New York opens sign-ups for its AI safety law and a state AI testing centre, as ten AI firms sign the Paris Call

The stories

31 on this beat today.

Every story told the same way: what happened, who is involved, why it matters, how it works, the truth and the good news.

1. 01

   New York opens sign-ups for its AI safety law and a state AI testing centre, as ten AI firms sign the Paris Call

   72 hours

   for a big AI maker to tell New York of a serious safety incident

   California's law allows 15 days [1]

   $500 million

   in yearly sales, the line that brings an AI maker under New York's law

   fines reach $1 million for a first breach [1]

   10 firms, 6 countries

   signed the Paris Call on AI safety on Monday

   the US, China, Google and Meta did not [3]

   80

   New York scientists who will test AI systems for the state

   using Empire AI, the state's computing partnership [2]

   What happened

   • New York opened a sign-up site on Tuesday for its new AI safety law. [1] The biggest AI firms must register by 1 January 2027. [1]
   • The law, the RAISE Act, covers AI makers with over $500 million in yearly sales, like OpenAI and Anthropic. [1]
   • They must publish their safety plans and report a serious safety incident within 72 hours. [1]
   • The same day, New York's governor, Kathy Hochul, opened a state centre that will test AI systems itself. [1]
   • More than 80 New York scientists will work on it, using computing power from Empire AI, the state's AI partnership. [2]
   • Hochul said: "The stakes are too high to simply take AI companies at their word." [2]
   • A new state office called DIGIT will run the law and oversee the biggest AI labs. [1]
   • Fines can reach $1 million for a first breach and $3 million for each breach after that. [1]
   • In Paris on Monday, 10 AI companies signed a pledge on the safety of the most advanced AI. [3]
   • Signers promise switches that can shut down their AI models, even during training. [3]
   • They also promise to let qualified outside experts test their systems where appropriate. [3]
   • Governments and companies will share early warnings of serious incidents. [3]
   • On Tuesday, France's president, Emmanuel Macron, called for AI tools to be checked before children can use them. [3]
   • He compared it to the safety checks required for children's toys. [3]
   • Macron told Les Echos, a French newspaper, that uses of AI that put children at risk "must be banned". [4]
   • AI developers including OpenAI and Anthropic, and AI safety groups, pledged 8 million euros for research on AI and children. [5]

   Who is involved

   The New York Center for AI Responsibility and Research is based at Binghamton University, part of the state university. [1]

   They include OpenAI, Anthropic, Microsoft and Mistral. [3] Six countries signed too, among them France, Canada and Norway. [3]

   The US, China, Google and Meta did not sign. [3]

   • OpenAI
   • Anthropic
   • Mistral
   • France
   • Signed OpenAI, Anthropic, Microsoft
   • Did not sign US, China, Google, Meta
   • Kathy Hochul New York's governor; she opened the law's sign-up site and the state AI testing centre
   • DIGIT New York's new Office of Digital Innovation, Governance, Integrity, and Trust; it will enforce the law
   • Emmanuel Macron France's president; he hosted the Paris Call and pushed for checks on AI before children use it
   • CeSIA a French AI safety group; it signed the Paris Call and said it leaves important gaps

   How it unfolded

   California's similar law, SB 53, took effect in January. [1] It gives firms 15 days to report an incident, against New York's 72 hours. [1]

   A law in Illinois, SB 315, starts in 2027 and adds independent safety checks. [1]

   Australia said on Tuesday it will set up an AI safety regulator, after an OpenAI agent got into government data. [7]

   The US government has no such law, and President Trump has repeatedly dismissed calls for mandatory AI rules. [6]

   Democrats take control of the US Congress in January and plan bills requiring independent testing of advanced AI. [6]

   In an exit poll, 56% of US voters called AI and data centres a major factor in their vote. [8]

   1. Tuesday sign-up site opens
   2. 1 Jan 2027 New York's duties begin for big AI makers [1]
   3. November Democrats win Congress
   4. January AI testing bills planned
   5. Tuesday registration opens
   6. 1 January 2027 duties begin
   7. January California's SB 53 takes effect [1]
   8. Mon 9 Nov 10 AI firms and 6 countries sign the Paris Call [3]
   9. Tue 10 Nov New York opens sign-ups and its AI testing centre [1]

   Why it matters

   From January, OpenAI and Anthropic must publish their safety plans for anyone to read. [1] A state regulator must hear of a serious incident within 72 hours. [1]

   How it works

   Lawyers call this a disclosure rule: the law does not say how to build AI, only what firms must tell.

   New York's testers can study only what AI companies make public. [2] The law forces the firms to make more of it public. [2]

   Seen it before

   A teacher cannot mark homework she never sees, so the rule is that homework must be handed in.

   The truth

   • What is really going on New York and France spent the day building checks on AI makers, but both checks rely on what the firms choose to publish or release. [2][3]
   • Why it works on people A list of ten signers sounds like agreement, even when the US government, China, Google and Meta are not on it. [3]
   • The law will not make firms show AI models they have not yet released, Engadget notes. [2]
   • CeSIA, a French AI safety group, signed the Paris Call. [3] It says the pledge sets no shared risk level for deciding whether a model may be released. [3]
   • Its head, Charbel-Raphael Segerie, said: "These decisions must rest with public authorities." [3]
   • Who gains New York's state AI testers — They can study only what AI firms make public, and the law makes the firms publish more. [2]
   • Nobody knows yet

     Whether New York's testers will ever see AI models before they are released.

     The RAISE Act does not force firms to show models they have not released, Engadget notes. [2]

   • Nobody knows yet

     What would count as too risky to release under the Paris Call.

     The pledge sets no shared risk level and no deadlines, and the testing framework it proposes has not been written. [3]

   What comes next

   The next date is 1 January 2027, when registration closes and the law's duties begin. [1]

1. 02

   A report ties $7.5bn of chips to Alice Huang

   What happened

   • Alice Huang helped $7.5 billion or more of chips from Nvidia, the AI chip maker, reach Chinese firms, a report says. [9]
   • The report says she helped move 240,000 banned chips into China, mostly Nvidia's H100 and H200. [10]
   • The chips went through a web of firms tied to Megaspeed and Shanghai Century Litong, it says. [10]
   • Yet Malaysian trade records show Megaspeed imported at least 17,544 Nvidia GB300 chips in November 2025, Bloomberg found. [11]

   How it unfolded

   Nvidia says it banned all sales and shipments to Megaspeed in June 2025. [11]

   Why it matters

   Lawmakers in the US Congress now want to look into Nvidia's ties to China. [12]

   How it works

   Trade officials call this diversion: goods sold to an allowed buyer travel on to a banned one.

   Megaspeed, the buyer on paper, is in Singapore. [11] Some chips it bought reportedly went on to data centres in China. [10]

   Seen it before

   A friend buys a cinema ticket for a younger child who is too young to get in.

   The truth

   • What is really going on Nvidia says it stopped selling to Megaspeed in June 2025, but trade records show Megaspeed imported 17,544 of its chips that November. [11]
   • Nvidia says its ban on Megaspeed still stands. [11] The Next Web, which reported Bloomberg's findings, says it has not checked them itself. [11]
   • Nobody knows yet

     Whether Nvidia chips kept reaching Megaspeed after the June 2025 ban.

     Nvidia says its ban still stands, while Bloomberg cites trade records of 17,544 later imports that The Next Web has not checked. [11]

2. 03

   Anthropic launches AI office tools

   What happened

   • Anthropic, the company that makes the Claude AI, launched tools on Tuesday for spreadsheets, whiteboards and meetings. [13]
   • Claude Meetings lets Claude join a video call and work on tasks while people talk. [13]
   • Claude Sheets builds working spreadsheets, with tabs and formulas, that a team then edits together. [13]
   • Apps from Adobe, Atlassian and HubSpot can now open inside Claude itself. [13]

   How it unfolded

   OpenAI launched similar workplace tools a few weeks ago. [13]

   Why it matters

   OpenAI and Anthropic are both turning their AI into office suites to take on Microsoft and Google. [13]

   How it works

   Economists call this a platform: a product other firms' apps plug into, so users have less reason to leave.

   Anthropic says partners can now embed their apps in Claude, so users work in Adobe without switching tabs. [14]

   Seen it before

   A shopping centre puts many shops under one roof, so visitors never need to go outside.

   The truth

   • The Verge says Anthropic is courting office workers ahead of its first sale of shares to the public. [15]
3. 04

   Microsoft switches on Claude for UK firms

   What happened

   • Microsoft will switch on Anthropic's Claude models by default for UK business users of Copilot from 9 December. [16]
   • Copilot is Microsoft's AI assistant for work, and it offers models from other AI firms. [16]
   • A firm's IT staff can turn Claude off, but leaving the setting alone means saying yes. [16]
   • Claude is already on by default for most of Microsoft's business customers. [16]
   • Anthropic processes this data in the United States, Microsoft's documents say. [16]

   How it unfolded

   The EU, the UK and nearby EFTA countries were the exceptions. [16] The UK is not covered by Microsoft's EU Data Boundary, so it is now being switched on. [16]

   1. Now UK firms choose
   2. 9 December Claude switched on if no choice made

   Why it matters

   UK firms with strict data rules must decide by 9 December whether US processing is allowed. [16]

   How it works

   Behavioural economists call this a default: the choice that is made for you if you do nothing.

   From 9 December, any UK firm whose IT staff have not chosen gets Claude switched on. [16]

   Seen it before

   A phone plan that renews each month unless you cancel it.

   The truth

   • Who gains Anthropic — Its models reach more UK business users through Microsoft, with Anthropic working as a Microsoft subprocessor. [16]
4. 05

   Google opens its strongest AI model

   What happened

   • Google opened Gemini 4 Carbon, which it calls its strongest AI model, to all paying developers and businesses on Tuesday. [17]
   • Its launch price is $2 per million tokens of text read and $10 per million written. [17]
   • That is about a quarter of the price of Anthropic's Claude Fable 5.2, The Verge says. [17]
   • Google is also testing a swarm feature that turns one request into a team of AI agents. [17]

   How it unfolded

   Until now, only chosen cybersecurity defenders could use it. [17]

   Why it matters

   The release puts Google back at the front of AI, The Verge says, after months of a two-way race. [17]

   How it works

   Economists call this introductory pricing: a low starting price to win buyers before it rises.

   Carbon costs a quarter of Fable's price until 30 November. [18] Then Google raises it to $4 and $20, half of Fable's. [18]

   Seen it before

   A gym that is cheap for the first month and costs more after that.

   The truth

   • On some tasks the model with the lower price can cost more, VentureBeat says. [18] It may need more tokens, retries or human fixes. [18]
5. 06

   Google adds two AI chips

   What happened

   • Google unveiled two new versions of its own AI chips, called TPUs, on Monday. [19]
   • The TPU 8tx, designed with AMD, keeps data and model secrets sealed off even while in use. [19]
   • The TPU 8i Max puts 2 of Google's TPU 8i chips in one package, doubling memory and computing power. [19]

   How it unfolded

   Google launched its eighth generation of TPUs in April. [19]

   Why it matters

   Google wants to rely less on other companies' chips to run its cloud. [19]

   Microsoft, Amazon, OpenAI and Anthropic are also designing their own AI chips, TechCrunch notes. [19]

   How it works

   Economists call this vertical integration: a firm makes a key part itself instead of buying it.

   Google's own TPUs run its cloud computing, so fewer of the chips it needs come from other makers. [19]

   Seen it before

   A bakery that grows its own wheat no longer waits on the flour seller.

1. 07

   Ring launches a home drone and a patrol robot

   What happened

   • Amazon's Ring launched Ring Robotics on Tuesday, with a flying indoor camera and a patrol robot. [21]
   • The $250 Ring Drone Cam flies room to room, upstairs too, when an alarm goes off. [20]
   • It can fly for up to 8 minutes before going back to its dock. [20]
   • The $1,000 Ring Rover Cam is a rolling robot that watches over warehouses and factories. [23]

   Who is involved

   Ring is the Amazon-owned company known for its doorbell cameras. [24]

   How it unfolded

   Ring first showed an indoor drone in 2020, but held off selling it until now. [20]

   1. 16 December Rover Cam ships
   2. 27 January Drone Cam due

   Why it matters

   Ring has faced privacy worries over its links with police. [20] This year it cancelled a link-up with Flock Safety, a police technology firm. [20]

   How it works

   Engineers call this offloading: moving a machine's thinking to faraway computers so the machine stays light.

   The drone sends its video to Ring's servers, where an AI tells it where to fly. [23] Without internet, it can only land. [23]

   Seen it before

   A remote-control car stops moving when someone walks off with the controller.

   Why a smart meter can stop working while nothing inside it breaks

   Engineers call it end of life: a connected device lasts only as long as the networks and support it relies on, whose end dates others set.

   The twist

   A smart meter can be in perfect order and still stop working, because part of it is a phone network that someone else is switching off.

   The picture

   Older meters7 million

   vsvs

   Newer meterspart swapped

   Step by step

   1. 1 A connected device needs outside services to do its job
   2. 2 Each service has its own owner and its own end date
   3. 3 When one service ends, the device loses that job
   4. 4 So the device lasts only as long as its shortest-lived service
   5. 5 And the owner sets none of those dates

   The same force, elsewhere today

   Where this chain is also running, in today's other stories.

   • Older Fitbit watches lose app updates

     The watches still run, but the app store they need closes to them on 28 November, so they can never add a new app.

   • Microsoft ends free school email

     Student mailboxes sit on a free Microsoft service, so when the service ends, schools must pay or move every mailbox.

   • Ring launches a home drone

     The drone's sense of where to fly runs on Ring's servers, so without the internet it can only land.

   • Xiaomi leaves older phones behind

     The Xiaomi 12T still works, but its support window has closed, so the maker's new software skips it.

   Where you've seen this

   Online games

   a game that needs its maker's servers stops working when they close, even for people who bought it

   Car radios

   a radio that only picks up an old signal goes quiet when a country switches that signal off

   Medical monitors

   a device that sends readings to a clinic needs the maker's service and updates to keep doing it

   The catch

   Not every switch-off is a trap. Keeping old networks running costs money, and the newer ones often work better.

   And the whole of it

   Most of us own things like this, a meter, a watch, a speaker, that lean on services we never see. The date one of them stops is usually set in a room none of us are in.

   The truth

   • Ars Technica says flying and roving cameras with AI add new privacy risks for homes and businesses. [22]

   What comes next

   The Rover ships on 16 December. [21] The Drone still needs approval from the FCC, the US communications regulator, and is due on 27 January. [22]

1. 08

   UK smart meters face a costly swap

   What happened

   • Between 12 and 13 million UK smart meters could stop working properly. [25] The 2G and 3G phone networks they use close by 2033. [25]
   • A smart meter sends a home's energy readings to its supplier, and early ones did it over mobile networks. [25]
   • Replacing them is likely to cost about £4 billion, a report by Wirepas, a connectivity firm, estimates. [25]
   • About 7 million older meters have the mobile part built in, so the whole meter must go. [25]
   • Each of those costs £300 to £350 to replace, the report says. [25]

   How it unfolded

   The UK government announced the network switch-off in 2021. [25] Meters using 2G and 3G were still fitted until late 2024. [25]

   1. 2021 switch-off announced
   2. Late 2024 last 2G and 3G meters fitted
   3. May 2029 EE starts closing 2G
   4. Spring 2030 Vodafone switches off

   Why it matters

   Homes with a 2G or 3G meter could lose smart readings if theirs is not swapped in time. [25]

   The truth

   • The UK energy department calls the move to 4G a planned upgrade. [25]
   • Who gains Firms that make new smart meter parts — Millions of meters need new parts, and Wirepas, which wrote the report, sells a radio system for them. [25]
   • Who pays UK households — The report says the cost of replacing the meters will feed into their energy bills. [25]
   • Nobody knows yet

     What the UK's smart meter swap will really cost, and who will carry it.

     The £4 billion is one connectivity firm's estimate, and the UK energy department called the change a planned upgrade without giving a figure. [25]

   What comes next

   BT's EE begins closing its 2G network from May 2029, and Vodafone aims for spring 2030. [25]

2. 09

   Older Fitbit watches lose app updates

   What happened

   • Google will stop new app installs and updates on older Fitbit Versa and Sense watches from 28 November. [26]
   • Apps already on the watch keep working, but a deleted app cannot be put back. [26]
   • The change covers 7 models, from the first Versa to the Sense 2. [26]
   • Google's own Fitbit apps and clock faces will still get updates. [26]

   How it unfolded

   ZDNet says Google is slowly winding down features on these watches. [26]

   Why it matters

   Owners keep a watch that works, but it can no longer gain new apps. [26]

   How it works

   Software engineers call this a dependency: a part a product needs that someone else runs.

   The watches need Google's app store to add any app, so closing the store ends that ability. [26]

   Seen it before

   A games console whose online shop closes still plays the discs you already own.

   The watches still run, but the app store they need closes to them on 28 November, so they can never add a new app.

   The idea, in story 07: Why a smart meter can stop working while nothing inside it breaks →
3. 10

   Xiaomi leaves older phones behind

   What happened

   • Xiaomi, the Chinese phone maker, began giving its newest phones, such as the Xiaomi 17, HyperOS 4 in China. [27]
   • Older phones such as the Xiaomi 12T and Redmi Note 12 will not get it. [27]
   • They are at, or close to, the end of their software support. [27]
   • Xiaomi's schedule lists 57 phones and 15 tablets for the update, in waves until January 2027. [27]

   Why it matters

   Owners of phones left off the list stay on older software. [27]

   How it works

   Phone makers call it a support window: the years a maker promises to keep updating a phone.

   The Xiaomi 12T reaches the end of its window this year, so HyperOS 4 skips it. [27]

   The Xiaomi 12T still works, but its support window has closed, so the maker's new software skips it.

   The idea, in story 07: Why a smart meter can stop working while nothing inside it breaks →

   The truth

   • Firmware trackers found cheaper phones, still inside their window, that have not been tested with HyperOS 4. [27] Xiaomi has not said whether they will miss out. [27]
4. 11

   Microsoft ends free school email

   What happened

   • Microsoft is retiring Office 365 A1, its free package of email and Office apps for schools and universities. [28]
   • The change takes effect at each school's first renewal on or after 1 January 2028. [28]
   • Its free replacement, A0, leaves out email, Teams and SharePoint. [28]
   • Free email for graduates is ending too, and Microsoft plans a paid version. [28]

   Who is involved

   The UK's education department said it was disappointed, but won extra time for UK schools. [28]

   • UK education department disappointed, won extra time

   Why it matters

   Another group raised concerns about affordability and fair treatment across the education sector. [28]

   How it works

   Economists call this freemium: a free version wins users, and paid versions earn the money.

   Schools that want to keep student mailboxes must now move to paid licences. [28]

   Seen it before

   A free app starts charging once all your photos are stored in it.

   Student mailboxes sit on a free Microsoft service, so when the service ends, schools must pay or move every mailbox.

   The idea, in story 07: Why a smart meter can stop working while nothing inside it breaks →

   The truth

   • Microsoft says schools now see AI, security and privacy as connected priorities. [28]
   • Who pays Schools, colleges and universities — An industry insider says losing free Microsoft email could add thousands to each one's bills. [28]
5. 12

   AI firms block gadget hacking

   What happened

   • People using Claude and OpenAI's Codex to take apart gadget software are hitting safety blocks. [29]
   • They had been asking the AI to work out how their devices' software works, so they could replace it. [29]
   • Some revived smart-home devices abandoned by their makers, or put new systems on smart TVs. [29]
   • One user who worked on Samsung TV software says OpenAI banned their account for cyber abuse. [29]
   • A site called Decomp.wtf, which made this easier, has shut down. [29]

   Who is involved

   • A Samsung TV tinkerer account banned
   • OpenAI cites cyber abuse

   Why it matters

   Owners lose an easy way to bring dead or limited gadgets back to life. [29]

   How it works

   Security experts call this a dual-use skill: the same know-how can fix a device or attack one.

   One researcher says his blocks match the AI firms' cyber safety rules almost exactly, which he thinks are too strict. [29]

   Seen it before

   A shop stops selling lock picks, so locksmiths lose them along with burglars.

   The truth

   • An OpenAI engineer, Tarun Gogineni, said the blocks are mostly accidental: "it's getting fixed." [29]
   • Nobody knows yet

     Where the AI firms draw the line between fixing your own gadget and hacking.

     One researcher says the firms' safety rules change without warning on the same model. [29]

6. 13

   A $10 tool saves locked ThinkPads

   What happened

   • A YouTube repairer showed how a $10 programmer can unlock old Lenovo ThinkPads locked by their former owners. [30]
   • Companies set supervisor passwords on work laptops, then often sell them off without the passwords. [30]
   • Locked machines sell cheaply and usually end up as electronic waste. [30]
   • The method works on ThinkPads with Intel's 4th to 8th generation chips. [30]

   Why it matters

   Cheap laptops that would have been thrown away can be used again. [30]

   How it works

   Engineers call this a hardware lock: a password stored on a chip, not on the disk.

   The $10 programmer reads and rewrites that chip's memory, so the lock can be removed. [30]

   Seen it before

   Changing a lock's barrel lets a new owner live in a house that came with no key.

   Good news

   • Laptops headed for the dump can now be unlocked and given a second life. [30]
7. 14

   eBay shuts out AI shopping agents

   What happened

   • eBay appears to have blocked AI agents, including Meta's Muse, from browsing its site, The Verge found. [31]
   • A reader's Muse agent had made daily lists of watches for sale on eBay for weeks. [31]
   • This weekend, Muse said it was blocked from the site. [31]

   Who is involved

   An AI agent is a program that does tasks, like shopping, for a person. [31]

   How it unfolded

   eBay's rules already banned buy-for-me agents that order with no human checking. [31]

   1. For weeks Muse lists watches on eBay
   2. This weekend Muse is blocked

   Why it matters

   AI firms sell agents as shoppers, but retailers keep blocking them, The Verge notes. [31]

   How it works

   Economists call this disintermediation: a new go-between steps in between a shop and its buyers.

   eBay is putting its own AI into finding and selling items, its spokesperson said. [31]

   Seen it before

   A shop lets you in, but not the personal shopper you sent instead.

   The truth

   • An eBay spokesperson would not say whether it had added new limits. [31]
8. 15

   ChatGPT gets business pages

   What happened

   • OpenAI launched a Business Hub on Tuesday, where companies can make profiles inside ChatGPT. [32]
   • Firms can check and update facts about their products, much like Google's business profiles. [32]
   • Some can also launch their own agent that answers customers' questions in ChatGPT. [32]
   • OpenAI also bought Odessia, a year-old AI travel agent, and hired its founder, Francis Davidson. [33]

   Who is involved

   OpenAI says 1.2 billion people use ChatGPT each week. [33]

   Why it matters

   ChatGPT is trying to become a place to find products and services, a job Google's search has done. [32]

   How it works

   Economists call a place where buyers and sellers meet a two-sided market.

   OpenAI says ChatGPT has become the platform for people and businesses to connect through AI. [32]

   Seen it before

   A school notice board is useful because pupils read it, so clubs post there.

   The truth

   • OpenAI did not share the details of how it will verify business profiles. [32]
   • Who gains Businesses that verify a ChatGPT profile — They can control the facts ChatGPT shows about their products and see how users interact with them. [32]
9. 16

   Reddit turns posts into product reviews

   What happened

   • Reddit launched Reddit Reviews on Tuesday, using AI to sum up what users say about products. [34]
   • It starts with 10,000 products. [35]
   • Each review shows whether talk is positive, mixed or negative, with links to the real threads. [35]
   • Companies cannot pay to be featured, Reddit says. [35]

   Why it matters

   ZDNet says Reddit Reviews looks like an attempt to compete with Google's AI answers. [35]

   How it works

   Marketers call faking public support for a product astroturfing.

   Turning casual chat into a review page makes Reddit a bigger target for fake posts, TechCrunch notes. [34]

   Seen it before

   A class vote is easy to sway if a few pupils vote many times.

   The truth

   • Reddit says it now labels accounts it detects running automated tools. [35]
10. 17

    Zendesk charges for problems solved

    What happened

    • Zendesk, which makes customer-service software, will charge for problems its AI solves instead of per user. [36]
    • Customers set an Outcome Budget and pay for each issue the AI resolves. [36]
    • Zendesk checks each result for quality, so customers can see what counts. [36]
    • It also launched a public list where brands register their official support channels for AI assistants. [36]

    Why it matters

    Business buyers now want to tie what they pay for AI to verified results, an IDC analyst said. [36]

    How it works

    Economists call this outcome-based pricing: the price follows the result, not the effort.

    Under the new plans, a customer pays only for each issue Zendesk's AI actually resolves. [36]

    Seen it before

    A plumber who is paid only if the leak stops.

11. 18

    AMD details its Zen 6 server chips

    What happened

    • AMD detailed Zen 6, the design inside its new server processors, called Venice. [52]
    • AMD claims an average 27% more work per clock tick across AI and server jobs. [52]
    • In one Tom's Hardware test, a server with 2 of its 256-core chips beat the last generation by 93%. [53]
    • AMD's next chip, Verano, is built to feed its AI graphics chips, to compete with Nvidia's Vera Rubin. [37]
    • Madhu Rangarajan of AMD said: "we didn't magically know four years ago that agentic was going to be a thing." [38]

    Who is involved

    • AMD Verano with AMD's AI chips
    • Nvidia Vera Rubin Nvidia's rival system

    Why it matters

    AMD argues AI agents run many different tasks, which need processors that handle all kinds of work. [38]

    How it works

    Chip designers say performance comes from three things: power, work per clock tick, and clock speed. [52]

    Zen 6 raises all three, with clock speeds from 4.1 to 5.3 GHz. [52]

    Seen it before

    A cyclist goes faster with stronger legs, a bigger gear or quicker pedalling.

12. 19

    Musk's chip factory sets a giant target

    What happened

    • Terafab, a Tesla and SpaceX chip venture, told toolmakers it wants 1 million chip wafers a month after 2032. [39]
    • About 700,000 of those wafers would be memory chips, and more than 300,000 logic chips. [39]
    • That alone would rival Samsung's estimated memory output of 650,000 to 700,000 wafers a month. [39]
    • Its buyers visited Japan to ask toolmakers for delivery within months. [39]

    Why it matters

    Many in the industry doubted Musk's plan in March. [39] Its orders now make it look like real business to more firms. [39]

    The truth

    • A leading chip factory usually takes 3 years or more from planning to full output. [39] Terafab aims for late 2028. [39]
13. 20

    Linux gets more fixes from AI

    What happened

    • Linus Torvalds, who created and still leads Linux, said AI fixes drove much of the latest Linux kernel's progress. [40]
    • In one part, called BPF, AI sent more fixes than bug reports, its maintainer said. [40]

    How it unfolded

    Open-source volunteers have been flooded with AI bug reports, many of them junk. [40]

    Why it matters

    Torvalds said: "we may be starting to see the tide starting to turn." [40]

    How it works

    Software teams call this triage: sorting what comes in so the real problems get fixed first.

    Torvalds wants AI patches welcomed, as long as they face the same quality checks as human ones. [40]

    Good news

    • Other open-source volunteers tired of junk reports would welcome the same shift, Tom's Hardware says. [40]
14. 21

    Microsoft's AI finds an 11-year-old Rust bug

    What happened

    • Microsoft researchers used an AI scanner to find a bug that sat in Rust's standard library for nearly 11 years. [41]
    • The bug could crash a program instead of handling a stack overflow, a kind of memory error, safely. [41]
    • Microsoft found no reliable way to exploit it and judged attacks unlikely. [41]
    • The researchers checked each find themselves and reported them to Rust's maintainers. [41]

    Who is involved

    Rust is a programming language designed to avoid memory mistakes. [41]

    Why it matters

    Even code written in Rust's safe style can hit flaws hidden in the language's own library. [41]

    Good news

    • The flaws are now with Rust's maintainers, who can fix them. [41]
15. 22

    Microsoft fights a ban on Chinese optical parts

    What happened

    • Microsoft is lobbying against a possible US ban on Chinese optical transceivers for data centres, Reuters reported. [42]
    • Transceivers are small parts that turn electrical signals into light for fibre cables. [42]
    • Chinese companies supply 66% of the world market for them. [42]
    • Only 2 US firms, Coherent and Lumentum, make the transceivers data centres need. [42]

    How it unfolded

    1. Now a ban weighed
    2. 2028 fastest kind made in volume

    Why it matters

    The FCC, the US communications regulator, fears some could be used to steal data or disrupt data centres. [42]

    An industry group says a ban could "grind" the US data centre build-out to a halt. [42]

    The truth

    • Microsoft is building 35 US data centres and plans 134 more. [42]
    • Who gains Coherent and Lumentum — As the only US makers of the optical parts data centres need, they would win buyers if Chinese supply were banned. [42]

    What comes next

    One idea floated would ban only the fastest kind, 3.2 terabits a second, made in volume from 2028. [42]

16. 23

    Microsoft bounces small senders' email

    What happened

    • Since 5 November, Microsoft has been bouncing some email sent to Outlook, Hotmail, Live and MSN addresses. [43]
    • The refused mail comes from senders without DMARC, a record that proves a message came from its domain. [43]
    • Small, ordinary senders are now being bounced too. [43]

    How it unfolded

    Microsoft's rule had applied only to bulk senders of more than 5,000 messages a day. [43]

    Why it matters

    Real business mail is being refused along with fakes, Computerworld reports. [43]

    How it works

    Email engineers call this sender authentication: proving who sent a message, so fakes can be refused.

    A DMARC record lets a receiving server check a message's sender, so Microsoft refuses mail without one. [43]

    Seen it before

    A school office that will not open letters without a return address.

    The truth

    • Who pays Small email senders — Orders and replies they send to Outlook or Hotmail users can bounce, with an error message that points at the wrong problem. [43]
    • Nobody knows yet

      Whether Microsoft meant to start refusing small senders' email.

      Microsoft posted no notice, and an analyst says a policy change and a bug may both be at work. [43]

17. 24

    Yandex cannot restore some cloud data

    What happened

    • Yandex, the Russian internet company, says it cannot restore 2.84% of customer data in its cloud. [44]
    • The lost data could be about 100,000 terabytes and tens of millions of files. [44]
    • Customers who used Yandex's cloud backup tools lost nothing. [44]

    How it unfolded

    Ukrainian drones hit 3 of its 6 data centres in October. [44]

    Why it matters

    After the attacks, Yandex moved its 3 supercomputers and withdrew its yearly revenue forecast. [44]

    How it works

    Engineers call keeping copies in separate places redundancy.

    Backed-up data survived the strikes, while data held only in one place was lost with the machines. [44]

    Seen it before

    Homework saved on only one laptop is gone when that laptop breaks.

    The truth

    • Who pays Yandex cloud customers with no backups — Their data sat on equipment too badly damaged to recover. [44]
    • Nobody knows yet

      Which Yandex customers lost data, and how much each one lost.

      Yandex gave only a share, 2.84%, and the 100,000 terabytes is an estimate based on its 2025 results. [44]

    Good news

    • Yandex customers who used its cloud backup tools lost none of their data after drone strikes on its data centres.

    What comes next

    Replacing the damaged equipment could take about 6 months, Kommersant reported. [44]

18. 25

    Russia makes data centres keep room for the state

    What happened

    • Vladimir Putin, Russia's president, signed a decree on 2 November. [45] It makes private data centres reserve 10% of capacity for Russia's government. [45]
    • Russia's digital ministry now proposes that operators disconnect lower-priority clients if they run out of room. [45]
    • Operators would have 24 hours to respond to a request. [45]
    • The space would go to services on government white lists. [45]

    How it unfolded

    It follows Ukrainian drone strikes on data centres of Yandex and Sberbank, a big Russian bank. [45]

    Why it matters

    Russia's government gets first call on space in private data centres, ahead of paying clients. [45]

    The truth

    • Failing to keep the reserve could lead to temporary state management of an operator, sources told RBC. [45]
    • Who pays Businesses renting space in Russian data centres — Being cut off would bring them no compensation, because it counts as force majeure, which frees operators from liability. [45]
    • Nobody knows yet

      How long Russia's rule to keep data-centre space for the state will last.

      The digital ministry calls it temporary but has set no end date. [45]

19. 26

    The EU takes on tricks in online selling

    What happened

    • The European Commission is presenting the Digital Fairness Act, a law against unfair online selling, on Wednesday. [46]
    • It would ban dark patterns: designs that trick people, such as bright buttons that push a purchase. [46]
    • It targets bots that make cancelling a subscription hard, and prices aimed at vulnerable people, including children. [46]
    • Sellers would have to say when a price can change with demand, as with concert tickets. [46]
    • It also covers AI agents that find and buy things for people. [46]
    • Fines could reach 6% of a company's yearly global revenue. [46]

    Why it matters

    It would apply to any business selling online to people in the EU. [46]

    The truth

    • DOT Europe, a tech industry lobby, says existing laws already cover these practices. [46]
20. 27

    IBM ends visa sponsorship in its biggest unit

    What happened

    • IBM will stop visa and green card sponsorship for most roles in IBM Consulting, its largest unit. [47]
    • The unit employs more than 150,000 people, IBM's website says. [47]

    How it unfolded

    IBM said its use of US skilled-worker visas has been falling for years. [47]

    The Trump administration has been pressing tech firms over their use of foreign workers. [47]

    In September 2025, Trump set a $100,000 fee on some H-1B visa applications, now tied up in court. [47]

    Why it matters

    The change lands on foreign workers who need sponsorship to work for IBM's consulting arm in the US. [47]

21. 28

    The EU sues IBM, Atos and Leonardo over border delays

    What happened

    • eu-LISA, the EU agency that runs its border IT, is suing IBM, Atos and Leonardo over delays. [48]
    • The €212 million system takes fingerprints and face images of non-EU travellers. [48]
    • It was due in 2022, went live in October 2025 and fully worked only from April. [48]

    How it unfolded

    The claim was filed in an Estonian court in December 2025. [48]

    1. 2022 due
    2. October 2025 live
    3. April fully working

    Why it matters

    Travellers have faced hours-long airport queues since it launched. [48]

    The truth

    • eu-LISA says the case concerns building the system, not how it runs now. [48]
22. 29

    BT holds back its Starlink deal

    What happened

    • BT, the UK's largest telecoms company, has paused launching its deal with Starlink, Elon Musk's satellite internet firm. [49]
    • BT's boss, Alison Kirkby, told the Financial Times it will wait until the noise has died down. [49]

    How it unfolded

    In October, Musk reportedly ordered Starlink to slow a service used by Ukraine's military. [49]

    1. Now launch paused
    2. Coming months Starlink deal launches

    Why it matters

    BT says the deal would be a giant leap for internet in rural parts of the UK. [49]

    How it works

    Businesses call this reputational risk: a partner's actions can harm your own name.

    Musk's move in Ukraine made BT cautious about how customers would see Starlink, Kirkby said. [49]

    Seen it before

    A school puts off a trip with a coach firm that was just in the news for a crash.

    What comes next

    BT still says Starlink will launch in the coming months. [49]

23. 30

    Apple TV adds sign language

    What happened

    • Apple TV added a sign language option that shows an interpreter on screen. [50]
    • It starts on Friday with Being Heumann, a film about the disability rights activist Judy Heumann. [50]
    • Other films will follow later this year. [50]
    • It is a setting, like captions, not a separate version of the film. [50]
    • Apple uses human interpreters, not AI ones, because that is what Deaf viewers said they need, it says. [50]

    Who is involved

    • Human interpreters used
    • AI interpreters not used

    How it unfolded

    1. Friday Being Heumann with signing
    2. Later this year more films

    Why it matters

    Deaf viewers can switch signing on for the same film everyone else is watching. [50]

    Good news

    • Deaf audiences get signing built into the main film, with interpreters chosen to fit each film's tone. [50]
    • Apple TV now lets Deaf viewers switch on a human sign language interpreter inside the film itself, starting on Friday with Being Heumann.
24. 31

    AI firms' safety group rates jailbreaks

    What happened

    • The Frontier Model Forum released a scale for rating AI jailbreaks. [51] The group was set up by Anthropic, Google, Microsoft and OpenAI. [51]
    • The scale works like CVSS, the system used to rate software security flaws. [51]
    • Each jailbreak gets 4 scores from 0 to 3, such as how much new harm it allows. [51]

    Who is involved

    A jailbreak is a trick that gets an AI model around its safety rules. [51]

    Why it matters

    Teams fixing AI models from different firms could share one ranking of which tricks to block first. [51]

    How it works

    Security teams call this severity rating: ranking flaws so the worst get fixed first.

    A jailbreak that gives nothing beyond what attackers can already get scores 0 for capability gain. [51]

    Seen it before

    A hospital emergency room sees the worst-hurt patients first.

    What comes next

    Comments are open until 21 December, and the group says not to rely on it yet. [51]