Energy hunger of AI data centers: Are the power grids stable enough?
Energy hunger of generative AI is widely known; in addition to the expansion of power plants, electricity grids must also be able to cope with the higher load.
Electricity pylons in Bremen.
(Image: heise online / anw)
According to conventional wisdom in business and politics, the general use of generative AI with transformer-based large language models requires much more computing power and therefore much more electricity to operate the corresponding GPUs. It is still unclear where the electricity will come from, but the model providers Microsoft, AWS, xAI and Google are toying with the idea of expanding nuclear power. In addition to cheap electricity, the tech companies are hoping for a low carbon footprint for the greenwashing of their business ventures.
A report by the New York Times has now shown that data centers will account for four percent of US electricity demand in 2023. This is expected to rise to 12 percent in the next three years due to the increase in power required. This is partly due to the AI boom of the US tech giants. If the model providers now begin to commission their own power plants to meet their increased electricity requirements, this will have an impact on the electricity grid in the USA. Renovation work and protective measures against extreme weather have led to a 30 percent increase in the price of electricity since 2008. 2020. The newspaper assumes that the expansion of the power grid for tech providers' AI data centers will be passed on to smaller companies and consumers if the US government does not force tech companies to pay. In states such as Virginia, where many data centers are located, the price of electricity could rise by a further 25 percent.
Videos by heise
Energy issue also relevant for Europe
Parallel developments to the USA can also be observed in Europe, where 76 consortia are currently vying for EU funding to build five AI data centers known as gigafactories. As these data centers contain tens of thousands of AI accelerators, the energy requirements must be covered, but the European power grid must also be able to cope with the additional load. In Europe, too, the expansion of the grid could be passed on to taxpayers through subsidies.
(pst)