What does not recurse

A fabrication plant takes years to build. A grid connection is granted by a body with a queue and a committee. The constraints that set the pace of this loop are duller than intelligence: civil engineering, electrical supply, and the calendar.

Concrete, copper and permits

The equipment inside a fabrication plant is made by a handful of firms with order books. A data centre needs land, planning permission, water and a connection to a grid that somebody else operates. A substation is a civil engineering project. None of these compress at the rate the design work does, and all of them are needed before the design work can be tested at scale.

Electricity is where the constraint became visible. Data centres took about 1.5 per cent of world consumption in 2024 and are expected to double that share by 2030, which reads comfortably until it is read locally. The European Parliament’s research service puts EU data centres at [around 3 per cent of demand](https://www.europarl.europa.eu/RegData/etudes/BRIE/2025/775859/EPRS_BRI(2025) 775859_EN.pdf) and Ireland’s at over 20, because these facilities cluster rather than spread. The Commission’s summary of what follows is that the increase “is usually met with a lack of available capacity to connect to the grid”.

The experiment takes as long as it takes

A training run occupies specific hardware for a stretch of wall-clock time. So does an ablation, and the ablation is how anybody finds out whether the idea was any good. A better idea can shorten a run. It cannot remove the need to run it. Where the question is about the physical world rather than about software, the calendar belongs to chemistry, biology or a materials test, and those have never been persuaded to hurry.

A system that proposes ten times as many experiments in a year does not produce ten times the findings. It produces a queue, and the queue is scheduled by whoever owns the cluster.

This is the same limit that the instruments impose, arriving as physics rather than as epistemics. Confirmation is the expensive half of research. It was the expensive half before any of this, and the machinery of the last few years has been busy making the cheap half cheaper.

Efficiency does compound

Some of the physical layer does loop. Better systems make training cheaper, cheaper training buys more experiments, more experiments improve the systems, and the efficiency gains are real and compounding. Machine-assisted design has been ordinary in chip layout for years, and the chips themselves are the clearest case of an industry that improved its own means of production for decades.

What that loop cannot do is add a fab, a substation or a year. It shifts the constraint from arithmetic to infrastructure, where the feedback runs at the speed of construction and consent. The Commission’s proposed Cloud and AI Development Act, published in June 2026, sets out to triple EU data centre capacity in five to seven years, which is a fair statement of both the ambition and the units it has to be measured in.

The efficiency loop has a second consequence, which is that it walks the industry under any line drawn in compute. The same improvement that relieves the physical constraint erodes the measurement anybody was using to watch it.

Compute, power, chips and data are counted because somebody sells them. Two more inputs are consumed at least as fast and appear on no diagram.

Attention is one: the finite quantity of qualified reading available to check what the machinery produces, already the scarce input in most organisations before anything was generated. Consent is the other, held by the people who live next to the substation, the ones whose electricity bill moves, and the ones being paid to do the checking.

Both can be renewed in principle and are slow to acquire in practice, and neither responds to being asked for more.

Those two are scarce in the ordinary way, which is to say money buys more of them, slowly. What sits behind them is not scarce in that sense at all. An observation the loop did not produce cannot be produced by the loop however much it is given, and no quantity of compute, power or cleverness substitutes for one. It does not get cheaper as the machinery improves, and the machinery improving is precisely what puts more work in front of it.

The clerk’s brief

From the clerks, for the Patrician’s eyes

Compiled August 2026. Newest first; settled items sink into the assessment at the end. These entries concern the parts of the loop that have to be built rather than computed.

November 2025: The grid became the bottleneck

The Commission’s summary of data centre energy recorded 415 terawatt hours globally in 2024, 1.5 per cent of world electricity, projected to more than double by 2030, with 70 terawatt hours in the EU heading for 115, and observed that the increase is usually met with a lack of available capacity to connect to the grid. The clerks note that a connection queue is not a technical problem in the sense the industry is accustomed to solving, and that no model has yet been trained which can approve a substation.

July 2025: Parliament counted the same problem in local units

The European Parliamentary Research Service’s briefing on [AI and the energy sector](https://www.europarl.europa.eu/RegData/etudes/BRIE/2025/775859/EPRS_BRI(2025) 775859_EN.pdf) put EU data centre demand at around 3 per cent of electricity, above 20 per cent in Ireland, and recorded the tendency of AI facilities to cluster in particular places, with the resulting trade-offs against climate goals, land use and affordability. It also anticipated the Cloud and AI Development Act, which the Commission published on 3 June 2026 with the stated objective to “triple EU capacity in the next five-to-seven years and reach the needed capacity by 2035”. The clerks observe that the aggregate figure is comfortable and the local one is not, and that the objections and the permissions are both encountered locally.

What cannot be bought

Everything in this file behaves the same way: it can be bought, but not quickly, and not by being clever. The clerks’ standing assessment is that the recursive machine is bounded above by whatever it cannot manufacture for itself. That list has two halves. One is physical: fabrication capacity, electrical supply and experimental wall-clock. The other is neither physical nor purchasable: the qualified attention that checks the work, the standing to refuse it, and the patience of the people living near the transformers. The clerks add that the loop runs at the speed of the slowest thing in it. At present the slowest thing in it is pouring concrete.