This Week in Voltage
The U.S. government just placed a very large bet on abundance.
A new federal initiative will provide low-cost financing to increase output from the existing nuclear fleet by 5 GW by 2029 — targeting surging power demand from data centers and manufacturing. Not new reactors. Not a decade-long construction cycle. Squeezing more electrons out of plants already running. That's the kind of pragmatic, build-now thinking that actually moves civilizational needles.
Meanwhile, FERC confirmed it will issue large-load interconnection rules by the end of June 2026 — a hard deadline for the framework that will govern how data centers and industrial loads actually connect to the transmission system. I've covered the interconnection bottleneck in previous issues and again last week. The June deadline is the moment we find out whether FERC's urgency is real or rhetorical.
Both moves point the same direction: build more, connect faster, supply the demand. Which makes this a good week to bury the alternative argument once and for all.
Deep Charge: Efficiency Gains Are Real. They're Also Not Enough.
The conservation case for AI energy goes like this: AI is getting more efficient per task, so maybe the demand surge isn't as bad as feared. And the efficiency numbers are genuinely impressive. The IEA's latest analysis finds that energy use per AI task has dropped by at least an order of magnitude annually in recent years. Simple text queries now consume less electricity than running a television for the same duration.
If you stopped reading there, conservation looks like a viable strategy. Efficiency improves, demand stabilizes, problem managed.
But the IEA doesn't stop there — and neither should you. The same report documents that new, far more energy-intensive AI applications are being launched constantly: video generation, reasoning models, agentic tasks. The efficiency gains per query are real. The aggregate demand trajectory is still sharply upward. This is Jevons' Paradox running at civilizational scale: make something cheaper to use, and people use vastly more of it.
Meanwhile, the IEA reports that capital expenditure from just five major technology companies exceeded $400 billion in 2025 — larger than global investment in oil and gas production — and is expected to jump another 75% in 2026. AI factories have more than tripled in capacity over the past 18 months. These are not efficiency-constrained numbers. These are abundance-seeking numbers.
The conservation argument assumes a fixed or slowly growing demand ceiling that efficiency improvements can meet. The actual demand curve is being set by the most capital-intensive buildout in the history of computing. Efficiency is a tailwind. It is not a substitute for supply.
By the Numbers
- 5 GW — additional nuclear output targeted by the U.S. government through low-cost financing, to be delivered by 2029 from the existing fleet (Reuters)
- $400B+ — capital expenditure from five major tech companies in 2025 alone, exceeding global oil and gas production investment (IEA)
- 75% — projected jump in that capex figure for 2026 (IEA)
- 3,500+ pages — comments FERC reviewed in its large-load interconnection docket before committing to June action (JDSupra/Troutman Pepper Locke)
- >3x — growth in AI factory capacity over the past 18 months, per IEA satellite tracking (IEA)
What We're Fighting For
The conservation-first crowd isn't wrong that efficiency matters. They're wrong about what efficiency is for. Every efficiency gain in AI inference is fuel for the next expansion — more queries, more users, more applications, more demand. That's not a bug. That's how civilizational technologies work. Electricity didn't liberate humanity by being used sparingly.
The 5 GW nuclear financing initiative and FERC's June interconnection deadline are two data points in the same argument: the institutions that matter are, however slowly, choosing supply over scarcity. The question isn't whether to build — it's whether we build fast enough to stay ahead of the demand curve that's already here.
Watch June. FERC's interconnection rules will either clear the path or add another layer of procedural friction to a queue that's already years deep. That ruling is the most consequential near-term decision for whether the grid becomes an accelerant or a ceiling.
The future is electric. The only debate is how much of it we're willing to build.