Rene Haas, CEO of ARM, said that by 2030, artificial intelligence could consume 25% of US electricity. Currently, AI data centers consume 4%. Without a significant increase in efficiency and/or a significant increase in government regulation, the current trend of AI development will not be sustained, he believes.
According to the January report The International Energy Agency, which says that ChatGPT consumes approximately 2.9 W*h of electricity per query, which is 10 times more than a standard Google search. Thus, if Google were to power its search entirely with AI, the company would consume at least 11 TWh of electricity per year, instead of the current 1 TWh.
The announcement comes ahead of an expected US-Japan partnership agreement on artificial intelligence, and against the backdrop of recent developments such as OpenAI’s Sora, whose current version is estimated to Factorial Fundsrequires at least one NVIDIA H100 GPU per hour to generate five minutes of video. A single 700W NVIDIA H100 consumes approximately 3,740 kWh per year.
The U.S. Energy Information Administration (EIA) said that in 2022, the United States produced 4240 TWh, with only 22% of that coming from renewable sources. Consumption amounted to 3900 TWh.
Sustainable development must also take into account the likely growth in demand from other industries and the scale of consumption of renewable and non-renewable resources. Given that the cost of electricity is almost doubled since 1990, more regulation seems justified.
Rene Haas expressed hope that an international partnership between Japan and the United States would help reduce these extremely high capacity estimates. However, corporate greed and demand for computing power are also international phenomena.
Source: Tom’s Hardware