AI power forecasts need better data: Training servers average far below rated maximum

By: Margaret Harding McGill

Date: December 8, 2025

While utilities scramble to build new generation capacity for the AI boom, new Carnegie Mellon University research reveals that even the most intensive AI training workloads draw only three-quarters of their hardware’s advertised maximum power. This discrepancy shows the need for greater transparency and precision in estimating AI’s energy demands.

Why it matters: Accurate measurement of AI energy use can prevent billions in wasted capital and infrastructure overbuilding. Without accurate data, grid operators cannot manage the unique, rapid-fire power fluctuations of AI workloads that threaten local grid stability. Right-sizing grid estimates will accelerate American AI leadership while protecting ratepayers from funding an AI power plant bubble.

Catch up quick: Current energy forecasts for AI data centers often rely on the manufacturer-rated thermal design power (TDP), which is the maximum theoretical power a server can draw before overheating. Utilities and policymakers use these numbers to plan capacity, but servers frequently idle while waiting for data from memory or other processors, meaning they rarely hit that theoretical ceiling.

Action: Researchers at Carnegie Mellon University, including Alex C. Newkirk, Emma Strubell, and Costa Samaras, set out to discover how much power the industry-standard NVIDIA H100 chips actually consume when training massive AI models. They measured second-by-second power use across varying workloads, from image classification to Large Language Models (LLMs), and validated their models against open-source data from widely used training benchmarks.

Findings: Actual power demand during training consistently averaged at just 76% of the manufacturer’s rated power (TDP). This means these models use less energy than what utilities are planning for.

  • Standard methods that estimate energy use based on TDP were off by 27% to 37%. Our new, empirically calibrated model cut that error down to 11.4%.
  • AI model training draws power differently depending on model architecture. For example, transformer models (like those behind ChatGPT) exhibit a jagged power profile with rapid, high-frequency fluctuations, whereas image-processing models have a flat, steady demand.

Policy takeaways: AI training facilities use less power than their equipment ratings suggest, but inconsistent industry terminology makes it hard to know exactly how much. Clearer definitions and better data sharing would help utilities make smarter decisions, and would likely get AI data centers online faster.

  • Standard definitions for AI power consumption that rely on measured operational data, not just manufacturer nameplate ratings, can help prevent inflating demand forecasts.
  • If data centers share granular, anonymized power data, utilities can approve connections faster, knowing the actual load is significantly lower than the theoretical max.
  • The rapid power swings of LLM training can harm local power quality. Policymakers should support grid modernization efforts that deploy fast-response technologies (like battery storage or solid-state transformers) to smooth out these fluctuations.

Faculty experts