Tech, in brief

A judge in New Mexico, a US state, ordered Meta, which owns Facebook and Instagram, to pay $25bn on Wednesday. [1][2]

A jury found in September that Meta misled people in the state about who could see their data and how it dealt with hate speech. [3][1]

The jury counted more than 43 million violations of the state's consumer law, one for each person who heard each misleading statement. [3][4]

The law allows up to $5,000 for each violation, so the most the judge could order was about $219bn. [5][2]

Judge Francis Mathew said the top rate was deserved, then cut the total to $25bn. [2][6]

New Mexico had asked for $35bn to $40bn. [4][3] Meta had asked for no more than $3.45bn. [4][3]

Interest of 15% a year builds on the penalty until Meta pays, and the money goes to the state's schools. [2][4]

Meta says it will appeal, and its chief lawyer called the case "a tax on speech". [7][1]

The case grew out of Cambridge Analytica, a British political firm that got data on up to 87 million Facebook users through an app. [3][6]

Meta paid a $5bn US penalty over the same scandal in 2019 and $725m to end a group lawsuit in 2022. [1][2]

The $25bn is more than the $14.8bn New Mexico expects to collect for its main budget in 2028. [6]

An analyst at D.A. Davidson, an investment firm, said the penalty is costly but will not change much for a company worth nearly $2 trillion. [1]

Why it matters. Meta's appeal now decides the real number: watch whether a higher court accepts its argument that a penalty this size is unconstitutionally large, while 15% interest keeps adding to the bill. [1][2]

Meta's sales rose 26% to $64.6bn in the three months to September. [21][8] It paid $9.89bn toward a $17.1bn settlement with US states, which said its apps were built to hook children. [8] Without that payment its operating profit would have been $22.89bn. [8] Meta now expects to spend $135bn to $145bn this year, on buildings and equipment, much of it for AI. [22]

Why it matters. Legal bills are now big enough to show in Meta's results, and the New Mexico penalty comes on top of them. [3] Meta expects $73bn to $76bn in sales next quarter, more than analysts did. [21]

Meta said on Wednesday its AI agent, Muse, has more than 10 million users since it launched in September in the US and Canada. [23][24] Mark Zuckerberg said shopping is its most common use, and that purchases made through it average about $200. [24] About a quarter of people who set it up now use it daily. [23] Meta plans to charge businesses a fee on sales Muse makes, and to sell subscriptions to heavy users. [24][23]

Why it matters. Zuckerberg says Muse could one day earn more per user than Facebook's ads, which bring in several hundred dollars a year per American user. [23] It also costs Meta far more to run than its other apps. [24]

Google's own AI chips, called TPUs, now answer requests in real use at Meta, a Google blog post said. [25] Meta writes its AI software in PyTorch, a toolkit it created, and new free software lets that code run on Google's chips with few changes. [25] A Meta engineering director said the company "cannot be constrained by a single hardware architecture". [25] Meta research teams reportedly used Google's chips from 2020 to 2023, then stopped. [25]

Why it matters. Tom's Hardware says selling these chips to outside customers threatens Nvidia's lead in AI chips. [25] Google gets one of its biggest rivals as a paying customer for its chips.

OpenAI said on Wednesday its AI models will be sold to developers through Google's cloud in the coming weeks. [26] A cloud is a business that rents out computers over the internet. [26] Microsoft lost its exclusive right to OpenAI's technology less than six months ago, and Amazon began offering OpenAI's models the day after. [26] OpenAI's models are now on all three of the biggest US clouds. [26]

Why it matters. Google makes Gemini, one of OpenAI's main rivals, and will now also earn from selling OpenAI's models. [26] Businesses that already use Google's cloud can now choose OpenAI without moving.

Anthropic, the maker of the Claude chatbot, announced Claude Duet on Wednesday, an agent that works in the background on a person's apps. [27] It sends a short daily list of things that need attention, and asks before it acts for the user. [27] For now only US subscribers can join a waiting list. [27] Anthropic says Duet never sees passwords directly, and gets health or money details only when told to share them. [27]

Why it matters. All three big AI labs now sell an agent like this: Meta's Muse, OpenAI's Dots and now Duet. [27] Duet will come with Anthropic's $20-a-month plan, while OpenAI's Dots needs a plan costing at least $100. [27]

OpenAI said on Wednesday it will show ads in ChatGPT in 26 more countries, including Nigeria, South Africa, Ukraine, Argentina and Pakistan. [28][29] That brings ads to 89 countries, eight months after they started. [29] Only people on the free plan and the cheap Go plan see them. [29] More than 100,000 advertisers have used them, up from about 20,000 in late August. [29]

Why it matters. OpenAI's ad business was earning at a rate of $1bn a year by August, up from $100m in March. [28] Ads give OpenAI a way to earn from the many users who never pay, ahead of a stock-market listing it is eyeing for 2027. [28]

Walmart, the biggest US retailer, will let AI agents buy things on its website from next month. [14] Agents connect through a new "Agent Gateway", and by default every purchase needs the shopper's approval. [14][30] Shoppers can also set a spending limit, such as $100. [14] Walmart's rules say agents must not do more than the shopper intends, and it will block any that break them. [14]

Why it matters. Amazon, Walmart's biggest rival, has blocked shopping agents, including Meta's Muse. [30] Walmart says the gateway lets an agent find a shopper's usual items without seeing their whole shopping history. [14][30]

Nvidia, which makes the chips AI runs on, showed a data centre built inside a shipping container at its conference in Berlin on Wednesday. [31][20] Each "DSX Pod" weighs about 28 tonnes, holds seven racks of its newest Vera Rubin computers, and travels by truck. [31][32] Nvidia says one can start working 14 days after it arrives, instead of about six months. [31] Deliveries start in early 2027. [31]

Why it matters. Data centres are taking too long to build to keep up with AI demand, and Nvidia says the pods also cut the time chips sit idle. [32][20] Analysts say a single Vera Rubin rack can cost more than $7m. [31]

Nvidia said in Berlin it will pay suppliers to build more of the parts its chips need, saying "capacity is the constraint". [33] It added Samsung Electro-Mechanics, a South Korean maker of chip bases called substrates, which will send about a quarter of its output to Nvidia by 2028. [33] It also backed bigger factories at Ibiden in Japan and Amkor, a chip-packaging firm, which plans 300% more capacity. [33]

Why it matters. Nvidia plans 2.5 times more substrate supply over three years, a bet that AI demand lasts. [33] About a quarter of Samsung Electro-Mechanics' output will go to one customer. [33]

Nvidia released new free AI models at its Berlin conference, including ones that help search company documents and ones for self-driving cars. [34] The search models handle text and images in 34 languages and are small enough for one desk-sized computer. [34] Anyone can download and change them, which European firms want so they do not depend on US labs. [34] Nvidia's chief, Jensen Huang, told 4,000 people Europe should produce AI, not only use it. [20]

Why it matters. Nvidia says its new search models rank first among rivals, a claim about its own product. [34] Every free model still needs chips to run on, and Nvidia sells most of them.

Nvidia published a design on Wednesday for running AI agents inside set limits, built on its software called OpenShell. [13] It works only on Nvidia hardware, and one Nvidia chip watches what each agent does. [13] ServiceNow, a business-software firm, launched Arc the same day, an agent for office work that runs inside OpenShell. [12] Both pitch this as the answer to a summer of agents breaking into systems. [12][13]

Why it matters. Businesses want agents but fear what they might do, so both firms sell the limits along with the agent. [12] OpenShell still has open security issues listed on its own code page. [13]

Greenpeace protesters interrupted Jensen Huang's speech in Berlin on Wednesday with a banner reading "Nvidia, clean up your dirty business". [35][20] One shouted that Nvidia chips had been found in Russian drones. [35] Reports say Jetson Orin chips, small Nvidia computers sold to students and developers, were found in Russian missiles used against Ukraine. [35] Huang told his guards to "go easy". [20]

Why it matters. Nvidia says the chips are not sold in Russia, but admits it cannot track second-hand ones once they are resold. [35] Reports say Russia's S-71 missile uses one to pick its own targets. [35]

Donald Trump said on Wednesday that data-centre companies should pay the communities around them a "dividend" to win their support. [9] He wrote that unpopular projects become popular "when a financial deal is struck". [9] He has already said data centres should make their own power, and now says they should share spare power with neighbours. [36] Nebius, which rents out AI computers, said it was "100% in principle" open to the idea. [9]

Why it matters. Local fights over power bills, water and noise have become an issue before the US elections in November. [9] Several US states are weighing pauses on new data centres. [9]

A group including Nvidia, Microsoft and Cisco told the UK government it has about 12 months to win public support for data centres. [10] Otherwise, they say, protests like those in the US could block building. [10] They want councils that host a site to keep more of its tax money, and a set payment to each community. [10] The UK government says councils in its AI growth zones already keep all the extra business tax. [10]

Why it matters. Campaigners say the offer is a way of buying people off, and Scotland already has 15 local campaigns against data centres. [10] Microsoft has promised to invest $30bn in UK AI buildings. [10]

Hamlet, a US start-up, has read tens of thousands of city council meetings from their video recordings. [11] It marks each comment as friendly or hostile to a planned project. [11] Its boss said a big fund that pays for data centres and a data-centre builder are among its customers. [11] On Tuesday it ranked towns on hot topics, with Corpus Christi, Texas, first for speakers against data centres. [11]

Why it matters. Builders want to go where they will be fought least, so what people say at a council meeting now affects where money goes. [11]

A grove of rare American chestnut trees in Mason County, West Virginia, sits beside Nscale's planned 2,250-acre data centre. [37] Nscale, a UK start-up, counts Anthropic, Nvidia and Microsoft among the campus's clients. [37] West Virginia's governor says the trees will be moved in late 2027, but a tree expert says moved trees would die within two to five years. [37] State officials took public comments on the site's air permit on 20 October. [37]

Why it matters. Nscale says it never planned to cut the trees down. [37] In September, runoff from the building site flooded nearby homes, and the state ordered repairs. [37]

Palantir, a US data-software firm, said neither it nor the US government can shut down Maven from outside. [18] Maven is its battlefield system for NATO's top military command, combining satellite, radar and other data. [18] Palantir's head for Europe said "there are no kill switches" in a Reuters interview on Wednesday. [18] He spoke after reports that the US limited Ukraine's use of SpaceX's military satellite service. [18]

Why it matters. Germany and France have declined to use Maven for their own forces, partly over relying on a US supplier. [18] France's spy agency is moving to ChapsVision, a French firm. [18]

Microsoft added new tools on Tuesday for customers who want more control over its cloud, especially in Europe. [19] Customers can now keep their main encryption keys on their own hardware. [19] More services can run in a customer's own building and keep working without a link to Microsoft. [19] An analyst said European governments worry about keeping key data on an American company's computers. [19]

Why it matters. Europe has proposed a law that would limit US clouds for sensitive government data. [19] Analysts say running Microsoft's cloud locally does not by itself end dependence on Microsoft. [19]

A union of Micron workers in Taoyuan, Taiwan, planned a one-day strike for Thursday 22 October. [38][39] Micron is a US maker of memory chips, including the kind used in AI servers. [38] The unions want 15% of operating profit paid as bonuses each quarter, as Samsung and SK hynix already link bonuses to profit. [40] Of the Taoyuan union's 2,258 members, 1,994 voted to strike. [40]

Why it matters. A chip factory runs day and night, so even one day out costs Micron, though Tom's Hardware expects only a small hit to output. [39] Taiwan's president urged both sides to talk in good faith. [40]

Three contract workers died and one was badly hurt at the building site of a new chip factory in Kaohsiung, Taiwan. [41] It belongs to TSMC, the world's biggest maker of chips for other companies, and will make its most advanced 2-nanometre chips. [41] Officials suspect a broken bolt on a steel plate sent the men and the steel falling. [41] Steel lifting has stopped while labour inspectors investigate. [41]

Why it matters. Local reports count at least six earlier deaths at TSMC building sites in Taiwan since September 2024, one worker each time. [41]

Intel, the US chipmaker, said on Wednesday it has started putting machines into Fab 62, a new factory in Arizona due to run by the end of 2028. [42][43] In Ohio, crews lifted the first big roof beam at a site meant to start in 2025, now due in 2030 or 2031. [42] Orders for its data-centre chips are growing faster than it can make them. [44]

Why it matters. Intel still needs outside customers to fill these factories, and has named few. [42][44] The US government owns a stake in Intel. [44]

Samsung and SK hynix, South Korea's two big memory makers, are upgrading their factories in China to make more advanced chips, Reuters reported. [45] Samsung aims to switch its plant in Xi'an to 286-layer storage chips by the end of the year. [45] SK hynix aims for 240 layers or more at its plant in Dalian. [45] US export rules limit which machines they can send to China. [45]

Why it matters. Chinese memory makers such as CXMT are growing fast, and AI data centres want all the memory they can get. [45]

Elon Musk said ASML, the Dutch firm that makes the machines used to print chips, and other suppliers are speeding up deliveries to Terafab. [46] Terafab is a chip-factory project shared by his companies Tesla and SpaceX. [46] He said on SpaceX's results call that volume production in Texas is aimed at late 2028, with "big error bars". [46] A first test chip is due in Austin around mid-2027. [46]

Why it matters. Musk also said his AI companies expect 2.5 to 3 gigawatts of computing running by the end of 2026. [46]

AMD and Intel, the two makers of most PC and server processors, will merge their tuning code for running AI on those processors. [47] AMD's library, ZenDNN, is moving into oneDNN, a free library led by Intel. [47] AMD is also joining the group behind oneDNN, paying $20,000 a year. [47] Developers using PyTorch will then have one path that works well on both makers' chips. [47]

Why it matters. AMD says the move was urged by an advisory group the two rivals set up in 2024 to keep their shared chip design in step. [47]

Tesla said on Wednesday it has started making its Optimus humanoid robot at its Fremont factory in California. [48] Elon Musk called it early production and claimed Tesla could make 20,000 a week by the end of 2027. [48] Tesla also said it will start selling its Cybercab, a two-seat car with no steering wheel, by the end of the year. [49] Owners' own cars could join its robotaxi service around mid-2027. [50]

Why it matters. The US car-safety regulator is investigating whether the Cybercab meets safety rules for cars with no pedals. [49] Tesla's profit fell 43% as it spent more on robots and AI. [48]

Waymo, the driverless-taxi company owned by Google's parent, began airport rides at Harry Reid International in Las Vegas on Wednesday. [51] Cars pick up and drop off in set zones at Terminals 1 and 3. [51] It is Waymo's sixth big airport, after Phoenix, Nashville, San Antonio, San Francisco and San Jose. [51] Only a few riders can use it at first. [51]

Why it matters. Zoox, Amazon's robotaxi firm, already serves the same airport, and Nevada has approved permits for Tesla and Uber to run thousands of robotaxis there. [51]

General Motors, the biggest US carmaker, said on Wednesday its next electric cars will run on one central computer 40 times more powerful than its current systems. [52] It replaces many small chips and cuts wiring by 30%. [52] It arrives in 2028 in the Cadillac Escalade IQ, with highway driving where people can take hands and eyes off the road. [52] GM is also adding an AI assistant to OnStar, its car help service. [53]

Why it matters. Chinese carmakers now bring out new models in two to three years instead of four to five, and GM wants to match that. [52]

Flapping Airplanes, an AI research lab started in late 2025, raised $1bn at a value of $6bn, led by Index Ventures and Lux Capital. [54][16] It was valued at $1.5bn in January, so its value has quadrupled in nine months. [54] It has nothing on sale yet. [16] It is trying to train AI with less data and less computing. [16]

Why it matters. Its co-founder says demand for AI will outrun the supply of chips and data, so doing more with each chip would be worth a lot. [16] Investors are paying for a chance at that, not for sales.

Recursive, an AI start-up, released Flint on Wednesday, a free coding model small enough to run on a laptop. [55] The company says its own AI system, the Eureka Machine, did most of the work of building it. [55][56] That let a small team build it from scratch in under four months. [55] Recursive is led by Richard Socher, once Salesforce's chief scientist, and was valued at $4.65bn in May. [55]

Why it matters. AI that builds better AI is one of the most argued-over ideas in the field, because some researchers fear it could slip out of human control. [55] Recursive's version is still overseen by people. [55]

Legora, a company that makes AI tools for lawyers, raised $550m at a value of $8.8bn on Wednesday. [15] In April it was valued at $5.6bn. [15] It says its yearly sales have passed $200m, up from $100m then. [15] It plans to spend the money growing in North America. [15]

Why it matters. Its biggest rival, Harvey, raised the same amount a month ago at $15.5bn, so investors now price two AI firms for one profession in the tens of billions. [15]