I've been working on a longer piece about what AI is actually going to cost us, and I keep landing on the same uncomfortable answer. The cost of AI isn't the cost of AI. It's the bill coming due on fifty years of underinvestment in the power grid and the water systems that quietly hold up everything else. Compute is the headline. Infrastructure is the story.
So I'm putting this research up on its own ahead of that larger article. The position needs to be in the open now, because there isn't a serious conversation about future prosperity that doesn't run straight through power generation and water. Skipping that conversation is how we got here.
A lot of smart people have already given up on solving it on the ground. The sense of futility runs so deep that billionaires are openly proposing to put data centers in orbit, as if the answer to permitting friction in Virginia is to launch a kilometer of cooling loops above the atmosphere. I'll say it plainly. Neither space nor Narnia is going to save us. We solve this on Earth, with practical tools, on the grid we have.
What follows is what I've found about that grid and what we can actually do about it.
Summary
The U.S. electric grid is the product of a 1945 to 1975 build-out that has spent the last fifty years aging in place. Roughly 70% of large power transformers are over 25 years old, the system earned a C-minus from ASCE, and it's now feeding a country that needs two to three times more capacity to absorb AI data centers, reshored manufacturing, and electrified heat and transport.
The fastest path forward isn't new lines, which take 10 to 15 years to permit. It's reconductoring the existing rights-of-way with advanced composite-core conductors. A Berkeley-led PNAS study finds reconductoring alone could double transmission capacity by 2035 and yield about $85 billion in cost savings. Combine that with grid-enhancing technologies, small modular reactors at retired coal sites, and virtual power plants, and the path to a much bigger grid actually pencils out.
Diagnosis: a grid built for a different country
Most of the grid you depend on every day was built between 1945 and 1975, during the high-investment "Great Compression" era. A lot of it is now 40 to 70 years old, well past its intended lifespan. That's not just a maintenance budget problem. It's a structural truth about how reliable the system can be.
The component-level decay is concrete. Roughly 70% of large power transformers are over 25 years old and increasingly failure-prone, and substations are vulnerable to cascading failures where one bad component takes down a wide area. The U.S. now experiences more outages than most peer developed nations, and weather events that used to be one-off problems (hurricanes, wildfires, winter storms) now routinely take down chunks of the system because there's so little margin left.
The American Society of Civil Engineers gave U.S. energy infrastructure a C-minus in its most recent report card. The 2021 grade was a D+. So things are improving on paper, but the headline isn't really capacity. It's integrity.
There's a quieter problem layered on top. Older grid components were never designed for modern attack surfaces. A lot of SCADA systems and substation controls predate the threat environment we live in now, which means cybersecurity isn't an add-on concern. It's baked into the age of the equipment.
Why governance makes the problem hard
The U.S. grid isn't a grid in the way most people picture one. It's a patchwork of more than 3,000 independent utilities, layered under fragmented state regulators that often actively block interstate interconnection. Unlike most European grids, no single authority is in charge of "the grid" as a coherent thing.
Permitting drag is the dominant friction. Building a major new transmission line in the U.S. now takes an average of 10 to 15 years. NEPA-era environmental review and citizen-voice litigation are the largest drivers of that timeline.
The downstream effect of those two facts is staggering. Over 2 terawatts of generation capacity, overwhelmingly wind, solar, and storage, currently sit in interconnection queues waiting for permits to plug in. This was already a crisis in 2019, before any AI-related demand surge. Capacity sitting in queue is functionally non-existent for the grid.
The demand collision: why "now" feels acute
The current emergency is the collision of an expected maintenance failure with an unexpected demand explosion. Three forces hit at the same time and broke an assumption planners had quietly relied on for two decades.
| Factor | 2021 projection | 2026 reality |
|---|---|---|
| Manufacturing onshoring | "A distant goal" | Massive surge in battery and semiconductor megafabs, each consuming as much power as a small city |
| Electrification pace | Moderate, slow | Aggressive uptake of heat pumps and EVs, shifting peak loads to winter in many regions |
| AI and data centers | "Stable growth" | Exponential surge; AI servers use 3x to 5x more power than standard servers, doubling local demand overnight |
From 2005 to 2020, U.S. electricity demand was essentially flat. Efficiency gains from LEDs and better appliances quietly masked the fact that no new baseload was being built. Planners assumed that offset would carry through 2030 and chose not to build. AI and electrification ended that 20-year flat-demand era roughly five years ahead of schedule. Data centers are responsible for about half of the immediate demand increase. Reshoring and electrification supply the other half.
This is the part I want people to understand clearly. The AI boom isn't the cause of the capacity crisis. It's the trigger that exposed it. The crisis was already loaded.
The fast-path solution: reconductoring
A first-of-its-kind Berkeley-led modeling study (Callaway, Phadke et al., PNAS 2024) found that the cheapest and fastest way to expand transmission capacity is not to build new lines but to replace the conductors on the 53,000 miles of existing U.S. transmission lines with advanced composite-core (ACCC) materials.
The numbers are striking. Doubling transmission capacity on the existing grid is feasible by 2035 with reconductoring alone. The study estimates $85 billion in system cost savings by 2035, and $180 billion by 2050. Wholesale electricity costs drop 3 to 4% on average. And critically, none of this requires fighting through the 10-year permitting bottleneck, because the lines and rights-of-way already exist.
The technology has already been deployed in Belgium, the Netherlands, and on U.S. river crossings. It just hasn't made it to the bulk of the overhead lines that feed residential and commercial customers. The Department of Energy has since committed hundreds of millions of dollars to reconductoring projects in direct response to the modeling work.
The full 10x toolkit
Reconductoring is the largest single lever, but it isn't the whole answer. A 10x grid is a portfolio.
| Strategy | Mechanism | Scaling potential |
|---|---|---|
| Advanced reconductoring | Replace old aluminum lines with advanced composite-core conductors on existing towers | 2x to 3x capacity on existing rights-of-way without new permitting |
| Grid Enhancing Technologies (GETs) | Dynamic Line Rating plus AI to adjust real-time power flow based on weather and conductor temperature | About a 40% increase in existing grid capacity, deployable in months |
| Small Modular Reactors (SMRs) | Factory-built nuclear units sited at retired coal plants, using the existing grid interconnections already in place | Roughly 10x faster deployment than traditional large-scale nuclear; first U.S. construction permits active in 2026 |
| Virtual Power Plants (VPPs) | Aggregate home batteries, EVs, and smart thermostats into a dispatchable resource | Can offset dozens of traditional peaker plants |
| Distributed Energy Resources and microgrids | Localized solar and storage; islanded microgrids for hospitals, bases, and critical facilities | Reduces strain on the bulk grid and improves resilience to weather and attack |
| Smart-grid and predictive maintenance | Smart meters, automated sensors, AI-driven fault prediction | Reduces outages, enables condition-based maintenance and real-time adjustment |
The three-pronged 10x roadmap
If I had to compress this whole picture into three actions:
- Reconductor every major transmission line to double existing flow.
- Deploy SMRs on federal lands and at retired coal sites, bypassing state-level zoning delays.
- Nationalize interconnection standards so state-level NIMBY litigation cannot stall multi-state energy transfers.
None of these are exotic. None of them require waiting on a breakthrough. They require deciding to do the work.
The obstacle is the way
There's a line from Marcus Aurelius that I find equal parts inspiring and terrifying. "What stands in the way becomes the way." Ryan Holiday turned it into a book brand. I think it's actually a load-bearing observation about civilizations.
We tend to talk about power generation as a dependency for wealth. Something that has to be in place before the real economy happens. That framing undersells it badly. Power generation isn't a dependency for wealth. It is wealth. The lines, the substations, the reactors, the distribution networks, all of it is wealth in physical form, and the act of building it is a generational transfer of jobs, trades, apprenticeships, and engineering capacity that doesn't come from anywhere else.
Fixing the grid isn't an expense we'd love to avoid on the way to prosperity. It is the prosperity, and it pays out for decades. That's the part of the conversation that's missing when people frame infrastructure cost as a tax on innovation. The bricks and the wires and the welders are the innovation. They always were.
Actionable takeaways
For grid decisions, the binding constraint is permitting and interconnection, not engineering or capital. Solutions that reuse existing rights-of-way (reconductoring, GETs, SMRs at coal sites) clear faster than greenfield builds and should be prioritized when speed matters.
Treat data centers as half the problem, not the whole problem. Reshored manufacturing and electrification together equal AI's grid impact. Plans that focus only on hyperscaler demand will underbuild.
Watch the interconnection queue, not nameplate capacity. Around 2 TW of mostly-renewable capacity is already built or planned but cannot connect. That's invisible capacity for the people who depend on the grid.
Cybersecurity is now load-bearing. Modernization plans that don't replace pre-internet SCADA components are deferring rather than solving the risk.
The Bipartisan Infrastructure Law (2021) is necessary but not sufficient. The dollar figures move the needle on early-stage projects, but the regulatory and interconnection bottlenecks consume most of the schedule.
Closing thought
I don't think the people pitching orbital data centers are stupid. I think they've looked at the U.S. permitting environment, run the math, and concluded that escaping the planet is genuinely cheaper than fixing it. That's not a flex. That's a confession.
We can do better than that, and we should. The tools are sitting on the shelf. The economics work. The labor force exists. What's been missing is the will to treat the grid as something worth being seriously good at again.
If you're working on this from any angle, utility side, vendor, policy, regulatory, civic, I'd love to hear what's actually moving and what's still stuck. The companion piece on water is coming, and the larger article on AI's real cost will follow that. The position needs to be visible while it's still useful.
Notes
- This article focuses on the electric grid specifically. A companion piece is in progress on the parallel water story (different physics, similar political economy), and a third on the cross-cutting regulatory and economic causes that produced both.
- The "AI and data centers as the grid villain" narrative is in productive tension with Andrew Ng's framing of data-center load growth as electricity-price-reducing under Lawrence Berkeley analysis. Both can be true depending on the time horizon.
- Reconductoring as the cheap-fast-easy answer is the kind of high-ROI infrastructure intervention that historically gets adopted only after the slower, more expensive options have visibly failed. Worth tracking how quickly utilities actually move on the DOE-funded projects.
- Open question: do GETs scale linearly with grid size, or do they hit a coordination ceiling once enough lines are dynamically rated? The literature is thin.
Citations
- "Aging Electric Infrastructure in the United States," Shalini Bhat, UW–Madison Interdisciplinary Professional Programs (May 2025). Diagnosis of grid age, outage costs, and modernization options. Source for the 70%-of-transformers, ASCE grading, and DER and microgrid framing. https://interpro.wisc.edu/aging-electric-infrastructure-in-the-united-states/
- "Advanced Conductors Provide Path for Grid Expansion," Berkeley Research summary of Callaway, Phadke et al., Accelerating transmission capacity expansion by using advanced conductors in existing right-of-way, PNAS 2024. Source for the reconductoring economics ($85B and $180B savings, 3 to 4% wholesale cost reduction, interconnection backlog). https://vcresearch.berkeley.edu/news/advanced-conductors-provide-path-grid-expansion / https://doi.org/10.1073/pnas.2411207121
Further reading
- "The U.S. Urgently Needs a Bigger Grid. Here's a Fast Solution." (New York Times)
- "How a simple fix could double the size of the U.S. electricity grid" (Washington Post)
- "There Is a Stupidly Easy Way To Expand the Grid" (Heatmap News)