Meta’s October 19 spending guide is the kind of number that makes a product announcement look small. Capital expenditure for 2026 is penciled in at $130 billion to $145 billion. More than $100 billion of it is for the machines that train and serve AI.
The comparison that matters
In 2025 the same line was $72 billion. The jump is not a nicer office. Meta described data centers and custom silicon, the two things you buy when you have decided rented GPUs are not the whole plan. Two days later the company reported $64.6 billion in third-quarter revenue and said it was creating a non-voting Class C share. The revenue says the ads business can currently carry a frightening capex line. The share class is a separate argument about who votes.
Prometheus, Meta’s training cluster, is already in the story the company tells. It says that cluster trained Muse, and that the build is heading toward a million GPUs. The capex guide is how you pay for a sentence like that.
What a reader should take from it
You will not see this money as a new button in Instagram tomorrow. You will see it as more AI features that assume a huge, always-on model, and as a company that has to keep ad margins fat enough to justify the bill.
If you sell power, land, or chips, Meta just published your next two years. If you use the apps, the only honest reaction is to watch which features actually ship. A capital plan is not a product.