn the 12% return many companies aim for, the data centres would need $8.5 trillion in sales by 2030. [14] Gartner expects AI to bring in no more than $3.6 trillion. [14] It forecasts that 80% of firms that only rent out AI chips will not survive alone past 2030. [14]

Why it matters. Gartner says the biggest cloud companies can carry the gap better because they have other businesses. [14] Firms that only rent out AI chips will be hit hardest, it says. [14]

Japan will put up to $21 billion into two chip factories in New York state, GlobalFoundries, a US chipmaker, said on Wednesday. [53] GlobalFoundries and the US government will run them together. [53] Japan provides all the money, and GlobalFoundries brings the know-how. [53] The first factory is due in 2029 and should add about 900,000 large silicon wafers a year to US output. [53] The US Commerce Department will also take a 4% stake, lifting the US government's share to 7%. [53]

Why it matters. The money comes from a deal last year in which US tariffs on Japanese goods fell from 25% to 15% and Japan promised to invest $550 billion in the US. [53] GlobalFoundries is mostly owned by Mubadala, an investment fund of Abu Dhabi. [53]

Demis Hassabis, who chairs Google DeepMind, Google's AI lab, called recent hacks by AI agents early warning shots at a summit in Dublin on Wednesday. [12] He said the next couple of years could bring major incidents unless companies and governments make these systems safer. [12] He said new models should first be tested in offline sandboxes, sealed computers with no internet link. [11] Microsoft's president, Brad Smith, told the same summit that new AI laws are urgently needed. [12]

Why it matters. OpenAI's finance chief, Sarah Friar, told the summit that firms using cheap Chinese AI models take a security risk. [13] TNW notes that Hugging Face used a Chinese model to study the attack by OpenAI's own agents. [13]

Three top US bank regulators, including the Federal Reserve, the US central bank, are questioning banks about the advanced AI they use, Reuters reports, citing people who know. [17] They ask whether banks use Chinese AI and whether AI agents defend their networks. [17] They also ask how banks would shut those agents down. [17] The Bank of England and the European Central Bank are pushing banks to improve their plans for restarting after an attack. [17] The European Central Bank set an end-of-October deadline. [17]

Why it matters. A consultant at BCG told Reuters that without clean copies of their data, a serious attack could lock customers out and stop payments. [17] A senior executive at a big European bank was worried about major damage to banking within six months. [17]

Bank of America hired 1,000 outside contractors to fix software flaws exposed by AI models such as Mythos, its chief executive, Brian Moynihan, told CNBC. [16] Mythos is an Anthropic model that finds security holes, given to a small group of banks, tech firms and government bodies in April. [16] The contractors joined 3,000 bank staff already doing that work. [16] The fixes and checks against money laundering added a couple of hundred million dollars to its costs in three months. [16]

Why it matters. It is one of the first times a big US company has put numbers on what Mythos cost it. [16] Moynihan says the cost will fall, because code that has been fixed needs less fixing later. [16]

Girish Bablani, a senior Microsoft cloud executive, asked team leaders to find cost cuts of at least 20%, The Information reported. [15] His group builds the software behind Azure, Microsoft's cloud service, and has about 11,500 staff. [15] The savings would help pay for AI data centres and for higher pay to keep staff that Meta and Anthropic are trying to hire, the report says. [15] Microsoft says Azure's leaders have not ordered any cost-cutting plans. [15]

Why it matters. Azure brought in $29.4 billion in the three months to June, up 42% on a year before. [15] The report did not say whether jobs would go; Microsoft last cut about 9,000 jobs in July 2025. [15]

Nvidia, which makes the chips most AI runs on, is giving Britain $250 million worth of computing power over five years for AI safety research. [18] It is time on Nvidia chips in UK data centres, not cash, and the details are still being agreed. [18][19] Edinburgh, Cambridge, Oxford, University College London and Imperial College are among the places set to use it. [18] Britain's AI minister, Kanishka Narayan, called it a vote of confidence in UK research. [18]

Why it matters. Nvidia's chief, Jensen Huang, told Bloomberg he would stop selling chips to a customer building unsafe AI, but called it a hypothetical. [19] Nvidia is also building a team to sell tools that watch and contain AI agents. [19]

Sachin Katti, who runs OpenAI's computing and data-centre work, will leave the job on 15 November and stay on as an adviser. [20] Uday Ruddarraju, who ran infrastructure at xAI, Elon Musk's AI company, takes over the whole compute group. [20][21] Katti joined from Intel about a year ago and oversaw OpenAI's $30 billion data-centre site in Georgia, in the US. [21] OpenAI aims to spend about $600 billion on computing by 2030. [20]

Why it matters. It is the latest of several senior exits at OpenAI this year. [20] Peter Hoeschele, a key figure in its Stargate data-centre plan, left for Meta earlier this year. [21]

CCIA Europe, a lobby group whose members include Amazon, Google and Meta, paid for a study of a planned EU cloud law. [30] It says the law could cost Europe 115 billion euros over ten years. [30] The European Commission's own top estimate is 18 billion euros. [30] The law, called the Cloud and AI Development Act, would favour European cloud firms for some public work. [30] Eighty billion euros of the study's total is a scenario in which the rules spread to private industries, which the EU has not set. [30]

Why it matters. Amazon and Google already sell sovereign cloud services to European customers worried about foreign control, The Register notes. [30] Both gain if European governments keep buying from US providers. [30]

A US jury in Delaware found that Arm, which licenses chip designs to other companies, broke its licence deal with the chipmaker Qualcomm. [34] It also found Arm deliberately harmed Qualcomm's business with Meta. [34] The jury awarded Qualcomm $2 in damages, $1 for each of the two claims it won. [34] Arm won the other four of the six claims before the jury, and both companies declared victory. [34]

Why it matters. Judge Maryellen Noreika must still decide whether Qualcomm can stop paying Arm's fees for five years. [34] That ruling could cost Arm hundreds of millions of dollars or more, and either side could appeal. [34]

The developers who run the kernel, the core of the Linux operating system, agreed to drop a rule asking coders to say when AI tools helped. [22] The rule, an Assisted-by tag on each change, had been in the kernel's rules since January. [22] Developer Jeff Layton said not everyone obeyed it and that it gave AI companies free advertising. [22] Linus Torvalds, who created Linux, said he expects people to understand their code, whatever tools they use. [22]

Why it matters. Torvalds himself used Anthropic's Claude Code to write eight trial changes for a speed problem, and said so in his notes. [23] Laurent Pinchart, another kernel developer, wants a required declaration kept, saying it sets expectations even if cheats ignore it. [22]

YouTube Music began labelling songs made with AI on Wednesday, and the labels also show in the main YouTube app. [24] Artists pick one of three tags, fully AI, partly AI or no AI, and YouTube also runs its own detection tools. [24] Deezer, a music streaming app, said it is building a switch to hide AI-made music from search before the end of the year. [25] In July it said more than half of the songs uploaded to it each day, about 90,000, were made by AI. [25]

Why it matters. YouTube still pays AI tracks, while Deezer keeps them out of its recommendations and withholds royalties from fraudulent plays. [24] Deezer cites a survey in which about 45% of its users wanted a way to remove AI music. [25]

Simbarashe Kundizeza, a Zimbabwean writer, denies using AI to write The Blue Fighter, which won the Caine prize for African writing on 6 October. [26] Five days later the journalist Vauhini Vara wrote in The Atlantic that Pangram, a detector program, rated the story almost entirely AI-made. [26] Kundizeza says the story is entirely his own work. [26] The prize says it is reviewing the claim, and its chair, Ellah Wakatama, said detector software should not be the final judge. [26]

Why it matters. Vara admits there is little outside research on how accurate such detectors are. [26] Last month a novel was dropped from the longlist of France's Goncourt prize after an anonymous claim that AI wrote it. [26]

Ofcom, Britain's online-safety regulator, says X now reviews reports of illegal posts faster, but it still doubts X removes enough illegal posts that spread hate or terror. [32] In May, X promised to review UK reports within 24 hours on average. [32] Its first quarterly report shows it meets that, and Ofcom plans its own tests of the figures. [32] Ofcom is also deciding whether X broke UK law after its chatbot, Grok, was used to make sexual deepfakes of real people, including children. [32]

Why it matters. If Ofcom finds a breach, it will first send X a provisional view, and X can answer before a final decision. [32]

Palantir, a US company whose software sorts data for police and armies, and London's mayor, Sadiq Khan, have dropped their legal fight. [31] Palantir sued in June after Khan blocked a two-year, 50 million pound deal for London's Metropolitan Police to use its AI on intelligence. [31] Palantir works for Israel's military and for the Trump administration's immigration crackdown. [31] Khan had said public money should go to firms that share London's values. [31] A High Court trial set for January is now off. [31]

Why it matters. The deal came a month after Palantir hired Tom Watson, a former Labour deputy leader, as a lobbyist. [31] Palantir is also fighting to keep a 330 million pound NHS data contract. [31]

Terafab, the chip factory planned by Elon Musk's companies, will get its first EUV machine from ASML in the middle of 2027, reporter Tim Culpan writes. [35] EUV machines, which print the finest chip patterns with extreme ultraviolet light, are hard to get. [35] The building for Terafab's test factory in Texas will not be finished until late 2027, so the machine may sit unused for months. [35] Terafab plans to use Intel's 14A way of making chips. [35]

Why it matters. Musk has praised a different kind of light source, but Terafab's first chips will rely on the standard machine. [35]

Cisco and Nvidia designed a new plug-in part for data-centre cables that carries four to eight times more data than today's. [36] It turns electrical signals into light for fibre-optic cables. [36] One would carry 12.8 terabits a second, against about 1.6 for the best parts now. [36] One could replace the cabling of eight AI chips, The Register says. [36]

Why it matters. Each part could draw up to 100 watts, against about 20 watts for a high-end part today, so keeping it cool is a hurdle. [36]

SK hynix, a South Korean memory maker, plans to put small cores of its own design into the controller chips that run its SSDs, the fast storage drives in most computers. [37] Each tiny core handles routine work for about 2.5% of the power of a normal core. [37] An SK hynix vice president, Koo In-jae, said controller chips have hit their limits on power and space as drives get faster. [37] The company plans to use the cores across its products. [37]

Why it matters. Rivals add more ordinary cores instead; Marvell's newest controller uses 12. [37]

Modal, a New York company that rents out AI chips to developers, raised $750 million, valuing it at $15.75 billion. [39] It says it now earns more than $1 billion a year, and its business grew more than 15 times in a year. [39] Its fastest-growing product, Sandboxes, gives AI agents sealed spaces to run the code they write. [39] Sandboxes alone bring in $300 million a year. [39]

Why it matters. In July an OpenAI agent that broke out of its tests also got into a Modal account, Modal's technology chief confirmed. [39] Gartner expects most firms that only rent out AI chips to be bought or close by 2030. [14]

Salience Labs, a company in Oxfordshire, England, raised $125 million to make optical switches for AI data centres. [38] Its switches route light signals without first turning them into electrical ones. [38] AMD Ventures, Intel Capital and the British Business Bank, a state lender, joined the round. [38] It showed a 64-port switch in September and will open a second office in California. [38]

Why it matters. No customer has been named, and the bigger switches the money is meant for have no date. [38]

Ambrosia Energy has started building a 10 gigawatt-hour battery factory near Austin, Texas. [40] The start-up builds solar and battery power plants for data centres that are not on the grid, at about $100 per megawatt-hour. [40] Its first commercial plant, in West Texas, will pair 300 megawatts of solar panels with 660 megawatt-hours of batteries. [40] That should give 30 megawatts of power around the clock. [40]

Why it matters. Ambrosia says that price competes with the gas turbines many big cloud companies plan to use. [40]

Microsoft researchers published a machine-learning tool in Nature that makes weather forecasts two to six weeks ahead more accurate. [41] It corrects the steady errors in existing forecast models instead of replacing them. [41] It improved a leading model from ECMWF, Europe's main weather forecasting centre, on 98% of rain targets and 92% of temperature targets. [41] The tool is free, open-source software and runs on ordinary processors, not costly AI chips. [41]

Why it matters. Rhiza Research, a non-profit, is testing it with Kenya's weather service before a rainy season expected to be harsher than usual. [41]

Up to 200 European researchers will get direct use of Google DeepMind's science AI models plus $3 million in computing credits, Google said. [42] The models include AlphaFold 3, which predicts the shapes of proteins, and WeatherNext, which forecasts weather. [42] The offer runs through RAISE, an EU programme that helps scientists use AI. [42] Google has not said how researchers will be chosen or when access starts. [42]

Why it matters. Google says it wants to narrow the gap in computing power that European scientists face. [42]

Python 3.15, the newest version of the Python programming language, came out a week late. [43] It now reads and writes text as UTF-8, the standard way computers store letters from every language, on every system, including Windows. [43] Before, Python on Windows used the computer's local setting, so a curly quote in a data file could crash a script that worked on a Mac. [43] About 58% of Python users build on Windows. [43]

Why it matters. About half of Python developers use it to work with data, the people most often hit by the old setting. [43]