This is kind of headline bait. I work in energy forecasting so I look at these numbers all day. Facebook's 350k H100s might draw around 1kW each, let's say 2kW

This is kind of headline bait. I work in energy forecasting so I look at these numbers all day. Facebook's 350k H100s might draw around 1kW each, let's say 2kW with the rest of the rack and cooling. That's 700MW for what is probably the largest AI cluster in the world. You know what's as big as that? Riot Blockchain's Rockdale mining facility (700 MW capacity). Which one do you think provides more value, training Llama 3 or mining bitcoin? The reason I'm not worried is because solar is now dirt-cheap, and compute is quite interruptible. If it costs you $50/MWh to train your LLM during the night but $5/MWh to do it during the day, maybe you're ok taking 2.5x as long to save $45mm on wholesale electricity costs. I'm not saying it's trivial or even easy, but industrial energy usage is huge already. Freeport LNG is also ~700MW. EV-ification and crypto are already huge stresses on the grid, I don't think AI data centers are the only story.