By Peter Zanatta
The AI industry has developed a curious habit. Whenever power becomes scarce or expensive, the response is not to use less of it, but to go hunting for more.
New data centres are announced. Grid upgrades are demanded. Power purchase agreements are paraded like trophies. Even nuclear reactors are now spoken about as if they were just another line item in a procurement plan. This is all presented as bold thinking.
It isn’t. It’s avoidance. Because there is a simpler question nobody seems keen to ask: if power is genuinely hard to come by, why are we so relaxed about wasting it?
AI workloads are famously energy hungry. Training large models consumes eye-watering amounts of electricity. Running them at scale does the same. That much is not in dispute. What is rarely discussed is how much of that energy is actually doing useful work.
A great deal of it isn’t. Data is shuffled back and forth because systems are poorly designed. Hardware sits powered on but underused because utilisation is inconvenient to optimise. Capacity is provisioned permanently for peaks that happen occasionally, if at all. Cooling systems work overtime to compensate for inefficiencies elsewhere.
All of this is treated as inevitable. It isn’t. Efficiency has become the unfashionable word in the AI race. It sounds like restraint, or worse, like compromise. The industry prefers the language of scale, ambition and speed. Efficiency feels small by comparison.
But here’s the awkward reality, efficiency is the only lever that works immediately. New energy infrastructure takes years. It involves regulators, politicians, planners, lawyers and neighbours who suddenly discover a deep interest in local power usage. Improving efficiency, by contrast, can start tomorrow. And every gain compounds.
Yet much of the industry behaves as though electricity were still cheap, abundant and somebody else’s problem (perhaps it is in some countries).
On the hardware side, the focus remains on faster accelerators and bigger clusters. That’s understandable, but it ignores where much of the power actually goes. Moving data often consumes as much energy as processing it. Leaving systems idling burns power for no return. Over-engineered interconnects and poorly matched components leak energy constantly.
Buying more hardware without asking how well it is used is not progress. Software is even worse. Model size has become a status symbol, as if the number of parameters were an indicator of intelligence. Little attention is paid to whether the same outcome could be achieved with a smaller, better-optimised model. Inefficient code is allowed to scale because rewriting it would slow things down.
This is the kind of logic that looks clever right up until the electricity bill arrives. Running models in production, in particular, is quietly becoming an energy sink. Models that run acceptably in a lab become liabilities when deployed at scale. Every wasted operation is multiplied thousands or millions of times a day. That is not innovation. It is neglect.
Cooling, meanwhile, is still treated as a background concern. Facilities are over-cooled to cover design sins elsewhere. Air is pushed harder and harder through spaces it was never meant to serve. Liquid cooling is discussed earnestly, then postponed because it feels disruptive.
Removing heat costs energy. Producing less heat in the first-place costs far less. This should not be controversial.
The biggest blind spot, however, is workload design. AI jobs are often run as though resources were infinite. Everything is designed around peak demand, even when that peak is rare. Workloads are scheduled for convenience, not efficiency. Data is moved repeatedly because redesigning pipelines would require effort and coordination.
Running systems flat out, all the time, is not a sign of ambition. It is a sign that nobody is properly in charge. So why does this persist? Partly because efficiency has no champion. It doesn’t come with a product launch or a keynote. It doesn’t make share prices jump. It sits awkwardly between engineering, finance and operations, owned by everyone and therefore no one.
There is also a comforting fiction at work, that energy problems can be solved elsewhere. Someone will upgrade the grid. Someone will absorb the cost, etc.
Eventually, perhaps. But in the meantime, the industry is burning power it does not need to burn. The irony is that the same companies talking loudest about sustainability are often the ones most relaxed about inefficiency, provided it sits behind the data centre walls.
This is not a moral argument. It is a commercial one. Power costs money. Inefficiency multiplies that cost. And in a world where AI is supposed to deliver productivity gains, wasting energy to get there is a peculiar strategy.
The AI race will not be won by whoever secures the biggest power contracts or builds the largest models. It will be won by those who get the most value from every watt they consume.
Efficiency is not the exciting lever. It is the one the industry keeps stepping around.
Review