Letter 165 opened this series by arguing that Trust Capital — the market's confidence that an institution will do what it says — is priced into every asset whether or not anyone names it. Energy Capital is the second entry, and the more physically constrained: for the first time in the history of computing, the binding limit on how much intelligence the world can produce is not silicon. It is electricity, and specifically, how fast it can be delivered to a specific plot of land.
Every AI capex forecast published in the last two years assumed the constraint was chips. Get enough GPUs, the thinking went, and the rest follows. That assumption quietly broke sometime in 2025, and by 2026 it is simply no longer true: the harder constraint is delivery — sustained, high-wattage, grid-connected power at the specific location a data center needs it, not generation capacity somewhere in the abstract.
Worldwide data center electricity demand is projected to rise 27% in 2026 alone, from 104 gigawatts to 132 gigawatts, en route to roughly 290GW by 2030 — a scale-up that would have sounded implausible three years ago. Global data center electricity consumption sat at approximately 415 terawatt-hours in 2024, a figure already growing at 12% a year, more than four times the growth rate of total global electricity demand. By some estimates, 2026 consumption alone could approach 1,050 terawatt-hours — enough that if data centers were counted as a country, they would be the fifth-largest electricity consumer on Earth, sitting between Japan and Russia.
In the United States specifically, data center electricity demand is projected to hit 75.8 gigawatts in 2026, up from roughly 53GW in 2023 — and that expansion alone accounts for nearly half of all projected US power demand growth through 2030. The infrastructure build required to support this is being priced at up to $3 trillion by 2030, with grid and utility investment alone estimated at $720 billion through the same period.
The clearest signal of how binding this constraint has become is who is now paying for grid upgrades directly. In March 2026, seven of the largest AI infrastructure buyers — Amazon, Google, Meta, Microsoft, OpenAI, Oracle, and xAI — signed a White House-facilitated Ratepayer Protection Pledge, committing to directly fund the grid infrastructure improvements needed to support their own buildouts, rather than pass the cost to ordinary electricity customers. Texas, hosting a growing share of new capacity, passed Senate Bill 6, requiring any new energy load above 75 megawatts to participate in demand-response programs, with provisions for emergency disconnection during grid stress — a regulatory signal that the state is no longer treating hyperscale data center demand as a background variable.
The most direct evidence that power, not compute, is now the scarce input is that the largest AI buyers are becoming energy companies themselves — moving past power purchase agreements and into direct ownership of generation assets. Microsoft's $15.2 billion commitment to build data centers in the UAE was structured explicitly around the region's ability to guarantee renewable power capacity, not around chip availability. Capital is now following megawatts, not the other way around.
For three decades, capital chased wherever compute was cheapest. In 2026, for the first time, compute is following wherever power is guaranteed. That reversal is Energy Capital, and it now prices into every AI-adjacent asset whether the number appears on a term sheet or not.
The honest read on Energy Capital is not simply "own the power, win the trade." Grid-connected capacity is a long-duration, fixed asset, and the AI demand curve it is being built against is genuinely uncertain — a training-to-inference transition already underway could redistribute load from centralized clusters to smaller, distributed regional hubs, changing which specific sites hold the value. A hyperscaler or utility that overbuilds fixed generation against a demand curve that later shifts geographically carries real stranded-asset risk, not unlike the fiber overbuild of the early 2000s. Regulatory backlash of the kind Texas is already signaling with SB6 is a second live risk: political tolerance for data centers consuming disproportionate local grid capacity is not unlimited, and the entities with the least Energy Capital — ordinary ratepayers — are the ones absorbing the externality if the pledges of March 2026 don't hold in practice.
What is durable is the structural point: whoever secures firm, contracted, deliverable power at a specific site now holds a genuine moat that chip access alone can no longer provide, because chip access without power to run the chips is worth nothing.
Read against Letter 165's opening thesis on Trust Capital, Energy Capital is the physical-world twin: trust prices confidence in institutional behavior, while energy prices confidence in physical delivery. Both are inputs the market has always priced implicitly and is now beginning to price explicitly. The next letter in this series turns to a capital that is scarcer still because it cannot be built, only extracted or granted: Political Capital.
Energy Capital is real, measurable, and currently the single most binding constraint on AI infrastructure buildout — more binding than chip supply, which is what makes the March 2026 Ratepayer Protection Pledge and Microsoft's UAE power-first structuring rational rather than defensive. But firm power is a long-duration bet against a demand curve still finding its shape, and the regulatory tolerance for concentrated local power draw (Texas SB6 is the leading indicator) is a genuine and underpriced risk sitting on top of it. The framework: value accrues to whoever holds contracted, deliverable power at the right site — not to whoever simply announces the largest gigawatt commitment.