There's lots of ways to read this. The way I read it, Meta is trying to keep Llama from getting "poisoned" with ITAR-controlled data.
Why? Because, oh boy, once a word or sentence or paragraph is classified as ITAR technical data, that little nugget is a bomb waiting to go off in your system. ITAR-classified data can't so much as look at a wire that's outside of the borders of the United States, which, somewhat obviously, throws a cinder block into most systems designs for cloud stuff. And if it's one of these newfangled LLMs, there's no way to take the system apart to get the nugget out. That is a LOT of compute wasted.
There's a reason that GovCloud and other ITAR clouds are awesomely crippled versions of their original incarnations - they have to audit every scrap of every environment and tool available with a fine-f#@king-tooth-comb. A lot of software vendors I know gave up on their GovCloud ports, and just send a team to do the install on-prem. It turned out to be a lot less work than maintaining a special GovCloud build that you have zero idea will work at rollout. If the customer is ok with Docker on-prem, this is an even easier decision. GovCloud storage and emails, yeah, that's fine. Anything more complicated than that, oof - it's going to be its own completely forked build, or else you're going to be going nuts with flags and such.
Now, ok, not going to name names, but there are some vendors providing "AI" in the ITAR space, but, IMHO, I'm extremely skeptical that they have gone through the right channels. I would tread with extreme caution here. Their public faces are quite familiar from some other DoD IT fiascoes from the last few years, and on their DoD LLM thing they've been, at best, disingenuous when questioned about the nature of the waivers they've "acquired". At worst they're outright lying. They seem to be working through informal channels most of the time with highly-placed connections, but that's a very brittle way to do this sort of business. You probably know who I'm talking about here. Anyway, state of nature right now, the safe way for experimenting with LLM in defense is still on-prem.