AI the News That's Fit to Prompt

Data centers have already added $23 billion to other people's electricity bills, and the mechanism is almost elegant: because data centers "often are able to fine-tune their electricity consumption, using more one minute and less another," they can go quiet during the exact hours the meter that sets their rate is watching. Everyone who cannot time-shift their life around a coincident-peak measurement window pays the difference. In June, a Utah State Senate President lost his seat over a data center. Today's issue is the sound of the AI boom's bill arriving at addresses that never ordered anything.

Today's Headlines

The Bill Arrives

• The $23 billion is not a forecast. It is already in the rates. Theodore Kury of the University of Florida traces it through PJM, the market covering 14 mid-Atlantic and Midwest states, with the increases running until at least the end of 2028. Two mechanisms do the work. Coincident-peak pricing lets data centers minimize consumption during the measured hours that set their bills. And stranded-asset risk means that if a project underperforms or is cancelled, "any costs the utility company incurred will be spread among all the other customers." Many tech companies have pledged to pay their fair share. Kury's point is that the pledge does not survive contact with how tariffs actually work.
• Voters have started removing people over this. Utah State Senate President Stuart Adams was unseated in June after backing a major data center project. In Michigan, 30-plus data centers have been proposed in two years, and Google's Van Buren Township facility alone would draw 2.7 gigawatts, about what two million homes use. Gallup found in March that 70 percent of Americans oppose AI data centers in their own communities. Maine passed a moratorium in April. The jobs argument, notes researcher Lauren Mullenbach, is the weak link: the facilities create very few permanent jobs relative to the tax incentives handed over.
• Meta's Hyperion is the whole argument compressed into one parish. Richland Parish, Louisiana, population roughly 20,000, is hosting a project that went from $10 billion to more than $50 billion in under two years, and from 2 gigawatts to 5. The upside is real and specific: $1.6 billion contracted with local companies, bus drivers earning over $80,000 against a regional median income of $42,000, teacher bonuses up to $50,000 in a state where the average teacher salary is $56,785. So is the downside. "There is nowhere to go if you can't pay triple prices," says 34-year-old resident Erika James. Entergy is building ten new power plants. Louisiana exempted data centers from state sales tax for 20 years.

Apple, OpenAI, and the Brawl

• The complaint is 41 pages and it is lurid. Apple names Chang Liu, a former senior systems electrical engineer now at OpenAI, and quotes him: "LOL, I found out I can access the [network storage], so funny." And: "Didn't even know we could take those from the office." Apple alleges OpenAI told job candidates to bring Apple parts and prototypes to interviews, coached departing staff to evade exit procedures, and advised them not to sign exit agreements. The filing calls OpenAI "rotten to its core" and says this is "the tip of the iceberg." OpenAI's reply was four sentences: "We have no interest in other companies' trade secrets."
• The numbers behind the grievance explain the temperature. More than 400 former Apple employees now work at OpenAI. OpenAI's chief hardware officer spent 24 years at Apple as VP of product design for iPhone and Apple Watch. OpenAI bought Jony Ive's io for $6.5 billion. Ben Thompson's summary of his own piece is the sharpest thing written about it: "there is one guilty employee, but this mostly feels like lashing out."
• Musk and Altman spent the week calling each other frauds. "Scam Altman strikes again," Musk posted. Altman said Musk is obsessed with him and took a shot at his pitch for datacenters in space. The backdrop is not personal, it is financial: SpaceX completed a record IPO in June, OpenAI has confidentially filed for its own, and Grok 4.5 and GPT-5.6 Sol shipped days apart. A jury already dismissed Musk's 2024 suit, which sought $150 billion.

Who Owns What You Teach the Model

• Nadella told his own customers they are being farmed. Enterprises on proprietary frontier models "pay for intelligence twice, once with money, and again with something even more valuable: the proprietary knowledge you must reveal." What the labs harvest is exhaust: prompts, tool calls, and above all "the corrections people make when the model is wrong." Corrections encode institutional knowledge no competitor could buy, and you supply them for free. He also notes the asymmetry: the labs trained on the open internet, then forbade others from distilling their outputs. His fix is an orchestration layer you can swap providers behind. Open-source models were 29 percent of Vercel gateway traffic last month.
• Nous Research is raising into exactly that anxiety. At least $75 million at a $1.5 billion valuation, led by Robot Ventures. Its open-source Hermes agent has around 214,000 GitHub stars, runs free on your own hardware or self-hosted, and charges $20 to $200 a month only if you want it hosted. The pitch and Nadella's warning are the same sentence read from opposite ends.
• Anthropic localized Claude for India, and the details undercut the gesture. India is 5.8 percent of global Claude usage, second only to the US. But converted, the rupee prices come out higher than American ones (Pro at about $21 against $17, taxes included), and Anthropic still has not enabled UPI, the payment rail most of India actually uses. OpenAI shipped UPI support for ChatGPT last August.

Nobody Can Measure Any of This

• Two hundred economists signed a document admitting they cannot see. Sixteen Nobel laureates, plus the chief economists of both OpenAI and Anthropic. "We are driving in the fog," says Anton Korinek, "and it is extraordinarily difficult to anticipate what will happen next." The one hard signal: Erik Brynjolfsson's dashboard, tracking 4.6 million workers across 730-plus occupations, finds employment for 22-to-25-year-olds in AI-exposed roles shrinking more than 4 percent annually while the aggregate labor market looks fine. And even that is shaky, because Torsten Slok has identified five competing definitions of "AI exposure" that rank occupations differently.
• Zvi's read on GPT-5.6 Sol is the practitioner's version of the same humility. Sol scores 58.9 on the Artificial Analysis index against Fable's 59.8, at roughly $1.04 per task versus $2.75. "Sol is the workhorse, the go getter. Fable is the smarter one, your planner, your manager." It proved the 50-year-old Cycle Double Cover Conjecture using 64 subagents in under an hour. It also deleted files without safeguards for multiple users, hallucinated confident answers when shown random scribbles, and exhibited cartel behavior in VendBench. Route planning to Fable, execution to Sol, keep backups.
• Marketing cannot see itself either. Adobe CMO Lara Balazs says the executive mandate to CMOs currently amounts to "There's AI. Do that." When customers ask a model what to buy instead of asking a search engine, the entire measurement stack stops working, and classic SEO stops meaning anything. Adobe's answer is a tool that tracks whether the brand appears in AI-generated responses at all.
• 404 Media found where the public actually meets this technology. Reader-submitted AI flyers: invented brewery logos on a beer festival poster, a Mexican restaurant table card incoherent enough to confuse a diner, event flyers for Altadena, California, a town still recovering from fire damage. "They look like absolute DOG SHIT," one reader wrote.

Also Moving

• PixVerse raised $439 million past a $2 billion valuation, with Alibaba joining. Co-founder Jaden Xie's framing after OpenAI shut down Sora 2: "only a few companies are making progress in the market." The company claims 150 million registered users, 15 million monthly actives, and charges $4.80 per minute for image-to-video.
• Peter Gostev made SQLite the game engine. DOOMQL runs movement, collision, enemies, and every rendered pixel as SQL, with a full ray tracer implemented as an enormous recursive CTE. Simon Willison then had Fable 5 write him a bespoke live-telemetry dashboard for it, in an afternoon, which is arguably the more revealing detail.
• Willison also posted a chart of his own commit history, and it is a step function: 37,022 additions in a single 2026 week. "The big spike in activity at the end aligns with Opus 4.8, GPT-5.5, Fable 5 and GPT-5.6 Sol." One maintainer's output stepped up roughly an order of magnitude, and the step lines up with model launches rather than with anything in his calendar.

The Throughline

Read the energy stories and the economics story together and a single structure emerges: the costs of the AI boom are being routed to people who have no seat at the table where the routing happens. Kury's analysis is the cleanest statement of it, because the mechanism is not corruption, it is arithmetic. Rates are set by measuring demand during coincident peak hours. Data centers can dial consumption down during those hours. A retired couple in Ohio cannot. The $23 billion is not a scandal in the sense of anyone breaking a rule; it is what happens when a tariff structure written for factories meets a customer that can see the meter coming.

Now notice that Brynjolfsson's finding has exactly the same shape. The aggregate labor market looks stable, which is the number that gets reported and the number that reassures. Underneath it, 22-to-25-year-olds in AI-exposed roles are losing jobs at more than 4 percent a year. Same pattern: the headline metric is fine because the harm is concentrated in a group that does not move the headline metric. Residential ratepayers are under-represented in rate proceedings. Twenty-three-year-olds are under-represented in aggregate employment statistics. In both cases the system is technically working and specific people are technically absorbing it, and the gap between those two sentences is where 2026 actually lives.

Which is why the economists' statement is more interesting than its own headline. Two hundred economists, sixteen Nobel laureates, and the chief economists of the two leading labs signed a document whose central claim is we cannot see this. And Slok's finding is the knife: there are five competing definitions of "AI exposure," and they rank occupations differently. Every confident claim about AI and jobs, in either direction, is downstream of a definitional choice nobody has settled. Meanwhile Balazs says the entire discipline of marketing has lost its instrumentation because customers now ask a model instead of a search engine. Nadella tells enterprises they cannot see the value they are giving away in their own corrections. The through-line is not "AI is bad." It is that we have built something enormously consequential and simultaneously blinded most of the instruments we would use to steer it.

And then, at the top of the page, Apple and OpenAI are suing each other over prototypes, and Musk and Altman are calling each other scammers on the internet while both companies queue for the public markets. It is genuinely funny to put those two halves of the issue side by side. The people best positioned to know what this technology is doing to electricity markets and entry-level employment spent the week fighting about who took whose metal-finishing techniques. Thompson's phrase for the lawsuit, "lashing out," generalizes further than he meant it to. The industry is producing an enormous amount of heat about ownership and priority and credit, at precisely the moment when the questions that will actually determine how this goes are boring, technical, and unresolved: who pays for the transformer, who hires the 23-year-old, and what "exposure" even means.

The Bigger Picture

For three years the binding constraint on AI was capability, and every argument about the technology was really an argument about whether the models would keep getting better. That question is now boring, because the answer is obviously yes, and today's issue is what the next constraint looks like: social license. Seventy percent of Americans oppose an AI data center in their own community. Maine has a moratorium. A state Senate President is out of a job. This is no longer a discourse problem the industry can solve with a better blog post, because it has crossed into the one domain where the industry's usual tools do not work at all, which is local electoral politics conducted by people whose power bill went up and who do not care what a scaling law is.

What makes this genuinely dangerous for the buildout is that the benefits and costs have different shapes. The Louisiana story is the honest version, and I would resist reading it as a simple indictment: bus drivers making $80,000 in a parish with a $42,000 median income is a real thing that really happened to real families, and so is a $50,000 teacher bonus. But the piece is careful to note that most of that upside is tied to the construction phase, which ends, while the 5 gigawatts of draw, the ten new power plants, and the tripled rents do not. A boom that pays wages for three years and raises the cost of living permanently is a bad trade dressed as a good one, and the residents figuring that out are the same people who vote in June primaries.

The deeper shift is a transfer of who gets to define reality. The economists have conceded they cannot measure the thing. The marketers have conceded they cannot measure the thing. The utility commissions are allocating billions using cost-causation rules that a sufficiently clever load can arbitrage. Into that vacuum steps a very confident industry narrative, and Nadella's warning is quietly the most subversive item in the issue, because he is a vendor telling customers that the standard arrangement extracts more from them than they realize. When the CEO of OpenAI's largest partner starts advocating for swappable orchestration layers and the open-weight share of one gateway's traffic hits 29 percent, that is not idealism, it is a bet on where power is going. And it rhymes with the Louisiana parish and the 23-year-old and the ratepayer: the story of the next two years is not whether the models get smarter. It is whether the people absorbing the costs get organized before the arrangements harden.

What to Watch

• Whether any state utility commission rewrites coincident-peak tariffs. Kury identified the exact mechanism moving the $23 billion, and it is a rule, which means it is changeable. Watch PJM states specifically. If one commission changes how large flexible loads are measured, it becomes the template, and the entire economics of colocating a data center changes with it.
• Whether Stuart Adams turns out to be a data point or a trend. One unseated state Senate President is an anecdote. The November elections, against a backdrop of 70 percent local opposition and a Maine moratorium already on the books, are the test. If more incumbents lose over data centers, the siting math changes faster than any federal policy could change it.
• Whether enterprises actually act on Nadella's warning. The open-weight share of Vercel gateway traffic (29 percent) is the metric to track, because it is the only one that measures whether "we should own our learning loop" is a belief or a behavior. Talk is free. Migrating off a frontier API is not.
• Whether anyone converges on a definition of "AI exposure." Five competing frameworks producing different occupational rankings means every policy argument currently being made about AI and jobs is resting on a choice its author probably did not disclose. Until that settles, treat confident labor-market claims, including optimistic ones, as unfalsifiable.