It is often said that the human brain is much more energy efficient than AI systems because the brain, while doing similarly complex cognitive tasks as AI, uses globally orders of magnitude less energy than the best AI supercomputers. It seems that the same common belief is shared by in their target article, which also lacks a direct comparison of energy efficiency between neural and AI systems. The goal of this comment is to challenge this notion by presenting specific hard data on the costs of computations in mammalian brains and AI machines. Using “the energy per operation” as a measure of computational efficiency, I show that the issue is more complex and level dependent and that the human brain is more expensive in this regard than the best AI machines.

The human brain consumes about 20 W of metabolic energy (Aiello & Wheeler, Reference Aiello and Wheeler1995), contains 86×10⁹ neurons (Azevedo et al., Reference Azevedo, Carvalho, Grinberg, Farfel, Ferretti, Leite, Filho, Lent and Herculano-Houzel2009), and neurons fire action potentials at the rate of 1 Hz (Karbowski, Reference Karbowski2009, Reference Karbowski2012). Assuming that an action potential corresponds to a single neural operation, this means that the human brain uses about 200 pJ/op. A much more accurate cost of a single action potential, based on detailed theoretical analysis and simulations, yields (Attwell & Laughlin, Reference Attwell and Laughlin2001; Karbowski, Reference Karbowski2009): 80–130 pJ/op for human cortex and 40 pJ/op for mouse cortex. (For a mouse neuron, an action potential requires removal of about 15×10⁸ sodium ions from the neuronal interior that costs 5×10⁸ ATP molecules, and 1 ATP uses 0.083 aJ energy. A human neuron uses 2–3 times more energy because its volume is 2–3 times larger).

Let us compare these numbers to the corresponding figures in AI semiconductors. AI devices can be broadly divided into two classes: fast supercomputers with von Neumann architecture and slower neuromorphic systems without von Neumann architecture serving specific purpose computations. The first group has a very different structure from mammalian brains (with spatially separate memory and decision modules), whereas the second tries to mimic the brain architecture (spatially close memory and CPU). The fastest supercomputer in the first group, known as “El Capitan,” contains 6×10¹⁵ transistors, performs 2.8×10¹⁸ operations/sec, and dissipates 35 MW of energy (Science & Technology Review 2024), which gives 12.5 pJ/op. This implies that this “brute force” machine is 6–10 times more energy efficient per operation than neurons per action potential in the human brain. This is the first sign that biological information processing is costly on a microscale.

The second class of AI devices with neuromorphic design is much slower than the above supercomputer but also more energy efficient. The biggest neuromorphic system in the world, known as “Hala Point” with 10⁹ artificial neurons (128 artificial synapses per neuron) and 2.6×10¹² transistors, uses 2600 W and performs 2×10¹⁶ operations/sec (Intel Newsroom 2024). This gives the efficiency of 0.13 pJ/op (10⁷ op/sec per artificial neuron), which is 100 times better than for the fast supercomputers and 600–1000 times better than the cost of operation in a human neuron. This means that current neuromorphic systems can achieve much better performance on a microscale in terms of energy/op than the human brain. However, on the other hand, the energy used per artificial neuron is 13×10⁵ pJ/sec, which is 10000 times larger than in a human neuron. This is another type of “energy efficiency” (per second), and that one is won by the biological neurons.

The last numbers suggest that one has to be careful with the notion of energy efficiency because the context, i.e., energy per what, matters. Consequently, let us analyze the efficiency on a more fundamental, smaller scale, corresponding to transistors (in AI) and synapses (in brains).

The El Capitan supercomputer and Hala Point neuromorphic computer are assembled with the same type of 7 nm transistors (FinFET technology) that use about 1 aJ/op (Larsh et al., Reference Larsh, Siddique and Banad2025). Thus, on a nanoscale, the efficiency of computation increases dramatically by a factor of 10⁵–10⁷, and it is the same in both computers. These numbers indicate that 1.8×10⁷ transistors are involved in one operation in El Capitan and 10⁵ transistors in Hala Point. On the other hand, one synapse uses 2×10⁵ ATP molecules per synaptic transmission (Laughlin, de Ruyter van Steveninck, & Anderson, Reference Laughlin, de Ruyter van Steveninck and Anderson1998; Karbowski, Reference Karbowski2012), which is 0.017 pJ/op. This is 4700 times more energy efficient than in a neuron, but still 10⁴ less efficient than in a transistor. However, one can argue that a typical synapse is about 1000 times larger than the 7 nm transistor and contains roughly 10⁴ proteins (Sheng & Hoogenraad, Reference Sheng and Hoogenraad2007), which also consume energy and take part in synaptic plasticity, responsible for learning and memory (Lisman, Yasuda, & Raghavachari, Reference Lisman, Yasuda and Raghavachari2012). This suggests that the efficiency per synaptic protein is probably similar to the efficiency of a modern transistor.

What are the implications? First, a single action potential in the brain is much more costly than a single operation in AI hardware, and thus it is more valuable. That is why mammalian brains are ultraslow in comparison to AI machines and employ a “sparse coding” strategy (to do things using as little energy as possible). Second, efficiency becomes comparable for brains and AI hardware on the nanoscale, on a level of synaptic proteins and transistors. Despite these similarities for the energy cost of information processing in the nanoscale, the global “cognition” of brains and AI machines differs significantly, suggesting that other factors also matter. These might be related to differences in, e.g., connectivity architecture (denser connections in the brain; Karbowski, Reference Karbowski2014) and information carriers (much more flexible synaptic proteins in the brains than electrons in transistors; Lisman et al., Reference Lisman, Yasuda and Raghavachari2012).