The end of intelligence is energy

AI is not just code; it's a factory. Large models are like a perpetual motion machine that converts electricity into intelligence—they occupy space, generate heat, and consume power. Essentially, they are no different from steel mills.
Training a model is like using vast amounts of electricity to compress thousands of years of human knowledge into parameters. This is not algorithm optimization; it's energy compression. Ultimately, all computing power is powered by solar energy conversion—coal, hydro, wind—all originating from that star.
OpenAI's Stargate project consumes 1-2 GW of power, equivalent to half a nuclear power plant. Data centers' electricity usage has approached the levels of Germany or France. This is no longer a budget issue but a civilization-level energy management challenge.
In the 20th century, electricity powered industrial production; in the 21st century, electricity itself is the raw material of cognition. Whoever can convert light into tokens with the lowest energy consumption will control the boundaries of intelligent evolution.