The human brain consumes about twenty watts, what a monitor draws at rest. Training GPT-3 cost around 1,287 megawatt-hours, the electricity an average U.S. household uses in over a century. The efficiency gap between the biological brain and the silicon that aims to imitate it is measured in thirteen zeros. Talking about "neurons that compute" as if they were transistors was a useful metaphor for a while. Taking it seriously was the mistake.
The obscene figure
Twenty watts. That's what a waking adult brain consumes, reading this, anticipating the next sentence, recalling something from twelve years ago while another part of the same organ regulates breathing without asking permission. What a monitor draws at rest, a bedside lamp, a router. On that budget, eighty billion neurons sustain a chemical conversation in parallel, maintain an identity across decades and produce, as a strange byproduct, the subjective experience of being here.
The figure is so obscene in its modesty that it's worth repeating. Twenty watts.
On the other side, the figures discomfort. Patterson and his coauthors calculated in 2021 that training GPT-3, with its hundred and seventy-five billion parameters, consumed around 1,287 megawatt-hours. The electricity of a hundred and twenty U.S. households for a year, or of one alone for over a century. And that's only the training. Daily inference —every time the model answers a question—, fine-tunings, global deployments. The bill grows at every link. The IEA estimated in 2025 global datacenter consumption at 415 terawatt-hours for 2024, with a projection to 945 by 2030. Stanford's AI Index recorded in 2026 some 29.6 gigawatts of capacity dedicated to AI. The efficiency gap between brain and silicon is measured in thirteen zeros.
Thirteen.
The usual reaction is to breathe deep and think the hardware improvement will come. Maybe. Before taking comfort, it's worth examining why the gap is that size. The answer isn't in evolutionary cunning or any mystery. It's that the two machines don't do the same thing, and never have. Calling what happens in a brain "computation" was a working metaphor in 1943 and has been taken seriously for eighty years. Taking it seriously was the mistake.
The metaphor that slipped in as a description
McCulloch and Pitts published in 1943, in the Bulletin of Mathematical Biophysics, a foundational text: A Logical Calculus of the Ideas Immanent in Nervous Activity. The proposal was elegant and limited: if you modeled the neuron as a binary threshold unit, a network of such units could compute any logical function. The text didn't claim that real neurons were that. It claimed that a model that poor was enough to found a calculus.
On top of it grew all the cybernetics of the fifties, then connectionism —the current that in the eighties returned to using networks of simple units as a model of mental processing—, then today's artificial neural networks. At each jump, the model got simpler, not more complex, because the simplification was productive in engineering. Deep networks reduce the neuron to a multiplication of weights followed by a nonlinear function. It works very well for recognizing patterns and predicting the next token (the minimal unit of text the model emits, a word or fragment).
What happens is that along that path the discourse started reading it backward. If the model captures part of neuronal behavior, the neuron is a computer. If the artificial network reproduces brain tasks, the brain is a network of that kind. The metaphor went up a rung, stopped being a tool and became a thesis. And the thesis is wrong.
What the neuron really does
The biological neuron isn't a binary unit. It's a cell with thousands of dendrites, an axon that can be several centimeters long, chemical and electrical synapses, ion channels that open and close in milliseconds, neurotransmitter vesicles that get recycled, enzymes that adjust permeabilities in real time. Each synapse is modified with each use. Each neuron adjusts its excitability according to accumulated activity, glucose, hormones, sleep. The signal isn't a zero or a one. It's a firing pattern, with variable frequency, with phase synchronizations that matter. A transistor changes state in nanoseconds; a neuron, in milliseconds, and never holds it. Six orders of magnitude slower. And even so, in parallel, with eighty billion units each connected to several thousand, it produces something no GPU cluster reproduces.
Massive parallelism changes the problem. A GPU runs in parallel, yes, but its parallelism is regular: the same operations over tensors. The brain does something else. Each region has its own cytoarchitecture (the particular cellular organization of each cortical area), long and short connections that integrate modalities with no central clock. There's no clock. There's no instruction set (the set of basic operations a processor knows how to execute). There's no memory separate from processing. The synapse is at once store and operation. Synaptic plasticity —that the connection between two neurons strengthens or weakens with use— means the infrastructure itself changes with each experience, without a separate process of "writing to memory." That fusion of hardware and software has no equivalent in silicon. Not for lack of ingenuity. For architectural incompatibility.
Twenty watts against thirteen zeros
The figure returns. The interesting question isn't only how much each one consumes. It's how much useful work each watt delivers.
An H100 GPU dissipates between 350 and 700 watts depending on load. A serious training rack groups dozens. A hyperscale datacenter adds up to megawatts. And even so, for tasks a five-year-old solves without thinking —recognizing a familiar face in poor light, inferring intent from posture, deciding whether a sentence is ironic— the system stumbles or requires amounts of compute that scale brutally with difficulty. The child's brain does all that at once, while digesting, while regulating its temperature, while storing the whole thing in episodic memory. For twenty watts.
The explanation isn't magic. It's evolutionary engineering on a chemical substrate. A neurotransmitter molecule in a confined volume of a hundred nanometers transmits information at a cost electronics doesn't come close to matching. Silicon pays in each operation the cost of moving electrons through long channels, at very high frequencies, generating heat that must be dissipated with active systems. The brain pays in molecules that are already there, in a temperature it regulates for other reasons, in an architecture where compute and storage are the same thing.
Hardware that doesn't close the gap
It's not that the brain is better engineering. It's other engineering. One that silicon can't imitate without ceasing to be silicon.
Some insist the hardware will close the gap. It's worth looking at the trend honestly. Energy efficiency per operation has been improving at a decreasing rate ever since Dennard scaling —the empirical rule that for decades predicted each new generation of transistors would consume proportionally less as it miniaturized— broke down in the mid-2000s. The industry compensates by stacking more units, not by making each unit more efficient. The absolute consumption of datacenters grows. The IEA and AI Index projections aren't alarmist. They're accounting. The gap isn't closing. It's opening.
Consciousness, that layer that doesn't appear in the datasheet
So far the brain has been compared with silicon in quantitative terms. Watts, synapses, parallelism, plasticity. It's a comparison already generous to the computational metaphor, because it accepts its rules. But the uncomfortable element is missing. The one that appears in no benchmark (the standard set of tests measuring a model's performance), in no architecture paper, in no product announcement.
Consciousness.
Antonio Damasio has spent thirty years explaining, in Descartes' Error (1994) and The Feeling of What Happens (1999), that human consciousness isn't an abstract cognitive attribute floating over the brain. It's an emergent property of the whole organism. It's tied to the body, to the autonomic nervous system —the network that regulates breathing, heart rate and digestion without voluntary intervention—, to the continuous feeling of being alive. The substrate isn't only the cortex. It's the brainstem, the body sending signals upward constantly, the loop between viscera, emotion and cognition that constitutes the elementary sense of being someone. Consciousness, so described, isn't the result of a computation. It's the result of an organism regulating itself and representing that regulation. There's no possible computation without an organism. Disassembling the organism in silicon doesn't produce consciousness, because the critical operation —bodily regulation with its biological urgency— doesn't exist in silicon.
The hard problem
David Chalmers formulated in 1996 what's called the hard problem of consciousness. There's an easy problem, in a relative sense. Explaining the neural correlates (the patterns of brain activity that systematically accompany a concrete mental state). That's difficult but tractable. The hard problem is another. Why is there something it's like to be the brain running those processes. Why does the integration of information produce subjective experience. Why does red feel like red and not like nothing.
That problem hasn't been touched. Not in neuroscience, not in philosophy of mind, not in artificial intelligence. It isn't pending. It isn't even clear what kind of answer would resolve it.
What won't be captured by an algorithm
Roger Penrose, in The Emperor's New Mind (1989), argued from another direction that certain aspects of human thought —in particular, the grasping of non-formalizable mathematical truths— are incompatible with any algorithmic process. The argument, supported by Gödel's theorem (the logical result that in every sufficiently powerful formal system there are truths that can't be proved within the system itself), is disputed. The underlying question persists. If what the brain does isn't only more efficient than silicon but qualitatively distinct, no quantitative progress equals it.
This is where the computational metaphor, already fragile on the energy front, breaks beyond repair. A system that computes can be faster or slower, bigger or smaller. But the difference between a system that computes and a system that computes with consciousness isn't one of degree. It's one of kind. And no silicon has crossed that border, nor is the road it's taking heading in that direction. Contemporary generative models aren't a millimeter closer to consciousness than the networks of the eighties. They process more data.
The silicon that tries to resemble it
Not everything ignores the asymmetry. Neuromorphic computing —the design of chips whose operation imitates the structure of the nervous system— proposes chips that operate by pulses, asynchronous, with simulated synapses that modify locally. IBM with TrueNorth, Intel with Loihi, academic projects like SpiNNaker and BrainScaleS, the European Human Brain Project, recent work from Texas A&M. All trying to close the gap by the architectural route.
For certain concrete tasks —sensory processing, recognition of temporal patterns—, neuromorphic chips consume several orders of magnitude less than an equivalent GPU. It's a real success. Where they get stuck is in generality. What a neuromorphic chip does well is what the brain does well by architecture. Integrating asynchronous inputs, learning locally and continuously. What it doesn't do yet is scale to the tasks contemporary AI pours resources into. Massive language, generative models, broad symbolic reasoning. Those tasks prefer the classic architecture of dense matrices, which is exactly where the brain is bad and silicon is good. The paradox of the moment is that the two architectures are complementary for different tasks and nobody has found a way to make a single substrate do both with the efficiency of either.
And, even more relevantly, no neuromorphic chip aspires to generate consciousness. It doesn't even appear on the list of objectives. The quantitative gap may shrink in the coming decade. The qualitative one isn't being touched.
What follows when the analogy fails at the base
If the metaphor is defective on the energy, architectural and experiential fronts, the conclusions built on top inherit the defects. Three circulate with an authority they don't deserve.
The first, comparability. Public discourse treats the brain and the models as systems comparable on a single functional plane. They aren't. They share no architecture, no substrate, no energy economy, no relationship with the body. Comparing them in terms of "general intelligence" is like comparing a kidney with a treatment plant because both filter. The operation is nominally the same. What matters, in each case, is radically different.
The second, imminent superiority. The narrative that the models will "reach and surpass" the brain soon assumes the cognitive dimensions are well defined on both sides. They aren't. Benchmarks measure what they can measure. Tasks with evaluable answers. The part of human behavior that doesn't reduce to that —the bulk of what constitutes a conscious life— isn't compared, because nobody knows how to measure it. Saying the model surpasses the human on benchmarks says nothing about general superiority. It says the model is better at what's been decided to measure.
The third, existential threat. If the models aren't architecturally comparable to the brain, the idea that they could develop consciousness, agency or a will of their own by mere quantitative scaling has no foundation. That doesn't mean there are no real threats in the deployment of AI. There are, and they're operational, economic, political, labor-related, manipulative, about the concentration of power. But the metaphysical threat, the model that "awakens," is built on the metaphor taken as literal description. If the metaphor is false, the derived threat is rhetorical, not technical.
The layer that doesn't replicate
Twenty watts. Eighty billion neurons. A continuous subjective experience. A capacity to regulate emotions, to have intuitions, to err in productive ways, to get bored, to need sleep, to dream. A memory that rewrites itself every time it's evoked, that confuses, that invents, that holds a self coherent enough for the word "I" to have a referent.
What's truly extraordinary about the brain isn't that it "computes." It's that it's the organ of an organism that feels itself alive. That layer, consciousness, isn't an unimportant epiphenomenon (a byproduct with no causal function of its own). It's the condition of possibility of everything we call individual and collective progress. The scientific, artistic, moral and technical advances of the species aren't products of an abstract computing capacity. They're products of conscious minds solving problems that matter to them because they feel them, because they have an affected body, because they live in communities of other conscious minds with whom they negotiate meaning. No system without consciousness does that. Not for lack of computing capacity. For lack of the engine.
Damasio wrote it decades ago. Human reason doesn't oppose emotion. It operates on an emotional substrate without which it doesn't decide, doesn't prioritize, doesn't attend. Patients with lesions in the ventromedial prefrontal cortex —the frontal region of the brain involved in the emotional valuation of options— keep their abstract intelligence and lose the capacity to make everyday decisions. The reasoning doesn't fail. What fails is the feeling that orients the reasoning. That substrate is bodily. And the body doesn't go into silicon in any reasonable version of the future.
What's sold, what there is
The fashionable line says we're creating artificial intelligences ever more like us. What's being created is something else. Systems that reproduce, at colossal energy cost, a subset of tasks the human brain does in passing while it gets on with living. It's a tool. A good tool. And it's here to stay. But calling it "comparable intelligence" is an imprecision that has turned productive for selling.
Dirty question. When was the last time a model got bored, felt gratuitous curiosity, decided something mattered to it without anyone asking, recalled something twenty years later and felt a muscle in its face move.
Twenty watts. A consciousness. Thirteen zeros of distance.
Definiciones
Token. The minimal unit of text a language model processes or emits. It doesn't always coincide with a word. It can be a whole word, a fragment or a punctuation mark.
Inference. Each time an already-trained model produces a response from an input. Distinct from training, the prior phase in which its parameters are adjusted.
Synaptic plasticity. The capacity of the connection between two neurons to strengthen or weaken depending on use. It's the basic biological mechanism of learning and memory.
Cytoarchitecture. The particular cellular organization of a concrete brain region. Which types of neurons compose it, how they're arranged in layers, how they connect to one another.
Instruction set. In computer architecture, the set of basic operations a concrete processor knows how to execute. Processors are designed around a specific instruction set.
Dennard scaling. An empirical rule formulated in 1974 that predicted that, as transistors miniaturized, power density would stay constant. More transistors per unit area without increasing total consumption. It stopped holding in the mid-2000s.
Benchmark. A standard set of tests used to measure and compare the performance of AI models. It serves to compare them against one another, not to measure what the model can't do.
Neuromorphic computing. A line of chip design whose operation is inspired by the structure of the nervous system. Asynchronous pulses, simulated synapses, local learning, instead of the classic paradigm of dense matrices and a central clock.
Neural correlates. Patterns of brain activity that systematically accompany a concrete mental state. Identifying them doesn't explain why the mental state feels the way it feels. It describes what happens in the brain when it occurs.
Hard problem of consciousness. David Chalmers's formulation (1996) distinguishing between explaining the brain mechanisms associated with consciousness and explaining why those mechanisms are accompanied by subjective experience. The second question remains unanswered.
Autonomic nervous system. The part of the nervous system that regulates, without voluntary intervention, vital functions such as breathing, heart rate or digestion. A central piece in Damasio's explanation of consciousness.
Epiphenomenon. A phenomenon that accompanies another but has no causal function of its own. Whoever holds that consciousness is an epiphenomenon claims it doesn't influence behavior. Damasio rejects that reading.
Referencias
McCulloch, W. S. and Pitts, W. (1943), A Logical Calculus of the Ideas Immanent in Nervous Activity, Bulletin of Mathematical Biophysics 5, pp. 115-133. Origin of the computational metaphor applied to the neuron and of the idea that a network of simple units can compute any logical function.
Patterson, D. and others (2021), Carbon Emissions and Large Neural Network Training, arXiv:2104.10350. Source of the figure of 1,287 megawatt-hours attributed to training GPT-3.
International Energy Agency (2025), Energy and AI Report, https://www.iea.org/reports/energy-and-ai. Origin of the figures for global datacenter consumption: 415 terawatt-hours in 2024 and a projection to 945 terawatt-hours in 2030.
Stanford Institute for Human-Centered AI (2026), AI Index Report 2026, https://hai.stanford.edu/ai-index/2026-ai-index-report. Source of the figure of 29.6 gigawatts of datacenter capacity dedicated to AI.
Damasio, A. (1994), Descartes' Error. Emotion, Reason, and the Human Brain, Putnam, and (1999), The Feeling of What Happens, Harcourt. Theoretical basis for the description of consciousness as an emergent property of the whole organism and for the argument on the emotional substrate of reason.
Chalmers, D. (1996), The Conscious Mind, Oxford University Press. The formulation of the hard problem of consciousness, central to the stretch distinguishing computation from subjective experience.
Penrose, R. (1989), The Emperor's New Mind, Oxford University Press. The argument, supported by Gödel's theorem, against reducing human thought to algorithmic processes.
Stiefel, K. M. and Coggan, J. S. (2023), The energy challenges of artificial superintelligence, Frontiers in Artificial Intelligence 6:1240653. DOI: 10.3389/frai.2023.1240653. A technical reference on the energy constraints of scaling AI models.
Texas A&M Research (2025), AI that uses less energy by mimicking the human brain, https://stories.tamu.edu/news/2025/03/25/artificial-intelligence-that-uses-less-energy-by-mimicking-the-human-brain/. A reference for recent work in neuromorphic computing.
Human Brain Project (2023), Learning from the brain to make AI more energy efficient, https://www.humanbrainproject.eu/en/follow-hbp/news/2023/09/04/learning-brain-make-ai-more-energy-efficient/. European outreach on bio-inspired chips and their limits.
Para profundizar
Hofstadter, D. (1979). Gödel, Escher, Bach. An Eternal Golden Braid. Basic Books. A classic exploration of the relationship between formal systems, self-reference and mind; a natural resonance with Penrose's argument and with the question of consciousness.
También te interesa
- Quantum computing and thought. The god of the gaps in a physics edition
- The brain as a non-deterministic system. The noise was ours, not theirs
- Recognition as the basis of everything. What happens before you know you've thought
No comments yet
No comments yet. Be the first.
Leave a comment