Nobody is refusing to build generation. Almost nobody can connect it.
The power problem is not a shortage of generation. It is a shortage of connection, and every solution in this case is priced against that fact.
Generation gets built. What does not happen is the connecting: 77 per cent of everything that entered a United States interconnection queue between 2000 and 2019 was withdrawn, and the median project that did connect waited nearly five years. The five routes that follow are each a way of buying a megawatt without joining that line, and each is sold on a different number.
Named sites only, drawn from the route files below. The clustering is the finding: geothermal follows the rock, gas follows the absence of a queue, and small modular reactors follow the regulator. Switch the view for the United States and Europe close up, filter by route, and select any node to open its file.
Every capacity figure in this case, and in the chapters around it, should be read against these. Seven are measured. One is contested, and it is marked as contested rather than dropped.
Each is a different answer to the same question, and each is sold on a different number. Take any one.
Firm power at a capacity factor no intermittent source matches, sited where the rock allows it. The drilling curve is real and transferring from shale. The binding constraint is grid size, not heat.
Contributor seat open on this route
The cheapest unit on the board and still falling, 78 per cent off since 2015. It defers the queue rather than avoiding it, and the merchant arbitrage that pays for it compressed 48 per cent in a year.
Contributor seat open on this route
The fastest route to first power and the only one with a mature supply chain. It is also the only one getting harder: the backlog now exceeds a decade of production, and near population the permitting risk has reached a federal court.
Contributor seat open on this route
The largest gap on the board between what has been ordered and what exists. Eleven gigawatts of backlog against a single 320 MW phase in permitting, priced against a forty-year record of overruns.
Contributor seat open on this route
Cheapest per megawatt-hour and mature everywhere. It solves the price of power and does nothing for its arrival, which is the reason the other four exist.
The five routes are answers to the question of where the next megawatt comes from. This is the question underneath it: how many of them are actually needed. The installed base runs at a power usage effectiveness of about 1.52 while the best sites run at 1.02, and closing even part of that gap across a fleet drawing tens of gigawatts frees load for compute rather than overhead. It is the cheapest capacity in this case and it does not enter the interconnection queue.
Two things are true at once and the argument usually picks one. Per unit of work, this industry is getting radically more efficient: ten times the tokens per watt in a single hardware generation, a thousandfold collapse in the price of a fixed capability. In aggregate, it is drawing more power every year, because each efficiency gain lowers the price of compute and a lower price buys more of it. Between 2010 and 2018 efficiency absorbed the growth entirely. Since 2024 it has not come close.
Which is why this section sits beside the readiness index rather than on it. Efficiency is not a route to power and it should not be scored as one. It is the only lever on this board that is already installed, already connected and already paid for, and the reason it stays unpulled is that nobody is paid to pull it. The routes have vendors. The megawatt you do not need has nobody selling it.
The claim that the AI buildout is unprecedented is doing a lot of work in this industry, and it is not quite true. Grids have been built this fast before, more than once, on a larger relative scale. What is unprecedented is the conditions under which it is being attempted.
Score the five against the three conditions that made them possible and the pattern is not subtle. Every fast buildout had at least two. The two fastest had all three.
The AI buildout has none of them. There is no single buyer: there are five hyperscalers bidding against each other and against every other large load on the same grid. There is no standardised design: each campus is bespoke, and the rack inside it changes generation every eighteen months. There is no state direction: in the United States the state is currently moving in the opposite direction, with the first statewide permitting moratorium signed in July 2026 and at least fifteen more states weighing one.
Which leaves one honest historical analogue, and it is the uncomfortable one. US gas from 1998 to 2005 is the only case on this board built at speed, by a market, with none of the three conditions, and it ended with $150bn to $200bn of stranded value and plants handed to lenders. That is not a prediction. It is the base rate for building this fast without a buyer, a standard or a plan, and it is the only base rate available.
Everything in this file is graded Contested or Estimated, and the reason is uniform: none of these technologies has reached commercial operation, so every number is a company or investor disclosure rather than an audited outcome. They are here because three of the four already have signed offtake contracts against them, which means capital is being allocated on these numbers whether or not they hold.
Sort them by how far the demonstration is from the claim and the order is clear. Long-duration storage is not really frontier any more: it is in production in West Virginia, deployed for Georgia Power, and priced within a factor of two of its target. Superhot rock has a physics claim, a field site and a fiftyfold gap between what has been drilled and what the resource needs. Fusion has two contracts, neither of which pays out until a plant that does not exist is connected. Space-based solar has a genuine first, power beamed from orbit and measured on a rooftop, attached to an efficiency chain that delivers about an eighth of what it collects.
The discipline this section is trying to enforce is simple. A signed offtake is not a megawatt. Google has contracted 200 MW of fusion and Microsoft 50, and the combined delivered total to any grid, from any fusion device, anywhere, remains zero. That is not an argument against the technology. It is an argument against putting any of it in a 2030 capacity plan, which is the decision this case exists to inform.
A cost curve says where a technology is. It does not say when it arrives. These are the three crossings that matter to a buyer, projected at three rates, each anchored to a buildout that actually happened rather than an assumed pace. No route improves on all three axes at once, and one gets harder on two of them.
Read the demand chart carefully. It is drawn on the medium case for every route simultaneously, which is already an optimistic reading, and the combined contribution still runs behind demand until the early 2030s. That gap is filled today by the existing grid, which is what the queue figures at the top of this case describe. The routes here do not remove the grid from the equation; they change how much of the gap has to pass through it.
Every route scored 0–10 on the five things a buyer weighs when choosing between them. Pick a route to read its shape against the field; switch the dimension to re-rank the bars; filter by type. No route fills the shape, and the two that come closest do so from opposite ends of the risk.
Darker is weaker, brighter is stronger. Click any column to sort by it, click a route to open its read, and switch the lens to re-weight the overall. A buyer optimising for speed and a buyer optimising for certainty do not choose the same route.
The most useful finding in this case is not on the readiness board. It is that a problem of exactly this shape has been solved five times in the last hundred years, and every time it looked impossible until it was routine.
France took nuclear from 8 per cent of its electricity to nearly 80 in sixteen years, on 58 reactors built to one design. Rural America went from one farm in ten with power to nine in ten inside two decades, and the binding constraint was never the technology, which had existed since the 1880s. It was capital and organisation. China moved 340 gigawatts from where it is generated to where it is used. Solar fell from $6.36 a watt to under 40 cents on a cost curve that held for twenty years. None of those were easier problems than this one. They were differently organised ones, and organisation is a choice rather than a piece of luck.
The routes on this board are further along than the debate usually allows. Two of them have already crossed on cost, today. Batteries are 78 per cent cheaper than they were in 2015 and are the most siteable thing here by a distance. Geothermal's drilling curve is transferring out of shale and the evidence is measured rather than promised: the same operator cut 70 days to total depth down to 21, then held those 21 days while drilling five thousand feet deeper and sixty degrees hotter. That is a productivity gain, not a one-off. The frontier is closer than its reputation too, and the clearest case is the least glamorous: long-duration iron-air storage is in production at a former steel mill in West Virginia and priced within a factor of two of its target.
And the number all of this is measured against may not hold either. The installed fleet runs at a power usage effectiveness near 1.52 while the best sites run at 1.02, a liquid retrofit costs about $2m per MW against $11m for the same capacity built new, and silicon is delivering something close to ten times the tokens per watt in a single hardware generation. Closing even part of that gap frees real capacity on connections that already exist, which is the cheapest and fastest megawatt available anywhere in this case. The demand line is a projection, and projections of this kind have been wrong in both directions before.
Nothing here suggests one route takes the load. The honest reading of the crossover chart is that the grid that already exists does the heavy lifting into the early 2030s, while gas and batteries carry the near term, renewables and geothermal come up behind them, and the reactors and the frontier arrive late enough to matter for the decade after. Old sources and new ones, running together, is not a failure of the new ones. It is what every buildout in the precedent file above actually looked like from the inside.
The one caution the record insists on is that the fast buildouts had at least two of a single buyer, a standardised design and state direction, and this one currently has none of the three. That is worth holding, because it is the difference between the French programme and the American gas bust. But it describes a set of conditions, and conditions can be assembled. A standard can be adopted. A queue can be reformed, as Great Britain has already shown by replacing first-come with first-ready. Coordination is the missing input, not capability.
So the question this case leaves is not whether the power arrives. On the evidence above, most of it probably does. It is whether it arrives in the order and at the cost the people paying for it were told to expect, and that is decided less by the technology than by who is willing to organise it.
The Cost of a Megawatt is a Sociogencia case inside The Next Hotspot, an interactive read on where the world actually builds AI infrastructure. The full edition is live, and the remaining drops are dated below.
The Next HotspotWhere the build actually happens, and what is in the way.
