In September 2024, Constellation Energy announced it would restart the undamaged reactor at Three Mile Island, the Pennsylvania plant best known for the partial meltdown of its other unit in 1979. The entire output, about 835 megawatts, was contracted to Microsoft for twenty years.
Amazon had already moved. Earlier that year it bought a data centre campus adjacent to Talen Energy’s Susquehanna nuclear station, with arrangements eventually covering up to 1.92 gigawatts of supply running to 2042. Meta subsequently signed a twenty-year agreement supporting a nuclear plant in Illinois.
Consider what these companies were actually buying. Not chips, not engineers, not model architectures. They were buying decades of electricity, at fixed terms, from generating stations that in some cases had to be brought back from retirement to supply them. Technology firms with among the highest margins in the economy chose to lock up power the way an aluminium smelter would.
That is the thesis of this piece, arriving as an event rather than an argument. The scarce resource in artificial intelligence is drifting away from the things the industry talks about and towards the things it used to take for granted.
The bottleneck moved from invention to deployment
The first phase of this boom was about invention. Research groups produced new architectures and training methods, and the public met the result through a browser window. Intelligence appeared weightless.
It was never weightless. Every response rests on a chain of physical assets: data stored, parameters held in memory, calculations performed, machines connected to one another at high speed, heat removed, and electricity arriving continuously rather than when the weather permits.
Training makes this obvious. A model is built by adjusting an enormous set of numbers in response to examples, distributed across processors that must constantly exchange results. Delay between them wastes the whole cluster, so the machines are packed together and wired tightly, producing a facility engineered around sustained intensity rather than a room with servers in it.
Inference creates a different pressure, and ultimately the larger one. A finished model has to answer whenever it is asked, which means capacity must be installed before the requests arrive and must sit ready during the hours when they do not. The economics resemble an electricity grid more than a software product: the operator provides for the peak and lives with the trough.
The distinction matters because training attracts attention while inference determines whether any of this becomes an ordinary part of the economy. A model earns nothing when it is born. It earns when it answers, and answering is a physical act with a bill attached.
As adoption spreads, the problem changes from acquiring a small number of exceptional machines to operating a large and dependable fleet of ordinary ones, for years, with contracts, maintenance, spare capacity and a power supply that does not fail. The industry starts to resemble a hybrid of telecommunications, energy and industrial property.
Why electricity became the constraint
Power was always present in computing and was usually treated as an operating expense. For facilities of this kind it can determine whether a project happens at all.
A large data centre does not simply need electricity in the annual sense. It needs a connection to the transmission network with sufficient capacity, a substation able to transform and distribute that power, and the ability to absorb sudden changes in load without destabilising anything around it. Once customers depend on a service being available continuously, backup generation and storage stop being optional.
The quantities involved explain why this stopped being routine. The reactor Microsoft contracted produces around 835 megawatts, which is the scale of a substantial conventional power station and enough to supply a mid-sized city. The arrangement Amazon pursued runs to nearly two gigawatts. These are not procurement line items. They are the sort of numbers that appear in national energy planning documents, committed to single corporate customers, for periods longer than most of the technology inside the buildings will survive.
Location therefore becomes an electrical question before it is a property question. Cheap land is useless if the local network cannot support a new connection of the required size. Abundant renewable generation nearby does not help if it is intermittent and nothing balances it. Proximity to customers competes with land cost, permitting, and the availability of transmission. Every site is a compromise between power, connectivity, cooling, regulation, construction time and distance from demand.
This is why the race cannot be understood as a race to buy processors. Processors in a warehouse earn nothing. They have to sit in a functioning building, connected to a network, supplied with power, operated by people who can keep them running. The genuinely scarce asset is frequently the right to connect a site to the grid rather than the site itself.
Cooling imposes its own limit. Chips convert electricity into computation and heat in roughly equal enthusiasm, and air stops being adequate as equipment grows denser, which pushes operators towards liquid systems that consume their own power and water and constrain which buildings can be used at all. The computer is not an isolated object. It is one component of a thermal system that has to be designed around it.
What makes this a moat
Scarcity in this industry is usually discussed in terms of semiconductors. The more durable scarcity sits one layer beneath them.
A chip can be redesigned, second-sourced, or made less necessary by better software. A high-voltage connection cannot be improvised. It requires planning permission, engineering, equipment with long lead times, cooperation from a network operator, and frequently the consent of people who live nearby. Those frictions are not romantic, and they are exactly what gives existing assets protection that a technical lead does not have.
The defining advantage in this industry is the ability to convert electricity into reliable computation, in the right place, at sufficient scale, for long enough to earn back the capital.
This does not mean utilities quietly become the winners. It means that whoever controls the tightest constraint in a given region gains negotiating power, and the tightest constraint is increasingly not the chip. A developer with an excellent model and no facilities must rent capacity from someone who may ration it, raise the price, or serve a larger customer first. Building instead means acquiring land, permits, grid access, hardware, engineering expertise and financing, on a schedule measured in years.
The provider faces the mirror image: customers ready to pay, and insufficient power or equipment to serve them. Its position was largely determined by procurement decisions made years earlier. In an industry that prizes rapid iteration, infrastructure introduces a long memory, and a company that secured capacity early can hold an advantage over a competitor whose model is just as good.
Capital is part of the mechanism rather than a footnote to it. These facilities require enormous investment against demand that is difficult to forecast, which is precisely the profile that has drawn private lenders into the sector, in the same migration of credit out of banks that has reshaped corporate lending more broadly. Tangible assets and contractual cash flows make the loans legible. They do not make them safe. A building optimised for one generation of hardware is difficult to repurpose, and a long power contract signed at the wrong moment becomes an expensive obligation rather than a strategic asset.
The institutions in the way
There is a tendency to assume that a sufficiently well-capitalised company can buy its way through a physical constraint. The record suggests otherwise.
When Amazon and Talen arranged for the data centre campus to draw power directly from the Susquehanna plant, federal regulators rejected the interconnection arrangement. The objection was not about artificial intelligence. It concerned what happens to everyone else on the network when a very large customer is connected in an unusual way, and who pays for the consequences.
That is the shape of the problem. Electricity networks are regulated. Communities absorb the effects of construction, land use, water consumption and noise. New generation and transmission take years to approve and build. An AI company can pay for power and still be unable to determine when a substation gets built or how a transmission corridor is permitted, which makes its expansion dependent on institutions that answer to different constituencies entirely.
The way these disputes resolve will shape the industry’s structure more than any product decision. If connection rules favour incumbents, incumbents gain time. If governments accelerate construction, new entrants get access. If local conditions tighten, facilities migrate to jurisdictions with different politics. None of that is determined by the quality of anyone’s model.
The strongest objection
The argument above assumes demand grows faster than technology reduces the cost of serving it. That assumption may be wrong, and the case against is worth stating at full strength.
Models can become smaller and more specialised. Work that once required a large general-purpose system can increasingly be done by a compact model on modest hardware. Chips perform more operations per unit of energy with each generation. Software eliminates unnecessary computation. Workloads can be scheduled into hours and regions where the grid has room. If those gains compound, the current scarcity is a transitional condition, and facilities built around today’s assumptions will look overbuilt rather than strategic.
Technology has dismantled apparently permanent bottlenecks before. Storage got cheap, networks got fast, and computing moved out of specialised rooms into ordinary objects. Treating present energy intensity as a fixed characteristic of machine intelligence would be a straightforward error.
There is a demand-side version of the same objection. Not every application will justify its infrastructure cost. Pilots stall. Enterprise deployments underdeliver. If productivity gains arrive slowly, customers will resist the prices required to support the capacity being built for them, and the industry will have constructed ahead of durable usage.
What weakens the objection is that efficiency has rarely reduced total consumption of a useful thing. It changes the economics of use. When the price of an answer falls, applications that were previously uneconomic become viable, and volume expands to fill the space that efficiency created. A smaller model can support a much larger number of interactions, which is the ordinary rebound that follows almost every improvement in industrial efficiency.
The second weakness is that AI is not one workload. Training frontier models, adapting them for narrow tasks, generating video, processing live information and serving millions of small text requests have different constraints, and an improvement that relieves one may do nothing for another. Efficiency in text generation does not solve the requirements of continuous video processing.
The honest conclusion is not that power demand must rise indefinitely. It is that efficiency becomes a competitive weapon inside a system whose overall scale still matters. The best operators will use less power per task and may still need more power in total, because they serve more tasks.
Optionality is the asset
If the constraint is physical, the strategic question changes from how large a model is to how cheaply and reliably it can be turned into a service. That reframes what is worth owning.
The strongest position is not the largest facility or the biggest power contract. It is the ability to shift workloads between sites, upgrade equipment without rebuilding the room, serve different kinds of customer, and stay viable across a range of energy prices. A building that can host more than one generation of hardware outlasts one designed to a narrow specification. A connection that can expand in stages is worth more than one that demanded a decade’s commitment before demand was proven.
This is also why the boundary between technology and infrastructure is dissolving as an investment category. The choice is not between software companies and utilities. It is between combinations of intellectual property, physical capacity, contracts and regulatory access, and the entity that ends up capturing the value may be whichever partnership, acquisition or financing structure joins a good model to a good site.
The comparison to earlier industrial build-outs holds with one difference that matters. Railways and power stations were built for operating lives measured in decades. AI hardware becomes obsolete far faster, which means the durable asset is not the equipment but the envelope around it: the land, the connection, the cooling, the permits and the ability to replace what is inside without starting again.
The same logic reorders what national ambition in this field actually requires. Governments pursuing a position in artificial intelligence tend to concentrate on research funding, talent and semiconductor supply, and those remain necessary. They are not sufficient. A country can produce excellent researchers, secure allocation of advanced chips, and still find that its companies cannot deploy at scale because generation is tight, transmission is congested, or connection queues run for years. Industrial policy for this technology looks less like a science budget and more like an electricity plan, which is an awkward realisation for administrations that have spent a decade treating energy infrastructure as a decarbonisation problem rather than a growth one.
It also creates an unfamiliar competition between uses. Electricity committed to computation is electricity not available for heating, transport or manufacturing, at a moment when several countries are trying to electrify all three. The political question of who gets the next gigawatt has no technical answer, and the industry has not yet had to argue its case in those terms.
What to watch
The thesis is confirmed if power availability and grid access increasingly decide which companies can expand, irrespective of model quality, and if long-term energy arrangements keep moving from procurement detail to strategic asset. Watch whether more regulatory decisions of the kind that blocked the Susquehanna arrangement appear, since those mark the point where the constraint becomes institutional rather than merely physical.
It is weakened if efficiency gains materially reduce total infrastructure demand, if adoption stalls at the pilot stage, or if new generation and transmission arrive faster than expected. The decisive evidence will not be another impressive demonstration. It will be whether the industry can convert demonstrations into a service that stays available at a price customers keep paying.
