That's only true for the US though. Electricity costs in China are a third of what they are in the states, and open models are now becoming very competitive. We can also see here that DeepSeek is already close to breaking even. The other factor is that models continue to get more efficient, and even local models like Qwen 27b can do serious work. Especially, when combined with good tooling. A coding harness that uses Qwen with five different roles scores as well as Anthropic Fable on the benchmarks.