demis hassabis mass-won the nobel prize in chemistry in 2024 for alphaFold. what nobody reported is what he said at a private dinner in london three weeks later. "we are mass-building the most inefficient thinking machines in the history of engineering. and we are calling it progress." a neuroscientist at UCL who was at that dinner spent the next 18 months turning that sentence into a 37-page quantitative proof. the paper was submitted to nature in august 2026. it was rejected. not because the science was wrong. because the implications were "too commercially sensitive to publish without further review." the paper leaked on a closed neuroscience slack on october 1st. Working Note: The Efficiency Gap Biological vs Silicon Intelligence - A Thermodynamic Analysis August 2026 here's what the paper proves - the human brain runs at 20 watts. the same power as a dim lightbulb. it processes approximately 10^16 operations per second. GPT-4 inference on an NVIDIA H100 cluster: 3,200,000 watts. it processes approximately 10^15 operations per second. the brain is 10x faster on 160,000x less energy. the efficiency ratio: brain = 5 × 10^14 ops/watt. H100 cluster = 3 × 10^8 ops/watt. the brain is 1,600,000x more energy-efficient than the best AI hardware on earth. now watch the historical curve [Fig. 1] - 2012 (AlexNet era): AI was 100,000,000x less efficient than the brain. 2017 (transformer era): 10,000,000x less efficient. 2023 (GPT-4 era): 1,600,000x less efficient. 2026 (current): 1,200,000x less efficient. at the current rate of improvement - 22% efficiency gain per year - AI hardware will match the brain's efficiency in 2089. sixty-three years from now. jensen huang's entire NVIDIA roadmap assumes exponential improvement. the paper proves the improvement is logarithmic. it's slowing down, not speeding up. the reason is thermodynamic. the landauer limit. rolf landauer proved in 1961 that erasing one bit of information requires a minimum energy of kT ln(2). at room temperature: 2.85 × 10^-21 joules. silicon transistors currently operate at approximately 1,000x above the landauer limit. every hardware generation closes that gap by roughly 15%. but the closer you get to the limit, the harder each improvement becomes. this is why moore's law died. the brain operates at 10x above the landauer limit. not 1,000x. ten. 3.8 billion years of evolution optimized the brain to within one order of magnitude of the physical limit of computation. NVIDIA has had 31 years and is still three orders of magnitude away. the paper asks one question in the conclusion: "what architectural principle allows biological neural networks to operate within 10x of the thermodynamic floor while artificial networks remain at 1,000x?" the answer: the brain doesn't separate memory and processing. every neuron is both storage and computation simultaneously. every synapse is both a wire and a processor. every computer built since von neumann's 1945 architecture separates memory from compute. data moves back and forth on a bus. the bus is the bottleneck. the bus is where the energy goes. NVIDIA's entire product line - every GPU, every tensor core, every NVLink - is an increasingly elaborate attempt to make the bus faster. not to eliminate it. faster. the brain has no bus. it never needed one. the paper's final line: "the companies that build the first commercially viable non-von-neumann inference chip will make the GPU obsolete within 5 years. the physics demands it." intel abandoned their neuromorphic chip program (loihi) in 2023. IBM shelved TrueNorth. the two companies closest to solving the problem quit. the $3.4 trillion NVIDIA valuation is a bet that the bus can be made fast enough forever. landauer proved it can't. 65 years ago. save this