Training a single large AI language model requires roughly 10 gigawatt-hours of electricity, about what a thousand households consume in a year. By 2030 that is set to almost triple. Every chatbot conversation, every AI-generated image, every automated search is drawing from a grid increasingly strained to keep up.
This is the part of the AI story that rarely makes the headlines alongside the productivity gains and the dazzling demos. The infrastructure underneath it all is expensive, scarce, and getting dirtier by the quarter.
Which is why a paper published in Science Advances this month deserves more attention than it got.
Researchers at the University of Cambridge have developed a new type of nanoelectronic component, a so-called memristor, that mimics the way human neurons simultaneously store and process information. Current AI systems rely on conventional chips that constantly shuttle data back and forth between memory and processing units. This back-and-forth is where enormous amounts of electricity disappear. The Cambridge approach called “neuromorphic”, or brain-like computing, eliminates that gap. The result could reduce AI energy consumption by up to 70%.
Lead researcher Dr Babak Bakhit spent three years on largely unsuccessful experiments and a huge number of failures, but by late 2025, the first genuinely promising results appeared.
It’s worth being clear-eyed here. Semiconductor manufacturing is difficult. You have to mine the materials, custom-build and employ staff for the factories, and this just adds a new complex process into the mix. Only a handful of countries, mostly in Asia, have the infrastructure for this, and they are often forced to collaborate internationally, to complete different stages. This is not a chip you can buy next year. It is a proof of concept, but a significant one, with a patent filed and a clear industrial rationale already in place.
Why does this matter for New Zealand? We are not building AI chips here, and we are unlikely to start. But we are a country with abundant renewable electricity that the rest of the world increasingly wants for exactly this kind of computing. This makes our country an attractive option for tech companies striving to meet net zero targets. The question of how much of our grid we hand over to offshore data centre demand, and on what terms, is one our government has barely begun to answer.
A reduction in AI energy consumption would not just be a win for tech companies’ bottom lines. It would change the calculus of whether AI infrastructure is compatible with a liveable climate. It might also mean that the tools reshaping medicine, science, and public services become genuinely accessible to smaller nations, rather than remaining luxuries rationed by whoever can afford the power bill.
