Notes
The AI Data-Center Buildout
On this page
The AI data-center buildout is the physical side of the intelligence explosion: the conversion of land, power, water, and capital into compute at a pace that is reshaping grids today and, in the essays' projection, eventually leaves the planet. It is the place where the energy transition and the Singularity turn out to be the same story.
The power problem, 2026
Global data centers consumed about 415 TWh in 2024 — roughly 1.5% of world electricity — surged about 17% in 2025, and are projected to nearly double to ~945 TWh by 2030, with AI workloads driving the growth. U.S. data center demand is forecast to jump from about 31 GW in 2025 to 66 GW in 2027. Hyperscalers are signing massive renewable PPAs and resurrecting nuclear plants (the Three Mile Island deals), but interconnection delays and the spiky, high-density load profile of AI training clusters make the grid the gating factor.
Batteries are becoming central to the bridge: large BESS arrays as seconds-to-hours backup that is cleaner and faster than diesel, co-located renewables-plus-storage serving campuses directly, and behind-the-meter systems that let data centers act as flexible loads — even sell ancillary services — as stationary packs fall below $100/kWh. The pragmatic path to 24/7 carbon-free operation runs through storage, efficiency, advanced cooling, and a diverse generation mix including eventual small modular reactors.
Nuclear at factory scale
The essays' forward map has the buildout going nuclear in the 2030s: a major wave of terrestrial hyperscale construction powered by factory-produced micro-nuclear reactors and thorium systems, dense and reliable enough to sit beside the load. Power, water, and land become political issues in places that never thought of themselves as energy infrastructure hubs. By the early 2040s the terrestrial wave reaches full speed — and its constraints become binding. Power, cooling, and siting limits for the largest projects are exactly what makes the next step economically rational rather than romantic.
The leap to orbit
Once heavy lift gets cheap — routine high-cadence Starship-class operations collapsing the cost of moving mass — the frontier of compute starts leaving Earth. First experimental orbital testbeds in the late 2030s, then large-scale orbital platforms for the biggest training runs in the 2045–2055 window: vast constellations of interconnected computing nodes powered by uninterrupted solar, cooled by vacuum, networked at light speed. By the 2075–2100 period, orbital data centers are mature core infrastructure, fed by lunar helium-3 and scaled fusion.
The strategic reading is stark. A nation with orbital compute superiority can run more simulations, design more iterations, and react faster than any ground-based adversary; controlling orbital compute, the essays argue, will matter the way controlling oil fields did in the 20th century. Destroying an enemy's orbital constellations would be the equivalent of eliminating their officer corps, intelligence apparatus, and weapons-design capability simultaneously — making them at once the most tempting target and the most heavily defended asset in existence. The buildout that begins with substation queues in Virginia ends, on this map, with the most strategically important infrastructure in the solar system.
Related concepts
Battery Technology & Grid Storage
Appears in
Details
- Section:
- Notes
- Updated:
- 2026-06-12