The human brain consumes roughly 20 watts of power. The world's fastest supercomputer, LineShine in Shenzhen, draws 42.2 million watts. While this comparison circulates widely on social media, the underlying data contains inconsistencies that have persisted for years.
The 20-watt brain figure rests on decades of metabolic research, with the brain accounting for approximately 2% of body weight and roughly 20% of resting oxygen consumption. However, viral posts often cite 12 watts, a figure from a 2023 paper in Frontiers in Artificial Intelligence that lacked citation support. That same paper estimated recreating a human brain digitally would require 2.7 billion watts, an extrapolation from a 10-million-neuron simulation that ran about 30,000 times slower than biological systems—a detail frequently omitted from secondary sources.
More reliable measurements exist. Epoch AI estimated a typical ChatGPT query at 0.3 watt-hours in early 2025, with a peer-reviewed study in Joule confirming 0.31 watt-hours. However, power consumption varies sharply with workload, and reasoning models can cost several times more.
Biological Efficiency Principles Adopted by AI
The human brain operates on sparse cortical activity, with average firing rates below 1 Hz and energy following change rather than constant clock cycles. Modern AI systems have independently reached similar conclusions. Kimi K2 activates only 3.1% of its parameters per token, while DeepSeek-V3 uses 5.5%—a dramatic decrease from Mixtral's roughly 28% in 2023.
AI systems also adopted low-precision computation from biology. DeepSeek trained a 671-billion-parameter model in eight-bit precision, while NVIDIA has pretrained a 12-billion-parameter model in four-bit. The largest efficiency difference between brains and digital systems remains unresolved: brains hold memory and computation in the same physical space, while digital machines separate them, consuming hundreds of times more energy fetching operands from memory than performing arithmetic.
Neuromorphic Hardware Struggles
Hardware explicitly designed to imitate neurons has not yet deployed frontier AI models in production. Intel's Hala Point research prototype contains 1.15 billion artificial neurons across 1,152 chips but remains a prototype at Sandia National Laboratories. According to Mike Davies, director of Intel's Neuromorphic Computing Lab, mapping large language models to such hardware remains technically unclear.
The commercial sector is thinner. BrainChip, the sector's flagship listed company, reported $700,000 in customer receipts against $5.3 million in operating outflows in its March quarter. Rain AI, which sought $150 million in funding, explored a sale in 2025. Researchers describe a circular problem: hardware companies await killer applications while developers cannot build applications without hardware to prototype on. A 2025 consensus paper in Nature signed by over 20 researchers concluded the field still lacks the necessary ecosystem.
Electricity Now Gates AI Expansion
The binding constraint on AI growth has shifted from chips to power. The International Energy Agency reported global data center consumption at 485 terawatt-hours in 2025, with AI-focused facilities growing 50% that year alone. The agency expects them to triple by 2030.
Grid access, not chip supply, now determines where facilities are built. Median time from an interconnection request to commercial operation exceeds five years, according to Lawrence Berkeley National Laboratory. In November, Microsoft chief executive Satya Nadella noted his company holds processors it cannot power. The shortage is energized buildings, not silicon.
Bitcoin Mining Pivots to AI Infrastructure
Bitcoin miners have spent a decade securing energized sites, signed power agreements, and interconnection rights. This infrastructure expertise has attracted AI companies. Public miners have signed AI contracts worth more than $70 billion in aggregate. Retrofitting existing sites costs roughly $3 million to $4 million per megawatt, while greenfield construction runs $10 million to $12 million.
Delivered capacity tells a different story than announced deals. Second-quarter 2026 filings show roughly 750 megawatts actually energized across the mining sector. Core Scientific accounts for approximately 437 megawatts, Galaxy's Helios campus 133, TeraWulf 102, IREN 50, and Riot 25. Hut 8 has contracted 949 megawatts but energized none.
The pivot has proven costly. Combined quarterly losses at miners MARA and CleanSpark reached $851 million in August. Preliminary Cambridge survey data from July indicated that about 10% of miners had allocated power to AI. Physical obstacles remain: mining tolerates interruption while AI tenants demand firm power, dense cooling, and fiber infrastructure rarely available at remote sites.
Core Scientific's revenue mix reflects the direction of travel. The company now earns 83% of revenue from colocation and just 13% from mining. Investors have priced this shift accordingly, with miners holding signed leases trading at higher multiples of their energized power.
Why Efficiency Gains May Not Constrain AI Growth
Global data center compute grew by 550% between 2010 and 2018 while energy use rose only about 6%, suggesting efficiency gains could absorb demand growth. The pattern broke afterward. United States data center consumption climbed from 58 terawatt-hours in 2014 to 176 in 2023.
Evolution optimized under hard constraints—a brain consuming 200 watts would have killed its owner. AI has never faced such biological ceilings. Instead, it has faced capital constraints, which stretch more easily than electricity does. That relationship is now changing, raising questions about what AI becomes once power availability, rather than capital, determines what gets built.
