Day Lila

Information Technology · Tuesday, 13 October 2026

01 Briefing what happened

Meta releases Muse Wonder, the first model it says matches OpenAI's and Anthropic's best, and keeps the strongest version for paying users

Information Technology 47 sources

Meta's two new AI models score close to the top of an independent ranking, but neither can be downloaded and the stronger one comes only with a $50 or $200 monthly plan. Microsoft, Intel and AMD all showed hardware built to answer AI requests more cheaply.

56 of 58

Wonder Full's score, against the best model Artificial Analysis compared it with

OpenAI's GPT-6 Astra scored 53 and the smaller Wonder 52 [8]

$145bn

the most Meta expects to spend this year on its AI plans

Zuckerberg told the FT that getting here took a large amount of computing and money [5]

$50-$200

a month, the only way to use Wonder Full

it comes only with Muse Code; the smaller Wonder needs a $20 or $100 Muse plan [5][6]

The lead story — what happened

  • Meta, the owner of Facebook, Instagram and WhatsApp, released two AI models on Monday: Muse Wonder and Muse Wonder Full. [1][7]
  • It is the first time Meta has said a model of its own can compete with the best from OpenAI and Anthropic, the makers of ChatGPT and Claude. [7][3]
  • Artificial Analysis, an independent firm that scores AI models, gave Wonder Full 56 on its index. Among the models it compared, only Anthropic's Claude Opus 5.5 scored higher, at 58. [8]
  • The smaller Wonder scored 52, one point behind OpenAI's GPT-6 Astra. [8] Meta's own tests put Wonder Full ahead of rivals on some legal and software-building tasks. [2]
  • Programmers can buy Wonder at $4 for every million tokens it reads and $20 for every million it writes. A token is a piece of a word. [1][5]
  • Wonder Full is not sold that way. It comes only with Muse Code, Meta's tool for writing software, on plans costing $50 or $200 a month. [5][6]
  • People who use Muse, Meta's AI helper app that can work other apps for its users, will need a $20 or $100 monthly plan to get Wonder. [7][6]
  • Neither model can be downloaded. Meta's older Llama models could be run on anyone's own computers, which made them popular with researchers and start-ups. [8][2]
  • Meta says the models remember what matters to a user across conversations. It says they also try other ways when a step in a long task fails. [8][6]
  • The models come from Meta Superintelligence Labs, which Mark Zuckerberg set up about 14 months ago. It is run by Alexandr Wang, who founded the AI start-up Scale AI. [8][7]
  • Inside Meta the models were called Watermelon. In July, Wang told staff they used about ten times the computing of Muse Spark, Meta's earlier model. [9]
  • Meta expects to spend as much as $145 billion this year on its AI plans. [5] Its shares rose as much as 4.7% after the launch. [4]
  • Meta has not said whether it will publish the model's inner numbers for anyone to download. That choice decides how much it owes under the EU's AI law. [4]
Scores on Artificial Analysis's index, which combines ten tests of reasoning, knowledge, coding and work tasks. Higher is better.

Who is involved

  • Meta

    the owner of Facebook, Instagram and WhatsApp; it built the Wonder models and sells them

  • Alexandr Wang

    Meta's chief AI officer, hired in June 2025 from Scale AI, the start-up he founded; he runs the lab that made Wonder

  • Mark Zuckerberg

    Meta's chief executive; he rebuilt the company's AI work last year at a cost of billions

  • Artificial Analysis

    an independent firm that tests AI models and publishes one combined score for each

  • Anthropic and OpenAI

    the makers of Claude and ChatGPT; Wonder is measured against their best models

How it unfolded

  1. Jun 2025 Meta hires Alexandr Wang to lead its AI work [7]
  2. Apr 2026 Meta releases Muse Spark, the first model from its new lab [2]
  3. Jul 2026 Meta pauses one stage of training, which delays Wonder to October [7]
  4. Sep 2026 Muse, Meta's AI helper app, climbs to the top of Apple's App Store [7]
  5. Mon 12 Oct Meta releases Muse Wonder and Muse Wonder Full [1]

Where this points

Watch whether Meta lets anyone download Wonder, which decides what it owes under the EU's AI law, and when Wonder reaches people who use Muse for free. [4][3]

What is pushing on the whole day

The bar and the word are our reading of how hard each one is pushing today. The arrow is where it is heading. The evidence is in the stories below.

The bill for running AI Building↑

Microsoft says its new Maia 300 chip gives 40% more AI output for each unit of electricity than top graphics-chip systems. [11] Intel showed a chip built only to answer AI requests, cooled by air instead of water. [22] Meta charges programmers for every million tokens Wonder writes. [1]

Chip shares falling Building↑

SK Hynix, a South Korean memory-chip maker, fell more than 8% in Seoul on Tuesday. [14] Micron, an American memory maker, fell after China's CXMT showed a new memory design. [13] Broadcom's boss said money may soon limit AI building more than chips do. [15]

Officials checking AI and robotaxis Building↑

Britain's AI Security Institute, a government testing body, says it is testing the newest AI models again. [16] America's road safety agency is reviewing claims that Tesla's robotaxis sped in school zones. [23] A London Assembly report says it has not seen enough evidence to welcome driverless taxis. [24]

People staying away from AI Steady→

The Pew Research Center found that 51% of US adults never use AI chatbots. [25] In a YouGov survey for the London Assembly, 42% of Londoners opposed driverless taxis and 29% supported them. [24]

The rest of the day

30 more stories on this beat.

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

  1. 02

    Microsoft's own AI chip to take over

    Microsoft showed Maia 300 on Monday, the third AI chip it has designed itself. [10] From 2027 it will be the main chip Microsoft uses to answer requests to its AI services, such as Copilot, its AI assistant. [10][12] Microsoft says it produces 40% more tokens per megawatt of power than top graphics-chip systems, but did not say which ones. [11] TSMC, the Taiwanese firm that makes most advanced chips, will build it. [10] Training new models will still use chips from Nvidia and AMD. [12]

    Why it matters — Nvidia's chips fill Microsoft's data centres today, and every request Microsoft moves to its own chip is one Nvidia no longer serves. [12] Analysts at Evercore ISI estimate Maia will be about 10% of Microsoft's AI chip spending by early 2027. [12]

  2. 03

    China's CXMT shows a denser memory chip

    CXMT, a Chinese maker of computer memory chips, says it has made major progress on a design that fits more memory on each chip. [13] Each memory cell takes up less room because a new upright switch, which CXMT calls VSG, sits inside it. [13] Its president, Cao Kanyu, said a server memory product using the design should arrive by the end of 2026. [13] Samsung, SK Hynix and Micron, the world's biggest memory makers, are working on the same kind of design. [13]

    Why it matters — Shares in Micron, an American memory maker, fell 3.9% on Monday, CNBC reports, while Mint put the fall at 5%. [14][13] On Tuesday SK Hynix fell more than 8% and Samsung 6.7% in Seoul, their lowest since August. [14]

  3. 04

    Chip bosses say money may run short

    Hock Tan, the head of Broadcom, which helps companies such as OpenAI design their own chips, said money could become the biggest limit on building AI. [15][11] He said it could matter more than shortages of chips, memory or electricity. [15] Lip-Bu Tan, Intel's chief executive, said the cost of borrowing and doubts about returns are the main risks. [15] Both spoke on Monday at a conference run by DBS, a Singapore bank. [15] Hock Tan still expects demand for computing to beat supply through 2029. [15]

    Why it matters — Data centres take years to pay for themselves, so higher interest rates make each one harder to fund. [15] Hock Tan said demand could also slow if businesses do not find AI useful enough. [15]

  4. 05

    Britain gets to test new AI models again

    Britain's AI Security Institute, a government body that tests powerful AI before release, is now testing the newest models, including Google's Gemini 4 Argon, the AI minister said on Monday. [16] Reports last month said the Trump administration had asked OpenAI and Anthropic to hold new models back from British testers until US officials had checked them. [16] Anthropic had told the institute that nobody outside the US could use its Mythos 5.1 model. [16] Minister Kanishka Narayan said he was very confident the access would continue. [16]

    Why it matters — Britain has no big AI developer of its own, so these voluntary deals are its only early look at a model. [16] Narayan also set up a team to protect British government computers from attacks that use AI. [17] In one trial, AI models found 407 weak spots in nine government bodies' code in a month. [18]

  5. 06

    AMD designs a rack for AI helpers

    AMD, which makes computer processors and AI chips, showed a design called Helion on Monday at the Open Compute Project summit in San Jose, where firms share data-centre designs. [21] It is a blueprint for a full rack of computers holding up to 32,768 processor cores. [19] It is built to run AI agents, programs that carry out tasks for people, and not to train models. [20] Each agent gets its own slice of computer to work in, while the heavy AI answering runs on separate machines. [19]

    Why it matters — Meta's Muse and OpenAI's Dots agents already work this way, so every new agent user needs ordinary processor power as well as AI chips. [19] Cisco, Dell, HPE, Lenovo and Supermicro are among the firms AMD says will build it. [21]

  6. 07

    Intel names its answering chip

    Intel gave its Crescent Island chip a product name on Monday: the Intel GPU D-Series. [22] It is built only for inference, the work of running a trained AI model to answer requests. [22] It uses 350 watts and is cooled by air, so it fits ordinary data-centre racks. [22] On paper, each of Nvidia's newest Rubin chips does about 13 times as much work, but needs liquid cooling. [22] Intel is sending samples to customers and has not given a launch date. [22]

    Why it matters — Intel's last family of data-centre AI chips, code-named Ponte Vecchio, was a failure. [22] A chip that fits the air-cooled racks companies already own is a cheaper first step for buyers who cannot rebuild their data centres. [22]

  7. 08

    US reviews Tesla robotaxi speeding claim

    NHTSA, America's road safety agency, is reviewing a complaint that Tesla's driverless taxis broke the speed limit in school zones while the lower school-zone limits applied. [23] It covers about 1,000 vehicles, and no crashes or injuries are reported. [23] The complaint came with three video links. It arrived on 23 September, and the review opened on 9 October. [23] The review only decides whether to start a full investigation, and it has no deadline. [23]

    Why it matters — Speeding is also what is holding up Tesla's driver-assistance software in Europe. [23] Germany wants it kept to 10% over the limit, and an EU vote on approving it has slipped to December at the earliest. [23]

  8. 09

    London Assembly doubts driverless taxis

    A London Assembly committee, which checks the work of London's mayor, said on Monday it had not heard enough evidence to welcome driverless taxis. [24] Waymo, Google's robotaxi company, Wayve, a British start-up, and Baidu, a Chinese tech firm, all plan services in London. [24] The report found that Transport for London cannot suspend a robotaxi permit once a national agency issues one. [24] It wants the city to be able to cap the number of robotaxis and suspend permits. [24]

    Why it matters — London has about 105,000 licensed private-hire drivers, and their union, the IWGB, wants the rollout stopped. [24] The committee instead wants drivers who lose work retrained and paid for. [24]

  9. 10

    Half of US adults never use chatbots

    The Pew Research Center, a US research group, found that 51% of adults in the US never use AI chatbots. [25] About three in four people aged 65 and over never use them, and nor do 58% of those aged 50 to 64. [25] Even among 18- to 29-year-olds, a third never use them. [25] Six in ten non-users said they are simply not interested, and two thirds said they are unlikely to start within a year. [25]

    5 of 10

    about 5 in 10 US adults never use chatbots

    Pew's figure is 51%, drawn here as about half of ten people.

    Why it matters — Companies like Meta are spending up to $145 billion a year on AI. [5] This survey shows half of US adults have not started using it, and most of them do not plan to. [25]

  10. 11

    OpenAI adds 200 jobs in Dublin

    OpenAI, the maker of ChatGPT, said on Monday it will add 200 jobs in Dublin, home of its European headquarters. [26] That brings its planned hiring in Ireland to 450. [26] It is also opening a lab there where engineers help customers build with its models, its second after one in Singapore in May. [26] The news came two days before Dublin hosts an international AI summit, during Ireland's turn chairing the EU's Council of ministers. [26]

    Why it matters — OpenAI's Irish office is what makes Ireland's data regulator its main watchdog across the EU. [26] In March a court in Rome cancelled Italy's EUR15 million fine on OpenAI because Ireland's regulator had taken over. [26]

  11. 12

    AI tool finds homes from one room photo

    Graylark Technologies, a US company, has added a feature to its AI tool Raven that lets police find the exact home an indoor photo was taken in, 404 Media reports. [27] In a demonstration, a photo of a washer and dryer led to one Airbnb listing in Charlotte, North Carolina. [27] The tool appears to match photos against rental and property sites such as Airbnb, Vrbo and Zillow. [27] The Los Angeles police and the Miami-Dade sheriff's office have bought access to Raven. [27]

    Why it matters — Anyone whose home has been photographed for a rental or sale listing can now be found from one picture of a room. [27] Graylark's chief says the tool helps investigators in trafficking cases. [27]

  12. 13

    A researcher says Jev's idea was his

    Jev, from the start-up TypeSafe, is an AI model that answers yes-or-no and multiple-choice questions about data with a probability, instead of writing text. [28] Nandakishor Mukkunnoth, founder of ConvAI Innovations, says he designed the same approach a year earlier, in a paper published in March 2025. [28] He has since released his own free version, called Laya. [28] Within weeks of Jev, Microsoft, OpenAI, Cloudflare and Perplexity released similar tools. [28]

    Why it matters — These models are fast and cheap because they skip writing text, which is why so many firms copied the idea so quickly. [28] Mukkunnoth says his work was overlooked because it solved one sales problem, while TypeSafe sold a finished product with its own servers and prices. [28]

  13. 14

    AI agents rebuild a 2009 game's code

    Maurice Heumann, a developer, used AI agents from Anthropic and OpenAI to rebuild the source code of Call of Duty: Modern Warfare 2, a 2009 shooting game. [29] Source code is the readable version of a program, which its makers keep private. The work took three months and more than 500 billion tokens. [29] Almost all of the game's functions now exist, and 83% match the original byte for byte. [29] The agents' first month of code looked tidy but did the wrong things, until he added an automatic check. [29]

    Heumann's estimate of what he paid for three months of agent work, and what the same tokens would come to at per-token prices.

    Why it matters — He paid flat monthly plans, about EUR1,200, for work he says would cost $200,000 to $250,000 at per-token prices. [29] The game's copyright owner has already had one of his videos taken down. [29]

  14. 15

    Laptop sales may fall in 2027

    TrendForce, a market research firm in Taiwan, expects laptop shipments worldwide to fall by a few percent in 2027. [30] The processor, memory and storage now make up about 68% of the parts cost of a $900 laptop. [30] If makers pass those costs on to buyers, TrendForce says the fall could reach the high single digits. [30] It also expects the share of laptops made outside China to drop from 24% in 2025 to possibly under 20% in 2027. [30]

    • Processor, memory and storage68 · 68%
    • Everything else32 · 32%
    TrendForce's estimate of the parts cost of a $900 laptop, in percent.

    Why it matters — Memory stays short because AI data centres buy so much of it, and laptop buyers pay part of that bill. [30] Dearer parts also make it less worth moving laptop factories out of China. [30]

  15. 16

    GlobalFoundries returns to smaller chips

    GlobalFoundries, a US chipmaker, announced FDX Fusion, a new way of making chips that it plans to start using in Dresden, Germany, in 2028. [31] It fits more than twice as many transistors in the same space as its 2015 process. [31] It is meant for chips in cars, sensors and robots, which need low power more than top speed. [31] BMW, Bosch, Infineon, NXP, Siemens and Tesla have shown interest, but only Infineon, a German chipmaker, has confirmed it will work on it. [31]

    Why it matters — GlobalFoundries gave up on its most advanced chips in 2018. [31] One chip that can sense, compute and send radio signals could replace several separate chips in a car. [31]

  16. 17

    Intel to drop a chip security feature

    Intel said on Friday it will stop including SGX, a feature that walls off secret data inside a program, after its coming Diamond Rapids server chips. [32] It will focus instead on TDX, which walls off a whole virtual computer, a computer simulated in software. [32] That lets existing programs be protected without being rewritten. [32] Intel will keep fixing SGX on current chips into the early 2030s, and says AI agents handling company data are one reason for the switch. [32]

    Why it matters — Companies that built software around SGX now have several years to move it. [32] TDX is already offered by big cloud providers, including Alibaba Cloud. [32]

  17. 18

    Microsoft change locks admins out of Exchange

    Administrators of Exchange Online, Microsoft's business email service, could not reach its control page for several hours on Monday. [33] The trouble began at 09:37 UTC. [33] Microsoft said a recent configuration change caused the errors, and undoing it brought the page back. [33] It promised a first report on what went wrong within two working days. [33]

    Why it matters — While the page is down, companies cannot manage their staff's mailboxes or groups. [33] The Register lists similar Microsoft outages caused by settings changes in September, when SharePoint sites failed, and in April, when Outlook for iPhone broke. [33]

  18. 19

    Windows 11 setup shortcut closed

    Microsoft has closed a shortcut that let people set up Windows 11 without a Microsoft account, The Register reports. [34] Clicking 'Learn More' on the sign-in screen used to skip straight to making a local name and password. [34] A Windows watcher known as Bob Pony reported that the fix was made on Microsoft's own servers. [34] The Home and Pro versions for personal use require an account during setup. [34]

    Why it matters — An account ties a computer to Microsoft's online services, such as OneDrive storage. [34] People who do not want that now need harder, unofficial ways round it. [34]

  19. 20

    Records show how UCLA settled with Oracle

    UCLA, the University of California, Los Angeles, settled a dispute with Oracle, a big business-software company, over a system for running its finances and grants, records show. [35] The project, called Ascend, was budgeted at $120 million. UCLA stopped paying in September 2024, saying the software did not do what was promised. [35] UCLA will pay $350,000 of $1.66 million in unpaid bills, and Oracle drops the rest. [35] Oracle also gives $3.1 million in credits for future Oracle services. [35]

    • Paid by UCLA$0.35m · 21%
    • Dropped by Oracle$1.31m · 79%
    How the $1.66 million in unpaid bills was settled, in millions of dollars.

    Why it matters — The credits can only be spent with Oracle, and they expire in May 2028. [35] The deal was meant to stay confidential, and became public only through a request under California's public records law. [35]

  20. 21

    Copilot app now makes Office files

    Microsoft's Copilot app can now create a Word or other Office document from a request, and open it for editing while the chat carries on. [36] Other people can be invited to change the document by asking Copilot. [36] For now it reaches only customers in Microsoft's Frontier early-access group, and files are saved to OneDrive. [36] At the same time Microsoft is removing Themes by Copilot from Outlook, a feature it launched less than two years ago. [36]

    Why it matters — Businesses that keep their files in SharePoint, Microsoft's other storage system, may hold back because the feature saves to OneDrive. [36] The Register says Microsoft is quick to drop features once its attention moves on. [36]

  21. 22

    Nigeria's OPay files to list in New York

    OPay, a payments and lending company based in Lagos, Nigeria, has filed to sell shares in New York, probably next month, under the name OPAY. [37] Its revenue rose 137% to $467 million in the first half of 2026, and lending now earns more than payments. [37] Earlier reports valued it at around $4 billion. [37] Standard Bank, Africa's biggest lender by assets, agreed to invest up to $200 million. [37]

    Why it matters — It would be the first US listing by a finance-app company whose customers are mostly in Nigeria. [37] Its filing warns that shares later sold in Lagos may not be swappable for the New York ones. [37]

  22. 23

    US Air Force buys weather-balloon data

    WindBorne Systems, a California start-up, won a US Air Force contract worth up to $49 million for weather readings from its balloons. [38] It flies about 700 balloons around the world, and their readings also feed its own AI weather model. [38] The Air Force pays only for data it orders, and has ordered $1.7 million so far. [38] The contract was given without competing bids. [38]

    Why it matters — WindBorne says the US government chose it because no other company could gather global data the same way. [38] It is one of the first companies to feed its own measurements straight into an AI forecasting model. [38]

  23. 24

    Google's Fitbit Edge can be repaired

    Google launched the Fitbit Edge on Monday, a $179.99 fitness band with a small screen. [39] It sits between the screenless Fitbit Air and the Pixel Watch. [39] It is the first Fitbit Google has made to be repaired, and it costs $20 more than the similar Fitbit Charge 6, ZDNet reports. [40] Its AI health coach, built on Google's Gemini, needs a paid Google Health Premium plan. [39] It reaches shops on 27 October. [39]

    Why it matters — The battery lasts seven days, but only two with the screen always on. [40] The main new feature, the AI coach, needs a monthly fee on top of the $179.99 band. [39]

  24. 25

    Google takes a stake in Flo

    Google has invested in Flo Health, the maker of a popular period-tracking app, the two said on Monday, without saying how much. [41] From early 2027, Flo will connect to the Google Health app, and Fitbit readings such as sleep and skin temperature will feed Flo's cycle predictions. [41] They also plan joint devices, and research into early signs of menopause and conditions such as PCOS, a common hormone disorder. [41]

    Why it matters — The companies say data will move between the apps only with permission for each type, and will not be sold or used for ads. [41] Flo settled with the US Federal Trade Commission in 2021 over claims it shared users' health data with Facebook and Google. [41]

  25. 26

    Abu Dhabi works on a 'health age' score

    Ultrahuman, an Indian maker of smart rings based in Bengaluru, signed a two-year agreement with the Abu Dhabi Public Health Center. [42] Together they will build a model that estimates how fast a person is ageing, from medical tests and data from wearable devices. [42] The aim is to offer it to people across the United Arab Emirates through government health apps. [42] Ultrahuman has sold about 800,000 rings. [42]

    Why it matters — It is the company's first try at running a health formula for a whole country. [42] Ultrahuman says the model will take data from any ring, watch or strap, not only its own. [42]

  26. 27

    OnePlus 16 launches only in China

    OnePlus, a Chinese phone maker, launched the OnePlus 16 in China on Monday. [43] It is its first top phone since the company left the US and European markets this summer. [43] It has a 9,000mAh battery, one of the largest in any top Android phone, and a screen that refreshes 185 times a second. [43] Its main camera rises to 200 megapixels. [43]

    Why it matters — Buyers in the US and Europe can no longer buy OnePlus's best phone from the company. [43] The big battery and fast screen are aimed at people who play games on their phones. [43]

  27. 28

    Sorkin's Facebook film flops

    The Social Reckoning, Aaron Sorkin's film about the Facebook whistleblower Frances Haugen, took $5.5 million in its first weekend in US cinemas, Variety reports. [44] It cost $30 million to $42 million to make. [44] Sorkin's The Social Network, about Facebook's start, opened with $22 million. [44] Reviews have been mixed, with a score of 55% on Rotten Tomatoes. [44]

    First-weekend takings in US cinemas for Sorkin's two Facebook films, in millions of dollars.

    Why it matters — Columbia Pictures, part of Sony, stands to lose money on it. [44] A horror film, Other Mommy, took $27.5 million the same weekend. [44]

  28. 29

    Chart of vanishing Russian bots misread

    After Ukrainian drones knocked out Yandex data centres, a chart citing Cloudflare data spread online showing a sharp fall in bot traffic. [45] Many read it as Russian propaganda bots going quiet. [45] Antibot4navalny, a group that tracks Russian troll networks, says it measured Yandex's own web crawlers, programs that read websites for its search engine. [45] Another chart, shared by a European Parliament member, may be AI-made: it shows data for 15 October, a date that had not come yet. [45]

    Why it matters — Most Yandex services were still down, including its payment service, Yandex Pay, the head of Yandex's finance arm said. [45] A computer scientist, Raphael Wimmer, asked politicians and journalists to stop sharing unchecked images. [45]

  29. 30

    Sandi Toksvig backs a Wikipedia rival

    Sandi Toksvig, a British comedian and writer, is raising money for Mappa Mundi, a digital atlas of knowledge she is building with the University of Cambridge. [46] She told the Cheltenham literature festival this weekend that Wikipedia is mostly edited by white, middle-aged European men. [46] By the Wikimedia Foundation's own estimate, only 19% of Wikipedia's biographies are about women. [46] Wikipedia holds 65 million articles in more than 300 languages. [46]

    Why it matters — Toksvig points to Donna Strickland, a physicist whose draft Wikipedia page was turned down until she won a Nobel prize. [46] Wikipedia is the largest store of knowledge ever put together, so any rival starts far behind. [46]

  30. 31

    Raspberry Pi's desktop returns for old PCs

    Raspberry Pi, the maker of cheap small computers, has released a new Raspberry Pi Desktop for ordinary PCs and older Intel Macs. [47] It is the first new version since 2022. [47] It is a light version of Linux, a free operating system, based on Debian 13. [47] Updates for the previous version ended on 31 August. [47] It runs only on 64-bit machines, and not on Apple's newer Mac chips. [47]

    Why it matters — It gives older computers a simple, free system to keep running on. [47] Existing users have to install it fresh, because the old version cannot be upgraded in place. [47]

02 Lesson why it matters

Why AI chips are now built for answering, not for learning

A model learns once but answers every time someone asks, so the answering bill grows with each user and now shapes the chips.

The twist

Training a model is paid for once, but every answer is paid for again, so a chip that gives 40% more answers for the same electricity saves money on every question for years.

The picture

Training is paid for once. Answering is paid for again on every question, which is why Microsoft judges Maia 300 by answers per megawatt.

How it works

  1. A model is trained once, on a big batch of chips
  2. Then it answers every time someone asks it something
  3. More users and longer agent tasks mean more answers every day
  4. So the cost of answering, in chips and electricity, keeps growing
  5. Companies build chips and racks just for answering, judged by output per watt
  6. And they charge per token, keeping the biggest model for those who pay most

The same force, elsewhere today

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

  • Microsoft's Maia 300

    Microsoft built it mainly to answer requests and judges it by output per megawatt, while it still trains models on Nvidia and AMD chips.

  • Intel's GPU D-Series

    Intel made a chip only for answering, cooled by air, so it can go into racks companies already own.

  • AMD's Helion rack

    AMD gives each AI agent cheap processor space for its own work and sends the heavy answering to separate machines.

  • Meta's Wonder prices

    Meta charges per million tokens written, and keeps the biggest Wonder for people paying $50 or $200 a month.

Where you've seen this

Taxis

a taxi firm picks cars by fuel used per mile, because a taxi drives all day, every day

Light bulbs

buildings that keep lights on all day switch to LEDs first, because LEDs use far less electricity every hour

Airlines

an airline chooses planes by fuel burned per seat, since it pays for fuel on every flight for decades

The catch

This only pays off while use keeps growing: if fewer people use AI than companies expect, chips built for a flood of questions sit idle.

And the whole of it

Each question typed into a chatbot is answered by a chip in a data centre that the person asking never sees. The person sees a reply in a second. The company sees one more answer to pay electricity for, among millions that day.

03 Truth what's really going on

What is really going on

Microsoft, Intel and AMD all showed hardware on Monday that is built to answer AI requests more cheaply, not to train bigger models. [10][22][19] Meta released its strongest models yet and charges for them, keeping the most powerful version for programmers who pay $50 or $200 a month. [1][5]

Why it works on us — A league-table number such as 56 against 58 feels like a plain fact about which AI is smarter, but it is one firm's average of ten tests it chose. [8]

Who gains

  • Meta — Charging $20 to $200 a month for Wonder gives it money from AI that does not depend on its advertising. [6]
  • Microsoft — Using its own chips has helped its profit margins over a decade, its finance chief Amy Hood said, and Maia moves more of its AI work onto them. [12]
  • TSMC — The Taiwanese chipmaker will build Maia 300, so Microsoft's move away from Nvidia still brings it the order. [10]
  • Police forces using Raven — The Los Angeles police and the Miami-Dade sheriff can now try to place a room photo at an exact address. [27]
  • Britain's AI Security Institute — Early access to new models keeps it a tester the biggest AI firms work with. [16]

Who pays

  • Nvidia — Each request Microsoft moves to Maia is one its graphics chips no longer answer. [12]
  • Shareholders in memory-chip makers — SK Hynix fell more than 8% and Samsung 6.7% on Tuesday after news of a Chinese rival's new design. [14][13]
  • Laptop buyers — Processor, memory and storage make up about 68% of a $900 laptop's parts cost, and makers may pass more of it on. [30]
  • London's 105,000 private-hire drivers — Waymo, Wayve and Baidu all plan driverless taxis in the city, and the drivers' union says their work is at risk. [24]
  • People with homes in rental or sale listings — Their rooms' photos can now be matched to an address by a police tool. [27]

What nobody knows yet

Open questions from across today’s stories — ours included.

  • 01

    Whether Wonder is really cheaper than its rivals.

    Meta's Alexandr Wang told CNBC it is priced well below OpenAI's and Anthropic's leading models. [6] TechCrunch says Wonder's $4 and $20 per million tokens is roughly the price of GPT-6 Sol or Claude Opus 5.5. [1] Neither side published a full price comparison.

  • 02

    Whether Meta will let anyone download Wonder.

    Meta has not said. [4] The answer decides how much it must report to the EU's AI Office, and Meta has declined to sign the EU's code of practice for the most powerful models. [4]

  • 03

    Which systems Microsoft measured Maia 300 against.

    Microsoft claims 40% more output per megawatt than top graphics-chip systems but did not name them. [11] An adviser, Patrick Moorhead, said nobody outside Microsoft has tested the chips yet. [12]

  • 04

    How many Maia chips Microsoft will actually use.

    Microsoft promises deployments measured in gigawatts from 2027, but a manager declined to give a timeline. [12] TNW says that could mean more than a million chips. [11] Evercore ISI expects Maia to be about 10% of Microsoft's AI chip spending by early 2027. [12]

  • 05

    When AMD's Helion racks can be bought.

    The Register and TNW say partner systems are due in November 2026. [21][20] Tom's Hardware says the rack will be available from the first half of 2027. [19]

  • 06

    Whether Britain's early look at AI models will last.

    Ministers never confirmed the access had been lost, and the minister himself said reliance on another country's technology makes any deal less permanent. [16] Anthropic had said no one outside the US could use its Mythos 5.1 model. [16]

  • 07

    Whether Tesla's robotaxis really sped in school zones.

    The evidence so far is three video links from one unnamed person, and the US safety agency has set no deadline to decide. [23]

  • 08

    Whether CXMT's new memory design works at scale.

    The claims come from the company's own president at an industry event; the first product is not due until the end of 2026. [13]

  • 09

    Who first came up with Jev's design.

    A researcher says his March 2025 paper had the same idea a year earlier. [28] Most of the copies are built on the same few small open models, which makes it hard to trace who borrowed from whom. [28]

04 Hope carry this

Britain pointed AI models at its own government computer code for a month and found 407 weak spots across nine government bodies, for 13,000 pounds in computing costs. Britain's government says two such trials have now found and fixed more than 1,000 possible flaws.

Also true today

  • Google's new Fitbit Edge is the first Fitbit it has made to be repaired.
  • Raspberry Pi released a free, light desktop system that lets ordinary PCs and older Intel Macs keep running.

Across the beats