The future of AI will be local. It will run on your phone and be smarter than you, and in fairly short order.

LLMs and the surrounding code that drives them ("harnesses") are so far from AGI that it will be decades before we see your prediction come true. We need not just one, but many more leaps forward in AI progress to achieve a level of AI that is capable of performing human-level reasoning and planning that will be "smarter than you" or me. If you have studied AI in the pre-LLM times you will have an understanding of the problems that need to be solved for general reasoning intelligence. We haven't solved any of those problems.

LLMs are a large step forward in the field of NLP, but that's it.

You may be correct from a technology perspective but you are forgetting how people work. These days people won’t even run their own cable modem or Wi-Fi access point. They would rather rent them from the ISP.

The human brain runs on 20 watts of power.

That's not at all correct. Or more appropriately, it's a true number used in a completely incorrect context.

It actually takes about 100 kilowatts to 1 megawatt to maintain a human brain, depending on your assumptions. You can't power your brain with electricity; you have to power it with food. And you can't power just your brain, you have to power your whole body. And producing food takes a ton of energy, way way more than people realize. It's just that most of it comes free from sunlight with crops grown in open fields. But if we're really talking about the energy needs of humans vs AI, the competition is between using those fields to grow food for humans to eat vs using those fields to hold solar panels to power AI.

I get this 0.1-1.0 MW figure from the YouTube and Nebula channel, "Science and Futurism with Isaac Arthur." There it was used in the context of deep space exploration and colonization, not terrestrial AI deployment. If you want to estimate the power requirements for life support for a hypothetical interstellar spaceship (just life support, not propulsion), you multiply the assumed crew number by 0.1-1.0 MW. If you're traveling between stars you have to bring your own sunlight with you. The vast majority of that energy is just for the huge amount of light needed to grow the food people eat. Interestingly, this life support requirement would still be a rounding error compared to the energy needs for propulsion. But if you want to estimate how big of a reactor you would need to sustain all of a person's needs without free sunlight, 0.1-1.0 MW is a decent figure.

For you as an individual deciding whether to pay to power an AI vs doing something yourself, sure, the 20 watt figure has some merit. You're going to have to eat enough to live no matter what you do. Maybe you burn a few extra calories while thinking hard on a problem, but your calorie needs are mostly fixed.

But when we're talking about whether it's cheaper to hire large numbers of humans or build large numbers of data centers to serve the needs of the ownership class, you have to look at the actual raw power requirements needed to keep a human being alive. Plant growth is an incredibly inefficient process. But computers can be powered extremely efficiently by just plopping some solar panels down in a field.