The AI Power Reckoning: Why Small Modular Reactors Are Back
Published:
AI doesn’t just eat compute—it eats reliable megawatts.
A Prophesy of the AI Future
In 2021, I put a slide into a deck that most senior government officials and private sector leaders politely ignored: the world was about to wake up to AI, and we were nowhere near ready for the power it would require. The math was obvious even then. Training modern models demanded data-center scale infrastructure, and inference would turn into a daily industrial workload.
Then ChatGPT launched in late 2022—and the world caught fire. Overnight, executives who had never asked a question about grid capacity were suddenly asking about GPUs, power purchase agreements, and “how soon can we build a data center?” The hard truth arrived fast:
The next frontier isn’t only algorithmic. It’s electrical.
Here is a simple example from December 2022 (just after the release of ChatGPT):
- ChatGPT Training: Consumed 1,287 MWh over a three-month training period, which is equal to the power consumed by 120 average sized homes in an entire year.
- ChatGPT Inference: 40M kWh per day! That is equal to the power consumed by 3,900 average sized homes in an entire year. In other words, ChatGPT uses the same power in one year as 1.4 million homes in the same year. Are we paying attention yet?? And these are 2022 numbers!
Wind and solar are essential, but they’re not always-on. Storage helps, but it’s not free, it’s not instant, and it scales slower than hype. When you’re trying to run large-scale AI systems – especially for national-security, healthcare, critical infrastructure, and continuous decision-support – you don’t get to “pause” because the sun went down.
That’s where Small Modular Reactors (SMRs) enter the conversation again – not as a buzzword, but as a practical response to a new constraint: reliable, low-carbon baseload power at scale.
What SMRs actually are (no marketing gloss)
The International Atomic Energy Agency describes SMRs as newer-generation reactors typically up to ~300 MW that can be shop-fabricated and transported as modules, deployed as demand rises—often with advanced or inherent safety features. (1)
This “build in factories, assemble on-site” concept is the whole bet: repeatability, standardization, and learning curves—more like aircraft production than one-off mega-projects.
Safety is part of the pitch – and some designs take it seriously
One of the most compelling arguments for certain SMR architectures is how they handle worst-day scenarios – station blackout, the kind of failure mode the world learned about at Fukushima. NuScale’s published safety concept emphasizes passive cooling and long-term heat removal without requiring external power or pumps, including a design approach that can support long-duration cooling under extreme events.(2)
Whether any specific design ultimately wins is still an open race—but the direction is clear: design for “walk-away safe,” reduce dependency on heroic interventions, and simplify the response to extreme events. (2)
Here’s the uncomfortable part: SMRs aren’t automatically cheap
If you want an honest conversation, you can’t skip the hardest issue: economics.
A widely cited “reality check” argues that SMRs face structural cost headwinds because smaller units often have higher costs per unit of capacity, and the expected savings from mass production require large, sustained demand and standardization – conditions that aren’t guaranteed. (3)
The same critique notes that meaningful cost reductions may require manufacturing in the hundreds or thousands, and that the large number of competing SMR designs makes true standardization difficult. (3)
That’s why serious SMR strategies don’t pretend the market will magically fix itself. They focus on the real levers:
- Demonstration units (FOAK) to prove performance and reduce investor risk (4)
- Regulatory harmonization so designs don’t get relicensed from scratch across jurisdictions (4)
- Supply chain rebuilding after decades of atrophy in many countries (4)
- Fuel cycle readiness, including potential needs like HALEU for some designs (4)
My takeaway and why I pushed this in government
I’ve watched how fast national priorities shift once constraints become visible. AI made the constraint visible. Not in a white paper – in executive dashboards.
And then I saw the same dynamic from the inside of government. While serving as Deputy Director of National Intelligence, I worked directly with Congress, the White House, and the Department of Energy to push on the policy and legal friction that was slowing nuclear deployment – especially SMRs. The goal wasn’t ideological. It was operational: reliable, low-carbon baseload power is now a national competitiveness issue, and the licensing/permitting timeline has become a strategic variable.
Here’s the leadership lesson:
When a system is stressed, the bottleneck becomes the agenda.
The SMR conversation is not “nuclear vs renewables.” It’s reliability + decarbonization + national competitiveness in an era where compute demand behaves like a new industrial revolution.
If we want the U.S. (and allies) to stay ahead – economically, scientifically, and strategically – then we must treat energy like the enabling infrastructure it is. That means modernizing licensing pathways, investing in testbeds and demonstrators, rebuilding supply chains, and being honest about timelines and tradeoffs. (4)
Because the AI era won’t wait for us to get comfortable.
Final question
If AI is becoming a 24/7 national capability, what are you doing to secure the 24/7 power it requires—before the constraint becomes the crisis?