Three Big Unknowns About AI's Hidden Energy Appetite

The elusive per‑response energy number

Earlier this year the search for a single key figure became a focus for researchers: how much energy does a leading AI model use to generate a single response. That number has been notoriously hard to pin down because only the companies running the models have access to the most accurate measurements. Reporters and researchers compared the task to estimating a car’s fuel efficiency without ever driving the car, relying on indirect clues and sparse data.

In recent months some firms have released rough figures. OpenAI’s CEO Sam Altman reported an average ChatGPT query consumes 0.34 watt‑hours, and Google said a Gemini answer uses about 0.24 watt‑hours. Those values roughly match earlier independent estimates for medium‑sized models, but they come with many caveats.

What the published numbers leave out

The numbers shared so far are limited and often vague. OpenAI’s figure appeared in a blog post rather than a detailed technical breakdown, leaving questions about which model was measured, how the measurement was taken, and how much results vary by query type. Google’s figure describes a median energy per query, which understates the tail of costly interactions when the model engages in heavy reasoning or produces very long outputs.

Crucially, these figures apply mostly to chat interactions. They do not capture energy use from other rapidly growing modalities like images, video, or large document processing, where inference and generation patterns can be very different. As Hugging Face’s AI and climate lead Sasha Luccioni notes, we need comparable numbers across modalities to understand the full picture.

Efficiency claims versus reality

Tech companies argue that AI will eventually deliver climate benefits by enabling efficiency gains in other sectors, for instance designing better HVAC systems or speeding materials discovery. But concrete, verifiable evidence that such benefits currently outweigh the emissions from the AI boom is scarce. Anecdotes exist, such as detecting methane leaks, yet transparency about the scale and net effect of these wins is limited.

Meanwhile, the number of AI data centers and the electricity they consume continue to grow. Big tech has reported rising emissions tied to AI investments, and even firms with ambitious carbon goals acknowledge that meeting them will be a long and difficult process.

The biggest unknown: future demand

Perhaps the most consequential unknown is not the energy per query but whether demand will scale to the levels companies are preparing for. OpenAI reports billions of daily prompts, and research forecasts suggest AI could consume as much electricity annually as a large share of households in the coming years if usage continues to surge.

But this projection depends on sustained uptake and monetization. Recent signs of slower momentum — from mixed reactions to new model launches to businesses reporting weak returns on AI investments — raise the possibility that demand may not materialize at the projected scale. If demand plateaus or collapses under unmet expectations, the current wave of data center buildout could turn out to be a short‑lived spike rather than a long‑term structural shift in energy systems.

What researchers want next

Researchers say the path forward requires more granular, application‑level data from companies, and better reporting across modalities and model behaviors. Only with that information can analysts move beyond back‑of‑envelope estimates and assess the true net climate impact of AI across deployments and over time. Until then, the three big unknowns — per‑interaction variability, realized efficiency gains, and long‑term demand — will keep shaping the debate on AI’s energy burden.