AI's energy consumption has become a topic where almost no one is calibrated. Alarming headlines compare individual queries to running a household for hours. Industry defenders dismiss the entire concern as overhyped. The truth, as always, is more interesting and more nuanced. Here is what the actual numbers say in mid-2026, and what they mean for the industry's future.
The per-query reality
A single chat-style query to a modern frontier model consumes roughly the same energy as a few seconds of streaming high-definition video — measurable, but not enormous. Image generation is more energy-intensive, on the order of running a desktop computer for a minute or two. Video generation is more expensive again. None of these numbers, on their own, justify the more dramatic headlines.
The aggregate picture is different. AI workloads are growing fast enough that even modest per-query costs add up to significant total demand. Major hyperscalers now report that AI workloads are the fastest-growing component of their power consumption, with some data centers projecting a doubling of demand over the next three years.
Training versus inference
Training a frontier model is genuinely expensive — the latest models consume electricity on the order of a small town's annual usage during their training run. That cost, however, is amortized across billions of subsequent queries. For most providers, inference now dwarfs training as a fraction of total energy consumption, and the gap is widening as model usage grows.
This is why inference efficiency has become the central engineering focus for every major lab. A ten-percent improvement in tokens-per-watt is worth more than almost any new feature, because it compounds across every query a model serves.
Where the power is coming from
The most consequential trend is the data center industry's aggressive shift toward clean and firm power sources. Long-term power purchase agreements for renewable energy, direct investment in nuclear restarts and new builds, geothermal pilots, and a wave of grid-scale battery deployments are all directly motivated by AI demand. In several regions, AI demand has been the catalyst that finally moved long-stalled clean energy projects forward.
The countervailing trend is the opportunistic use of existing fossil capacity, including a worrying number of coal plant life extensions in regions where clean buildout cannot keep pace. The net picture varies dramatically by region, and any honest assessment has to acknowledge both sides.
The efficiency curve
Per-token energy cost has fallen by roughly an order of magnitude over the past three years, driven by better hardware, better algorithms, and better software optimization. This curve shows no sign of plateauing. The most efficient inference setups in 2026 deliver capabilities at a fraction of the energy cost of equivalent capabilities in 2023.
Whether efficiency improvements outpace demand growth — Jevons paradox in action — is the central open question. So far, demand has grown faster than efficiency, which is why total consumption is rising. Whether that continues depends on factors no one can fully predict.
What builders can do
If you're shipping AI features, you can make meaningful efficiency choices. Use the smallest model that meets your quality bar. Cache aggressively. Batch when latency allows. Route easy queries to cheaper models. Avoid generating content speculatively when you don't have evidence users will see it. None of these are dramatic individually; together, they can cut a product's AI energy footprint by half or more.
Choosing providers and regions with cleaner grids is the other significant lever. Most major providers now publish carbon intensity by region; routing your workload to a low-carbon region is often a one-line change with real impact.
The longer view
AI's energy consumption is real, it is growing, and it deserves serious attention — but it is also tractable. Efficiency improvements, clean energy investment, and thoughtful product design can keep the trajectory manageable. The worst case is not inevitable. Neither is the best case. Which one we end up with depends on choices the industry is making right now, and the people making those choices should be honest about both the numbers and the stakes.