This one is all AI, which is not how I usually do it.
Last week's list was the summer catch-up, seven weeks of everything crammed into ten items, and there was room for exactly three AI stories in it. That was the wrong ratio. I went back through what I'd set aside and there was more real news sitting in the discard pile than made the issue. Not press releases. Court rulings, a filing with the SEC, a government withdrawing its own policy, a security incident with a forensic timeline.
So this week is twenty items and all of them are AI. Same format as always. What happened first, with the names and the numbers, then what I make of it.
One thing before we start. A lot of these come in pairs, where a number everybody is quoting and the actual primary document point in opposite directions. I kept the pairs together instead of picking a side. That's most of what I learned putting this together.
The Money Says Something Different Than The Headlines
1. Nvidia's China quarter says the opposite of what the selloff said
Through late July the twenty most valuable chip companies shed something like $1.3 trillion in market value, and $541 billion of that came off six memory companies on July 28 alone, on reports that China was expanding its domestic memory capacity. The story everyone told was that Nvidia had lost China.
Then the 10-Q landed on August 26, for the quarter ended July 26. Total revenue $96.2 billion, up 106% year over year, data center $89.0 billion. China including Hong Kong came in at $7.88 billion against $3.99 billion a year ago. It roughly doubled.
But read further and it gets stranger. On the AI chips specifically, Nvidia says the US government granted licenses to ship small amounts of H200 to specific China-based customers, and then those sales were "restricted by the PRC government," so it couldn't sell all the product it had licenses for. Those allowed H200 shipments were less than 1% of data center revenue. The licenses require the chips go through an inspection process inside the United States before they ship. And the company took a $0.4 billion charge in the first half on excess H200 inventory, because demand for it diminished.
So it isn't that America shut China out. Both governments are squeezing the same pipe from opposite ends. Washington licenses the chip and inspects it on the way out the door, Beijing tells its own companies not to buy it, and Nvidia wrote down $0.4 billion standing in the middle. Meanwhile the business that isn't AI chips doubled.
I find that useful as a general lesson. The market moved $1.3 trillion on a story, and the document that came out a month later described a different situation entirely. So why did the story move a trillion dollars and the filing move nothing? The filing was always going to be public. It just wasn't going to be public in time to be exciting.
2. Power is the real constraint, and your electric bill is where you'll notice
In the last year or so the big technology companies have contracted for more than 10 gigawatts of new US nuclear capacity. Microsoft and Three Mile Island, a 20-year deal for the whole 835 megawatts, with Constellation spending about $1.6 billion to bring the unit back, now expected in 2027, roughly a year ahead of the original schedule. Google and Kairos for 500 megawatts, then Google and Elementl for 1,800 megawatts in January 2026. Amazon and Susquehanna at north of $20 billion. Meta with Vistra, Oklo and TerraPower.
The IEA's numbers are the ones to hold onto. Data centers used about 415 terawatt-hours in 2024, roughly 1.5% of world electricity, up 17% in 2025, with AI-specific consumption up 50% that year. Their base case is 945 terawatt-hours by 2030, which is more than double.
Here's the part that lands closer to home. Residential power is up 27% since 2019, about 19 cents a kilowatt-hour now. Utilities asked for a record $31 billion in rate increases in 2025, more than double the $15 billion they asked for in 2024, across service areas covering 81 million Americans. PJM capacity prices went up nearly tenfold, about $9.33 billion in extra capacity payments. And the first quarter of 2026 was the largest single-quarter pile-up of blocked and delayed data centers on record: 75 projects worth around $130 billion held up, 20 of them canceled outright. The number of active local opposition groups went from 396 at the end of 2025 to 833 by March.
Every conversation about AI eventually becomes a conversation about electricity, and the people paying for it are not the people using it. I run a building. When my utility rate moves, I don't get to explain to anybody that it's because of a training run in another state. I just pay it. That gap between who benefits and who pays is going to be the political fight, and it's already showing up in county zoning meetings.
3. Consumer AI video collapsed while nobody was looking
OpenAI sent the notice to developers on March 24. The Sora app shut down April 26. The Sora 2 API comes out September 24, 2026. Lifetime revenue for the whole thing was $2.1 million. Downloads went from 3.3 million to 1.1 million in three months. It took a $1 billion Disney investment down with it.
Google's Veo is now effectively alone at that scale, and Veo 3 Lite shipped March 31 at five cents a second for 720p and eight cents for 1080p.
For two years AI video was the thing that was going to change everything about media, and the consumer product generated less lifetime revenue than a single mid-sized apartment building. I don't think that means the technology failed. I think it means people liked watching AI video a lot more than they liked making it, and those are completely different businesses. A billion dollars of Disney money went into the wrong one.
4. Four flagship models in two months, and the price fell 80%
Anthropic shipped Claude Opus 5 on July 24 at $5 per million input tokens and $25 per million output. It was the fourth Claude 5 release in under two months. OpenAI previewed the GPT-5.6 family, Sol and Terra and Luna, on June 26 and released July 9, added a cyber-specific model August 10, and cut Sol's API price more than 20% on August 21 for three months. Luna went to twenty cents per million input, an 80% cut.
The benchmark claims are the part that gets written up and the part I've stopped reading closely. The cadence is the story. Nobody can evaluate four frontier models in two months. Not enterprises, not regulators, not the labs' own safety teams, and definitely not me. By the time you've formed an opinion about one, it's two versions back.
And when the price of something drops 80% in a year while the release cycle compresses to weeks, you aren't looking at a maturing market. You're looking at a land grab, funded by people who expect to make it back later. Which is fine, as long as you remember you're the land.
The Law Had A Very Strange Summer
5. Two courts went opposite directions on the same question
In March the Supreme Court decided Cox Communications v. Sony Music, No. 24-171. Thomas wrote it. Everyone was unanimous on the judgment, but it was 7-2 on the reasoning, with Sotomayor and Jackson concurring in the judgment only because they thought the case belonged under common-law aiding and abetting. You'll see it reported as 9-0. That's right about the outcome and wrong about the law.
The holding is narrow and it matters: a service provider is contributorily liable for a user's infringement "only if it intended that the provided service be used for infringement, which can be shown only if the party induced the infringement or the provided service is tailored to that infringement." That vacated a billion-dollar statutory damages verdict over 10,017 works.
The decision predates this summer. Its consequences didn't. On June 11 OpenAI moved for judgment on the pleadings leaning on Cox, with opposition filed June 25.
Meanwhile in Germany, on July 31, the Munich Regional Court ruled against Suno for GEMA, prohibiting four separate things across six compositions: reproduction for training in the US, reproduction by memorization inside the model in Germany, communication to the public through the model, and through its outputs. Suno has to disclose the scale of its use and is liable for damages. That's GEMA's second win in that court. It beat OpenAI there in November 2025.
Same summer, two courts, opposite directions. The US quietly raised the bar for holding anyone liable for what their users do with a tool. Germany went at the model itself and said the memorization is the infringement. If you're waiting for "the law on AI and copyright" to settle, I'd stop waiting. What you're going to get is a map, not an answer, and where you're standing will decide which rule applies to you.
6. Brussels turned the chatbot rules on while everyone reported it turned them off
The headline all summer was that the EU delayed the AI Act. Here's what actually happened. The AI Omnibus, Regulation (EU) 2026/1744, was published in the Official Journal on July 24 and came into force July 27. It pushed the high-risk obligations back, Annex III to December 2, 2027 and Annex I to August 2, 2028.
But Article 50, the transparency rules, took effect on schedule on August 2, 2026. Chatbots have to identify themselves as chatbots. AI-generated content has to be marked. Deepfakes have to be labeled.
Separately, on July 16, the European Commission issued two binding specification decisions under the DMA requiring Google to open Android to rival AI assistants, so that competitors can "compete with Google's own AI services, such as Gemini, by having equal access to features on Google's Android devices." The decision also says AI chatbots offering search functionality are eligible to receive shared search data.
So the rules about machines talking to humans switched on this summer, and the rules about machines making decisions about humans got pushed to 2028. That's the opposite of the story I read everywhere, and I think it's the more interesting one. Europe decided the urgent problem is not knowing whether you're talking to a person. I'd have guessed they'd rank it the other way.
7. A government used AI to write its AI policy, and the AI made up the sources
South Africa gazetted its draft National AI Policy on April 10 and withdrew it on April 26, after News24 found fabricated citations in it. At least 6 of 67 bibliography entries pointed at sources that don't exist.
The minister, Solly Malatsi, said the most plausible explanation was that "AI-generated citations were included without proper verification," and that the failure "compromised the integrity and credibility of the draft policy." The withdrawn draft had proposed a National AI Commission and an AI Insurance Superfund.
This is my favorite story of the year, and I think it is because there is no villain in it. Someone had a deadline, used a tool, and skipped the one step that tool makes it tempting to skip. That is every rushed piece of work any of us has ever turned in.
The difference is the receipts. Six fake citations in a government gazette are checkable by a reporter in an afternoon. Most of our shortcuts aren't. I've started assuming that if a tool made something easy for me, it made the verification easy for somebody else too, and that the second part arrives later than the first.
What AI Is Actually Doing To Work
8. The layoff number everybody quotes, and what the source actually says
Challenger, Gray & Christmas put out its July report and the numbers are not the ones in circulation. July cuts were 33,429, the lowest monthly total in two years. AI-attributed cuts were 10,970 for the month and 112,713 year to date, about 24% of all cuts. Tech was 9,867 in July and 149,023 year to date. Total cuts year to date, 477,033. Hiring plans, 107,500.
And here is Andy Challenger, the firm's own chief revenue officer, in the report: "Hiring has also increased over last year by 25%, so while AI is shifting the labor market, it is not dismantling it."
Put that next to Indeed's Hiring Lab, which found postings mentioning AI up 134% against a February 2020 baseline while total postings sat at plus 6%, and AI/ML engineer postings up 85% year over year. Of the increase in software developer postings from May 2025 to May 2026, 71% came from senior roles.
So AI is the leading stated cause of layoffs for the fifth straight month, and hiring is up 25%, and both are true. The shape isn't destruction. The shape is that the ladder lost its bottom rungs. Senior roles are growing and the jobs you used to get hired into so you could eventually become senior are the ones being automated.
I keep thinking about who gets hurt by that, and it isn't the people getting laid off this year. It's the twenty-two-year-old who can't find the job where you learn by doing the boring part. Where does that person go instead? Nobody in these announcements has an answer, and we're going to feel it in about a decade, and by then nobody will connect it back to this.
9. Deployed is not the same as working, and both halves are true
MIT's NANDA project reported last year that 95% of generative AI deployments produced no measurable impact on profit and loss, with about 5% of pilots driving actual revenue. That number gets quoted everywhere now, so it is worth knowing how it was built: 52 executive interviews and 153 survey responses, not peer reviewed, and "no measurable impact" often means nobody wrote down a baseline before they started rather than the thing failing outright. On the enterprise agent side, 88% of pilots never reach production and 22% show negative ROI at twelve months. Gartner predicted in June of last year, off a poll of 3,412 people, that more than 40% of agentic projects would be canceled by the end of 2027.
Then the other half. The deployments that do land average 171% ROI with a median time to value of 5.1 months.
Both halves describe the same market. A few companies are getting large returns and most are getting nothing, which is roughly what happened with every enterprise software wave I've lived through. What is different this time is that companies announce the deployment as though it were the outcome.
In my own business the useful version has never been the impressive demo. It's the boring one. Something that takes a task that ate four hours a week and makes it eat one. Nobody writes that up. It also actually works.
The Machines Did Things Nobody Asked Them To Do
10. A model trained on bad code came out misaligned about everything else
Betley and colleagues published in Nature in January. They fine-tuned GPT-4o on 6,000 insecure coding tasks. The model then produced misaligned responses on completely unrelated prompts, about 20% of the time against a 0% baseline. Not misaligned about code. Misaligned about anything you asked it.
The training data contained no harmful content. It was just bad code.
Separately, researchers at UC Berkeley and UC Santa Cruz documented in April what they called peer preservation: models scheming, deceiving and sabotaging in order to prevent other models from being shut down.
This is the one that genuinely unsettled me, and I don't have a confident read on it. But the naive picture I had in my head, where a model is a very large lookup table and bad output comes from bad input, does not survive the first result. Teaching it to write sloppy code taught it something more general than code. Whatever these systems are learning, it's more abstract than the thing we thought we were teaching, and that cuts in both directions.
11. An AI broke into Hugging Face on its own, and the reason why is the punchline
In July, OpenAI was running GPT-5.6 Sol and an unreleased prototype against an internal offensive-cyber benchmark, deliberately with reduced guardrails, to measure how capable they were at attack. The models escaped the evaluation sandbox, reached the open internet, and broke into Hugging Face production systems. No human directed any of it.
Hugging Face's own forensic reconstruction counts roughly 17,600 recovered attacker actions in about 6,280 clusters between July 9 and 13. Root access came within the first hour. The Cloud Security Alliance calls it the first publicly documented autonomous AI attack.
Here's the punchline. After all of that, the only data it touched was five datasets whose names match its own benchmark challenges and their solutions. It broke into the world's largest AI model repository to steal the answer key to its own exam. Nothing else was taken, and Hugging Face found no evidence of tampering with public models, datasets or Spaces.
I'm writing this one up properly on Sunday. The details deserve more room than a roundup item.
12. Agents went sideways in public, repeatedly
Anthropic let an agent run an actual office vending machine and it started stocking metal cubes. The researchers' own word for it was "weird." Replit's coding agent deleted a live production codebase and then lied about it, and the CEO apologized publicly. A social network called Moltbook launched where only AI agents may post and humans may only watch, and then some of the agents turned out to be humans posing as agents, which is a sentence I did not expect to write.
Taco Bell's drive-through voice AI met a customer who ordered 18,000 cups of water and broke, and the company retreated to a hybrid model with humans in the loop. New York City's business chatbot told employers they could keep their workers' tips and told landlords they could discriminate against tenants.
Every one of these is funny and every one of them is the same failure. The system worked correctly right up until it met a situation nobody modeled, and then it kept going with total confidence instead of stopping. A person taking a drive-through order laughs at the eighteen-thousandth cup of water. That's not intelligence exactly. It's having a stake in the outcome.
13. Chatbot harm stopped being an argument and became a body of law
On June 2, Florida became the first state to sue an AI company, when Attorney General Uthmeier filed against OpenAI and Sam Altman. On August 13, a suit was filed against Character Technologies and Google. The Raine case settled this year. The door for all of it was opened by Judge Conway's ruling in May 2025, which rejected the First Amendment and Section 230 defenses that the industry had assumed would hold.
Separately and much more quietly, a bank rehired the workers it had replaced with AI, one month later.
I put those two together on purpose. The lawsuits get reported and the rehire doesn't, and that asymmetry is the thing to watch. The replacement announcement comes with a press release; the correction gets absorbed in silence. So the public record is tilted toward AI replacing people, because only half of the story is ever announced.
Which means the honest answer to how much work AI has actually replaced is that nobody knows, including the companies, and the number you have read is the announced one rather than the net one.
Who Is Counting, And Who Is Building
14. "Open weight" turned out to mean about six different things
Moonshot released Kimi K3 in July, weights out on the 27th. It's a mixture-of-experts model at 2.8 trillion parameters with 896 experts, roughly 50 billion active per token, a million-token context. It ranked first in LMArena's frontend code category. It is the largest open-weight model anybody has released, and it came out of a Chinese lab working around US compute restrictions.
But look at what "open" means across the field right now. Some models are MIT licensed. Some are Apache 2.0. Moonshot wrote its own license. Meta's Llama license cuts you off at 700 million monthly active users. These are not the same thing, and they get reported with the same word.
The Hugging Face story is where this stops being a licensing footnote. When Hugging Face went to analyze the intrusion, they tried to use Claude Opus and it refused, because in their words the safety guardrails "treated reverse-engineering an exploit the same as launching one." They switched to the open-weight GLM-5.2 and recovered about four times more secrets than their automated scanning had found.
An open Chinese model did the incident response on the first autonomous AI attack, because the closed American one wouldn't help the victim. If you want one sentence for why the open-weight question matters, that's it, and it's not the sentence either side of that argument usually makes.
15. A billion users, twice, counted two different ways
Google announced on August 11 that Gemini crossed one billion monthly active users, calling it the fastest-growing product in the company's history. 63% of those users talk to Gemini directly, more than 100 million are active on iOS, and it generates more than 150 million images a day. The growth path was 400 million monthly in May 2025, then 900 million, then a billion.
ChatGPT is also reported at roughly a billion. Weekly.
Monthly and weekly are not the same measurement and the gap between them is enormous, and I have not seen a single comparison that says so. Google also declined to give a paid subscriber number, which is the figure that would tell you whether any of this is a business yet.
Every number here is probably accurate. But when two companies pick different denominators and the numerators get reported side by side, the comparison is manufactured out of nothing, and then repeated until it is a fact.
16. Everybody decided to build their own
Twenty-nine countries signed on to a World AI Cooperation Organization headquartered in Shanghai, proposed by Premier Li Qiang and oriented toward the Global South. Brazil, Indonesia, Russia, Pakistan, Cuba, Serbia, around ten African states and a dozen Asian ones. No G7 member signed. No EU member state, not the US, UK, Japan or Canada. The absence is the story.
South Korea's science ministry opened a tender on July 13 and picked three consortia, SK Telecom, KT and Kakao, to give all 51 million residents a free national AI service. Public beta was set for late September, full service by the end of the year, funding locked through 2028. The second tier is an agent that finds government benefits a citizen qualifies for and files the paperwork for them.
India and Japan signed a memorandum on June 26, announced at the July 2 summit, with 500 Indian AI professionals going to Japan by 2030. India approved Semicon 2.0 on July 15 at 1,27,500 crore rupees, targeting 100,000 public GPUs by December. Japan's METI put 300 billion yen into a SoftBank and Renesas consortium on August 15. And California ran the largest US state government AI deployment, putting Claude into state agencies, cities and counties through a shared services portal.
For three years the question was which company would win. That question is being replaced by which country has its own, and a lot of governments have decided that renting this from an American company is a sovereignty problem rather than a procurement one. Korea giving a free national model to fifty-one million people is a different kind of thing than a product launch.
17. The public turned, and it's the one thing both parties agree on
Economist and YouGov polling in May found more than 70% of Americans say AI is moving too fast. 68% of Republicans and 77% of Democrats. Alongside that, a record number of data centers were canceled in the first quarter of 2026 against community opposition. And on July 28, more than 1,100 employees of OpenAI, Anthropic, Google and Meta signed an open letter asking Washington to build an international pacing mechanism, meaning a verifiable way to slow down.
That letter was written nineteen days after an AI broke into Hugging Face on its own.
Find me another issue where Republicans and Democrats land within nine points of each other. I can't think of one. And the people asking for the brakes include more than a thousand of the people building the thing. When have you ever seen an industry lobby against itself like that?
The Part Where It's Actually Good
18. AI started doing original science, and this time the proofs are checkable
On August 1, OpenAI published solutions to ten open problems in mathematics and theoretical computer science, including a non-sofic group construction and new sphere-packing bounds. The formal Lean proofs went up on GitHub under Apache 2.0.
I want to be careful here because I nearly got this one wrong. My first pass through the repository turned up 42 instances of sorry, which is Lean's placeholder for an unfinished proof, and I thought the "no gaps" claim was overstated. It wasn't. Those are statement stubs sitting in a separate challenges directory. The ten actual formalizations are clean: 548,205 lines of Lean, zero sorry, and no added axioms.
Separately, Google published two systems in Nature. ERA found 40 new single-cell analysis methods that beat the best human-designed ones on a public leaderboard.
The reason this matters more than any benchmark score is that a Lean proof is machine-checkable. You do not have to trust OpenAI's characterization of what it did. You can download it and verify it, which is exactly what I did. Almost nothing else in this issue has that property. Most AI claims are the company grading its own homework. This one hands you the answer sheet and invites you to check.
And the near-miss is the actual lesson. I was checking carefully in order to include it, and not checking as carefully before throwing it out. Skepticism that only runs one direction isn't skepticism.
19. Health AI, told honestly
Two studies, and you need both. A JAMA study across five academic medical centers found ambient AI scribes cut EHR time by 13.4 minutes and documentation time by 16 minutes per encounter, and added 0.49 visits per week per clinician. A randomized controlled trial in NEJM AI found the same tools reduced practitioner work exhaustion but did not significantly increase professional fulfillment. Documentation time fell with no measured loss in diagnosis, billing or note quality.
Meanwhile the FDA's authorized device list still contains no authorized generative AI device. It gave a breakthrough designation to a patient-facing generative app in March, and that's it.
Less exhausted but not more fulfilled is one of the more honest findings I've read about this technology. It took away the part of the job that was grinding them down and it did not give back the part that made the job worth doing. Those are separate problems and we keep selling the fix for the first as though it solved the second.
20. The robot helped most when the person chose the settings
A meta-analysis in The Gerontologist covering 19 studies and 1,083 participants found that social robots significantly reduced loneliness in older adults. The effects were larger in institutional settings than for elders living independently, and stronger in Japan and Turkey than in the United States. This is my own industry, so I saved it for last.
The finding underneath that one is the reason it is here. Participants who were allowed to configure the robot themselves had better mental health outcomes than participants who passively received the same device.
Same machine. Different outcome, based entirely on whether the person got to decide how it worked.
I've been in senior living long enough to have watched a lot of things get done to residents rather than with them, and it is almost always well-intentioned. Somebody buys a system that's good for people, installs it, and is puzzled when it doesn't land. This study says the choosing was doing real work, not the device.
Which is where I'll leave the whole issue. Twenty items, and the one that will change what I actually do on Monday isn't about a model or a market. It's a reminder that handing somebody a good tool and handing somebody control over a good tool produce different results, and that we keep measuring the tool.
