A 2026 Microsoft Research study found that popular estimates of AI energy consumption are dramatically inflated, with a typical chatbot query actually using about 0.31 watt-hours, not the tens of watt-hours frequently cited in media reports. The research, published in the journal Joule, reveals that the gap between published figures and real-world deployments comes down to what each measurement actually includes, not disagreement about physics.

Why Do Energy Estimates Vary So Wildly?

When researchers and companies report energy per AI inference task, they are measuring different parts of the system. Some measure only the graphics processing unit (GPU) chip itself, while others include the entire server node or even the full data center with cooling overhead. This fundamental difference in system boundaries explains why published estimates span more than two orders of magnitude, from roughly 0.1 watt-hours to tens of watt-hours for a single query.

The Microsoft Research analysis found that widely cited public figures are overstated by four to twenty times relative to production-optimized deployments. Independent academic teams at Carnegie Mellon and the University of Michigan have separately demonstrated that even the choice of measurement instrument, batch size, or output length can swing a reported energy figure by several-fold for the exact same model and GPU.

Google has published two different figures for a median Gemini prompt: a comprehensive estimate of 0.24 watt-hours and a narrower, chip-only figure of 0.10 watt-hours. OpenAI discloses an estimate of 0.34 watt-hours per average ChatGPT query, while an academic API-based estimate put a short GPT-4o prompt at 0.42 watt-hours. These threefold differences reflect different measurement boundaries, not errors in physics.

What Factors Most Dramatically Change Energy Per Query?

Eight measurable variables each independently swing per-task energy by 25 percent or more. Understanding these factors helps explain why energy claims require precise specifications to be meaningful.

• Model Size and Architecture: Larger models with more parameters consume more energy per inference task, though efficiency improvements in newer architectures can partially offset this.
• Numeric Precision: Quantization, a technique that reduces the precision of numbers used in calculations, is one of the most effective levers for reducing energy consumption per task.
• Batch Size: Processing multiple queries together in larger batches improves energy efficiency compared to handling requests individually.
• Context Length: The amount of text or tokens the model must process as input affects total energy consumption per query.
• Output Length: Reasoning models that generate far more output tokens per response are the largest driver of higher energy per task, with reasoning-style responses using roughly thirteen times more energy than typical chatbot exchanges.
• GPU Generation: Newer GPU generations like Nvidia's Blackwell B200 are significantly more efficient, achieving median energy reductions relative to the H100 at matched latency.
• Parallelism Strategy: How the computation is distributed across multiple processors affects overall energy efficiency.

The Microsoft Research study found that reasoning models generating far more output tokens per response are the largest driver of higher energy per task, while quantization, larger batch sizes, and newer GPU generations are the most effective reduction levers.

How to Evaluate Energy Claims About AI Systems

• Check the Hardware Specification: Verify which GPU or accelerator is being measured. Nvidia specifications show thermal design power ranging from 300 to 400 watts on older A100 and H100 PCIe generations to up to 1,000 watts per GPU on the newer Blackwell B200.
• Identify the System Boundary: Determine whether the measurement includes only the GPU chip, the full server node, or the entire data center including cooling overhead. Facility overhead varies significantly; Google discloses a global fleet power usage effectiveness (PUE) of 1.09, while Microsoft reports 1.16 to 1.17, both well below the broader industry average of 1.54.
• Understand the Task Definition: Confirm what type of query is being measured, whether it is a short chatbot exchange or a longer reasoning-style response that generates substantially more output tokens.
• Review the Measurement Instrument: Assess whether the measurement uses Nvidia's Management Library (NVML) driver-level interface, cumulative energy counters, or wall-outlet methodology, as each approach can produce different results for the same workload.
• Account for Idle Power: Verify whether the measurement subtracts idle power draw, which occurs even when the GPU is not processing requests, to avoid over-attributing energy to individual tasks.

What Does This Mean for Global Data Center Energy Use?

The International Energy Agency estimates that data centers consumed 415 terawatt-hours (TWh) of electricity in 2024, representing about 1.5 percent of world electricity use. This figure is projected to more than double to around 945 terawatt-hours by 2030 as AI-optimized servers grow electricity draw roughly 30 percent annually.

Despite the scale of this growth, standardized and audited disclosure of energy per inference remains immature. MLPerf Power's wall-outlet methodology and Hugging Face's fixed-batch AI Energy Score are the most rigorous public benchmarking efforts to date, but neither yet publishes routine, segment-broken-out joules-per-task figures comparable across vendors.

The key takeaway is straightforward: any energy-per-inference claim is only as trustworthy as its stated hardware configuration, measurement instrument, system boundary, and task definition. Without these specifics, published figures can mislead by orders of magnitude, making it essential for researchers, regulators, and the public to demand transparency about how energy claims are actually measured.