The U.S. and China Are Racing Ahead — at the Planet’s Expense
The development of AI supercomputers is reaching a new orbit: in terms of scale, cost, and energy consumption, it now rivals mega-infrastructure projects. A study prepared by analysts from Georgetown, Epoch AI, and RAND provides insight into where the industry could be by 2030.
From 2019 to 2025, spending on equipment and energy consumption at advanced AI centers doubled every year. If this pace continues, by the end of the decade a single AI supercomputer could host two million specialized chips, cost $200 billion, and consume 9 gigawatts of electricity — equivalent to the output of nine nuclear reactors. That’s about as much energy as 7–9 million households would use.
Efficiency improvements aren’t keeping pace with the growth in scale. Performance per watt has improved by 34% annually, but total energy consumption has risen by 100% each year. Researchers emphasize that even with technology optimizations, the overall load on power systems continues to grow.
Currently, the world’s most powerful AI supercomputer is Colossus, built by xAI. It was constructed in 214 days, cost $7 billion, runs on 200,000 AI chips, and consumes 300 megawatts — about as much as 250,000 residential homes.
The AI supercomputing sector is rapidly commercializing. In 2019, the private sector owned only 40% of such computing power, but by 2025, that share has risen to 80%. Over six years, commercial AI systems have grown 2.7 times faster per year than government systems (1.9x). This reflects a paradigm shift: supercomputers have evolved from research tools into industrial assets generating profits.
Major projects confirm this trend: OpenAI is pursuing the $500 billion Stargate initiative, and NVIDIA has a similarly ambitious plan. In total, Epoch AI’s database lists over 500 major AI centers built over the past six years.
The U.S. retains leadership in AI supercomputing capacity, controlling about 75% of all AI supercomputers. China follows with 15%, while traditional leaders like Japan and Germany have fallen behind. Physical location, however, doesn’t always correlate with usage — many data centers are accessed via the cloud.
Yet technological progress comes with significant costs. According to Good Jobs First, at least 10 U.S. states lose over $100 million in tax revenue annually due to data center tax incentives. Moreover, these facilities consume vast amounts of water and land, putting pressure on local ecosystems.
In early 2025, analysts from Cowen reported early signs of a market "cooldown." Major cloud providers, including AWS and Microsoft, began scaling back or pausing some projects. It remains unclear whether this represents a strategic pause or the beginning of an industry slowdown. But one thing is certain: the AI era is also the era of energy-hungry machines.
