AI Raises Energy Demand, Reshapes Climate Impact

The piece reframes the AI-climate debate from whether a single ChatGPT call matters to how aggregate AI demand grows and what powers it. Global data centers consumed 415 TWh in 2024, and AI-specific servers used roughly 93 TWh in 2025, or about 0.3% of global electricity. A short, single-turn query is roughly 0.3 Wh, but reasoning models and agentic workflows raise that by 10-100x, with o3 at 33 Wh, GPT-4.5 at 30 Wh, and Claude 3.7 Sonnet at 17 Wh in measured benchmarks.

The shift driving higher per-query cost is movement from lightweight inference to sustained, multi-step reasoning and agent orchestration. Key technical points practitioners should note:

• •o3, GPT-4.5, Claude 3.7 Sonnet show order-of-magnitude higher energy per query than short-turn models.
• •Agentic workflows multiply inference steps, I/O, and state management overhead, raising energy and tail-latency concerns.
• •Aggregate impact depends on user behavior, model defaults, and how providers meter and route workloads.

Comparing sectors clarifies scale: residential air conditioning uses more than six times AI's footprint today, industrial motors about forty times, and global streaming sits around 100-120 TWh. The crucial contrast is growth rate. If reasoning models and agents become default for broad classes of apps, AI could shift from a niche consumer to a major grid actor. That influences datacenter sizing, regional grid stress, procurement of renewables, and corporate reporting on Scope 2 emissions.

Track per-query energy trends, deployment defaults (reasoning vs. short-turn), caching and routing optimizations, and hardware efficiency gains like quantization, sparsity, and accelerator improvements. Policy and corporate procurement (time-of-use, renewable contracts) will shape how that demand maps to emissions.