Voters turned on data centers, the flat-rate AI plan started to crack, and a 10-person company showed what spending 30 times payroll on agents looks like

The US midterms handed Congress and most big states to candidates who campaigned against data centers, in the same week one AI provider cut usage limits without warning and said the service cost too much to run. For a business, both point the same way: the cheap, unlimited phase of AI is ending, so know what you use, what it is worth, and how you would cope if the price or the rules changed.

• Democrats won Congress and the governorships of nine of the ten largest states, and every one of 41 local data center measures looks set to pass
  My take: AI itself was not the top issue, but data centers were a clear vote loser. In one national poll 61 percent of voters said they dislike candidates who support building them locally, and only 5 to 14 percent said they like them. Texas elected a governor who promised a pause on new data centers until rules are written and an end to their tax breaks, and Ohio and Georgia flipped too. I would not expect a sweeping federal AI law, because the president has promised to veto. What looks likely is smaller rules that both parties can accept: whistleblower protection, incident reporting, testing before release, and more legal liability for the companies that build models. Two things follow for a business. First, the supply of computing power will grow more slowly than the AI companies planned, so do not build a budget that assumes prices keep falling and limits keep rising. Second, new duties placed on AI vendors tend to arrive in your contract as new terms, so read the next terms update from your AI provider instead of clicking accept. If you sell an AI product yourself, start keeping a simple log of failures and complaints now, because incident reporting is the rule most likely to reach you.
• Grokbot users hit new usage limits with no notice, and Elon Musk said the cost of serving them was unsustainably high
  My take: Users of the Grokbot personal agent ran out of usage days early, and the explanation was blunt: it costs more to serve than it earns. Two weeks earlier the same product was described as the fastest growing thing its maker had seen. This is the clearest sign so far that flat monthly AI plans are priced below cost. One analysis this summer found that a full month of heavy use on a 200 dollar plan would cost about 14,000 dollars at normal pay-per-use prices. I do not think this makes AI a bad buy. For work that replaces hours of skilled labour it is still cheap at the real price. But a plan that changes under you with no warning is a business risk. Do three things. Find out which of your team's tasks run on flat subscriptions and which on metered billing. For the two or three workflows you depend on most, estimate what they would cost at pay-per-use rates and decide whether they are still worth it. And move routine, high-volume jobs to a cheaper model now, while it is a choice and not an emergency. On one new benchmark of valuable professional tasks the second-place model scored about one point lower than the leader at a third of the cost.
• Google took the top spot on the main AI ranking two weeks after its last model, and its search ad growth slowed to 12 percent
  My take: Google announced Gemini 4 Carbon only two weeks after its previous model, and it now leads the most widely used independent ranking, one point ahead of Anthropic's Opus 5.5 and about 23 percent cheaper per task. It is not on general release yet, so treat the scores as a promise until people have used it. The ranking itself was rebuilt this week around agent work and private tests that are harder to game, and the gap between the best American models and the best open Chinese model widened from 3 points to 12. The other half of the story is in Google's results. Cloud revenue doubled, but search growth fell from 19 percent to 12 percent in two quarters, and Google said its AI answers earn less per query than normal search results. Two actions. If you rely on Google search for customers, check how your traffic and ad results have moved over the last six months, because more people now get an answer without clicking. And do not sign a long contract tied to one model. The leader changed twice in a month, and Google says the pace will increase.
• A 10-person game company spent 3.1 million dollars on AI in a month, 30 times its payroll, and its agents shipped 77,000 code changes
  My take: Spawn runs a permanent team of AI agents, each with a name, a memory and a clear area it owns, and the humans set direction. The agents made 123,000 changes over the past year and only about 1,000 had to be undone. Most businesses should not copy this. It needs a product where problems show up within minutes, and a tolerance for agents breaking things in production. But one idea in it works for any company. The founder woke up to 700 questions from his agents and realised he had answered most of them before. The team turned two years of past decisions into written principles, and now an AI judge applies them and passes only new situations to a human. In the first eight days the judges handled more than 11,000 decisions and the founder made 27. Output rose from 1,000 changes a day to 7,000. You can do a small version of this without any agents. Take the decisions you make again and again, such as refunds, discounts, exceptions and approvals, and write down the rule and the reason. That document makes your staff faster today and it is exactly what an AI assistant needs to act for you later. The founder's other lesson is worth keeping too: the agents never asked for anything ambitious, so the big goals still had to come from people.
• About 40 percent of desk workers received AI-written filler from a colleague last month, and each piece cost the reader close to two hours
  My take: A study of US desk workers found four in ten had been sent work in the past month that looked finished but was not, and it cost the person on the receiving end almost two hours to sort out. In a separate survey of 6,000 workers, almost 70 percent admitted sending work to someone else without reading it. That is the real problem with AI-written content inside a company. The sender saves ten minutes and the reader pays for it. The fix is not to ban the tools. I'd suggest three rules for your team. Decide what you think before you open the AI tool, even as rough spoken notes, because the model gives you the average answer unless you give it your own view. Never send anything you have not read in full and could not defend line by line. And match the effort to the stakes: a routine internal email can be fully AI-written as long as it is short, but anything a customer sees needs a person's judgement in it. If you manage people, send weak work back and ask for the point in five sentences. Once everyone can produce a polished 40-page report in a minute, the people who stand out are the ones whose work is short, specific and clearly their own.
• An AI legal startup bid to buy a bankrupt 95-year-old Wall Street law firm for its name and its archive of past work
  My take: Crosby, a 50-lawyer firm that does legal work with AI, offered to buy the brand and internal data of a top-20 law firm that collapsed after its senior partners left. A law firm without its partners used to be worth close to nothing. The bid exists because the firm's templates, precedents and step-by-step methods for complex bankruptcy work are what an AI system needs to learn that work. Google recently bid 10 million dollars for the operating data of a failed airline for the same reason. Critics say much of that material belongs to clients and cannot be sold, and the deal may not happen. The lesson holds either way. The written record of how your business does its best work is now an asset with a price, and most small companies keep it in people's heads and scattered inboxes. Start collecting it: your best proposals, your checklists, how you handled the hard cases. It makes your own AI tools far more useful, it protects you when a key person leaves, and it adds value on the day you sell the business.