The more energy-efficient the AI, the more kilowatts it consumes.

Artificial intelligence infrastructure continues to require electricity at unprecedented levels, and new organizational solutions and technologies are seen as the solution, according to S&P Global. However, in absolute terms, data center electricity requirements are increasing year over year.

The implementation of more efficient machine learning approaches, along with advanced AI optimization techniques, has dramatically reduced the energy consumption of each model output. For example, Google claims that in May 2025, Gemini models consumed 33 times less energy per text query than the previous year.



Artificial intelligence models have become significantly more cost-effective. Research teams have found that the computing resources required to achieve a given level of performance during the training phase have halved every eight months since 2012.

Other experts also note that the efficiency of AI infrastructure is increasing exponentially. For example, semiconductor manufacturer Applied Materials, cited by S&P, has revised its long-term performance-per-watt target to orders of magnitude—from 1000 to 10,000 times by 2040—thanks to advances in materials science.

Historically, GPU generations have been released on a two-year cycle, but starting in 2024, NVIDIA and AMD, under pressure from unprecedented demand for AI infrastructure, have switched to a one-year refresh cycle for their flagship architectures.
NVIDIA claims its new "Vera Rubin" platform, which began testing with customers as early as February 2026, will deliver 10 times better performance per watt than the current-generation platform.

However, despite all these advances, overall power consumption continues to increase with each new generation of GPUs. As performance per watt increases, so does the absolute power consumption required to meet the extreme demands of large-scale AI models.