As AI development and usage surge, so does the energy demand associated with it. According to a recent study, AI can use up to 82 terawatt hours<\/a> of power in 2025\u2013equivalent to Switzerland\u2019s yearly energy consumption. \u201cAi\u2019s energy demand won\u2019t go down, but we can work on slowing such an increase,\u201d adds Burr. He will be one of the speakers at the upcoming Watt Matters in AI<\/a> conference, happening in Eindhoven on November 26.\u00a0<\/p>\n Watt Matters in AI is a conference that aims to explore the potential of AI with significantly improved energy efficiency. In the run-up to the conference, IO+ publishes a series of articles that describe the current situation and potential solutions. Tickets to the conference can be found at wattmattersinai.eu.<\/p>\n View Watt Matters in AI<\/a>3D optical computing: minimizing AI energy usage<\/p>\n Graphics processing units (GPUs) are the hardware cornerstone of AI development, enabling complex computations. The American NVIDIA is currently the GPU market leader\u2013 and the world\u2019s most valuable company at the moment. At the heart of AI\u2019s large language models (LLMs) processing lies a mathematical operation that is being calculated repeatedly, called Matrix-vector multiplication (MVM). Essentially, the data streams are multiplied many times.\u00a0<\/p>\n In 3D optical computing, this data is encoded in a light beam that passes through a lens, which effectively copies data across the matrix. \u201cAs a result, you can perform many operations in the light domain, where an insignificant amount of energy is used,\u201d explains Lumai\u2019s head of product.\u00a0<\/p>\n Companies working with silicon technology are adding more silicon to get faster, more powerful AI solutions, yet needing more power to operate. For example, an NVIDIA H200 GPU can consume up to 700 watts of power. OpenAI\u2019s first European datacenter, Stargate Norway<\/a>, will have 100,000 operating GPUs.\u00a0<\/p>\n Burr: \u201cTherefore, the costs for building a datacenter ramp up: the more power you use, the bigger the power supplies and the cooling equipment you need to operate it. Furthermore, performance can\u2019t be simply increased more and more, because then you hit thermal issues. Getting heat out of chips is challenging, and designing chips that perform good thermal management can get expensive.\u201d<\/p>\n P<\/p>\n Philipp Burr<\/p>\n Head of product at Lumai<\/p>\n After a long career in the semiconductor industry, he joined Lumai fascinated by their innovative approach towards AI computing. <\/p>\n Targeting AI inference\u00a0<\/p>\n The British startup, employing 20 people, targets AI inference in datacenters. When developing a large language model (LLM), there is an initial training phase, when the model is fed with information. Then comes the inference phase, where the acquired knowledge is being tested, with the model making decisions and predictions on new, unseen data.\u00a0<\/p>\n According to MIT Technology Review journalists James O\u2019Donnell and Casey Crownhart, the bulk of the AI\u2019s energy consumption happens in inference<\/a>, rather than in initial training. Researchers revealed that up to 90% of the electricity consumed by AI models powers this second phase of the AI model creation.\u00a0<\/p>\nWatt Matters in AI<\/p>\n
![\"Philipp](\"https:\/\/www.newsbeep.com\/au\/wp-content\/uploads\/2025\/08\/Phil_Burr.jpg\")
<\/p>\n