Information Technology · Thursday, 3 September 2026

01 Briefing what happened

Google was ruled an illegal ad monopoly in April 2025. A US judge has now decided it sells nothing.

Information Technology 1 min 55 sources

A US federal judge accepted that Google's advertising business is an illegal monopoly and still refused to break it up. The rules Google must follow instead are sealed for 14 days.

20%

of every ad sale, paid to Google by the publisher using AdX

AdX is the exchange the US government asked the court to make Google sell [1][4]

3 in a row

US attempts to break up a large technology company, all refused

Meta kept Instagram and WhatsApp, Google kept Chrome, and now Google keeps AdX [1]

14 days

that the order stays sealed while both sides ask for redactions

so the outcome is public today and the obligations are not [2][3][6]

4.1%

of Google's revenue came from Ad Manager in 2020, the newest public number

later figures were blacked out of the court file [1]

The lead story — what happened

A US federal judge, Leonie Brinkema, refused to make Google sell AdX, the auction at the centre of its advertising business. [1][2]
The same judge ruled in April 2025 that Google holds illegal monopolies in two advertising markets, and that it unlawfully locked publishers into AdX. [1]
AdX runs the instant auction that fills the ad slots on a web page. Publishers pay Google a 20% fee on every sale. [1][4]
The US justice department and a group of states sued in 2023 and asked for the sale. They argued Google cannot be trusted to run AdX. [1][4]
Google argued a forced sale would be technically difficult and cause a long, painful transition that would hurt its customers. [1]
The judge instead adopted most of the behaviour changes the two sides had proposed, with her own modifications. [3]
Nobody outside the case knows what those changes are. Her order is sealed for 14 days so both sides can ask for redactions. [2][3][4][6]
Reported possibilities include stopping Google favouring itself in its own auctions, and giving rival tools the live data Google sees. [3]
This is the third time in a row a US judge has refused to break up a large technology company. [1]
In the other two, one judge let Meta keep Instagram and WhatsApp, and another let Google keep Chrome, pointing to new competition from AI firms. [1][2]
The newest public figure for the size of the business is old. Google's Ad Manager was 4.1% of revenue and 1.5% of operating profit in 2020. [1]
Google's vice-president for regulatory affairs, Lee-Anne Mulholland, said the company was pleased the court rejected breaking apart its tools. [4][5]
Once the final opinion is published, Google can appeal the underlying finding that it is a monopoly at all. [3]

Who is involved

Leonie Brinkema

the US federal judge in Alexandria, Virginia who tried the case; she found the monopoly and refused the sale [1][3]
Google

the world's largest seller of online advertising; it keeps AdX and must change how it runs it [2]
The US justice department

the US government's law-enforcement arm; it sued in 2023 with a group of states and asked for the sale [1][4]
Web publishers

the news sites and blogs that sell ad space through AdX and pay Google 20% on each sale [1][4]
Lee-Anne Mulholland

Google's vice-president for regulatory affairs; she said the company was pleased with the ruling [4][5]

What is pushing on this

US courts' appetite for breakups Easing↓

three refusals in a row, this one included [1]

Google's legal exposure Easing↓

it can now appeal the monopoly finding itself [3]

Pressure on the US justice department Building↑

it won the argument about the law and lost the fix it asked for [2][5]

Publishers' position Steady→

they still pay the 20% fee and cannot read the order [1][4]

How it unfolded

2023 the US justice department and a group of states sue Google over its advertising technology
April 2025 the judge rules Google holds illegal monopolies in two advertising markets
Last year a separate trial is held on what should be done about it
Wednesday the judge refuses the sale and adopts behaviour rules instead
In 14 days the sealed order becomes public, and Google can begin an appeal

Where this points

The next real test is what the unsealed order actually requires - watch whether it forces Google to hand rival tools the same live auction data it can see itself.

The rest of the day

40 more stories on this beat.

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

02
Anthropic ships a cheaper, less restricted model

Anthropic released Fable 5.1 and Mythos 5.1 on Tuesday, the same model in two forms. Reading from cached context costs 75% less, and customers can now keep monitoring data on their own machines. [7][8]

Why it matters — Cheap cached context is what makes an assistant that runs all day affordable to leave running.
03
Google's third Flash model in six weeks

Google released Gemini 3.8 Flash on Wednesday. It costs $0.75 per million tokens of text put in - a token is about three-quarters of a word - until the end of the year, against a list price of $1.50. [9]

Why it matters — Google has not shipped a top-tier Pro model since early 2026, and the introductory price outlives most models.
04
Meta releases Muse Spark 1.3

Meta released Muse Spark 1.3 and has not decided whether to publish its weights, the numbers that let anyone run a model themselves. Its chief AI officer, Alexandr Wang, called it competitive with Anthropic's and OpenAI's best. [10][11]

Why it matters — In Europe the weights decision is also a legal one: the AI Act excuses genuinely open models from part of its rules.
05
OpenAI says its next model can break in

OpenAI said Astra is the first of its models to reach what the company calls a critical level of cyber ability - it can find and exploit unknown software flaws without a person guiding it. Those abilities go only to a private group of testers at launch. [12][13]

Why it matters — OpenAI graded itself against its own framework, and did not name the testers or say how they were chosen.
06
Microsoft will publish Azure's revenue

Microsoft said on Wednesday it will report quarterly sales for Azure, its cloud computing business, for the first time, and will cut its reporting segments from three to two. [14]

Why it matters — Amazon has published its cloud revenue since 2015; Microsoft chose this quarter to start.
07
Microsoft 365 outage ran into a second day

Email, search, file sync and Teams calendars failed across Microsoft 365 from 31 August into 1 September. Microsoft traced it to a misconfiguration in a shared sign-in system used by several services. [15][16]

Why it matters — One shared sign-in component took down services that look entirely separate to the people using them.
08
Telstra names the cause of its July outage

Australia's largest telecoms company said an outside investigation found a fault in its network timing system caused the mobile outage in early July. [17]

Why it matters — Timing is the invisible layer a phone network runs on; when the clocks disagree, calls fail.
09
Chinese court freezes Nexperia's own units

A court in Dongguan froze Nexperia's shareholdings in four of its Chinese subsidiaries, worth about $300 million, until August 2029. Wingtech, the Chinese group stripped of control of the Dutch chipmaker last year, brought the case and wants 8 billion yuan. [18][19]

Why it matters — Nexperia makes the basic chips inside cars, and it cannot sell or restructure those units while the freeze holds.
10
Opera loses its EU case over Edge

The EU's General Court rejected Opera's challenge to the European Commission's decision not to treat Microsoft's Edge browser as a gatekeeper under the Digital Markets Act. [20]

Why it matters — Edge passed the user-number threshold and was excused anyway, on Microsoft's argument that it is not an important gateway.
11
ChatGPT falls under the EU's strictest rules

The European Commission designated ChatGPT a very large online search engine, and Reddit and Roblox very large online platforms, under the Digital Services Act. Each passed 45 million monthly users in the EU, and ChatGPT has four months to comply. [21][22]

Why it matters — It is the first AI chatbot brought inside the EU's toughest tier of online-harms rules.
12
Lawsuit over secret US AI review rules

The group Protect Democracy sued to force the US government to publish how it reviews frontier AI models before release. It says the framework's text, its participants and its legal basis are all unpublished. [23]

Why it matters — A safety review nobody can read tells you only that somebody with the power to approve did.
13
New US rule aimed at renting chips remotely

The US is preparing an export rule to stop Chinese firms renting advanced AI computers remotely through Thailand and Singapore, which are not covered by the current controls. It could be shown to industry as soon as September. [24]

Why it matters — Rules about shipping chips do not cover logging in to them from somewhere else.
14
Taiwan's six-year hunt for hidden chip labs

Rest of World traced Taiwan's investigations into Chinese chip firms operating behind locally registered front companies. A 2021 raid hit one whose Taiwanese and Nanjing arms shared an English name; an engineer there was paid double the industry average. [25]

Why it matters — What is being recruited is the know-how inside people, which no export control reaches.
15
South Korea's chip exports triple

South Korean semiconductor exports rose 209% from a year earlier to a record $46.65 billion in August, 47.5% of the country's goods exports that month. One analyst put chips at nearly 80% of August's export growth. [26]

Why it matters — An economy that grows on one product has to keep hoping that product does not stall abruptly.
16
Micron's Taiwan workers move toward a strike

A preliminary survey found 80% of Micron's Taiwan union members backing a strike over pay. They want a profit-sharing scheme like the bonuses paid at Samsung and SK hynix. [27]

Why it matters — The memory shortage made the companies rich, and the people running the plants want a share of it.
17
China's chip printers put 22 years behind

A UBS analyst wrote that China's lithography tools - the machines that print circuit patterns onto silicon - are at roughly where the Dutch maker ASML stood in 2004. [28]

Why it matters — It is one analyst's estimate, and it is the clearest public figure anyone has offered on the size of the gap.
18
Texas cannot tell which data centres are real

Texas holds 474 gigawatts of grid connection requests and no way to know how many are genuine. Where utilities started charging, demand vanished: Exelon's pipeline of likely projects fell 40% after it began asking developers to put money down, and an Ohio utility's fell by more than half after study fees of up to $100,000. [29]

Why it matters — Grid operators plan power plants years ahead against numbers that include projects nobody intends to build.
19
Pennsylvania's governor turns on developers

Governor Josh Shapiro said too many data centre developers do not care about the communities they build in. He signed an order on 18 August requiring stricter permits above 25 megawatts and details of who will actually use each site. [30][29]

Why it matters — Pennsylvania has more than 100 proposed data centres and 20 permit applications.
20
$31.6 trillion, and it does not last

PwC estimates AI data centre spending could reach $31.6 trillion by 2050, more than railways, electrification or the internet each needed. Unlike those, the hardware is replaced every four to six years. [31]

Why it matters — A railway built once carries trains for a century. This is a bill that arrives again and again.
21
Equinix signs a deal with Nvidia

Equinix, which has rented data-centre space since 1998 and now houses equipment for more than 10,500 customers, struck a deal with Nvidia on Wednesday. [32]

Why it matters — The oldest layer of the industry is the one nearly everyone building AI has to come through.
22
SpaceX starts casting its own turbine blades

SpaceX began making turbine blades and vanes at a foundry in Bastrop, Texas, to speed up the gas generators powering its AI data centres. Elon Musk said the move cuts generator delays by about 18 months. [33][34]

Why it matters — Waiting for a grid connection is now slower than building your own power station.
23
An Australian operator's water use climbs

NextDC said it used 2.40 litres of water per kilowatt-hour in the year to June, up from 2.25, while profit beat forecasts. [35]

Why it matters — Cooling is where a data centre's demand stops being electricity and becomes water.
24
The suppliers nobody names are running short

Reuters found a shortage of transformers is now limiting how fast data centres can be delivered, lifting makers of power and cooling equipment alongside the chip firms. [36]

Why it matters — A transformer steps grid voltage down to what machines can use, and very few companies make them.
25
Cursor loses OpenAI's coding models

OpenAI will cut Cursor's customers off from its coding models in November. It cited concerns about whether SpaceX, which now owns Cursor's maker, will keep to its terms of service. [37]

Why it matters — A tool you rely on can lose its engine because of a fight between its owner and its supplier.
26
A two-year-old startup worth $3.2bn

AfterQuery, which pays skilled professionals to produce training data for AI labs, raised at a $3.2 billion valuation five months after a round at $300 million. Its founders are 22 and 23. [38]

Why it matters — A Y Combinator partner called it the fastest any of its startups has reached a billion-dollar valuation.
27
Wonderful doubles to $5bn in six months

Wonderful, which sells software that coordinates AI agents inside companies, more than doubled its valuation to $5 billion. Insight Partners led again, and Salesforce invested for the first time. [39][40]

Why it matters — It began as a customer-service tool for non-English-speaking markets and now works in more than 35 countries.
28
Polymarket raises $1bn at $21bn

The betting-market company Polymarket raised $1 billion in a round led by 1789 Capital, valuing it at $21 billion. [41]

Why it matters — Bloomberg had reported it in early talks earlier in August at a valuation above $20 billion.
29
$50m to check what AI agents install

AIR raised $50 million to find the AI agents running inside a company, vet the tools and add-ons they use, and block the ones that fail its checks. [42]

Why it matters — Agents install software the way an operating system does, without the oversight an operating system gets.
30
A startup to predict outages

Empirik launched with $21 million, incubated at Sequoia by two people who previously ran infrastructure at Rubrik and VMware, to forecast system failures before they happen. [43]

Why it matters — It launched in the same week as a two-day Microsoft 365 failure.
31
Thyme Care raises $125m

Thyme Care, which guides cancer patients through treatment remotely, raised $125 million led by Morgan Health, with Humana and CVS Health Ventures, at a valuation above $2 billion. [44]

Why it matters — That roughly doubles what it was worth less than a year ago.
32
India's old PCs rent new power

Reliance Jio opened JioPC to anyone with an internet connection, not just its own broadband subscribers. It streams a virtual computer with up to eight processors and 16GB of memory, from about $11 for two months. [45]

Why it matters — India had more than 65 million PCs in use last year, many of them too old for current software.
33
$35m more for battery swapping in India

Magna, a car-parts maker, put another $35 million into Yuma Energy, which swaps batteries for electric two- and three-wheelers in India. Yuma has done more than 60 million swaps with about 100,000 batteries. [46]

Why it matters — Swapping sidesteps charging time, which is the thing that keeps a delivery rider off the road.
34
Reform UK would scrap the UK's data law

Reform UK pledged to abolish the UK's version of GDPR, the data-protection law, and replace it with a lighter one modelled on New Zealand's, where the maximum fine is NZ$50,000. [47]

Why it matters — The party is currently being sued under the law it wants to remove.
35
Meta's India chief joins OpenAI

Sandhya Devanathan, Meta's vice-president for India and Southeast Asia, is leaving after more than a decade. She will run OpenAI's consumer growth, enterprise sales and regulatory work across Southeast Asia and Australia. [48]

Why it matters — She was inside Meta's decision-making in India while the company faced growing scrutiny there.
36
Models know more than they can say

Google Research and Technion found frontier models hold 95-98% of tested facts internally but fail to retrieve them while answering. Up to 65% of the missed facts came back when the model was given longer to think. [49]

Why it matters — If recall rather than knowledge is the limit, then a bigger model is the wrong fix.
37
OpenClaw rebuilds itself

The OpenClaw agent project shipped version 2.0, rewriting installation, memory, plugins and permissions at once, and adding sandboxing, approval controls and auditing. It said the pace of work had outgrown its foundation. [50][51]

Why it matters — A Gartner analyst said agent security cannot be handled layer by layer, because those layers share one identity and one authority.
38
China's robots dance but cannot work

Reuters found China's humanoid robots are too slow and error-prone for most factory work, with manufacturing capacity far ahead of real demand. One researcher put the hard part at the last centimetre, sometimes the last millimetre. [52]

Why it matters — Capacity built for a market that does not exist yet usually ends in a painful consolidation.
39
US tariffs land on drones and robots

The US imposed tariffs on imported drones and their parts, taking effect in September with more on components in 2027, and tightened rules on foreign-made advanced robots. Both cite national security. [53]

Why it matters — China still holds the manufacturing scale that the tariffs are trying to price out.
40
Vietnam asks China for better investments

Vietnam's prime minister, Le Minh Hung, asked Chinese firms investing in the country to move toward AI, 5G and semiconductors rather than low-value assembly. [54]

Why it matters — Vietnam has absorbed a lot of relocated manufacturing and wants the part of it that pays more.
41
A $399 open-source duck

Hugging Face's robotics arm began selling Microduck, a $399 robot with a camera, a lidar sensor and motion sensors. Its behaviours can be trained in simulation and loaded onto the machine, and the training code is public. [55]

Why it matters — Robotics research has mostly needed hardware costing more than a car.

02 Lesson why it matters

Whoever loses the case is the only expert on the cure

A court can prove what a company did. What would happen if you took it apart is a question about the future, and the company is the only one who has ever been inside it.

The twist

Proving what a company did tells you almost nothing about what will happen to it, because the second argument is about a future only the company can describe.

How it works

A court answers two questions, in two separate trials
First: did you break the law? That question is about the past
Second: what should be done now? That one is about the future
Only the company has ever been inside the thing being fixed
So it is the only witness on what the fix would break
The finding stands. The fix shrinks to what the loser says is safe

Where you've seen this

A divorcing couple's business

only the spouse who runs it can say what splitting it would do to the staff

A hospital found at fault

the inquiry hears from the hospital which ward would close if it reorganised

A landlord ordered to make repairs

what is possible, and by when, comes from the landlord's own builder

A school told to change admissions

the school is the one explaining what its timetable and buildings allow

The catch

Often the company is right. A badly designed fix really can hurt the people it was meant to protect - and from outside, being right and being the only voice look exactly the same.

And the whole of it

Everyone who has ever complained about something and been told by that same thing how hard fixing it would be is standing inside this. Each of us can explain our own corner well enough to say why it cannot change, and none of us can see the rest of it.

03 Truth what's really going on

What is really going on

Google was found to be an illegal monopoly, and the argument that decided its punishment was one only Google had the information to make.

Why it works on us — A one-word outcome - no breakup - is public today, while the pages saying what Google must actually do stay sealed for two weeks, so the version everyone hears is the half that favours the winner.

Who gains

Google — It keeps the exchange and the 20% fee on it, and once the opinion is public it can appeal the monopoly finding itself. [1][3][4]
Large technology firms with US antitrust cases pending — Three breakup requests refused in a row sets what the next court is expected to do. [1]
Anthropic's business customers — Cached context costs 75% less, and monitoring data can now stay on machines they control. [7][8]
South Korea's exporters — Chips were 47.5% of August goods exports at a record $46.65 billion, and nearly 80% of the month's export growth. [26]
Makers of transformers and cooling gear — A shortage of the boxes that step grid voltage down is now a limit on data-centre delivery, and it is lifting their orders. [36]

Who pays

Web publishers — They keep paying Google's 20% fee, and cannot read the order that is supposed to help them. [1][2][4]
Rival advertising-technology firms — The remedy they were promised is a set of behaviour rules whose text they have not seen. [3]
Micron's workers in Taiwan — 80% backed a strike survey over bonuses while colleagues at Samsung and SK hynix receive profit-sharing. [27]
Cursor's paying customers — They lose OpenAI's coding models in November because of a dispute between OpenAI and Cursor's new owner. [37]
People living near proposed US data centres — Grid operators plan and price years ahead against request queues that include projects nobody intends to build. [29][30]

What nobody knows yet

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

01

What Google actually has to change.

The judge adopted most of the proposed behaviour rules and sealed the order for 14 days so both sides can ask for redactions. [2][3][4]
02

How big AdX is now.

The newest public figure is 2020's - 4.1% of revenue and 1.5% of operating profit - and later numbers were redacted from the court file. [1]
03

Whether a forced sale really would have been technically hard.

Google said so and the US justice department disputed it. No independent assessment has been published. [1]
04

What OpenAI's Astra model can really do.

OpenAI graded it against its own framework, named no outside testers and did not say how they will be chosen. [12][13]
05

How the US government decides which AI models may be released.

A lawsuit says the review framework's text, its participants and its legal basis are all unpublished. [23]
06

Whether Meta will publish Muse Spark 1.3's weights.

Meta has not decided, and in Europe that choice also changes which parts of the AI Act apply to it. [10]
07

How much of Texas's 474 gigawatts of grid requests is real.

Nobody has counted. Where utilities charged for a place in the queue, the pipeline fell by 40% or more. [29]
08

How good Meta's new model actually is.

The comparisons come from Meta's own chief AI officer, and the reasoning mode used for its benchmark scores was not available at launch. [10][11]

04 Hope carry this

A computer eight years old in India can now borrow the power of a new one over the internet, for about eleven dollars for two months. Reliance Jio opened the service to anyone with a connection this week, in a country with more than 65 million PCs already in use.