Planning delays are the new chip shortage

The limit on AI right now is people, not chips.

Sunday. Three pieces from yesterday that look unrelated: a Reuters column on delayed data centres, a poll about where people will accept one, and TechCrunch's running list of dead AI products. Each one is somebody outside the industry declining to go along, a council, a voter, a paying user.

In 30 seconds:

1. 72% of Americans don't want a data centre in their own town. Several US states have halted new construction, and a 50 MW project in Melbourne was shelved after local objections.
2. Democrats are keeping AI out of their own ads. More than 200 AI-made political ads are running, and they're betting the Republican ones cost votes.
3. Character.ai spent $13.8 million a month on compute and earned $2.8 million. It's the newest name on a list where about 42% of company AI projects get abandoned.

Advertisement

1. States and neighbours are stopping data centre plans

Several American states have stopped new data centre construction outright, and in Australia local opposition pushed NextDC to shelve a 50 MW project in Melbourne, which its share price noticed.

Here is the build-out those decisions sit in front of. Stijn Van Nieuwerburgh of Columbia Business School puts American capacity at 183 GW by 2032, roughly five times the pre-2025 level, costing more than $10 trillion and over 3.5% of GDP across 2022 to 2032.

An ABC News, Washington Post and Ipsos poll of 2,545 US adults has 66% against new data centres and 72% against one in their own community.

NextDC did not stop at one site. On 23 October it paused all uncontracted development and land acquisitions pending a review of its capital programme.

My read is that the planning queue has quietly replaced the chip queue as the thing that decides 2028 capacity, and almost nobody is pricing it. A council hearing and a grid connection are slow, local and public, and no amount of private credit raised this month makes either move faster. So when a supplier promises me capacity for 2028, I want to know which of its sites hold planning permission and a signed grid connection, and I read the rest as intent.

Read the source →

2. Campaigns are betting that visible AI costs votes

More than 200 AI-made political ads have run in this American midterm cycle, with over $100 million behind them out of more than $10 billion of total campaign ad spending. Thirty-one states have laws covering AI-generated political advertising. Nothing has passed Congress or the Federal Election Commission, and Tennessee is one of the states that wants a notice at the beginning and the end of the ad.

The part worth reading twice sits on the other side of the fence. Democrats are keeping AI video out of their own ads and betting that the Republican ones backfire with voters.

The thing I am watching is not the ads, it is the disclosure rule that follows them. Thirty-one state laws and nothing federal is the shape of a market where the cost of an AI mistake lands on whoever ran the thing rather than on whoever built it, and a candidate works that out faster than a procurement officer does. My expectation is that the first labelling requirement with real teeth comes out of a race somebody lost, not out of the FEC.

Read the source →

Advertisement

3. The dead list now runs to a page

TechCrunch published a running list of AI products and companies that did not make it. The number sitting under it, from S&P Global Market Intelligence: about 42% of AI initiatives get abandoned by their corporate parents.

Three entries, by money burned. Humane raised $230 million for the AI Pin, and HP bought most of what was left for $116 million. Character.ai filed for bankruptcy five years after launching. The Browser Company is switching Arc off in stages, with Arc Search losing its AI features on 23 November.

Character.ai's filings put the arithmetic in public: $13.8 million a month on compute against $2.8 million of peak subscription revenue, and on 20 October Disney agreed to buy selected assets for $20 million.

None of that is a model quality story. Somebody has to pay the monthly bill in the months before a product is loved, and $13.8 million against $2.8 million is a treasury problem wearing a product's clothes. Given the choice I take a dull AI feature inside something people already pay for over a beautiful standalone that needs a new habit first. My guess is the consumer half of that list grows faster than the enterprise half through next year.

Read the source →

Names to watch

Seen this week by Ins7ghts.

• Promenaut · an orchestration layer for enterprise agents
  Governance. A-Team Insight wrote up the platform on 24 October, and the company developed it in collaboration with HSBC.

Companies that later raised money had been in the news about twice as often in the quarter before. Worth a second look in three months.

Advertisement

Also today

• The UK committed £10 million a year to studying what AI does to employment. (2 min read)
• Hong Kong's monetary authority wants a named owner for every AI use case a bank runs, plus an inventory that includes the AI embedded in bought-in tools. (13 min read)
• Music publisher Reservoir sued Suno over 1,050 songs and Anthropic over the lyrics of 853, seeking more than $330 million. (9 min read)

Week AheadThe next three dates somebody else set

When I started bringing a PC along to play MP3s, there was a night the machine did not boot. I played that night with the vinyl I had brought with me.

Three dates in the next two months work the same way: they arrive whether or not you are ready, and none of them were set by the people who have to absorb them. The nearest is Tuesday 3 November, the American midterms, the first national vote with more than 200 AI-made ads inside it and no federal rule about labelling any of them.

Which of the systems you depend on switches off on a date you did not choose?

My call: before the end of November 2026, at least one US campaign pulls an AI-made ad over the public reaction to it, not over a legal complaint.

Read the deep dive →