AI Energy
August 6, 2025
Exploring the energy requirements of AI, reviewing growth predictions in light of current supply and known supply coming online. A review of Great Power energy policies and implications for their AI industry.
“Most investors spend the bulk of their time trying to forecast future demand for the companies they follow. The aviation analyst will try to answer the question: How many long-haul flights will be take globally in 2020? … Long-range demand projections are likely to result in large forecasting errors. … Supply prospects are far less uncertain than demand, and thus easier to forecast. In fact, increases in an industry’s aggregate supply are often well flagged and come with varying lags.”
— Edward Chancellor
Artificial intelligence is the most popular economic topic of the past few years, trumping tariffs, and world war three talk. Nothing has your everyday Joe, college student, and Wall St. professional more excited about the future than large language model (LLM) chat bots like ChatGPT and Grok. The ever optimistic techbro is preaching AI as a force for good. The Economist is predicting exponential growth. Many of those working on AI claim that Artificial General Intelligence (AGI) is inevitable and critical to national security, so we should rush to get there first.
Not only do these predictions feel biased, but they feel intellectually dishonest and/ or ignorant of roadblocks like energy capacity and plateauing LLM progress. Most obvious is the shortage of available and planned energy supply in many countries that seek to run more and more computationally expensive AI models.
Energy Needs
Processing billion of data points, training a model, and keeping it running requires energy. A whole lot of it. Current AI progress is dependent on juicing models with data and energy to create more intelligent AI, but it appears this is leading to diminishing returns. There’s a lot written on this by people smarter than me, I think Alexander Campbell’s explainer is a good start. The chart that captures the essence best:
While the chart is somewhat technical, it makes clear that exponential increases in computing power (x-axis) lead to marginal gains in model intelligence (y-axis). While the most visible LLMs like ChatGPT may be getting better slower, it seems likely new methods, models, and techniques will lead to advancement in the field.
Whether new AI models emerge or existing ones are scaled, more energy will be a requirement. This is where reality meets expectations. In the United States over the past 20 years energy supply growth has been non-existent (Chart 2). European supply has declined over the same timeframe (Chart 3). China on the other hand has steadily increased energy supply and shows no sign of slowing down (Chart 4).
Where is the energy going to come from for a large Western AI push? How will AI researchers in the US keep up as China continues to have more and more resources available to devote to the field?
Current AI predictions and hopes depend on large amounts of available energy, which the US and Europe do not have much of. Not only is there a lack of existing capacity, future planned and permitted capacity is marginal.
It seems impossible the US gets anything close to AGI without drastically changing our energy policies. If policies don’t change, AI predictions and expectations seem likely to fall short. Demand for AI means nothing is there isn’t enough energy supply.
Leave a Reply
You must be logged in to post a comment.