He said/implied that we would have shortage either way (without data center buildout, but everyone who wants access to LLM does private/local inference), which doesn't sound far fetched.

Single handedly? I'd say the the people using (or forcing others to use) the models, therefore justifying the need for them, play a part in it.