2026-10-07 · Gold Rush · markdown version
Google ordered nuclear reactors from Kairos Power in October 2024. Microsoft signed a twenty-year deal to restart Three Mile Island Unit 1. Amazon invested in X-energy's small modular reactor development and purchased a nuclear-powered data center campus from Talen Energy for $650 million. These are not pilot programs or sustainability theatre. The contracts commit to gigawatt-scale power delivery with penalties for non-performance. The patient on the table is the consensus narrative that AI capital expenditure flows primarily to semiconductor fabrication and cloud infrastructure, with energy treated as a manageable operating expense line. The contrarian read is that electrons, not transistors, represent the binding constraint on AI scaling over the next decade, and that commodity producers in uranium and electrical infrastructure are mispriced second-derivative AI plays because portfolio construction still classifies them as legacy industrials rather than enabling technology.
The consensus allocates AI exposure through semiconductor manufacturers, cloud hyperscalers, and software platforms. NVIDIA captured $22.1 billion in data center revenue in Q3 2024. Broadcom, AMD, and TSMC trade at valuations reflecting expectations of sustained AI chip demand growth. Energy costs appear in analyst models as a variable operating expense, typically three to seven percent of total data center capex. The International Energy Agency projects data center electricity consumption could double by 2026, reaching 1,000 terawatt-hours annually, but this figure is presented as a challenge for grid operators rather than a constraint on AI deployment. Utility-scale solar and wind additions are expected to meet incremental demand. The consensus treats power as abundant and fungible. Capital flows accordingly: semiconductor and cloud infrastructure funds saw $47 billion in net inflows during 2024, while commodity-focused vehicles experienced $8 billion in outflows. The market prices chips as scarce and energy as elastic.
The crazy read is that the hyperscaler nuclear deals reveal energy as the true bottleneck, and that the market has mispriced this constraint by at least one order of magnitude relative to semiconductor scarcity. Training GPT-4 reportedly consumed 50 gigawatt-hours. Frontier models in development could require ten times that figure. A single exascale AI cluster operating continuously draws 150 megawatts, equivalent to a small city. The United States added 6 gigawatts of battery storage capacity in 2023 but only 2.5 gigawatts of new nuclear capacity in the past decade. Intermittent renewables cannot support always-on training runs without storage costs that exceed compute costs. Natural gas plants face regulatory and emissions constraints. The hyperscalers are not buying nuclear power as a sustainability gesture; they are buying it because dispatchable, carbon-free baseload power at scale does not exist in sufficient quantity on the open market. Uranium spot prices rose from $47 per pound in January 2024 to $106 by December. Cameco, Kazatomprom, and Energy Fuels trade at valuations reflecting historical mining economics, not a scenario where AI training demand adds structural demand for hundreds of small modular reactors over fifteen years. The second-derivative play is uranium miners, nuclear fuel fabricators, electrical switchgear manufacturers, and transmission infrastructure operators. These entities sit outside the typical technology portfolio construction but could capture economic surplus if energy constraints bind before Moore's Law constraints. The reference hoard reallocation play shifts ten percent of equity exposure into broad commodities exposure as a proxy for this thesis, acknowledging that pure-play uranium vehicles carry concentration risk and that energy infrastructure spans multiple commodity inputs including copper, steel, and rare earths for grid upgrades.
Across 10,000 bootstrap paths (seed 562569), this play moves the reference hoard's 15-year median from $550,962 to $542,921 ($-8,041), and its goal probability from 6.7% to 5.4%. The median terminal value declines by approximately one percent. The fifth percentile outcome improves from $282,860 to $293,320, a gain of roughly four percent in the left tail. The ninety-fifth percentile outcome falls from $1,053,983 to $1,016,065, a decline of approximately four percent in the right tail. Maximum drawdown at the ninety-fifth percentile improves from 38.8 percent to 34.5 percent. The distribution compresses. The play sacrifices median growth and goal probability in exchange for reduced volatility and improved downside protection. The machine does not model uranium prices, nuclear construction timelines, or AI energy consumption trajectories. The simulation treats commodities as a generic asset class with historical return and correlation parameters. The output reflects diversification mechanics, not validation of the energy bottleneck thesis. The play could produce these results even if AI energy demand remains negligible and uranium prices collapse.
The thesis assumes hyperscaler energy demand translates into economic surplus for commodity producers rather than being captured by the hyperscalers themselves through vertical integration or monopsony bargaining power. Google, Microsoft, and Amazon are not buying equity in Cameco or Energy Fuels; they are signing long-term power purchase agreements with fixed pricing that could compress supplier margins even as volumes rise. The hyperscalers possess the capital and engineering resources to internalize reactor development, uranium enrichment, and fuel fabrication if margins justify it. Kairos Power and X-energy are majority-owned by entities that could become competitors to public uranium miners. The thesis further assumes algorithmic efficiency gains will not reduce energy intensity per unit of compute. Training efficiency improved by a factor of forty between GPT-2 and GPT-4 when measured in compute per parameter. Inference efficiency improved by a factor of one hundred with quantization and distillation techniques. If energy per token collapses by an order of magnitude, total energy demand could plateau even as model deployment scales. The nuclear deals could represent insurance against tail risk rather than anticipation of structural demand growth. The timeline for small modular reactor deployment remains speculative; no SMR design has achieved commercial operation at scale in the United States. Regulatory approval, construction delays, and cost overruns could push deployment beyond the fifteen-year simulation horizon. The commodity reallocation introduces exposure to oil, natural gas, and industrial metals that have no direct relationship to the AI energy thesis. Broad commodity indices could underperform even if the uranium thesis proves correct. The play also assumes current uranium prices do not already reflect AI energy demand expectations; if the market has priced in structural demand growth, the thesis offers no edge. The reference hoard simulation shows median underperformance, suggesting the market prices commodity exposure as a drag on growth rather than a source of alpha.
Specific falsifiable triggers would include: hyperscaler capital expenditure guidance showing energy costs remaining below five percent of total data center capex over four consecutive quarters, indicating energy has not become a binding constraint. Uranium spot prices falling below sixty dollars per pound for six consecutive months would suggest structural demand has not materialized or that supply expansion has outpaced consumption. Announcements from NVIDIA, AMD, or Broadcom of order cancellations or reduced chip production guidance due to insufficient power availability at customer sites would confirm energy as the binding constraint. Conversely, announcements of accelerated chip production to meet backlog demand would suggest chips remain the bottleneck. Published energy consumption figures for frontier model training runs showing less than ten gigawatt-hours per training cycle would indicate algorithmic efficiency has collapsed energy intensity faster than model size has grown. Commercial operation of five or more small modular reactors in the United States with capacity factors above eighty-five percent and construction costs below four thousand dollars per kilowatt would validate the deployment timeline. Alternatively, cancellation of two or more hyperscaler nuclear deals due to regulatory delays or cost overruns would invalidate the thesis. Cameco or Kazatomprom reporting long-term contract pricing below seventy-five dollars per pound would indicate monopsony power is compressing supplier margins despite volume growth. A scenario where hyperscalers announce in-house uranium enrichment or fuel fabrication capabilities would confirm vertical integration is capturing the surplus. Relative performance data showing commodity indices outperforming technology equity indices over a rolling thirty-six-month period would provide ex-post evidence that the market mispriced the constraint. The opposite outcome would falsify the thesis.