The economics of artificial intelligence have quietly shifted. Energy, not compute, is now the binding constraint on AI development, and the race to secure it is reshaping the energy sector in real time.
Microsoft's deal with Constellation Energy to reactivate Three Mile Island—the site of America's most serious commercial nuclear accident in 1979—exemplifies this pivot. The company will spend $1.6 billion to bring the facility online by 2028, securing all its output under a twenty-year power purchase agreement. This is not a side project; it is a statement that the company considers nuclear baseload power essential to competitive AI development. Meanwhile, Elon Musk's Colossus data center in Tennessee, now under construction, will consume 150 megawatts—enough to power a region of 4.5 million people—to train a single model. Amazon has already moved into nuclear-powered data centers through acquisitions. The pattern is unmistakable.
The scale is staggering. Goldman Sachs estimates that US data center energy consumption will jump from 3 percent of national demand today to 8 percent by 2030. Globally, AI data centers alone will approach half Argentina's total electricity consumption within three years. This is not hyperbole; it is the arithmetic of transformer model training at scale. Yet what makes the nuclear pivot rational is cost. A kilowatt-hour from nuclear generation undercuts fossil fuels and competes favorably even with renewables, all while running 24/7 without weather dependency. The technology is also, by the numbers, among the safest forms of energy production humans have developed. Perception lags reality: Three Mile Island shaped decades of nuclear anxiety, but the accident itself became the catalyst for modern safety protocols.
The capital concentration this creates, however, masks a deeper fragility. OpenAI is raising $6.5 billion at a $150 billion pre-money valuation—a figure equal to the entire US venture capital market in 1999, the year before the internet bubble burst. Add Elon Musk's $6 billion Series C for xAI, and two companies are absorbing venture firepower that historically funded entire ecosystems. If this wave of investment in foundational AI models does not generate proportional economic returns—if scaling data and compute hits genuine diminishing returns—the correction will be severe. What looks like disciplined capital allocation to professional investors may in retrospect look like concentrated FOMO. The nuclear energy infrastructure built to support it will remain long after the valuations correct.