Meta Is Running Two AI Companies

There is a version of the Meta question that sounds naive and turns out to be the right one to ask.

Meta reported that 3.60 billion people used at least one of its apps each day in June 2026. In the second quarter it generated $60.8 billion of revenue, including $59.4 billion from advertising. It also spent $31.1 billion on capital expenditures, including finance-lease principal payments, and now expects $130 billion to $145 billion for the full year. Those figures come from the company's latest quarterly release.

Why does an advertising company need to spend at the scale of a semiconductor manufacturer? Why not remain the best possible advertising platform and let someone else finance the compute?

The answer is that a meaningful part of Meta's AI infrastructure is not adjacent to the advertising business. It is embedded in the systems that choose content, retrieve ads, predict outcomes, and automate campaigns. But Meta is also funding frontier models, a consumer assistant, enterprise agents, model APIs, and the option to sell compute. Those activities may eventually reinforce the core business or become businesses of their own. Their economics are not yet equally observable.

The central claim: Meta's AI-related spending supports at least two programs with different evidence of return, but its disclosures do not let investors quantify the cost or return of either program separately.

That is narrower than saying Meta publishes only one AI number: it does not publish a clean “AI budget” at all. Capital expenditure is consolidated, operating costs run through several income-statement lines, and infrastructure supports more than AI. The problem is not a mislabeled total. It is the absence of a usable allocation.

The advertising engine is part of the AI budget

Meta's capital spending is not pure frontier-model spending. The company says its second-quarter outlay was driven by servers, data centers, and network infrastructure. Its 2025 annual report says technical-infrastructure costs are allocated by estimated usage and that most are assigned to the Family of Apps segment. That segment includes both the mature advertising business and many of Meta's newer AI efforts, which is precisely where the analytical problem begins.

The case that core AI investment has economic value is strong. The case that public investors can calculate its return on invested capital is not.

Signal loss raised the value of modeling

Apple's App Tracking Transparency requirement took effect on April 26, 2021. On devices where a user does not grant permission, apps cannot access the advertising identifier or track the user across other companies' apps and websites under Apple's definition. Those are the mechanics in Apple's developer notice.

The original draft of this argument went too far by treating every missing conversion as something Meta simply replaces with accelerator-based inference. Meta's actual response was broader: it rebuilt parts of its ads infrastructure, introduced aggregated measurement, invested in privacy-enhancing technology, and used more modeling where direct signals had weakened.

The financial impact also needs precise language. In February 2022, Meta estimated that the cumulative effect of Apple's iOS changes—not ATT in isolation—would be a roughly $10 billion headwind to its 2022 revenue. Management explicitly called that an imprecise estimate in the Q4 2021 earnings call.

The defensible conclusion is not that ATT explains Meta's present infrastructure budget. It is that reduced third-party signals increased the strategic value of first-party data, better prediction, and privacy-preserving measurement inside the ad system.

The interest graph is a larger ranking problem

Meta's shift toward recommended short-form video changed the scale of content selection. In its own infrastructure history, the company contrasts ranking material connected to a person's few hundred friends with ranking from the much larger corpus of uploaded content. Meta says that candidate universe is orders of magnitude larger and identifies GPUs and other accelerators as part of the solution.

The latest results show why this matters economically. Meta attributed double-digit growth in Instagram time spent and 9% growth in Facebook video time spent to recommendation and ranking improvements. Those are company-reported attributions, not independently audited causal estimates, but they connect model performance to inventory that Meta can monetize.

Advertising automation is already an AI product

The strongest version of “focus on the ad platform” imagines a system where a business supplies a product and objective, then software generates creative, selects an audience, chooses placements, and optimizes the budget and bid. Meta is already building that system.

In its Q2 2026 earnings call, Meta said its Advantage+ end-to-end products had surpassed a $75 billion annual revenue run rate, while more than 9 million small businesses had used at least one generative-AI creative tool. It also reported specific experimental gains from newer recommendation models, including increases in ad clicks and conversions.

Those disclosures demonstrate a fast feedback loop: Meta can launch a ranking change, observe engagement or advertiser outcomes, and decide whether to expand it. They do not establish the aggregate return on AI capital. The $75 billion figure describes revenue flowing through products that use automation; it is not incremental AI revenue, and Meta does not disclose the infrastructure and labor cost required to produce it.

That distinction matters. The core program has measurable product outcomes and an existing monetization engine. Public disclosure still stops short of an investment return.

The second program

Then there is the program aimed at creating new platforms: Meta Superintelligence Labs, larger frontier models, Meta AI, business agents, model APIs, and related research infrastructure.

This category deliberately excludes AI glasses and other wearable hardware. Meta reports those products in Reality Labs, which already has separate revenue and operating-loss disclosure. Combining wearables with the undisclosed frontier-model program would obscure rather than clarify the accounting.

The second program is not devoid of customers or monetization paths. Meta distributes its assistant through apps used by billions of people, has begun rolling out business agents, and says it plans enterprise APIs and productivity products. What is missing is a separately disclosed revenue line, cost pool, or capacity allocation that would let an investor test those opportunities against what Meta is spending to pursue them.

	Core ranking and advertising	New model platforms
Current economic channel	Engagement and advertising performance	Consumer, enterprise, API, and compute options
Public evidence	Product experiments and existing ad revenue	Usage, early products, and management forecasts
Feedback horizon	Weeks or quarters	Potentially multi-year
Principal downside	Spending outruns incremental ad value	New products fail to earn their infrastructure and research cost
Cost disclosed separately	No	No

The table is an analytical framework, not Meta's internal org chart. Shared models, researchers, data centers, and software make any exact division judgmental. That is a reason to ask for management's allocation method, not to pretend an outside observer can reconstruct one.

The no-buyer argument no longer works

It would once have been tempting to distinguish Meta from Amazon, Microsoft, and Alphabet by saying that Meta had nobody to whom it could sell spare compute. After the latest quarter, that statement is too categorical.

Meta now says it is evaluating model APIs, enterprise services, and direct compute sales. Management said it had received offers for compute at prices above its cost and described direct monetization as one possible return path. On the follow-up call, however, the company also said it had no definitive timeline to share.

An offer is not revenue, and an option is not a business segment. Meta still does not disclose external compute revenue, contracted capacity, or unit economics comparable to a cloud provider. But the possibility weakens the claim that every accelerator must earn its return only through advertising or engagement. Any analysis published after July 29, 2026 has to account for that change.

The disclosure gap is real, but narrower

Meta already reports two segments: Family of Apps and Reality Labs. In 2025, it recognized 82% of total costs and expenses in Family of Apps and 18% in Reality Labs. Reality Labs produced $2.2 billion of revenue and a $19.2 billion operating loss. Those figures, along with Meta's allocation policy, appear in the company's 2025 Form 10-K.

That disclosure is useful, but it does not solve the question in this essay. Frontier models, new consumer AI products, and much of the related infrastructure sit alongside the advertising machine inside Family of Apps. The filing does not separate core recommendation and ads investment from new-model-platform investment within that segment. It also says Meta's chief executive, as chief operating decision maker, evaluates the two existing segments using revenue and operating income rather than asset or liability information.

Reality Labs is therefore evidence that Meta can report a major long-duration bet separately, but not proof that accounting rules require frontier AI to be a third segment. Reportable segments follow how management organizes and evaluates the business. The stronger ask is for supplemental disclosure: management-defined spending or capacity ranges, workload mix, direct AI revenue, and the assumptions used to allocate shared infrastructure.

This distinction changes the interpretation of silence. The absence of a split is evidence of limited investor visibility. By itself, it is not evidence that management is hiding an uneconomic program.

The honest countercase

There are serious reasons even the narrower framing could be wrong.

The programs may not be economically separable

Meta is developing foundation models intended to power organic content and ads recommendations together. Architectures, training systems, inference optimizations, and researchers can serve both the core business and new products. A forced allocation might create false precision.

The frontier work may improve the core directly

Large language models are already being used for content understanding, recommendations, advertising retrieval, and engineering. The supposed second program may be partly an advanced research function for the first. If so, its return appears in engagement and ad performance even before a standalone AI product makes money.

The physical assets have multiple uses

Servers and data-center shells can support different workloads, and Meta says it is designing later capacity with flexibility over future server decisions. Redeployment is not costless—chips, networking, cooling, and software are workload-sensitive—but failed product demand need not produce wholly stranded facilities.

New revenue channels are arriving

Business agents, paid messaging, APIs, productivity software, and possible compute sales could turn today's option value into reported revenue. The second program's economics are unproven, not necessarily absent.

More disclosure can have real costs

Capacity and workload data would reveal strategic information to competitors and suppliers. And because Meta does not manage frontier AI as a separate reportable segment, a new number would depend on internal allocation judgments that could move over time.

What would change the assessment

Meta publishes a stable allocation of infrastructure capacity or spending among core recommendations and ads, frontier training, new-product inference, and third-party use.
Direct revenue from model APIs, business agents, productivity products, or compute becomes material enough to disclose with associated costs or margins.
Meta links reported product gains to incremental infrastructure cost, making a return calculation possible rather than merely reporting engagement and conversion uplifts.
Older accelerators are shown moving between training, ranking, and external workloads without material loss of economic value.
Capital expenditure growth slows while core engagement and advertising performance hold, suggesting that the mature program is approaching sufficient capacity.

Primary sources

Meta Q2 2026 results — financial results, users, capital expenditure, and segment figures
Meta Q2 2026 earnings call — recommendations, ads products, AI strategy, capacity, and potential compute sales
Meta Q2 2026 follow-up call — workload mix, enterprise plans, and limits on the compute-sales timeline
Meta 2025 Form 10-K — segment policy, infrastructure allocation, annual spending, and risk disclosures
Meta's infrastructure evolution — company account of recommendation and model-training workloads
Meta Q4 2021 earnings call — iOS impact estimate and ads-infrastructure response
Apple App Tracking Transparency notice — effective date and policy mechanics

This essay is analysis, not investment advice. Company-reported product lifts are management disclosures, not independent audits of causality or return on invested capital. The distinction between the two programs is analytical; Meta does not report them as separate businesses.