Chapter 5

AI and Reality Labs

Meta's reinvestment is not one bet but three, with three different return profiles. Reality Labs has lost about $88 billion since 2019 on revenue that has never cleared $2.3 billion a year, with no disclosed return milestone. AI infrastructure — capital expenditure of $69.7 billion in 2025, guided to $115–135 billion in 2026 — mostly serves the ad engine, where the return is measurable (Advertising Moat), but is now being locked into $237.7 billion of non-cancelable commitments. The mechanism that will test all of it, a depreciation wave, has barely begun.

Sources: FY2025 10-K, Segment Results [1]; Q4 & FY2025 earnings release [2]; Q1 2026 10-Q, Notes 6 and 8 [3] and Material Cash Requirements [4].

The report has established that the ad franchise is durable and that a founder controls it without an outside check (Control and Pay). This chapter sizes what that control is spending on, and how much of it can still be reversed if the return does not come. Meta itself flags the risk in plain terms: it expects expenses to rise as it invests in "new and unproven technologies, including AI and machine learning," and states that its AI and Reality Labs investments "have the effect of reducing our operating margin and profitability," warning that "if our investments are not successful longer-term, our business and financial performance will be harmed" [5].

Reality Labs: seven years, $88 billion, flat revenue

Reality Labs is the oldest and cleanest part of the bet, and the least productive. The segment houses Meta's virtual, augmented, and mixed-reality hardware, software, and content. Its operating loss has widened every year — from $4.5 billion in 2019 to $19.2 billion in 2025 — while segment revenue has stayed close to $2 billion the entire time, ending 2025 at $2.2 billion [6].

Sources: FY2025 10-K, Segment Results (2024–2025) [7]; FY2023 10-K, Segment Profitability (2021–2023) [8]; FY2021 10-K, Segment Profitability (2019–2020) [9].

Summed across 2019 through 2025, the operating losses total roughly $88.1 billion — $4.5B + $6.6B + $10.2B + $13.7B + $16.1B + $17.7B + $19.2B — against cumulative segment revenue of little more than $10 billion over the same span [10]. The 2025 loss alone equals 23% of consolidated operating income of $83.3 billion [11]. No filing discloses a revenue target, a break-even date, or a spending ceiling for the segment; the only stated discipline is the risk-factor caveat that the investment may not succeed [12].

This is the part of the spend that most cleanly fits the through-line's "no proven return." It is also the most reversible: Reality Labs is largely operating expense and internal development, funded out of the Family of Apps' $102.5 billion of operating income [13]. If the founder chose to cap it, the loss would compress within a year or two. The AI build, examined next, is a different matter.

The capex ramp, and a depreciation wave that has not landed

The larger number is infrastructure. Capital expenditure roughly doubled to $69.7 billion in 2025, and management guides 2026 spending to $115–135 billion, "with year-over-year growth driven by increased investment to support our Meta Superintelligence Labs efforts and core business" [14]. Unlike Reality Labs, most of this serves AI systems that already earn their keep on the ad side — the automated ad tools and ranking models detailed in Advertising Moat.

Source: derived from reported financials, FY2021–FY2025 Consolidated Statements of Cash Flows [15]; 2026 shown at the $125B midpoint of the $115–135B guidance range [16].

The gap between the two bars is the point. Capex has raced ahead of depreciation because most of the new spend sits in construction-in-progress and equipment not yet placed in service. Depreciation and amortization rose from $11.2 billion in 2023 to $18.6 billion in 2025 — real, but a fraction of the assets being built [17]. As that equipment switches on, the charge accelerates. In the first quarter of 2026, depreciation of property and equipment reached $5.68 billion, up from $3.84 billion a year earlier; the servers-and-network component alone rose 67%, to $4.38 billion [18].

Management has named this as the dominant cost pressure ahead. The CFO told investors that infrastructure would be "the single largest contributor to 2026 expense growth… driven primarily by a sharp acceleration in depreciation expense growth," and — importantly for how fast the charge builds — that Meta expects "a greater mix of our CapEx to be in shorter-lived assets than in prior years" [19]. Shorter asset lives mean the depreciation lands sooner and heavier. This is what makes AI infrastructure structurally different from Reality Labs: the spend converts into a fixed operating cost that persists whether or not the incremental revenue arrives.

The "superintelligence" framing sits on top of this. In mid-2025 the company created Meta Superintelligence Labs and Zuckerberg defined the goal as "developing superintelligence, which we define as AI that surpasses human intelligence" [20]. Beyond hardware, that ambition carries a compensation cost: the CFO cited a full year of expense for "the AI talent that we're hiring this year" as the second-largest driver of 2026 expense growth after infrastructure [21]. This is the least-proven slice: it shares the ad engine's compute, but its own return is a call option with no disclosed milestone.

From discretionary to committed

The most consequential change of the past eighteen months is not the size of the spend but its reversibility. Historically, capex was a year-by-year choice. Increasingly, it is a contract. Meta's non-cancelable contractual commitments — mostly third-party cloud capacity, servers, network infrastructure, and data centers — have climbed from $32.1 billion in September 2024 to $131.0 billion at the end of 2025 to $237.7 billion by March 2026 [22] [23] [24].

Sources: Q3 2024 10-Q, Note 12 [25]; FY2025 10-K, Note 11 [26]; Q1 2026 10-Q, Material Cash Requirements [27].

That figure understates the forward obligation. On top of the $237.7 billion, Meta disclosed $182.9 billion of operating and finance leases that had "not yet commenced" as of March 2026 — data centers, colocations, and network infrastructure with terms running as long as 30 years [28]. The pace has not slowed: in April 2026 the company added roughly $24 billion more in non-cancelable commitments, and it carries a contingent obligation to buy up to $14.7 billion of cloud capacity over five years [29]. Near-term, $42.3 billion of the commitments falls due in 2026 and $47.7 billion in 2027 [30].

Two smaller disclosures show how the build is being structured. Meta reclassified $5.0 billion of money-market funds to restricted cash tied to a multi-year purchase agreement [31], and in March 2026 it moved $1.48 billion of data-center assets — mostly construction-in-progress and land — to held-for-sale, to be contributed to a third party "for the purpose of co-developing data centers" [32]. The co-development structure keeps some of the capital off Meta's own balance sheet, but the capacity — and the obligation to pay for it — remains.

The practical consequence: a growing share of the AI build can no longer be dialed back by a change of heart. Reality Labs remains discretionary opex; a large and rising portion of the infrastructure spend is now legally committed for years. Set against Control and Pay, the accountability gap sharpens — the capital that outside holders cannot vote on is increasingly capital that management itself cannot easily unwind.

The read, the counter-case, and what would change it

The evidence points to a spend that is real, escalating, and — on the infrastructure side — hardening into multi-year contracts faster than its return is proven. Reality Labs is a self-funded, discretionary $88 billion write-off with no disclosed payback. AI infrastructure is more defensible because it powers the ad engine, but it is being committed at a scale ($237.7 billion contracted, plus $182.9 billion of pending leases) that will convert into a depreciation and cloud-cost load carried regardless of whether the superintelligence ambition pays off.

The strongest fact against a bearish read comes from management's own guidance: despite "the meaningful step up in infrastructure investment," Meta expects 2026 operating income to come in above 2025's [33]. If revenue growth stays ahead of the depreciation curve, the wave is absorbed and the build looks prescient rather than profligate. The spend is also fully self-funded — $115.8 billion of operating cash flow and a net-cash balance sheet (Financials and Estimates) — so this is a returns question, not a solvency one.

What would move the read, in either direction, is checkable in the filings:

Watch, alongside it, three lines that resolve the question before the income statement does: depreciation of property and equipment, now running at $5.68 billion a quarter and climbing [34]; the non-cancelable commitment balance, $237.7 billion and still compounding [35]; and the Family of Apps operating margin, which already slipped from 54% to 52% in 2025 as segment costs grew faster than revenue [36]. Whether the AI build is defending the moat or diluting it will show up there first.