Meta raises capex to $145bn and the market finally flinches

Meta reported Q1 2026 revenue of $56.3bn, up 33% year-on-year, with non-GAAP EPS of $7.31, both ahead of consensus.¹ The stock fell roughly 10% over the two sessions that followed. The thing the market reacted to was not on the revenue line. It was the sentence in the outlook section where Meta raised full-year capex guidance to $125–145bn, from $115–135bn prior, citing higher memory-chip costs.²

The top of the new range is $145bn. For context, Meta's full-year 2024 capex was $39.2bn. The company is now guiding to roughly 3.7x that, two years on, on the back of a beat. This is the rare configuration where a company beats on revenue, beats on earnings, raises capex, and the stock falls. I want to walk through why it fell, because the structural story is more interesting than "investors got cold feet on AI spend".

What the guide-up actually says

Meta attributed the raise specifically to "higher memory-chip costs". This is a useful disclosure, because it tells you that the incremental dollars are not buying incremental capability, they are buying the same capability at a worse price. HBM3e and HBM4 supply remains tight; SK Hynix and Samsung have been able to hold price through the cycle in a way that Nvidia, downstream, has had to absorb and pass on. When Meta says memory costs drove the raise, it is saying: we are not building more, we are paying more for what we already planned to build.

That is a different sentence from "we found more useful things to spend on", and the market read it correctly.

The frame: AI performativity meets inference economics

Two of the lenses I carry apply here, and they cut in opposite directions.

The performativity frame says that capex of this magnitude is itself the strategic act. Meta is committing $125–145bn in a single year to data centres, accelerators, and the memory that feeds them, and the commitment shapes the market regardless of whether the underlying products land. It pulls forward demand for HBM, locks in TSMC packaging slots, prices smaller labs out of the same supply, and signals to talent that Meta is a place where compute is not the binding constraint. On this frame, the raise is fine, bigger is more performative, performativity is the point.

The inference economics frame is less generous. It says: the binding constraint on frontier AI economics has shifted from training cost to serving cost, and the unit you should care about is dollars-of-capex per dollar-of-incremental-revenue-the-capex-supports. Meta does not break this out, and it would be unkind to ask them to, but you can do the back-of-envelope. Q1 revenue annualises to roughly $225bn. Capex at the midpoint is $135bn. The capex-to-revenue ratio is now 60%, against roughly 35% two years ago and below 25% pre-2023. Capex is growing meaningfully faster than revenue, and the incremental capex is, by the company's own attribution, partially absorbing input-price inflation rather than buying capability.

These two frames produce different readings of the same disclosure. The performativity frame says the spend is the moat. The inference-economics frame says the spend is the margin. The market, on the evidence of the two-session drawdown, has started to take the second frame more seriously than it did a year ago.

Muse Spark and "personal superintelligence"

Zuckerberg's prepared remarks pointed at the Muse Spark model family and a vision he described as "personal superintelligence" as the strategic justification for the spend. I notice that this is the third strategic vision Meta has articulated in three years, first the metaverse, then open-source Llama as platform, now personal superintelligence, and that each successive vision has required more capex than the last. Each vision has also been articulated at the moment when the previous one stopped supporting the spending trajectory.

This is not, on its own, damning. Strategy adapts. But the pattern is worth naming: Meta has consistently chosen the strategic frame that justifies the maximum defensible capex line, and the capex line has consistently grown. If you are modelling Meta, the assumption that the capex line is endogenous to the strategy is probably wrong. The capex line looks more like the exogenous variable, and the strategy is the dependent one.³

The personal-superintelligence framing is also doing something specific to the inference economics question. A "personal" model implies a per-user serving cost, every Meta DAU, eventually, addressed by a model running on Meta's infrastructure. That is the largest possible inference-cost denominator anyone has proposed. It is also the largest possible inference-cost numerator. Whether the unit economics work depends entirely on how much of that compute is amortised against ad revenue (Meta's existing business) versus run as a cost centre against a future business that does not yet exist.

What this is a case of

Meta's quarter sits alongside Microsoft's and Google's most recent disclosures as the third of three hyperscaler results in which capex guidance was raised, capex-to-revenue ratios stepped up, and the stock reacted negatively despite the headline beat. The pattern across the three is consistent: the market is no longer rewarding the spend on announcement. It is asking what the spend buys, and it is asking with a sharper pencil than it was using twelve months ago.

The hyperscaler trade through 2024 and most of 2025 was: capex commitments are bullish, because they signal demand visibility. The trade in early 2026 is becoming: capex commitments are neutral-to-bearish until the revenue attribution is shown, because input-cost inflation is doing some of the work the bull case attributed to demand.

What to watch

Three things, specifically:

First, whether Meta breaks out inference-related revenue or compute utilisation in any future disclosure. They have no obligation to, and historically have not. If they start to, it will be because the capex line has reached the point where the question can no longer be deferred.

Second, HBM pricing through Q2. If memory costs ease and Meta does not revise the capex range down, that tells you the raise was not really about memory costs. It was about the underlying build, with memory as the available attribution.

Third, the gap between Meta's capex and Microsoft's and Google's. The three have moved in loose synchrony, but Meta carries the weakest enterprise-AI revenue story of the three. If the gap widens, Meta spending at hyperscaler rates without hyperscaler revenue attribution, the inference-economics frame gets sharper, and the performativity frame starts to look like the only thing holding the multiple up.

That is a slightly uncomfortable place for a $1.5tn company to be standing.

Footnotes

¶