AI power consumption is growing at a rapid pace. To put this into perspective, let’s break it down and find out what’s really going on.
Power Consumption – USA
Currently the US is the undisputed hub for AI, hosting roughly 45% of the worlds data-center capacity. To start us off, let’s look at the breakdown on power being used by data-centers versus how much AI processing is of that whole.
- Total data-center Impact – Overall data-center power impact to the US grid at large is roughly 4 to 5%.
- AI specific workloads – AI specific workloads account for roughly .6 to 1.1% percent of the total data-center power usage. This is based on an estimated 15 to 25% of overall data-center power being used specifically for AI.
While the numbers above are not staggering, experts believe total data-center power usage will rise to 9 to 17% of the entire US grid by 2028-2030.
The EPRI (Electric Power Research Institute) has published their annual report. EPRI Executive Summary Report 2026.
Power Consumption – Globally
Globally the percentages are smaller, but the growth rate is steeper. This is due, in large part, to the size of the data-centers. Countries around the world don’t have the massive server infrastructure as the US currently has.
- Total data-center Impact – Worldwide data centers use about 1.5% of the total electricity.
- AI specific workloads – AI workloads account for roughly .2 to .4% of the total data-center power consumption.
The global usage is expected to double by 2030, which approaches 3% of the global demand. The AI specific portion is expected to triple within the same time period. The Brookings EDU published their findings titled, Global Energy Demands Within the AI Regulatory Landscape.
Approaches Being Taken Today
- Demand Reduction – We already covered immersion cooling with my previous article titled, AI Drinking the World Dry? This technology will help to reduce the overall power requirement by up to 40%. For further reading, check out Vertiv, Immersion Cooling Systems.
- Supply Orchestration – Predictive grid control, or smart grid systems do not reduce the power requirements. This technology performs real time monitoring, intelligent scheduling and optimal power management. For more information, check out Colocation America’s article titled, What Is a Smart Grid Data Center?
There is another approach which I should point out for completeness, but I’m not a fan of. That would be SMRs or Small Modular Reactors. This technologies has a fair amount of push back given it results in nuclear waste. I don’t see implementing a technology that removes or reduces one environmental impact only to pose another. However, for those interested, you can read more from the World Nuclear Association in their article titled, Small Modular Reactors.
Future Direction
The future direction is to reduce demand. This can be accomplished through the efforts already being made within the cooling technologies mentioned in my article titled, AI Drinking the World Dry. Power and water go hand in hand with reducing the overall environmental impact. The real holy grail is circular AI.
Circular AI
Let’s start off with a definition to clarify this approach. At a high level, Circular AI is the application of circular economy principles—reduce, reuse, recycle, and regenerate—to the entire lifecycle of artificial intelligence. Instead of a linear “take-make-waste” model, Circular AI focuses on making AI hardware and operations self-sustaining and closed loop.
To achieve above, we’ll need make advancements in the following areas.
| Hardware Lifecycle Redesign | Server components and hyper-dense accelerators will be designed from day one to be easily disassembled, upgraded, and modularly repaired, rather than being completely shredded and replaced every three to five years. |
| Algorithmic Recyclability | Instead of throwing thousands of megawatt-hours of electricity at training entirely new foundation models from absolute scratch, future AI architectures will “recycle” neural pathways, modularly adapting existing trained models to save immense amounts of computational energy. |
| Thermal Energy Circularity | The heat rejected by immersion-cooled servers won’t just be dumped into the atmosphere. True circular design will mandate that data centers serve as district heating hubs, channeling hot water directly into municipal grids, vertical farms, or manufacturing plants. |
Summary
The technology industry is currently reducing and optimizing its power requirements. This work also goes hand in hand with the reduction of its freshwater needs for cooling. Optimizing one aids in optimizing the other. The end goal, however, remains circular AI. A closed-looped data-center. Both freshwater and energy are considered a paradox. Requirements for both are increasing as AI becomes more popular. The paradox comes in where AI itself is assisting us in the management of these resources. This paradox actually helps peoples to look at the environment in a more sustainable fashion. This is an excellent example of using AI ethically, through making our resources more sustainable.
