By Mohamed Jameel, Managing Director, Votex Insights
Elon Musk has been remarkably clear about where AI is heading. In March 2024 he described the constraints on AI as almost predictable: first a shortage of chips, then of voltage step-down transformers, then of electricity itself. He is right about the symptom. I respectfully differ on the diagnosis.
My view is that the bottleneck isn't only about supplying more power. It's a sign that much of our research into information processing has been heading in an inefficient direction.
Musk sees the gap too
To his credit, Musk has noticed the same thing. On the Lex Fridman podcast in November 2023 he observed that a brain using about 10 watts for thinking "can still produce a much better novel than a 10 megawatt GPU cluster", and that what matters is "useful compute per watt". More recently, on the All-In podcast, he said human brains use about 20 watts, with only about 10 for higher brain function, calling it "a massive opportunity for improving the efficiency of AI compute because it's currently many orders of magnitude away".
Where we part ways is the remedy. His companies are answering the problem with scale. xAI is powering its Colossus data centres in Memphis with gas turbines, and SpaceX has asked regulators for approval to put solar-powered AI data centres in orbit. Neuralink, his brain company, is about connecting the brain to computers, not about learning from the brain how to compute with less energy. Rather than more gigawatts on Earth or in orbit, I believe the bigger prize lies in research that brings computing closer to how living systems process information.
The numbers behind the worry
The International Energy Agency estimates that data centres used around 415 TWh of electricity in 2024, about 1.5% of the world's total. It projects around 945 TWh by 2030, slightly more than all of Japan uses today, with AI the biggest driver. Training a single model is energy-hungry too. GPT-3's training was estimated at 1,287 MWh.
Then look at yourself
The adult human brain runs on roughly 17–20 watts, less than many light bulbs. Over 20 years of learning, seeing, feeling and deciding, that comes to about 3.5 MWh. Compare that with the 1,287 MWh for one training run above. The body manages its information with an elegance we have barely begun to copy.
To be fair, these comparisons aren't like for like. A synapse firing is not a digital calculation, and much of the brain's energy goes to keeping itself alive. Researchers caution that any ratio is only a rough order of magnitude. Even so, the gap is wide enough to be worth thinking about.
Where the research already points
What encourages me is that when engineers do learn from the brain, the gains are dramatic. IBM's brain-inspired NorthPole chip keeps memory and computing side by side, the way neurons do. In tests it was 25 times more energy-efficient than a comparable GPU at recognising images, and 72.7 times more efficient when running a small language model. Intel's Hala Point system packs 1.15 billion artificial neurons into a maximum of 2,600 watts.
Physics shows how far there is to go. Landauer's principle sets a minimum energy cost for erasing a single bit of information, and it was confirmed experimentally in 2012. Today's chips run far above that floor. Some researchers are exploring "reversible" computing to get closer to it.
Then there is the boldest idea: that the brain may use quantum effects. The Orch-OR theory proposes this, but it is strongly contested. Critics calculate that quantum states in the brain would collapse far too quickly to matter. I mention it as an open question, not a settled fact.
My argument
These are still side projects next to the hundreds of billions now going into conventional data centres. The IEA notes that data centre investment reached half a trillion dollars in 2024. Imagine if even a share of that went into the efficiency of the living brain, studied all the way down to its finest physical level. I believe we would be further ahead, and with a lighter footprint on the planet.
I am in good company in thinking the mind deserves more scientific attention. Max Planck, the father of quantum theory, told The Observer in 1931: "I regard consciousness as fundamental. I regard matter as derivative from consciousness." The philosopher David Chalmers put the puzzle sharply: "Why doesn't all this information-processing go on 'in the dark', free of any inner feel?" We have built machines that process vast amounts of information, yet the one processor that is aware of what it processes is still the least understood.
Even John von Neumann, whose design underlies almost every computer we use, is quoted as suggesting that mathematics may be "a secondary language built on the primary language of the nervous system". And Richard Feynman reminded us that "Nature uses only the longest threads to weave her patterns". Learn one of nature's designs deeply and it teaches you about the rest.
For me as a Muslim, this also echoes the Qur'an's repeated call to reflect: "And on the earth are signs for the certain [in faith]. And in yourselves. Then will you not see?" (51:20–21).
وَفِى ٱلْأَرْضِ ءَايَـٰتٌ لِّلْمُوقِنِينَ وَفِىٓ أَنفُسِكُمْ ۚ أَفَلَا تُبْصِرُونَ Surah Adh-Dhariyat 51:20–21, verified on Quran.com
The most efficient information processor we know of is within us. We have studied the machine more than the mind.
What this means for leaders
This isn't only a question for chip designers. Every leader runs on the same 20-watt system. Attention is our scarcest and most efficient resource, and we waste it on distraction just as we waste electricity on brute-force computing. Learning to direct it well is the theme of my book.
If this resonates, read The Attention Advantage, or get in touch to bring these ideas to your team.Zq
