The Energized Chip

In 1892, Wilhelm Ostwald proposed that everything that happens is fundamentally a change in energy. He called the theory Energetik and defended it for nearly two decades, holding that energy was more real than matter itself.¹ In Energetik, chemistry, motion, heat, life, and perhaps thought were all transformations of a single underlying quantity. Despite its early traction, the theory lost to atomism. Atoms existed, and energy could not replace particles, fields, and mechanisms as an explanation of the world.

The theory failed as ontology, but its method survived. Complex systems can often be understood through the energy available to them, the forms into which that energy is converted, the constraints on that conversion, and the waste produced along the way. There is no need to follow every molecule inside a turbine to understand how much useful work it can produce, and the same is true of a battery, a chemical plant, or an electrical grid. Energy doesn’t explain the internal structure that makes a system work. It defines what the system can do within a given budget.

More than a century later, the leading AI companies have begun to describe their ambitions in much the same terms, measuring what they can build by the power they can secure. Elon Musk has argued that the future currency will essentially be wattage.² Sam Altman asks what humanity might do with ten gigawatts of compute.³ Anthropic contracts the capacity to train and serve Claude in blocks measured not in GPUs but in gigawatts.⁴ NVIDIA is the most explicit: it describes the AI factory as a system that converts electricity into revenue-generating intelligence and measures its productivity in tokens per second per megawatt.⁵ As intelligence becomes something that can be produced repeatedly by machines, the industry now describes the capacity to produce it in the same language used for steel mills, chemical plants, and power stations.

The companies are right, and the implication runs deeper than the metaphor suggests. If intelligence is now produced at industrial scale, with the physical constraints that implies, then the AI company is becoming an industrial firm whose principal input is power. This essay is an attempt to follow the energy from the grid, through the facility, down to the chip, and back up to the economics of the firms being reorganized around it.

A power problem

Colossus 1, built by xAI and now contracted in full by Anthropic, has more than 220,000 NVIDIA GPUs and represents more than 300 megawatts of power capacity.⁶ Meta is building Prometheus as a one-gigawatt cluster and has described Hyperion as eventually reaching five gigawatts.⁷ OpenAI's Stargate program is organized around ten gigawatts of infrastructure.⁸ For scale, the Paducah uranium-enrichment plant in Kentucky, among the largest single electrical loads ever built in the United States, drew roughly three gigawatts at its Cold War peak. Some AI campuses now in development would exceed it.⁹

What the industry usually calls an energy problem is more precisely a power problem. A campus needs hundreds of megawatts of usable power, at a specific location and time, with the reliability and flexibility the workload needs. Electricity can be abundant across a grid and still scarce at the one substation where a campus needs it. There are two ways to ease that scarcity: new capacity, which expands the allocation, and coordination, which stretches how much computation an existing allocation can support.

The IEA estimates that the world's data centers collectively consumed roughly 415 terawatt-hours in 2024.¹⁰ Run at full power all year, Stargate's ten gigawatts alone would use more than a fifth of what the world's data centers consumed in 2024. This is why energy is becoming the binding constraint on the expansion of machine intelligence. Capital is available. Semiconductor production continues to grow. Algorithms diffuse rapidly. Firm power doesn’t scale in those ways.

Hundreds of megawatts cannot be copied, shipped overnight, or provisioned through an API. They must be generated, carried across a grid, and continuously balanced against every other load on the system. A delay in a transmission upgrade, transformer delivery, or interconnection study can leave billions of dollars of chips idle. Building an off-grid campus to bypass the interconnection queue doesn’t remove the problem. It transfers the problem to whoever operates the microgrid, with its own generators, storage, transformers, and controllers.

At the frontier, the scarce object is no longer the chip in isolation. It is the energized chip.

Inside the data center

Follow the megawatt through the fence line and it begins to divide. Some of it never reaches a chip: it runs the cooling, pumps, fans, power conversion, and lighting. The industry tracks this with power usage effectiveness (PUE), the ratio of total facility power to the power that reaches the computers. Lawrence Berkeley National Laboratory models about 1.14 for the newest AI-specialized facilities.¹¹ The industry average is closer to 1.5, while the best operators run near 1.08.¹² With overhead already below 10 percent at those sites, the building has little left to give. The larger gains now have to come from the machines inside it.

The thermal burden becomes more visible at rack scale. Hewlett Packard Enterprise rates its version of NVIDIA's GB200 NVL72 rack at 132 kilowatts, nearly all of which leaves as heat, most of it carried away by liquid.¹³ At these densities, thermal architecture is no longer a building service added after the computer is designed, but part of the computer itself.

Inside the chip

The same pattern repeats one level down. In frontier AI workloads, much of the energy inside a GPU goes to moving data rather than computing on it.

Bill Dally, NVIDIA's chief scientist, has made this argument for over a decade.¹⁴ A 32-bit floating-point multiplication on a seven-nanometer chip consumes roughly one picojoule. A single 64-bit access to HBM, the memory stacked alongside AI GPUs, consumes 250 to 450 picojoules in its input-output circuitry alone.¹⁵ Fetching an operand can therefore cost hundreds of times more energy than the arithmetic performed on it. Jeff Dean at Google has described the same physics from the TPU side: moving data from an accelerator's HBM to the processor can consume roughly a thousand times more energy than a single multiplication.¹⁶

NVIDIA's tokens-per-megawatt metric is therefore not primarily a measure of raw arithmetic. It also measures how efficiently the machine moves data. HBM stacks, interconnect fabric, memory hierarchy, and the placement of weights and activations determine how much useful work can be extracted from the megawatt.

This is where the engineering effort is concentrated, and it is working. Each hardware generation extracts meaningfully more tokens from the same electricity. The natural conclusion is that efficiency will eventually relieve the constraint. But I don't think it will.

A company that would not spend ten dollars to analyze every customer interaction may do so for ten cents. A scientist who could afford to test one hypothesis may test a million. An agent workflow that was uneconomical at last year's cost per token becomes routine this year. Each unit of intelligence becomes cheaper while the total quantity demanded grows faster. So far, efficiency has compounded demand rather than reduced it. Every picojoule saved inside the chip doesn’t relax the constraint at the fence line; it raises the value of the next megawatt brought through it.

The shadow price of a megawatt

If usable power limits how much a lab can produce, securing it stops being a procurement task and becomes a question of strategy. Other industries have been here before.

Airlines don’t organize themselves around jet fuel, even though it is one of their largest costs. They organize themselves around airport slots. Telecom companies organize around spectrum, semiconductor manufacturers around wafer capacity at the leading node, electricity markets around the transmission constraints reflected in locational marginal pricing. In each case, what shapes the firm isn’t its costliest input but the one it cannot buy in time, at any price. The scarcity of that input sets the value of everything else.

Economists have a name for this: the shadow price of a constraint, the amount by which output or profit would increase if one additional unit of the constrained resource were available. The constraints with the largest shadow prices identify a system's bottlenecks, because easing them produces the greatest improvement in performance. And shadow prices rank future bottlenecks, not just current ones.

When a single resource carries the highest shadow price for long enough, the firm begins reorganizing itself around securing it. The constraint stops being an operational problem and becomes the organizing principle of the business. Frontier AI is undergoing this transition. The industry is beginning to speak in megawatts rather than GPUs because power capacity has become the input whose shadow price now dominates the others.

Chips, capital, algorithms, and talent each have real frictions, but they can generally move across firms and regions faster than generation, transmission, and interconnection capacity can be built. That is what makes energy a binding constraint: not simply that it is expensive, but that it is inelastic on the timescale the industry needs. It is also why the fastest way to add usable power is often not to build it but to unlock it. Coordination across generation, storage, and compute, increasingly run by AI itself, can free capacity from existing connections in months, while new generation takes years.

This changes what a competitive advantage looks like. In a chip-constrained world, the company with the best access to advanced fabs captures the value. In an energy-constrained world, the company that can add gigawatts of firm power faster and at a lower marginal cost than its competitors captures the value.

AI labs as factories

Energy infrastructure is becoming part of the AI stack. Beneath the layers that used to define the industry (chips, networking, cloud systems, models, and applications) now sits another stack: generation, transmission, substations, transformers, cooling, batteries, and the software that coordinates them. Today, AI labs sign long-term power contracts, finance new generation, support nuclear restarts, buy gas turbines, reserve grid capacity years in advance, and design campuses around electricity that does not yet exist.

Meta has announced roughly six gigawatts across Prometheus and Hyperion, alongside nuclear agreements for as much as 6.6 gigawatts of new and existing capacity through 2035.¹⁷ AI labs are now helping decide which plants get financed, which reactors remain open, and where new generation is built.

Hyperscalers have signed long-term power contracts for more than a decade. What is new is the scale, the speed, and the willingness to take on construction and generation risk directly. An AI company does not have to become a regulated utility for the change to matter. It becomes an AI factory when energy development enters its production function: when it must choose generation technologies, finance projects, assume construction risk, secure interconnection rights, and coordinate the arrival of power with the arrival of chips.

The physical form of that factory is now visible on the ground, and Google's Goodnight campus in the Texas Panhandle shows it most completely. A buildout estimated at $29 billion pairs 531 megawatts of wind with a proposed 933-megawatt gas plant and roughly a gigawatt of batteries. Generation, storage, and computation are being planned as a single system. The pattern repeats elsewhere: at Colossus 2 in Memphis, xAI has installed roughly three gigawatt-hours of batteries, among the largest in the country, to absorb the power swings of training runs, and at GW Ranch in West Texas, Pacifico Energy, building for Amazon, holds what it calls the largest power-plant permit in the country, for 7.65 gigawatts of gas turbines.¹⁸ None of these is a data center with a power contract. Each is a power plant with a data center attached.

None was built by its tenant alone. Crusoe is building Goodnight for Google, Pacifico is building GW Ranch for Amazon, and Tesla supplied the batteries at Colossus 2. Taking on the energy problem has not meant building every part of the answer. Like any industrial firm, the AI factory is assembled from specialist suppliers.

Among the neoclouds I’ve been working with, the building happens at a speed with little precedent in industrial construction. Partners in power, networking, cooling, compute, and energy software are brought in from the first drawing, and the campus is designed as one machine around the electricity that can actually be delivered.

Models will improve. Chips will change. Architectures will be replaced. A permitted site with transmission access, firm generation, cooling capacity, and a path to several gigawatts can endure through all of them, which makes it the most durable asset an AI company can hold. How much useful work that site produces then depends on coordination, the argument I made in The Clay Jar and the Data Center.¹⁹

The measure

AI companies began as laboratories. The largest are becoming industrial systems that convert energy into intelligence, and their lasting advantage may come less from any single model or chip than from how quickly they can bring power online and how much useful cognition they can draw from each megawatt.

If this argument is right, it should show up in places that would look strange today. By 2027, I expect at least one independent frontier lab to register as a participant in a wholesale electricity market.²⁰ Before 2028, I expect a frontier lab to say publicly that a model or product arrived late because the power did, not the chips. The labs least likely to say it will be the ones that treated the coordination of power and compute as part of the plan from the start. By 2030, I expect at least one frontier lab to spend more each year on power, sites, and facilities than on the chips inside them.²¹ Two things would prove me wrong: efficiency outrunning demand, or firm power becoming abundant sooner than expected. Neither looks likely on the timescale over which today's capital is being committed.

Which brings the argument back to Ostwald. He was wrong that energy is what the world is made of. But for systems that produce something, he was right about what energy is: the budget that defines how much a system can do, whatever the mechanism inside. For a century that method belonged to engines, plants, and grids. It now belongs to intelligence too. However capable the models become, each one runs on an energized chip, and there will only ever be as many of those as there are megawatts to feed them.