What Meta actually reported
Meta posted second quarter revenue of $60.8 billion, up 28 percent year over year and ahead of the roughly $60.2 billion consensus. Earnings per share came in at $6.18 against an expected $7.14, and operating income fell 8 percent. The Family of Apps segment, the part of the business that funds everything else, saw operating income drop to $23.4 billion from $25.0 billion. The number that moved the stock roughly ten percent lower was further down the statement: free cash flow of $784 million, against $8.55 billion in the same quarter a year earlier. Capital expenditure reached $31.1 billion for the quarter and $50.9 billion for the first half.
Management then raised the floor of full year capex guidance to $130 billion from $125 billion while leaving the ceiling at $145 billion. Narrowing a range upward is a specific signal. It says the company has retired the scenarios in which it spends less, and that whatever uncertainty remains sits on the high side. For an operating business with Meta's margins, deliberately compressing free cash flow to under a billion dollars in a quarter is a statement about conviction. The market read it as a statement about discipline, which is a different thing, and priced accordingly.
The quote that matters to enterprise buyers
Buried in the call was the most useful piece of market intelligence any infrastructure buyer received this week. "We're getting a lot of offers for compute at a significant premium for what we paid for it," Zuckerberg said. He then explained why Meta is turning those offers down: "It would be foolish to basically just sell all of the compute and take a short-term profit." Set aside what this means for Meta's strategy for a moment and consider what it tells you about the market. Buyers are approaching a social media company, unsolicited, offering above cost for accelerated compute capacity.
That is a scarcity signal from a source with no incentive to exaggerate. Nvidia has reason to describe demand as insatiable. The neoclouds have reason to talk up utilization ahead of financing rounds. Meta gains nothing from telling the market that its own compute is worth more than it paid, other than justifying the capex line that just halved its free cash flow. When a company that does not sell cloud services confirms it is receiving premium bids for capacity, the shortage is real and it extends beyond the handful of vendors who benefit from saying so. Plan your 2027 accelerator access on that basis.
Hoarding compute is a competitive strategy now
Zuckerberg's reasoning was explicitly about time horizon. "My personal bet is that the people who invest in this are going to be rewarded and feel very good over time," he told investors. Meta has concluded that the option value of holding compute exceeds the cash it could realize by renting it, even at a premium. That calculus only works if you believe the capability frontier keeps moving and that access to training and inference capacity is the binding constraint on getting there. Meta is buying insurance against being compute limited at the exact moment a capability step change makes compute decisive.
Enterprises face a smaller version of the same decision every quarter, usually without framing it that way. The question of whether to commit to reserved capacity you may not fully consume is the same option value question at a different scale. We have generally advised against over provisioning, and in a normal market that advice holds. In this market, the cost of being short capacity during a strategically important launch is considerably higher than the cost of carrying some unused reservation. The largest and best informed buyers in the world are choosing to hold excess capacity rather than monetize it. That is worth weighing against the instinct to run lean.
Debt financing changes who carries the risk
CFO Susan Li told investors that Meta is moving toward a greater mix of debt to fund infrastructure with long term passive income potential. That is a material shift for a company that has historically funded expansion from operating cash flow. Debt financed data center construction spreads the cost across the useful life of the asset, which is defensible accounting for buildings and power infrastructure. It is a harder argument for the accelerators inside them, where the depreciation schedule is contested and the technology refresh cycle is measured in a small number of years rather than decades.
This matters to enterprise technology leaders for two reasons. First, debt financed capacity has to earn a return on a schedule set by lenders, which puts a floor under how cheaply that capacity can be offered to the market. Second, the entire sector is now converging on similar financing structures, from Meta to the neoclouds to the utilities building generation to serve them. Correlated financing across an industry creates correlated fragility. If AI revenue disappoints relative to the buildout, the correction will not be confined to one balance sheet. Any vendor risk assessment written before this quarter should be revisited with that in mind.
The cloud question Meta has not answered
Zuckerberg's comments hinted at something the company has resisted for years: selling compute to third parties. Fielding premium offers and declining them is a decision that can be revisited, and Meta now owns an infrastructure estate comparable to a major cloud provider. If ad revenue growth moderates while capex stays above $130 billion annually, the pressure to monetize that estate directly becomes hard to resist. A fourth serious hyperscale seller entering the market would be the single most useful development for enterprise buyers in this cycle, and the most disruptive for pricing at the incumbents.
We would not build a procurement strategy on that possibility yet. Meta has given no timeline, no product, and no commercial terms, and the current posture is explicitly to keep the capacity for internal use. What CIOs should take from it is a read on how the supply side may evolve over a three year horizon. The set of organizations capable of selling accelerated compute at scale is larger than the set currently doing so, and the economics that keep them out are shifting. That is a reason to keep contract terms flexible past 2027 rather than a reason to delay commitments today.
What this quarter should change in your planning
The concrete takeaway is about sequencing. Meta, Amazon, and the rest of the hyperscale buyers are placing multi year orders now for capacity that arrives in 2027 and 2028. Enterprise buyers who wait for prices to normalize before committing are effectively queuing behind everyone who committed earlier. If your AI roadmap has a material 2027 component, the reservation conversation belongs in this planning cycle, and it belongs with finance in the room rather than as an engineering line item.
The second takeaway concerns how you evaluate vendor claims about capacity. Every provider in the market will tell you their supply is constrained, because scarcity is a useful sales posture. Meta's disclosure gives you an independent reference point: a large, sophisticated, non selling operator confirming premium bids for compute. Use that as the benchmark when a vendor tells you the window is closing. The shortage is genuine at the level of leading edge accelerators and the memory that feeds them. It is considerably less genuine for general purpose compute, storage, and the older accelerator generations that will run most enterprise inference workloads perfectly well.
