Meta Q2: $3.6B of Retrospective Charges Against a Raised Capex Floor
Meta reported second-quarter revenue of $60.8 billion, up 28% year over year, family daily active people of 3.6 billion on average for June, up 3%, and guided third-quarter revenue below consensus. The stock fell more than 8% after hours. Underneath the headline sit three numbers that explain the reaction better than the guide does: $2.4 billion in charges related to legal proceedings, $1.18 billion in severance connected to the May layoff of about 8,000 staff, and free cash flow of $784 million, down 91%.
A 28% revenue quarter that generates $784 million of free cash flow is a company converting essentially all of its operating cash into infrastructure and obligations. The capex range for 2026 moved to $130 billion to $145 billion from $125 billion to $145 billion, which is a raised floor with an unchanged ceiling. That is a specific kind of guidance change: it removes the low scenario. Management is telling the market that the minimum commitment is higher than previously disclosed while declining to raise the maximum, which reads as confidence in the programme and reluctance to reprice the top of it in a quarter where the stock cannot absorb the news.
The $3.6 billion of combined legal and severance charges is retrospective spending, and it is the cleanest part of the print. Legal charges settle behavior that already happened. Severance pays for a headcount decision already made. Neither recurs, and both are the sort of item a patient holder should be happy to see cleared. The problem is the sequencing: the company took $3.6 billion of costs for the past in the same release where it raised the floor on spending for the future, and the free cash flow line absorbed both at once.
The user number is the one that constrains everything else. Family daily active people up 3% to 3.6 billion is saturation, and it has been for several quarters. There is no meaningful audience growth left to buy, so all revenue growth is now price and load, which means all revenue growth depends on ad ranking improving faster than advertiser willingness to pay deteriorates. That is precisely the workload the capex is being spent on, and it is the only honest justification for the number. Recommendation and ranking models improving conversion for advertisers is a measurable return on compute in a way that almost nothing else in the AI capex debate is.
Reality Labs remains the counterexample. Revenue of $431 million, up 16% and slightly ahead of the $423.4 million estimate, against a $4.62 billion operating loss that came in better than the $5.07 billion expected. Beating a loss estimate is not progress. This is a division that spends roughly ten dollars for every dollar it brings in, and it has been doing so long enough that the market has stopped modeling a path and started treating it as a fixed tax on the advertising business. A better-than-expected loss in the same quarter as $1.18 billion of severance elsewhere in the company is its own comment on where the cuts landed.
On competitive position, the moat is distribution to 3.6 billion people and the ranking stack that monetizes them, and it is intact. No competitor can assemble that audience, and the switching cost for advertisers is the measured performance of the auction rather than any contractual lock. The genuine risk is not a rival network. It is that the ranking improvements which justify $130 billion of annual spend are subject to diminishing returns that would show up in revenue per user long before they show up in any disclosed model metric.
The number that decides the next four quarters is incremental revenue per incremental dollar of capex, tracked against the depreciation schedule those assets carry. A raised floor with unchanged revenue guidance is only defensible if that ratio is holding.