The fish and the trap

The commons is not being abandoned. It is being out-waited.

Everyone is reading the data centre power story as an exit. The numbers say something stranger and more useful: almost nobody has actually left.

The energy analysts at Cleanview count 59 data centre projects planning to build their own power behind the meter, around 90 gigawatts of it, mostly gas (American Public Power Association). Then the rest of the number: about 2 gigawatts of that is actually operating. Roughly 1 per cent is under construction, about a third is permitted, and some 60 per cent exists only as an announcement. Ninety gigawatts declared, two built.

Why declare it at all? Because joining takes too long. Ember puts the wait to connect a data centre to the grid in an established European hub at seven to ten years (Ember). On the finance side, the attraction of going behind the meter is stated plainly by Jefferies: it lets developers avoid accountability for the effect they have on everyone else's electricity rates (Latitude Media).

The same shape turned up twice more in the same few weeks. In Michigan, 81 per cent of voters backed a proposal barring utilities from political giving; DTE and Consumers Energy filed suit in federal court two days later, and the law takes effect on 3 December unless a court stops it (Bridge Michigan). At COP31, pledged public climate finance fell by about 71 per cent, from roughly 40 billion dollars a year to 11.7 billion, with the United States contributing nothing and the EU climate chief calling fresh money "dead in the water" (Carbon Brief).

Three exits in a month: from the grid, from the rules, from the pot. But the honest version is more interesting than a story about powerful actors walking away, and Ember's own findings are what complicate it. In its Texas case, two large data centres running their own generation could take 2.7 gigawatts off the grid during a stress event, against 0.55 gigawatts from load shifting. Leaving can relieve the system you left. Amazon Web Services, meanwhile, says its goal is to get back to being grid-tied, and Schneider Electric reports that most large operators do not want to own power stations at all.

So this is not hostility. Ninety gigawatts announced against two built is a negotiating position, and it is working, because the alternative is a decade in a queue. The thing eroding shared systems is not that people have stopped believing in them. It is latency. A commons that takes seven years to join will be left by people who had no particular wish to leave.

Worth asking on Monday: where in your organisation is going around the process reliably faster than going through it? That is where your exits are being announced. They will be announced long before anyone builds one.

The connecting is finished. The deciding is not.

A machine has started doing the thing this publication claims as its edge, and the interesting part is the bit it could not do.

Matthew Schwartz, a Harvard physicist working as a visiting researcher at Anthropic, had Claude build a set of tools for mathematical physics. The tools turned out to travel: they applied to problems in fields they were never built for. The run produced 36 manuscripts in three months with 19 coauthors, across 18 fields by his own count and 22 on the project's site, and the harness was released openly under an MIT licence (Unite.AI, The Decoder).

One example stands for the method. A statistical tool doctors use to measure how the risk of dying changes with age was pointed at sunspot data (Singularity Hub). Borrowing a technique from a distant field and finding it fits is exactly what a cross-domain editorial practice sells. It is now available at volume, free, to anyone.

Here is the number that matters, though, and it is Schwartz's own. Around 400 candidate problems became 36 papers. Humans did the discarding. And his flattest observation about the system is the one worth keeping: Claude often produced results that were technically correct and that domain experts judged uninteresting.

That is a precise description of where the scarcity moved. Connection was never the bottleneck. Significance was. A machine can now generate a thousand true statements about how one field resembles another, and has no view on which of them anyone should care about. The mathematician Terence Tao has made the companion warning: mass-produced solutions can bypass the collaboration and the explaining that let a field actually understand what it has been handed.

Which lands awkwardly beside the other thing happening this month. England's higher education regulator now says AI has made assessing university coursework close to impossible (The Guardian). So at the same moment that producing plausible work became free, our main instrument for telling who understands anything stopped functioning. We are better than ever at generating candidates and worse than ever at telling who can judge them.

Worth asking on Monday: of the work you did last week, how much was connecting things, and how much was deciding what deserved attention? One of those just became abundant. Most job descriptions, including ours, still mostly describe the other one.

The curve is bending while the agreement collapses

Two well-evidenced positions that flatly contradict each other, held at full strength.

Position one: it is working. China's quarterly CO2 fell around 3 per cent, its largest drop since Covid, and global fossil-fuel CO2 is set to dip in 2026 (Carbon Brief). John Kerry's assessment at COP31 was that the United States and everyone else now need to follow China's lead on green energy (The Guardian). The curve people have argued about for thirty years is bending, and it is bending on manufacturing cost and industrial policy rather than on anybody's goodwill.

Position two: the apparatus built to deliver that is falling apart. Pledged climate finance is down about 71 per cent, the United States is contributing nothing, and 2026 is on track to tie 2024 as the warmest year on record. MIT Technology Review's verdict on the industry now setting electricity demand is blunt: the climate case for AI has not been made (MIT Technology Review).

Both survive scrutiny, and position one carries a detail nobody should enjoy: part of the 2026 dip comes from the Iran conflict suppressing oil use (Carbon Brief). An emissions reduction delivered by a war is not a policy, and nobody should want it on the scoreboard.

What has to be true for both? That the agreement was never the mechanism. It was the scoreboard. The actual mechanism turned out to be price, industrial strategy and, this year, disruption. Which cuts in both directions and should unsettle everyone equally: the collapse of the pledges tells us less about future outcomes than it feels like it should, and the falling curve tells us much less about our competence than we would like to claim. We have been grading ourselves on the wrong instrument for a long time, and this is the year it became obvious in both directions at once.

The teeth: if your results improved this year while your governance fell apart, which number goes in the report? And which one do you actually believe?