The first modular nuclear reactor project in the United States was cancelled in November 2023 after its estimated cost jumped from $5.8 billion to $8.9 billion — a 53% increase that killed the economics before a single concrete pour for the reactor building. NuScale Power, the company behind it, saw its stock price fall over 90% from its IPO peak. Yet venture capital continues to flood into nuclear startups, driven by a single narrative: AI’s insatiable power demand needs clean, 24/7 baseload electricity, and only advanced nuclear can deliver it. Here is the error: treating a capital-intensive, long-cycle infrastructure bet as a software-growth narrative.
Context: The AI Power Panic
The story is compelling. Data centers for training large language models and running inference require hundreds of megawatts of continuous power. A single AI cluster can consume 50 MW or more. Hyperscalers like Microsoft, Amazon, and Google have signed power purchase agreements (PPAs) for existing nuclear plants, even buying the output of reactors slated for decommissioning. Silicon Valley’s investors see a structural gap: renewable sources like solar and wind are intermittent, battery storage is still too expensive for multi-day gaps, and natural gas, while cheap, carries carbon liabilities. Nuclear — specifically Small Modular Reactors (SMRs) and fusion — offers the promise of clean, dense, always-on power. The result is a gold rush of capital into startups like Terrapower, Commonwealth Fusion, Helion, and Oklo.
But the gap between narrative and physics is wide. AI power demand is exploding now (2024–2027). Commercial SMRs are not expected to deliver power before 2030. Fusion is at least a decade away. The deployment timeline mismatch is the first crack in the story.
Core: The Economics of Nuclear Hope
Let me walk through the numbers, because in energy, as in smart contract security, the details in the arithmetic are where the exploits hide. Tracing the gas leak where logic bled into code — in this case, the logic of modular cost savings bled into financial models that ignored first-of-a-kind risks.
Current Levelized Cost of Energy (LCOE) estimates for SMRs range from $100 to over $200 per MWh, depending on the design and financing assumptions. Compare that to combined-cycle natural gas at $40–60/MWh, or solar-plus-storage at $50–80/MWh. Even with the 30% Investment Tax Credit from the Inflation Reduction Act, SMRs remain uncompetitive without either a high carbon price or below-market financing from technology buyers. NuScale’s cancelled project required a special rate agreement with utilities that would have paid roughly $89/MWh — and even that proved insufficient when construction costs escalated.
Fusion is even more speculative. Commonwealth Fusion’s SPARC tokamak is targeting Q>1 (net energy gain) by 2025, but commercial electricity is projected for the mid-2030s. Helion claims it can power Microsoft by 2028, but its science has not been peer-validated by the broader fusion community. Based on my experience auditing complex financial instruments for DeFi protocols, I recognize the pattern: over-optimistic timelines paired with insufficient stress-testing of failure modes. Every governance token is a vote with a price — here, every fusion investment is a bet with a timeline that may never materialize.
Beyond cost, the supply chain is broken. Many advanced SMR designs — Terrapower’s Natrium, Oklo’s Aurora — require High-Assay Low-Enriched Uranium (HALEU), enriched to between 5% and 20% U-235. Currently, the only commercial HALEU producer is Russia. The U.S. has exactly one pilot facility, operated by Centrus Energy, which began producing small quantities in 2024. To fuel even a handful of SMRs by 2030, the U.S. needs to build multiple centrifuge cascades — a process that itself takes years and billions of dollars. The uranium spot price has already tripled since 2021, from $30/lb to over $90/lb, driven by supply deficits and reactor restarts. If AI-driven demand for new reactors materializes, the fuel cost will rise further, crushing already marginal economics.
In the silence of the block, the exploit screams — here, the silent exploit is the cumulative effect of cost overruns, fuel constraints, and regulatory delays. The NRC’s licensing process for a first-of-a-kind SMR takes 40–60 months. Congress has proposed reforms to shorten this to 24 months, but those bills are stalled. Even if passed, the NRC must build new expertise for non-light-water designs. The regulatory timeline alone pushes SMR commercial operation past 2030.
Contrarian: The Blind Spots the Narrative Ignores
The gold rush narrative assumes that nuclear is the only baseload solution. But two competing technologies are advancing rapidly. Hydrogen-capable gas turbines (GE’s 7F series can already burn a 30% hydrogen blend, with 100% on the roadmap) offer a dispatchable, lower-carbon alternative without the 10-year construction horizon. Long-duration energy storage — specifically iron-air batteries from Form Energy, targeting $20/MWh — could allow solar-plus-storage to cover multi-day gaps, eroding nuclear’s baseload value proposition.
Then there is the water problem. Both nuclear reactors and data centers are water-intensive. A typical 1 GW nuclear plant uses 30–60 million liters of cooling water per day. A hyperscale data center can use 1–5 million liters daily. In drought-prone regions like the American Southwest, competing for water with agriculture and municipalities creates a permitting nightmare that few investors have modeled.
The biggest blind spot, however, is the assumption that AI power demand grows linearly forever. Historically, each generation of AI chips has delivered dramatic efficiency gains. NVIDIA’s next-generation Blackwell architecture claims a 25x improvement in energy efficiency for inference workloads per dollar. If chip-level efficiency outpaces load growth — and it has for most of computing history — the projected electricity demand surge may peak earlier than expected, deflating the urgency to build costly new nuclear capacity. Governance is just code with a social layer — but here the governance of energy planning is just a social layer over hardware physics, and the hardware is improving faster than the reactors.
Takeaway: A Long-Dated Option, Not a Gold Rush
Silicon Valley’s nuclear investments are best understood as long-dated call options on regulatory reform, cost reduction, and sustained AI growth. The probability of hitting strike price — actual commercial power delivery before 2030 — is low. Investors who treat these as near-term energy solutions will be disappointed.
The real signal to watch is not another startup funding round, but the first SMR to receive a Combined Operating License from the NRC and begin construction without cost overrun. Until then, the gold rush remains a narrative with fragile optics. Optics are fragile; state transitions are absolute. When the next project cancellation hits, the capital flow will pivot as fast as it arrived.