Why does your brain run on 20 watts while AI models need megawatts? During this year's AI House Davos fireside chat on embodied AI, Yann LeCun (Founder and Executive Chairman, Advanced Machine Intelligence) explained the fundamental hardware limitation holding back artificial intelligence. In biological brains, each synapse and neuron exists as a physical computing element. But in silicon neural networks, we reuse the same hardware to compute outputs for multiple neurons. This means constantly shuttling data between memory and processing, and that's where all the energy goes. Even advanced GPUs can't match biological efficiency because current fabrication technology makes it impossible to dedicate one computing device per neuron. The gap isn't just about efficiency. It's about scalability. The AI industry optimizes what's measurable with current tools rather than solving fundamental problems. Throwing more GPUs at neural networks delivers diminishing returns because we're hitting physical limits of data movement. Watch the full conversation on the AI House Davos YouTube channel. Link in the comment section.
This highlights the real bottleneck of today’s AI: data movement, not compute. Biological systems fuse memory and processing at the physical level, something our silicon architectures still can’t replicate. Scaling with more GPUs brings diminishing returns unless we rethink hardware and system design. The next leap in AI will come from brain-inspired, energy-efficient architectures, not just bigger models.
This reframes the whole "just add more GPUs" conversation. The real gap is architectural, not computational. Loved Prof. LeCun's line about the AI industry "digging the same trench." The energy bottleneck is just one symptom of a deeper limit
I'd be careful with the 20W comparison. Last week a researcher used 16 parallel Claude agents to build a C compiler in Rust -> 100,000 lines of code, compiles the Linux kernel. Took days. The human equivalent? A team of 4 engineers, 8 years, in a heated office building with AC, monitors, commutes, coffee machines, and a CI/CD pipeline running 24/7. The brain runs on 20 watts. The human system around it doesn't. Add lighting, HVAC, transportation, food production -> a single engineer's true energy footprint is closer to 10,000 watts sustained. 4 engineers × 8 years × 10kW vs 16 agents × 3 days × whatever GPU cost you want. Per unit of useful output, I'm not sure the biology wins anymore
The challenge isn’t just power but how we design hardware for true neural efficiency. Excited to see where innovation takes us next.
I believe we Piris Labs (YC W26) are close to breaking the data movement barrier. If data movement bandwidth increases by 10× using optical interconnects, while both power consumption and scaling costs decrease, the compute node and its associated memory begin to behave as a coherent system.
Until we approach the brain’s level of energy efficiency, we won’t have anything truly comparable to human intelligence. It's obvious: complexity and energy efficiency at max! The brain is the result of millions of years of evolution. What we call the “reptilian brain” encoded core properties of the world long before conscious reasoning: gravity, balance, light, spatial orientation, threat detection: a form of evolutionary pre-training. As Yann LeCun noted in Davos, many species are born with built-in priors. A human doesn’t need to fall off a cliff to learn it’s dangerous. Core survival systems regulating breathing, heart rate, blood pressure, reflexes, and instinctive responses are already embedded at birth. Today’s AI systems are mostly trained on discrete datasets. That’s why emerging approaches like “world models” aim to build internal representations of physics and cause–effect structure. The world is continuous and analog. Purely discrete models have limits in approximating physical reality. Future AI may require more analog-like computation and evolutionary-style pre-training grounded in physics. Energy efficiency, embodied priors, and world modeling may matter more for intelligence than raw parameter count or GPUs :)
The data movement bottleneck LeCun describes maps directly onto a problem we face in educational AI deployment. Schools running local inference on modest hardware hit the same wall: the computation is feasible, but shuttling embeddings between memory and processor makes real time interaction with students impractical. The 20 watt brain benchmark is not just a curiosity; it is the design target for any AI system meant to operate at the edge, in classrooms, clinics, or field research stations where megawatt infrastructure does not exist. If the next architectural breakthrough comes from rethinking the memory hierarchy rather than scaling transistor counts, the implications for democratising AI access in education are enormous.
Why are we comparing the physics of living organisms and non-living organisms as if the science is the same? The hamiltonian is not the same thermodynamics of the human brain/body as another closed system such as a nueral [computer] network when measuring potential and kinetic energy.
The comparison between a human brain and an AI model is fundamentally flawed because it ignores the astronomical capital expenditure of evolution. Citing the brain's 20 watt running cost as a benchmark for efficiency commits a massive system error; it disregards the terawatts of energy consumed over billions of years of mutation, trial, and death to refine that biological architecture. Nature has already paid the energy debt for that structure through eons of selection, whereas we are attempting to replicate the result in real-time using silicon. It is not a like for like comparison of efficiency; it is comparing a system where the R&D costs were fully amortised by the pre Cambrian era against a prototype that is still paying for its own construction. Also the operational mechanics are so distinct that comparing their wattage is scientifically incoherent. As Yann admits, we are forced to use hardware multiplexing, shuttling data back and forth at high energy costs, because we lack the physical technology as biology does. Criticising a GPU vs a brain does on 20 watts is ignoring the fact that the GPU is fighting against its own physics to emulate a biological layout that simply doesn't exist in silicon. Thx