ArmInfo. Every time we scale AI, we increase energy consumption and worsen environmental impacts. This was stated by Olena Bura, Managing Partner and CEO of the Swiss consulting company Enzentra GmbH (Switzerland), during a panel session on the topic: "Application of modern digital solutions and artificial intelligence technologies for environmental issues within the framework of RES 2026."
She noted that training AI models requires approximately 12 tons of gas emissions, which is equivalent to approximately 15 round-trip transatlantic flights. "And this is just the training phase," the CEO of Enzentra GmbH emphasized.
According to the expert, no company can yet answer the question of how much AI costs and what value is created per unit of emissions. Many organizations have already developed responsible approaches to using AI and are building the corresponding infrastructure. They are also addressing risk management, data protection, and security issues.
However, according to Olena Bura, this list of measures is not exhaustive and does not fully answer the question of what resources and in what quantities are required for AI to operate.
She identified three main sources of global pressure that must be considered. First, Bura pointed to the lack of infrastructure, which is becoming a strategic bottleneck. Related to this is pressure on capital allocation, high infrastructure costs, and reputational risks.
"Clients and investors are already aware of the environmental consequences of AI. Organizations that ignore this factor will inevitably face risks of infrastructure instability, as well as environmental and economic shocks. Long-term sustainability requires recognizing that infinite scaling is impossible and has consequences," she stated. To move from simply responsible use of AI to environmentally responsible use, the expert said, it's necessary to think not only about maximizing model accuracy but also about efficiency. Energy consumption must be considered and optimized.
Furthermore, Bura noted, it's important to evaluate not only the economic return on investment, but also the environmental one. "We need to move from static management to a more 'metabolic' one-flexible and responsive to constant change. It's important to understand the costs of creating and operating AI infrastructure.
Furthermore, scaling shouldn't be blind-it needs to be done intelligently. Not every model improvement requires retraining, so it's important to distinguish between what's truly necessary and what can be abandoned," she emphasized.
Regarding environmental factors, Bura noted that it's worth considering existing 'accelerators'-approaches and practices that help reduce impact. She highlighted several strategies based on examples from large tech companies such as OpenAI, DeepMind, and Meta. Their primary focus is increasing capacity and rapid growth, which is a classic approach. The second model, according to the expert, is optimization. Companies strive to improve the efficiency of existing solutions and reduce losses, taking environmental factors into account, but so far without a major redesign. Bura cited Microsoft and NVIDIA as examples. "Here, the primary focus is on efficiency," she noted.
However, she identified the most sustainable model as AI development, taking environmental constraints into account as an integral part of the architecture.
