The Nuclear Gold Rush Has a 53% Overrun Problem

The fork wasn't a fork. It was a 53% cost hike that killed the first U.S. small modular reactor (SMR) project before a single watt could power a server. In November 2023, NuScale scrapped its Carbon Free Power Project in Idaho after the estimated cost ballooned from $5.8 billion to $8.9 billion. The reactor design was certified by the NRC. The regulatory path was cleared. The economics were not. Yet Silicon Valley keeps writing checks to nuclear startups, spurred by the narrative that AI's insatiable demand for clean baseload power has ignited a new energy gold rush. I've spent five years dissecting crypto narratives that sounded just as compelling. This one smells the same. Context: The premise is real. Data centers require 24/7 high-density electricity, and hyperscalers like Microsoft, Amazon, and Google are desperate to decarbonize their supply. Nuclear offers a high capacity factor with near-zero operational carbon. The article that sparked this analysis (Crypto Briefing, August 2025) correctly identifies the intersection of AI demand and investor attention. But it treats “nuclear” as a monolith, lumping together large reactors, SMRs, and fusion startups as if they share the same timeline and risk profile. They don’t. The difference between an SMR and a fusion reactor is the difference between a submarine and a starship. The source article also ignored the supply chain, the competitive landscape, and the most inconvenient truth: cost. Core: Let’s dissect the economics first. SMRs like NuScale’s VOYGR design have a levelized cost of electricity (LCOE) estimate of $100–$150 per MWh—some scenarios exceed $200. Compare that to natural gas at $40–$60 or solar-plus-storage at $50–$80. The only way SMRs compete today is through government subsidies or corporate PPAs that effectively tax the buyer’s ESG budget. That’s not a market; it’s a patronage system. Historical data from 116 nuclear construction projects shows an average cost overrun of 117%. SMRs were supposed to escape this curse through factory fabrication. NuScale proved otherwise. The first-of-a-kind penalty remains alive and well. Then there is the time mismatch. AI load is exploding now—2024 to 2027. The earliest operational SMR in the U.S. is Terrapower’s Natrium reactor in Wyoming, targeting 2028. Fusion is at least a decade away, if ever. The U.S. Energy Information Administration expects 30 GW of new natural gas capacity and 100 GW of solar by 2026, with essentially no new nuclear. So the immediate solution to AI’s power hunger is not nuclear startups—it’s gas turbines and batteries. The “gold rush” narrative implies we are digging nuggets today. We are not. We are buying lottery tickets for 2035. The supply chain adds another layer of fragility. Uranium prices have tripled since 2021 to over $90/lb, driven by supply deficits and reactor restarts. SMRs like Terrapower and Oklo require HALEU (high-assay low-enriched uranium, 5–20% U-235). The U.S. has no commercial HALEU production today. Centrus Energy’s demonstration plant won’t deliver meaningful quantities until 2025 at the earliest. Relying on Russian imports is geopolitically toxic. One feedstock bottleneck can chain the entire sector. The source article never mentioned this. Regulatory friction is another blind spot. The NRC approval timeline for advanced reactors is 40–60 months. Proposed reforms could cut that to 24 months, but legislation remains unpassed. Meanwhile, water consumption—both for reactor cooling and for data center cooling—creates regional constraints. In arid states like Arizona or Nevada, a nuclear-powered data center would strain already scarce water resources. The article ignored this. Competing technologies also threaten nuclear’s baseload value. Hydrogen-capable gas turbines (GE 7F already blends 30% hydrogen) and long-duration storage (Form Energy’s iron-air batteries targeting $20/MWh) could erode the nuclear advantage faster than expected. In Texas, solar-plus-storage already supplies over 60% of grid demand during some hours. The nuclear thesis assumes these alternatives won’t improve fast enough. That’s a bet, not a certainty. Contrarian: But the bulls have a point. The sheer scale of AI power demand—projected to reach 10% of total U.S. electricity by 2030—is a structural shift. Regulatory tailwinds are real: the Inflation Reduction Act offers investment tax credits up to 30% for advanced nuclear, and bipartisan infrastructure bills have allocated billions for demonstration. Some investors, like Bill Gates with Terrapower, are betting on proven technology with a credible path. And Microsoft’s virtual PPA to restart a unit at Three Mile Island shows that tech giants are willing to pay a premium for nuclear’s attributes. Assets don’t have feelings; investors do. The nuclear revival feels inevitable because the narrative is emotionally resonant. But the unit economics remain cold and unresponsive. The contrarian insight is that the real value may not sit in the equity of startup builders, but in the PPAs and utility assets that bridge the gap. Companies like Constellation Energy and NRG, which own both nuclear and gas assets, offer a hedged play on the energy transition without the binary risk of a single reactor design. Also, if AI chip efficiency improves faster than expected—NVIDIA’s next-generation GPUs already show a 30% reduction in power per Teraflop—demand growth could plateau by 2027, undercutting the nuclear investment thesis entirely. Takeaway: Cold hands dissect the heat of a hype cycle. The nuclear-AI gold rush is a story about capital allocation, not energy physics. Until a reactor design clears the NRC’s construction and operating license in less than four years, and delivers electricity at or below $60/MWh, it remains a science project with a valuation multiple. We audit the funding rounds, but we’ll mourn the investors who forget that in energy, scale takes time and time is the one thing hype cannot compress. Yield is a sedative; volatility is the needle.