Ideas?

• CO2 tax - if it'd be high enough to completely pay for the damage, this shit would stop pretty fast. But even less than that would help. Alternative: Certificates without loopholes. Some use would survive, e. g. an IT professional would still use $ 50 worth of energy per day if it gives a 10 % productivity boost, but models would start consolidating and use all tricks to keep it efficient, rather than push out whatever they can. Only works when imports from regions that refuse to participate are taxed when imported, or outright banned.
• Huge advantage of machine learning: The "when" is completely flexible. Could just use excess power from renewable peaks, or even nuclear & coal nightly production. But as long as it's cheap enough to just make more power around the clock: Why should they? They won't do it voluntarily. Solutions could start with a "green" label for consumers, but that would probably not do that much. It also won't help when we force them to use 100 % renewables and nuclear, and then they just buy all solar panels and wind turbines off the market leaving us with higher costs and trouble switching to net 0
• Evaluate the market and identify the bubble. Does an AI focussed company make conservative use of existing capabilities, without overhyping them, or put their money on likely near-future developments, or depend entirely on optimistic future capabilities?
• With such measures in place, we'd still have the models they trained so far. They'd eventually plateau anyway (or already have). When training of new models stops, as we make it too expensive to spend a lot of power for a tiny improvement, a good part of the power waste stops.