AGI's hidden bottleneck: power grid can't handle 100GW demand
The most daunting obstacle to artificial general intelligence may not be algorithmic breakthroughs but an unglamorous infrastructure problem: the global power grid cannot physically support the compute required to achieve it.
A viral analysis from a UK grid engineer reframes the AGI debate from software to hardware: electricity consumption. The author calculates that a single NVIDIA GB200 AI rack consumes 120kW continuously, totaling 1,050,000 kWh per year—equivalent to 389 average UK households. Even before cooling, each H100 GPU draws 700W under load.
Scaling to a hypothetical AGI deployment with 1 million accelerators would push GPU power alone to 700MW. With networking, cooling, and infrastructure losses, total demand could hit 2GW continuously, or 17.5 TWh annually—the same as 6.5 million UK homes. If global AGI infrastructure reaches 100GW, it would consume 876 TWh per year, matching Japan’s entire national electricity output. The argument: AGI cannot be treated as just a software problem when physical power constraints are this stark.
- The grid bottleneck could cap GPU sales: hyperscalers facing multi-year delays in power connections may limit their NVIDIA deployments, creating an opening for more efficient chip designs.
- Efficiency improvements (20% per generation) are falling short of compute growth (~4x per year), making dedicated power generation or novel compute paradigms essential for continued scaling.
- The 100GW figure is speculative but underscores a real tension: AGI may require a national-scale electricity allocation, forcing companies to invest in nuclear, renewables, or radical efficiency leaps.
Why It Matters
AGI's viability depends as much on power grids and energy breakthroughs as on AI research and scaling laws.