Experts and pundits have debated the feasibility of small modular reactors for years—largely because they misunderstand the economics of factory production. AI’s endless demands on power are forcing a reasessment
Background
NVIDIA’s latest deal with OpenAI is measured not in dollars—but in gigawatts. Their plan to deploy 10 GW of NVIDIA systems marks a turning point in the energy race behind artificial intelligence. By 2030, AI servers could consume up to 20% of the total electricity demand on an already overstressed U.S. power grid. This is forcing a fundamental rethink of where that power will come from.
One emerging answer is small modular reactors (SMRs): compact nuclear units roughly 10–30% the size of conventional reactors. More importantly, SMRs are factory-built rather than custom-constructed on site, as has traditionally been done. Major AI infrastructure firms—hyperscalars such as Meta, Amazon, Microsoft, Google, and now OpenAI—are investing in SMR developers to power the next generation of AI and cloud computing. Their capital and long-term commitments could finally move SMRs from prototype to commercial reality.
But have they made a mistake? Pundits—and even many experts—have debated for years whether the economics make sense. Advocates tout SMRs as the affordable savior of clean energy. Critics, citing cost overruns at projects like Georgia’s Vogtle plant, see them as an economic disaster. What both sides fail to grasp is that the economics of factory-built products are fundamentally different from those of traditional, site-built reactors.
Why It Matters — The Economics of Factory Production
Learning curves are a game-changer. Dozens of studies show that every doubling of production in factories cuts costs by 10–30%. For complex technologies like reactors, the rate is likely lower—5–15% per doubling, as Idaho National Lab estimates.
Pro-SMR pundits are right: learning curves most likely will make nuclear far more affordable once production lines scale. Some compelling evidence exists to back advocates. Virginia-class submarines, which use factory-built small nuclear reactors, fell 20% in real cost between 2004–2014. If submarines—far more complex and mobile—can realize learning effects, stationary reactors should too. However…
Critics are also right: the first SMRs are likely to be more expensive, probably much more expensive, than conventional reactors because their costs include setting up the factories.
The Catch — Designs Must Remain Stable
Learning curves only work if designs don’t constantly change. And nuclear faces three problems:
Engineering drift: Insiders say engineers often tinker too much, chasing improvements instead of sticking with proven models—constantly resetting the learning curve.
Regulatory volatility: Environmental and safety rules frequently shift mid-project, delaying builds and scaring off investors..
Abandoned projects: When companies scrap an SMR initiative, they lose hard-won experience and reset the curve entirely.
Unless both industry and regulators commit to stability, SMRs won’t achieve the factory-driven cost reductions they need.
What To Do About It
Both sides have it wrong. SMRs are neither guaranteed saviors nor doomed failures. Their economics depend on what government and industry do now:
SMR companies must freeze their designs. Constant redesigns reset the learning curve and erase progress.
Regulations must remain fixed once construction begins. Midstream changes delay builds and deter investment.
AI hyperscalers must invest in factory production and stick with it. Initial pilot efforts should be expanded, not scrapped, to fully exploit the learning curve.
Governments must back the hyperscalers. Public programs should reinforce private investment and accelerate the transition from prototype to production.
