Indeed. Anyone familiar with public company financials and budgeting (like Wall Street analysts) know any number like this is only a broad estimate at best. For example, is that number counting the full value of capital expenditures (such as GPUs) immediately amortizing that cost over the asset's life? Most accounting methods would say capex should be depreciated over time but in many kinds of purchase transactions NVidia may book the full revenue in the quarter they deliver the card, so in that sense, the money is "in the economy" which may be more appropriate given the context.
And that's just a simple example. It gets much more complex and there are lots of judgement calls about how multi-use assets are booked and employee costs are apportioned across business units. It's likely no one at Meta even has a credible budget number for overall Meta-wide "AI spend over the next three years." If you ask the CFO to come up with a number, they're going to assign a team to make a best estimate under "given assumptions" - and then that team is going to ask you 20 or 30 questions about how you want to count. So... it's understood that numbers cited at a photo opp like this are SWAGs at best. Plus, even if finance penciled out an estimate, that estimate would certainly change substantially during the yearly budget cycle and change meaningfully quarter to quarter. So, such a snapshot in time is pretty worthless in terms of actual predictive value.