The conversation about the limits of AI usually focuses on intelligence — how smart can these systems get, what can they not yet do, where reasoning breaks down. But the more I look at where the genuine constraints lie, the more convinced I am that the binding limit in the near term is not cleverness at all. It is electricity.
This is not a metaphor. Training and running large AI systems consumes staggering amounts of energy, and the demand is growing far faster than the infrastructure to supply it. The computing clusters that power frontier AI are enormous physical facilities that need vast, reliable supplies of power and the means to dissipate the heat they generate. The frontier of AI is increasingly a frontier of energy and real estate, not just of mathematics.
I find this clarifying because it reframes the whole competition. We tend to imagine the AI race as a contest of ideas and talent. Those matter. But ideas and talent are abundant compared to the thing that is becoming truly scarce: the ability to actually power and cool the machines. You can have the best model in the world and still be throttled by your inability to build enough capacity to train and serve it. Increasingly, whoever can marshal the most energy and infrastructure can run the most AI, and that is a very different kind of advantage from having the cleverest researchers.
This explains a lot of behaviour that otherwise looks strange. The intense interest in securing power supply, in locating facilities near abundant energy, in long-term deals for electricity, and even in speculative ideas like putting data centres in space to tap solar power and escape earthly cooling limits — all of it makes sense once you see energy as the real constraint. These are not eccentric side quests. They are the main event, dressed in unglamorous clothing.
There is a sobering implication here that the optimistic AI narrative tends to skip. If AI's growth is gated by energy, then AI's expansion collides directly with everything else that needs energy — and with the environmental cost of generating it. The dream of intelligence becoming effectively unlimited runs into the very physical reality that intelligence, at this scale, has a power bill, and that bill is paid in real generation capacity that has to be built, fuelled, and cooled somewhere.
I think this also changes who the important players are. The story has been dominated by the companies that build models. But the companies and countries that control energy, grid capacity, and the physical buildout may end up holding more leverage than we currently assume. When a resource becomes the bottleneck, power flows to whoever controls it. In an energy-constrained AI world, the energy providers and infrastructure builders are not supporting characters. They are protagonists.
There is a hard question lurking under all of this about priorities. As AI's appetite for power grows, societies will face genuine choices about how to allocate scarce energy, what to build, and what trade-offs are acceptable. Those are not technical questions. They are political and environmental ones, and they will be decided by people who may not have signed up for AI to reshape their energy systems.
So when I hear that the only real limit on AI is how clever we can make it, I am sceptical. The limit I keep running into is far more mundane and far harder to wave away. The question that will shape the next phase is not how smart the machines can become. It is how much power we are willing and able to give them — and what we are prepared to give up to do it.
The above reflects my personal views only and is intended for informational and discussion purposes. It does not represent the position of any employer or organisation.