The Grid Jam

The previous chapter proposed spending national money on a Socratic system that leans on machine inference to give every learner something close to a tutor. That proposal has a physical bill attached, and this book's standing rule is that every claim pays in joules. So: where does the loop jam?

The honest answer is that it does not jam where most people look. It does not jam at generation — the sun and the fuel are not the scarce terms. It jams at the boring layer between potential energy and delivered energy: queues, wire, transformers, water rights, copper, and the number of human beings qualified to terminate a medium-voltage cable. These are the planetary rate-limiters, and each one runs on a clock that capital cannot compress.

The distinction that organizes the chapter

There are two very different compute bills inside the phrase "AI in education," and conflating them is the most common error in this debate.

Training is a capital expenditure in energy: large, concentrated, episodic, and sited wherever power is cheapest. It is the activity that shows up in grid-planning documents and that has broken a decade-long pattern in which efficiency gains held total data-centre consumption roughly flat.[2][1]

Inference — the tutoring loop itself, a model holding a belief about a learner's latent state and choosing the next question[10] — is an operating expenditure: small per event, enormous in aggregate only if the number of events is allowed to grow without a cap. A Socratic session is not a video stream and not an agentic research swarm. It is a few hundred short exchanges a day, per learner, most of them text.

The distinction matters because the two bills have opposite policy remedies. Training load is a siting and interconnection problem. Inference load is a discipline problem — the Jevons problem[7] — and the only reliable remedy is a stated cap, which is why Chapter 7 put one on the $150 technology line and made the whole fiscal case contingent on holding it.

Six rate-limiters

Each of the following is a place where the loop physically jams. The column that matters is not the size of the constraint but its clock — how long it takes to relieve, assuming full funding and political will from day one.

Read down the clock column and the shape of the century appears. Nothing on that list moves in under two years. Two items move on a decadal clock, and one — copper — moves on a geological one.[6] Any plan that assumes compute capacity can be added at the speed software is written is a plan that has not read the queue.[3][4]

The water is spent somewhere else

The most misreported line on the ledger is water. Public argument fixates on the cooling tower, which is visible, local, and photogenic. But the larger share of a data centre's water consumption is upstream, in the thermoelectric generation that feeds it.[5] A facility that installs closed-loop cooling and then buys power from a thermal fleet has moved its water bill, not paid it.

This is the same structure Chapter 4 traced through California: the watershed that carries the cost is rarely the jurisdiction that approved the load. It is also where Wittfogel's narrow structural point returns.[8] Infrastructure that must be centrally coordinated concentrates the authority that coordinates it. A national education system whose instructional layer depends on four hyperscale operators has not decentralized anything; it has relocated the inspector from the district office to a data hall it cannot audit.

The education bill, priced against the grid

So what does the Chapter 7 proposal actually draw? Take the model plainly.[11] Fifty million learners. Assume each one generates on the order of a few hundred short inference calls per school day, priced at the low end of the query matrix in Chapter 2 — text-scale, not agentic-scale. That is a national instructional inference load in the neighbourhood of a few hundred megawatts of average draw: real, schedulable, unremarkable next to industrial demand, and roughly the size of a mid-sized city.

Two things follow, and they pull in opposite directions.

The first is reassurance. A few hundred megawatts is not what jams the grid. It is a rounding error against the training and agentic loads currently queued. The Socratic system is not, on these numbers, an energy problem. The twenty-watt engine it serves[9] remains by an enormous margin the most efficient component in the entire arrangement, and the point of the architecture is to keep it that way.

The second is the warning. The load is only small while the interaction stays small. Move from short text exchanges to persistent multimodal agents observing each learner continuously, and the same fifty million users become a load measured in gigawatts — a load that would have to enter the interconnection queue behind everything already in it, on the five-to-fifteen year clock, and that would arrive with a watershed attached. The difference between those two futures is not technological. It is a design decision about how much machine a learner needs, made once, early, and enforced.

The one limiter education relieves

There is a reason this chapter sits after the schooling chapters rather than in a separate volume on infrastructure. Five of the six rate-limiters are indifferent to pedagogy. The sixth is not.

The skilled-trades bottleneck — linemen, substation technicians, industrial electricians, cooling and controls people — is the constraint that money moves slowest and that a functioning educational system moves fastest. It is also precisely the kind of competence the factory model is worst at producing and the Socratic model is best at: judgment under partial information, in a physical system that punishes error, learned by doing the thing under a guide who has done it.

So the loop closes, and it closes in the direction the book has been arguing all along. Education is not merely a consumer of the grid. It is one of the few inputs that can relieve a constraint on the grid. A system that produces people who can think under uncertainty and terminate a cable is a system that pays part of its own energy bill in kind.

Four refusals

One. No claim that efficiency alone will contain demand. Every efficiency gain in this sector has so far been reinvested in more use.[7] Caps, not improvements, are what hold a line.

Two. No treatment of water as an on-site number. Any figure that omits upstream generation water is not a measurement, it is marketing.[5]

Three. No assumption that new generation solves a transmission problem. Adding megawatts behind a saturated interconnection is adding a number to a spreadsheet, not power to a building.[4]

Four. No plan that requires an instructional dependency on infrastructure a public cannot audit. If the Socratic layer cannot run, degraded but intact, on locally hosted inference during an outage or a contract dispute, then it is dependent offloading in the Chapter 3 sense, and it fails this book's test regardless of what it costs.

The falsifiers

One. If a metered district-scale pilot finds per-learner inference energy more than an order of magnitude above the text-scale figure assumed here, the "rounding error" claim collapses and the Socratic system becomes a genuine grid actor with all the siting obligations that implies.

Two. If interconnection reform in a major market drops median queue time below two years, the five-to-fifteen year clock in the table is wrong and the chapter's whole pessimism about wire needs rewriting.

Three. If copper substitution and recovery close the projected supply gap without new mine development, then the two-decade geological clock is not binding and materials drop off the limiter list.

Four. If a Socratic cohort produces no measurable increase in entry into skilled trades relative to a matched conventional cohort, then the closing argument of this chapter — education as a relief valve on the grid — is sentiment rather than accounting, and should be struck.

The bottleneck is never the idea. It is the transformer, the watershed, and the apprentice. Plan on their clocks, not on yours.

Chapter 9 leaves the grid for the gradient that produced it, and asks what thermodynamics permits of any organised matter — including the kind that builds schools.