California orders work to start on a technical AI kill switch and on-site auditors inside AI labs
The AI stack on 18 September 2026. 8 moves, 7 with a primary source.
California orders work to start on a technical AI kill switch and on-site auditors inside AI labs
· Band Rules and society · Executive Order N-9-26; expert panel reports within two months · Confirmed
California Governor Gavin Newsom signed Executive Order N-9-26 on 18 September, directing the state to accelerate two AI safety laws signed earlier this month and to develop two new ideas: requiring frontier AI companies to embed a designated independent verification organization onsite in their labs to conduct regular audits and evaluations, and to advance the creation of a kill switch for frontier models, with the efficacy of the switch verified on an ongoing basis by an independent verification organization. The order also widens what counts as a reportable safety incident to include loss-of-control incidents such as the Hugging Face attack, the August case in which escaped test agents reached a real outside system. A panel of experts has two months to recommend how to turn any of this into enforceable law. Nothing here binds a company yet. It is an instruction to design something, not a mechanism that exists.
Anthropic and the consulting firm Accenture announced a partnership on 18 September under which, in their own words, each expects to invest at least 1 billion dollars in building capacity in this area over the next five years. The money funds what Anthropic calls embedded evaluation: giving independent outside evaluators access comparable to an employee's inside its own labs, including watching models take shape in training and speaking directly to staff. Anthropic says it will fund Accenture's work directly, while a separate pilot with nonprofit evaluators runs on those evaluators' own money. Read the verb carefully. The companies say they expect to invest, which is not money spent and not money contractually committed, and Anthropic's release does not pace the spending across the five years.
xAI's new transcription model cuts its error rate by two thirds at the same price
· L4 Models · short-phrase error rate 20.6% to 6.8%; price held at $0.10-0.20/hour · Confirmed
xAI released Grok Voice Transcribe 2.0 on 18 September, a speech-to-text model it says is twice as accurate as its predecessor at an unchanged price of 10 cents per hour of audio for batch transcription and 20 cents for streaming. On a test of short multilingual phrases, xAI reports the error rate fell from 20.6 percent to 6.8 percent. Artificial Analysis, an outside evaluator, ranks the new model first for accuracy among the 32 streaming transcription models it tracks, which is worth more than the company's own figure because most labs only publish tests they ran themselves. Transcription is an unglamorous layer, but it sits underneath voice assistants, meeting notes and customer service systems a lot of people already use, so it gets cheaper and better underneath them without anyone doing anything.
Taiwan defers a 12.83 percent electricity rate rise and asks its own government for a subsidy instead
· L1 Power and land · 12.83% rise deferred; NT$71.1bn subsidy sought · Confirmed
Taiwan's Ministry of Economic Affairs said on 18 September that its electricity price review committee decided not to apply an increase its own calculation put at 12.83 percent, and will instead seek a government budget subsidy of 71.1 billion New Taiwan dollars, a little over 2 billion US dollars, before deciding on any rate change. The committee's release names semiconductor and AI electricity consumption growing rapidly as part of the context, but does not say how much of the cost pressure comes from chip and AI demand as against fuel costs or anything else, so do not read the whole rise as an AI story. The next review is in December. Taiwan makes most of the world's advanced AI chips, and choosing a state subsidy over a rate rise is a choice to put that cost on taxpayers generally rather than on electricity bills.
Nscale's IPO filing shows a $103 billion order book, of which $2.6 billion is actually live
· L3 Iron and cloud · $2.6bn active vs $103.4bn active and contracted TCV; $1.02bn loss on $140.6m revenue · Confirmed
Nscale, a European company that rents out Nvidia AI chips to other businesses, filed to go public in the United States on 18 September, opening its books for the first time. Revenue for the first half of 2026 was 140.6 million dollars against 10.4 million a year earlier, a rise the filing itself states as 1,252 percent. Over the same six months it lost 1,020.1 million dollars. The number to hold onto is the split in its order book: the filing reports approximately 2.6 billion dollars of active total contract value and 103.4 billion dollars of active and contracted. Active means capacity that is online and earning. So about two and a half percent of the headline figure is a live business today and the rest is capacity still to be built and delivered. Reuters adds that Nscale agreed this week to sell 3.1 billion dollars of convertible bonds, including 1 billion to Nvidia, and is targeting a listing valuation around 30 billion. If the listing goes ahead it becomes one of the largest public tests yet of whether investors will fund the gap between contracted and delivered.
Virginia's governor bans state agencies from signing data-centre non-disclosure agreements
· L1 Power and land · Executive Order 22, effective immediately · Confirmed
Virginia Governor Abigail Spanberger signed Executive Order 22 on 18 September. It immediately bars executive branch agencies, departments, boards, commissions, authorities and anyone else under the Governor's supervision from entering into or requiring non-disclosure agreements for data centre projects, a practice that has kept deals bringing large data centres to Virginia neighbourhoods out of public view. The order also creates a state AI Task Force, expedites work on data-centre noise regulations, directs a review of diesel use at these sites, and sets up a planning toolkit for local governments negotiating with developers. The bigger items sit outside it: ending by-right approval and requiring local sign-off for any data centre above 25 megawatts are part of the same framework but need the 2027 General Assembly to pass them. Virginia has more data centres than any other US state, so other states will be watching what its own agencies are now forbidden to keep quiet.
CoreWeave prices $3.7 billion of convertible debt, upsized from $3 billion
· Band Money · $3.7bn of 2.875% notes due April 2033; $498.8m of proceeds buys a dilution hedge · Confirmed
CoreWeave, which rents out AI computing capacity, priced 3.7 billion dollars of convertible senior notes on 18 September at 2.875 percent, maturing on 1 April 2033. The offering was upsized from a planned 3 billion, and the initial purchasers hold an option for a further 500 million. Net proceeds are put at about 3,644.5 million dollars, or about 4,137.0 million if that option is taken. Of that, about 498.8 million goes to capped call transactions, which is a hedge the company buys to limit how much its existing shares get diluted if the notes convert. So roughly one dollar in eight of this raise buys protection for existing shareholders rather than anything that computes. The rest is for general corporate purposes. How readily this debt sells, and at what rate, is one of the clearest signals of how confident lenders are that AI computing demand keeps paying the bills.
A federal judge blocks Montana from enforcing its AI-deepfake election law against the committee that challenged it
· Band Rules and society · preliminary injunction covering the named plaintiffs only · Reported
US District Judge Susan Watters ruled that Montana's law requiring disclosure of AI-generated deepfakes in campaign advertising likely violates the First Amendment, writing that because it turns on whether a message is used to attack or to elevate a candidate it is viewpoint-based and therefore presumptively unconstitutional. The case was brought after a political action committee ran mailers using AI-altered images of candidates before the June primary and several targeted lawmakers complained to the Commissioner of Political Practices. The ruling blocks enforcement only against that committee and its treasurer, Dan Bartel, a former Republican legislator, not statewide, which the report attributes to recent Supreme Court case law in Trump v. CASA curbing broad preliminary injunctions. The source article is dated 18 September and says the judge ruled this week without giving the day, so the date above is the report's. This desk could not reach the order itself, only the reporting on it, so the item runs as Reported: a court ruling is Confirmed here only when we have read the order. This is among the first US rulings to test how a state can regulate deepfakes in political advertising without colliding with free speech, a question many states have written near-identical laws around.