agentic

For a time, one of the controversies surrounding the consumption of artificial intelligence had to do with water. Some time later we learned that the calculations (even consuming a lot) They were not accurate and we begin to look at something so worrying: the tremendous amount of energy What data centers, and AI, need to function. It is something that is endangering the energy integrity of some countriesbut as AI approaches adolescence and agentic stagewe will enter a new phase. That of obscenely multiplied consumption. In short. The Korea Advanced Institute of Science and Technology, or KAIST, has conducted a study in which he has quantified the energy cost of AI agents. Unlike a chatbot, which is a system to which we make a request, it gives us a result and that’s it, an agent is a chain of operation in which the software performs different actions autonomously. That, obviously, implies that the hardware that is moving that software spends more time doing things In the study, they measured the energy consumption of both chatbots and agents and concluded that, using a large-scale language model comparable to current commercial AI services, a single complex request to an agent consumed 348.41 Wh of electricity. depending The agent and model you use will also consume more or less. For example, a framewoprk called LATS used 62.1 times more energy in tests compared to an AI chatbot, while one on Meta’s Llama-3.1 Instruct 70B model consumed a peak of 136.5 times more per query. GPUs waiting (and consuming). There is another key, and it is time. Those queries consume more energy because more resources come into the picture and have estimated that agents take up to 153.7 times longer than conventional chatbots to process responses. During that time, GPUs remain idle more than half the time, but still consume electricity. They are not at full capacity, but they are “on guard” waiting for the response from external tools and websites so, when that response arrives, they start processing the data and performing the corresponding action. This issue of stopped GPUs is not new and a few weeks ago it was noted that the vast majority of the equipment that the hyperscalers had purchased They were doing nothing most of the time. The electrical network. And the problem is projection. Currently, this technology is in the era of chatbots, but the industry is moving towards agents. Right now platforms like Nvidia’s Vera Rubin just for that, and when they come into play, the study’s projection is that energy demand will reach an equivalent to half of the electricity consumption of the entire United States. It is estimated that in 2023, US data centers “barely” consumed 4.4% of the national total and will double by 2030, but KAIST’s estimates far exceed previous forecasts. Redesign. While those data centers continue to grow, the power grid cannot say the same. Renewables are not enough to satisfy the voracity of AI and we must resort to nuclear, gas and even coal and, in Europe, already we are seeing reactions of some countries that either move new data centers away from their large cities or reject them outright. The reason? Saturated networks and data centers that would have a demand greater than that of the population itself, generating problems in the energy infrastructure. It doesn’t look like the hyperscalers are addressing these issues because, as Jensen Huang himself, CEO of Nvidia, pointed out some time ago, it remains more than five years of wild investment in infrastructure for AIbut the South Korean study commented that a solution to energy demand would come from a redesign of the entire network. From microchips to AI models and the electrical infrastructure of data centers themselves. As if that were simple, and even more so now that Big Tech is in the race to deploy AI agents in both business and consumer applications. We’ll see if the power grid can keep up with that pace. In Xataka | Talking about artificial intelligence is talking about energy, and the fashionable term is ‘bragawatts’