In general I see absolutely no problem with AI "capturing" some of the data center power that was used for mining. The difficulty mechanism works as always. For mining companies, it may also be a good way to diversify if the Bitcoin price gets too low and difficulty hasn't adjusted enough yet. Bitcoin's security is high enough, it may even be too high actually, due to the mining bubble of the last 2-3 years which made big mining companies invest too much in mining hardware. A continuation of the slight difficulty decrease we've seen since late 2025 would help existing miners a bit, even if in the longer term the relation between profits and difficulty would return to an equilibrium.
However, I see also a high likelihood for the AI boom at least to cool down a bit. I don't know if it's really a "bubble", because LLMs and diffusion models (graphics/music) are a powerful technology. And thus I expect their usage actually to grow some years more.
But the extreme "takeover" scenarios ("we will not need to work anymore in 2028" or even "AI bots will kill all humans in 2030"), which are of course obvious marketing stunts, are even further away from reality as many think.
LLM power increase is for example slowing down (see articles like
this one). Their basic flaws like hallucinations are not easy to remove. And thus, even if they can replace some work, the growing suspicion that they will "not" replace most work in the next decade will make lose the AI boom steam, perhaps already in 2026. And then profit margins for AI datacenters will decrease, and it will make Bitcoin mining more attractive (in relative terms) again.
