AI Models Cut Power Use as Scale Explodes

A February 2026 study by TRG Datacenters reveals a decisive shift in the AI race: efficiency is becoming as important as raw capability. As demand for generative AI accelerates, leading models are significantly reducing electricity consumption per query—an essential move as power grids strain under surging data center loads.

The standout performer is Grok 4.1, which reduced energy use per query by 38%, from 0.55 to 0.34 watt-hours, while handling 134 million daily requests. Gemini 3 follows with a 35% improvement, powering 850 million daily queries at the lowest cost per request in the study. Claude Opus 4.5, DeepSeek-V3.2, and GPT-5.2 also posted meaningful efficiency gains ranging from 19% to 27%.

While GPT-5.2 shows the smallest percentage drop, its impact is amplified by scale—processing 2.5 billion daily queries. A 19% efficiency improvement at that volume translates into massive energy savings, underscoring how incremental gains at hyperscale can rival larger percentage cuts at smaller volumes.

The deeper implication is structural. AI demand is projected to quadruple by 2030, and infrastructure expansion alone cannot meet that trajectory. Efficiency-first model design—through architectural optimization, hardware-software co-design, and smarter inference techniques—is becoming critical.

The findings suggest a maturing industry mindset: sustainable AI will not be built solely on larger models, but on smarter, leaner ones capable of delivering performance without proportionally escalating energy costs.