
Meta’s 2025 capex of $70-72 billion represents approximately 35% of revenue—unprecedented for a non-infrastructure company. Free cash flow is projected to compress from $54 billion to approximately $20 billion. These are not investment story metrics; they are stress indicators.
The Numbers
- 2025 CapEx: $70-72B (35% of revenue)
- Free cash flow: Projected drop from $54B to ~$20B
- 2026 outlook: “Notably larger” spending—analysts model $80-100B+
- Cumulative through 2028: $600B infrastructure commitment
The Infrastructure Constraint Framework
The AI revolution is not constrained by imagination or capital—it is constrained by atoms. This is the “bubble that bursts in slow motion”: demand exceeds supply at every layer of the stack, but the constraints are infrastructure, not capital.
Data centers take 2-3 years to build. Power plants take longer. TSMC can only fabricate so many advanced chips per quarter. Meta’s response follows the “buy to build” model: $14.3 billion for Scale AI, $30 billion in corporate bonds, $27 billion in private credit for Hyperion.
The Success Penalty
This is where the “success penalty” problem emerges: the better Meta’s AI products perform in market discovery, the more they are punished by capacity constraints. Infrastructure dependency creates a ceiling on scaling velocity regardless of product-market fit.
Meta admits its core business runs “compute-starved” because resources are redirected to frontier AI training. The company is simultaneously building for the future while cannibalizing resources from the present.
Financial Transformation
Meta has transformed from a “defensive” technology holding into an aggressive call option on the AI revolution. The capex explosion, FCF compression, and creative financing (SPVs, private credit, bonds) mean Meta is no longer a classic “cheap cash machine.”
Meta is now a high-risk, high-reward vertical integration bet, choosing to own atoms instead of renting them.
This is part of a comprehensive analysis. Read the full analysis on The Business Engineer.









