Artificial intelligence (AI) has untapped potential. It can break boundaries in health, productivity, education and in areas far beyond our own comprehension. Unlike humans, AI can read and analyze vast amounts of data quickly and efficiently. AI is more than generated cat videos on YouTube. The same AI tools can make our planet more resource efficient, give us better medicines and replace or streamline millions of boring jobs. But if we are to unlock AI’s potential, enormous amounts of energy are required. In a new research article, we have mapped the future power needs at KI, and we highlight the importance of stable access to power. Data centers are the cornerstone of the digital society. They currently use approximately 2 percent of the world’s power. But the forecasts indicate that this is only the beginning. In order to scale up and enable AI models to solve more and more complicated tasks, the development of data centers must accelerate. While there is currently a large gap between the potential of AI and what we are able to achieve with AI, there is a correspondingly large gap to cover the power needs of AI. Without enough power, there will be no further development in AI. An AI-supported Google search will increase the power required per search by almost thirty times. The previous limitation for AI has been infrastructure. Now we see that it is electric power that is becoming the big bottleneck. Next year, the AI giant NVIDIA’s servers alone will increase the power requirement by 20 terawatt hours (TWh) annually. Already in 2027, we can see an accelerated development more than the supply – if the power supply allows it. This means that in just four years, the annual increase for all artificial intelligence in the world will correspond to twice as much power as Norwegian hydropower produces in one year! NVIDIA’s production alone will be able to demand more power than the global forecasts of the International Energy Agency (IEA) indicate. This could change everything. It has been difficult enough to achieve the climate targets in the hope of no or low increase in total energy use. But no one has planned for such a sharp increase as AI can bring. Still, there is hope that the ongoing improvements in the efficiency of data centers will be able to help. But it can just as well cause a recoil effect: Streamlining gives even more incentive to make use of more computing power and can thus create an even higher power demand. This somewhat counterintuitive phenomenon is called Jevon’s paradox and has been known since the 19th century. In the past, we humans have dealt with final energy needs. For example, there is a limit to how much food we can eat, or how much steel we need for a specific building project. But computing can require an unlimited amount of power. We may always need even better, more and new data calculations. There is no natural final destination for the resource needs when it comes to AI and its need to process, store and transport data. If we are to keep pace in the AI race and utilize the untapped potential, we must ensure that energy production does not become a limiting factor. The question therefore becomes whether we are able to provide the emission-free power that is needed? Published 21.06.2024, at 15.29
ttn-69
An insatiable need for power – Expression
