AI's Next Bottleneck Is Energy, Not Chips
Even as Nvidia secures its chip supply, the next major constraint for scaling AI is emerging: energy. Analysts note that data centers could soon strain the U.S. electrical grid, potentially requiring new power plants or nuclear solutions to meet the immense power demands of future AI infrastructure.
The International Energy Agency projects that by 2030, data center electricity consumption will more than double to approximately 945 TWh. This is slightly more than the current total electricity consumption of Japan. The primary driver for this surge is the rapid growth of AI, alongside increasing demand for other digital services. Training a large language model is an energy-intensive process. For instance, training GPT-3 required an estimated 1,287 MWh of electricity, which is comparable to the energy consumed by an average American household over 120 years. The computational power needed to train the most advanced AI models has been doubling every 3.4 months on average. Beyond electricity, the water consumption of data centers is a significant environmental concern. A medium-sized data center can use up to 110 million gallons of water annually for cooling, equivalent to the water usage of about 1,000 households. Globally, data centers are estimated to consume around 560 billion liters of water each year. To meet these escalating energy demands, some tech giants are exploring nuclear power. In October 2023, Microsoft posted a job for a "Principal Program Manager, Nuclear Technology" to lead the integration of Small Modular Reactors (SMRs) and microreactors to power its cloud and AI infrastructure. SMRs offer the potential for constant, carbon-free energy in a compact format, ideal for the localized grids of data centers. The material requirements for AI infrastructure also have an environmental impact. The manufacturing of a single 2 kg computer requires 800 kg of raw materials, and the microchips essential for AI depend on rare-earth elements that are often mined in environmentally damaging ways. Furthermore, the boom in AI is projected to nearly double the demand for copper, a metal already facing global shortages.
