Also why talk about training not inference? That needs data centers too and could be what they're intending to do.
So this post is clearly not an effort to objectively work out the feasibility but just a biased list of excuses to support the author's unsubstantiated opinion.