Pose the question plainly: why does "doing AI" make a company's compute bill explode, when running a website or an app didn't? The answer was written into Meta's filings before the boom made it obvious.
An earlier Meta 10-K stated that "as we deepen our investment in new technologies like artificial intelligence, our computing needs continue to expand" — and that the company would "make significant investments" as a result. That foresight, surfaced via EdgarBeast, is the whole AI-infrastructure story compressed into one sentence.
The way this actually works: traditional software is mostly retrieval and bookkeeping — fetch a post, store a like, render a page. The compute per action is small and roughly constant. AI breaks that. Training a model is a one-time mountain of math, and inference — generating a recommendation, an image, an answer — costs orders of magnitude more compute per request than serving a static page.
So the more AI you put into products, the more your per-action compute cost rises, and the more hardware you need to keep latency acceptable. Meta's FY2025 report keeps AI at the center of its product roadmap — see the sec.gov filing — which is exactly why the "computing needs continue to expand" line from years earlier reads as a prediction that came true.
The takeaway for a general reader: the data-center construction frenzy isn't a fad bolted onto these companies; it's the mechanical consequence of moving from retrieval-shaped software to inference-shaped software. Meta told its investors this would happen. You can read the current framing in the sec.gov filing, and treat the compute buildout as the bill arriving for a decision disclosed long ago.