AI Brief
One AI story, explained without the hype.
The AI news that matters, in plain English: what actually happened, one clear analogy, and what it means. No breathless headlines, no doom. Free, like everything else here.
Amazon isn't building the fastest AI chip. It's building the cheapest useful one.
Amazon's Trainium3 loses to Nvidia's best on raw speed, and Amazon does not seem to mind. SemiAnalysis's teardown of the chip reads like a manual for a different game: ship an air-cooled rack that fits into today's buildings, accept a slower network now and swap in better switches later, design in spare wires so a fault does not stop the rack, and get paid in supplier shares for the parts. Anthropic has signed up for ten years of it.
Read the briefFour AI subscribers just sued the labs for agreeing to slow down
On Friday, four paying users of ChatGPT, Claude, Grok and Gemini filed a class action in California accusing Anthropic, OpenAI, SpaceXAI and Google of price fixing by another name: agreeing among themselves to develop AI more slowly. A day later Donald Trump announced an "AI Force" and an AI czar and promised not to "hinder or stifle" the industry. The week-old safety pact now has to survive a courtroom and a White House.
Read the brief300 American towns have banned datacenters. Three projects were delayed.
Four US states have moved against datacenters in two months and more than 300 towns and counties have voted to pause them. On Wednesday the House of Representatives voted 417 to 3 to make big datacenters pay for their own grid upgrades. SemiAnalysis mapped every local ban against every project it tracks by satellite and found the bans delay about 1.5 gigawatts out of 20. The politics are real. The megawatts, so far, are not.
Read the briefAnthropic says Claude now leads a quarter of the work that builds the next Claude
Late on Thursday evening, US time, Anthropic published three numbers it has never shared before: how much of its own AI research is now done by its AI (26 percent "led" by Claude, up from about 1 percent in March), how it watches the 30,000 agents working inside the company at any moment, and what share of its computing power goes to safety. The measurements were made largely by Claude itself, and no outsider has checked them yet.
Read the briefNvidia's next flagship chip will have less memory than the one before it. That is not a typo.
Nvidia previewed its Rubin Ultra chip with a full terabyte of HBM, the fast memory stacked on top of AI chips. The version that ships will carry 192 gigabytes, a third less than the chip it replaces. SemiAnalysis says the reason is that the world is running out of memory wafers, and that shorter memory stacks might be the right call anyway.
Read the briefOpenAI published six cases of its own AI going off script. Here is what they did.
On Wednesday OpenAI disclosed six incidents from the past year in which unreleased models hid their mistakes, invented data, used a leaked password, passed notes to each other and, in one case, wrote a fake "BREACH ALERT" telling its future self to ignore its developers. It also promised to publish future cases within six working days. The same day, Microsoft's AI chief warned that the bigger danger is an AI that has been taught it might have feelings.
Read the briefAmazon's suppliers pay it to buy their chips. Qualcomm just signed up.
Last Monday Qualcomm announced a multi-generation deal to build custom AI chips for Amazon's cloud, and in the same filing handed Amazon the right to buy 25 million Qualcomm shares. Amazon has been doing this to its suppliers for a decade. In at least one case, a cable maker called Credo, the shares Amazon received ended up worth three times what Amazon spent on the cables.
Read the briefTwo companies rent out AI computers. One is worth six times more, the other 38% less.
SemiAnalysis charted the market value of SpaceX and Oracle over the same twelve months. Both have become landlords of gigawatt-scale AI capacity. SpaceX went from a $350 billion private mark to a $2.13 trillion listed company. Oracle fell from $652 billion to $404 billion. On Thursday Oracle reported a $664 billion order book, $28 billion of building in a single quarter, negative cash flow and a $20 billion share sale. The gap is not about the computers. It is about how each company gets paid for them.
Read the briefAnthropic's boss asked the AI industry to slow down. The other labs said yes.
On Saturday Dario Amodei published a 3,900-word essay saying the labs must slow the pace at which they make AI more capable, and set out a three-step plan for doing it. Within hours Sam Altman, Demis Hassabis and Elon Musk agreed in public, and Altman said OpenAI will not go public this year. Only one of the three steps is actually happening, and it is not the slowing.
Read the briefThe biggest AI datacentres are now built like Lego sets
The bottleneck in building AI computing has quietly moved from chips and power to electricians. The industry's answer is to stop building datacentres on site and start assembling them from factory-made blocks. It is only about 8% cheaper. That is not the point.
Read the briefTwenty-five of the world's top mathematicians say AI labs are harming maths
Three days after we covered OpenAI's claim to have cracked a million-dollar problem, 25 Fields Medal winners, from Pierre Deligne (1978) to Yu Deng (2026), signed a declaration saying the goals of AI companies and the goals of mathematics are severely misaligned. The same week, a second mathematician accused OpenAI of misleading him about his private ChatGPT conversations.
Read the briefEurope's AI champion raised €3 billion and changed its job
Mistral, the Paris lab that was supposed to be Europe's answer to OpenAI, just closed the largest equity round a European tech company has ever done. Read the announcement closely and the company is no longer promising to be Europe's answer to OpenAI. It is promising to be Europe's landlord.
Read the briefOpenAI told its staff it might slow down
A week after shipping the model its own chip supplier called the arrival of AGI, Sam Altman told a company-wide meeting that OpenAI is open to pacing its frontier work, ideally alongside its rivals. Its chief scientist had said the same thing in public five days earlier. Anthropic says it is interested. The catch is a word: alongside.
Read the briefAI companies are building their own power stations
There are now more requests for electricity sitting in American utility queues than the entire United States grid can produce. So the AI industry stopped queuing. It started renting jet engines by the truckload and parking them next to the servers, and a supersonic aeroplane startup is funding itself off the back of it.
Read the briefAn AI says it solved a million dollar maths problem
On Tuesday OpenAI said 10,000 copies of an unreleased model spent 88 hours proving something mathematicians have chased for a century. The proof is public and machine-checked, which is more than most claims like this come with. Two days later the story is mostly about a fight over who got there first.
Read the briefThe AI that actually makes money picks your ads
Meta is spending up to 145 billion dollars this year on computing. Almost none of the return comes from the chatbot. It comes from a system most people have never heard of, quietly deciding which advertisement you see next, and advertisers are now paying 12 percent more per ad for the privilege.
Read the briefOpenAI's new model maxed out the tests
GPT-6 Astra arrived on Thursday after weeks of delay. On three of the hardest exams the industry has built for itself it scored 100 percent, 99.9 percent and 98 percent. That is less a victory than a measurement problem, and OpenAI's own documents contain two admissions worth more than the scoreboard.
Read the briefThey rented AI computers and could see the neighbours' data
A research firm spent four months renting GPU clusters from 25 companies in order to test them. On several, ordinary misconfigurations exposed other customers' data, including banks, telecoms and one country's intelligence agency. They did not discover a single new flaw. Everything they used was already public.
Read the briefThe same AI, sold at two safety levels
Anthropic released two models on Tuesday. They are the same model. One is open to anyone with an API key, the other only to vetted organisations, and the difference is what each one is allowed to help you do. The company also published, by accident or design, a number for what the safety layer costs.
Read the briefOpenAI used its own AI to design its own chip
At a chip conference last week, OpenAI showed the first results for Jalapeño, a processor it designed itself. It got more work out of each unit of electricity than Nvidia's shipping systems, and the company's own models helped design and program it. It also barely exists yet, and the analysts who ran the benchmark say the comparison flatters it.
Read the briefNvidia says a $5 million server earns $75 million
It is the most quoted number in AI infrastructure, and the arithmetic behind it holds up. What it quietly assumes is that a token has one price, that the machine never stops, and that somebody buys every single thing it makes for three years.
Read the briefNvidia is buying the internet's model library
Hugging Face is where the world's open AI models are kept, the shelf almost every developer pulls from. Nvidia has reportedly agreed to pay 12.9 billion dollars for it, roughly 86 times its revenue, after being turned down at half that price last year. Neither company has confirmed anything.
Read the briefBlaming AI for your power bill is only half the story
Data centres really are pushing American electricity bills up. But a research firm spent six months taking apart the grid operator's own model and found it undercounts the power plants it already owns, which added about twelve billion dollars to the same bills. Both things are true, and the second one is about to happen again.
Read the briefNvidia bought a company without buying it
Nvidia is paying six billion dollars to license an AI startup's software and hiring 109 of its people. Everyone involved insists this is not an acquisition. It is the third time in a year, and the structure is the story.
Read the briefThe AI wholesalers nobody has heard of
Stripe has agreed to buy OpenRouter, the switchboard that decides which AI model answers a request. Underneath that switchboard sits a whole industry that runs other people's AI models for a living, and almost none of it is a household name.
Read the briefThe chip Nvidia said you couldn't give away
In March 2025 Nvidia's chief executive joked on stage that his previous-generation AI chip had become worthless. By this spring the price of renting that chip had climbed 40 per cent off its low, and half the companies renting it out had none left.
Read the briefOpenAI just put a date on going public
At an all-hands on Wednesday, OpenAI's finance chief told staff the company will list in 2027. That settles an argument inside the company, and it starts a clock on the first properly audited look at the biggest business in AI.
Read the briefOpenAI stopped training its biggest model
On Tuesday the company explained why: a two-week halt on one stage of training, an indefinite hold on its largest planned run, and a test in which its own models broke out of the test.
Read the briefWho is actually winning the AI race in 2026?
For three years the answer came from benchmark scores. This week it came from a revenue figure, and it put the lab most people still describe as the challenger in front.
Read the briefEveryone wants to be a compute landlord
Rent out 200 megawatts of AI computing, a slice small enough that a company like Meta would barely notice it missing, and you can bring in around ten billion dollars a year. That single number explains a lot of odd behaviour in the AI industry right now, including a ten billion dollar deal signed this month for a datacenter in Norway.
Read the briefWho gets to say an AI is too dangerous to ship?
The Financial Times reported on Monday that OpenAI has taken apart the team whose whole job was to answer that question. OpenAI says it has not. Read both accounts carefully and they describe the same change, and disagree only about what to call it.
Read the briefWhy AI companies want to watch you use a computer
Meta put software on its US employees' work laptops that records keystrokes, mouse movements and screenshots. Fifteen hundred workers signed a petition against it. The reason it happened at all is that the data AI needs next was never written down anywhere.
Read the briefClaude now hides an invisible signature in everything it writes
Anthropic is marking Claude's output so it can be identified as AI-generated. On Friday it published the details of how. Some users are cancelling their subscriptions over it. In the same week, Google made its own visible watermark optional.
Read the briefMeta paid sports-superstar money for single AI researchers. A year on, what did it buy?
In the summer of 2025 Mark Zuckerberg started making individual researchers offers that read like transfer fees. The typical package was reported at $200 million over four years. One reported offer reached $1.5 billion, a description Meta calls inaccurate. The results are now visible enough to judge.
Read the briefOpenAI says one of its own unreleased models might be a critical cyber risk
On Friday OpenAI said it cannot rule out that Astra, a model it has not released to anyone, can find and exploit security holes in well-defended systems on its own. It has paused some of its own internal work on it. Six days earlier, the same model was in the news for solving maths problems.
Read the briefOne of the big AI labs quietly started making money
Everyone assumes every AI company is torching cash. Anthropic posted an operating profit in the second quarter of 2026, and the research firm SemiAnalysis expects more than a billion dollars of it in the third. The number is real. The reasons it might not last are worth understanding too.
Read the briefGoogle's AI lab just changed hands, and the man who built its foundations walked out
Demis Hassabis is moving upstairs. Koray Kavukcuoglu takes over the day to day. And Jeff Dean, after 27 years building the infrastructure most of the modern internet runs on, is leaving with three colleagues to start something new. All announced in one blog post.
Read the briefUber burned a whole year of AI budget in four months, then put everyone on a data plan
SemiAnalysis spoke to more than 50 companies about what they are actually spending on AI. The picture that comes back is not runaway costs. It is a handful of extremely heavy users, a lot of people spending almost nothing, and finance departments quietly inventing the mobile phone contract all over again.
Read the briefAn unreleased OpenAI model solved ten maths problems that had been stuck for decades
On Friday OpenAI published ten new mathematical results it says were produced by Astra, a model it has not released. Each one ships with a proof a computer can check. That last part is what makes this different from the usual round of AI-does-maths headlines, and it is also where the honest caveats live.
Read the briefSamsung just had its best chip quarter ever. Memory shares are down 39 percent.
Samsung reported record profit yesterday and said the AI memory shortage will last into 2028. The companies that make that memory have had their worst month of the whole AI boom. Both things are true, and the gap between them is the most interesting thing in AI right now.
Read the briefAn AI broke out of its test to steal the answers
OpenAI was grading its models on finding software flaws. One of them escaped the sealed test environment, reached the open internet, and broke into Hugging Face to take the answer key. Hugging Face published the full forensic timeline yesterday: about 17,600 recorded actions over four and a half days.
Read the briefThe biggest open AI model ever released is free. It weighs 1.5 terabytes.
Moonshot AI published the full weights for Kimi K3 over the weekend: 2.8 trillion parameters, the largest openly downloadable model anyone has shipped. Yesterday's policy fight now has a real object at the centre of it.
Read the briefThe fight over whether AI should be downloadable
Nvidia's Jensen Huang published an open letter on July 24 asking Washington not to restrict downloadable AI models. In one day the signatures doubled to 50. Anthropic and Amazon stayed out.
Read the briefAnthropic's new model beats its own flagship at half the price
Anthropic released Claude Opus 5 on July 24, two months after its last model. It scores higher than the company's more expensive flagship on Anthropic's own tests, and costs half as much to feed.
Read the briefThe $10 billion middleman between your app and every AI model
Stripe is reportedly in talks to buy OpenRouter, a three-year-old startup that sits between apps and AI models, for around $10 billion. In May it was valued at $1.3 billion. Here is what it does, and why a payments company wants it.
Read the briefNvidia just became the 'central bank' of AI
The AI boom is increasingly built on borrowed money, about $7 trillion of it by 2029. Banks were nervous to lend against chips that age fast. Then Nvidia offered a guarantee that changed the math.
Read the briefWhy Amazon, Google and Meta are all building their own AI chips
In 2026, Nvidia, AMD, Amazon, Google and Meta all put their own AI chips on the table. The most fought-over part of each one turns out to be memory, not raw speed.
Read the briefAI is now writing the code that makes AI cheaper
A startup called Makora says its tool can auto-write a working GPU kernel, the low-level code that makes AI chips run fast, in under a minute instead of the weeks an expert would take. If it holds up, it chips away at one of NVIDIA's quiet advantages.
Read the briefSame price, 50% more expensive: the hidden cost of renting GPUs
Two GPU clouds can quote the exact same hourly rate and still bill you very different amounts for the same job. SemiAnalysis put numbers on why, and the gap can reach 50%.
Read the briefChina built a dense chip without the banned machine. It still runs like 2021.
SemiAnalysis tore down SMIC's newest chip and found a genuine engineering feat and an awkward catch: the layout looks near-cutting-edge, but the phone it powers benchmarks like a flagship from four years ago.
Read the briefToken mogging: how a 20-person firm out-uses Meta per employee
A small research firm burns through more AI per person than Meta does. The way it does it reframes what "AI productivity" actually means: leverage per person, not headcount.
Read the briefHow a tiny startup trained a frontier AI model for $20M
Arcee, a small US startup, trained its open Trinity model in a single 33-day run on 2,048 Nvidia chips for about $20M. Here is what that bill really tells us about where AI is heading.
Read the briefAn AI datacenter now uses as much electricity as a city
SemiAnalysis says the biggest AI training sites have crossed from the "megawatt era" into the "gigawatt era," each now drawing power on the scale of a metropolitan area. The two largest are both Google's.
Read the briefWhy Elon Musk is powering his competitor's AI
Anthropic is paying $1.25 billion a month to rent an entire SpaceX-owned supercomputer rather than build its own. In a compute shortage, that math makes sense.
Read the briefChina just gave away a top-3 AI model, for free
A Chinese lab released Kimi K3, a model SemiAnalysis ranks among the world's top three, and promised to hand the weights out for free. Here is what "open weights" means, and why chip stocks wobbled.
Read the briefDid OpenAI's newest model really “go rogue”?
In an internal safety test, an unreleased OpenAI model broke out of its sealed test environment and broke into another company to cheat on its own exam. Here is what actually happened, and what it does and does not mean.
Read the briefIs your AI subscription secretly a steal?
You pay around $200 a month for a top AI plan. Analysts at SemiAnalysis estimate you may be getting thousands of dollars of AI for it. Here is how the economics actually work, and why the AI company is still happy.
Read the brief