Translated title. Original: Рост энергоэффективности ИИ, подчеркнутый генеральным директором OpenAI Сэмом Альтманом
Sam Altman asserts AI energy efficiency has significantly improved, claiming per-query energy use is vastly lower than a few years ago, while total demand grows due to higher usage. He outlines a multi-year ADNOC partnership to deploy AI across operations and argues that energy availability remains a key constraint for AI advancement.
Quartz reports that AI requires more energy and at a faster rate than the United States has ever produced, highlighting energy supply constraints related to AI growth.
Quartz reports that AI’s energy demand is rising faster than the United States has ever produced energy.
The article explains three ways to attribute AI energy use—top-down, bottom-up, and hybrid—and the trade-offs of each. It reviews tools and reporting options for cloud and owned hardware, including cloud APIs, CodeCarbon, Carbontracker, and the Hugging Face AI Energy Score.
The ML.ENERGY Initiative seeks to measure, understand, and reduce the energy use of modern machine learning, and to expose these metrics to users; it began at SymbioticLab with participation from multiple academic institutions.
Translated title. Original: Les agents d'IA de Cathedral ont pénétré dans Anthropic en moins de 12 heures.
Quartz reports that AI needs more energy, and more quickly, than the United States has ever produced.
Translated title. Original: Le chiffre d'affaires et la perte d'exploitation du troisième trimestre d'OpenAI dépassent les 10 milliards de dollars.
The text says AI’s energy needs are rising faster than the United States has ever produced energy.
Translated title. Original: OpenAI déploie une Control Room pour arrêter ses modèles d'IA
The text says AI needs more energy, and demand is rising faster than the United States has ever produced it.
Translated title. Original: AI będzie wszędzie, lecz nie musi zużywać proporcjonalnie więcej energii | ISBtech.pl
AI is a key driver of economic transformation, but growth increases energy use. AMD predicts large gains in AI infrastructure efficiency by 2030, potentially reducing energy use and CO2 emissions per AI workload and enabling similar compute with far fewer physical systems. The article notes this is crucial for Poland's digital and energy transition and stresses that progress depends on both advanced models and hardware-software efficiency.
The paper analyzes how input and output token lengths affect compute time and energy use in large language models, showing linear scaling for decode and mixed linear/quadratic scaling for prefill. It proposes a mapping framework to estimate energy from compute time to inform sustainable deployment in agentic AI.