RESEARCH MEMO · Investment Intelligence / AI Infrastructure
Meta Earnings: Advertising Grows 26%, but Is AI Capex Consuming Cash Flow?
Investment Intelligence | AI Infrastructure | Research Memo and Earnings Deep Dive | 23 October 2026
Coverage: Meta Platforms (META), Arista Networks (ANET) and the AI supply chain. All monetary amounts are U.S. dollars. Research as of 23 October 2026; not personalized investment advice.
Capital investment exceeded quarterly operating cash flow. That is a consolidated result, not an AI-only attribution. Advertising strength and shareholder cash returns require separate tests.
$63.061bn / $50.082bn − 1, rounded to 26%.[1–2]
$32.786bn operating cash flow less $33.942bn cash PP&E and $0.964bn finance lease principal.[1–2]
$7.40 / $7.25 − 1 after respective litigation and tax adjustments; not fully normalized earnings.[1,5]
Research view: stronger operating evidence; capital efficiency remains unproven. Existing ratings and real positions are unchanged.
01 · RESEARCH BRIEF
The one-minute brief
Meta Platforms (META), an advertising-led internet platform expanding into AI, reported $63.061bn of advertising revenue, up about 26%. Operating cash flow of $32.786bn fell short of $34.906bn of capex including finance lease principal, leaving negative $2.120bn of free cash flow.[1–2] The cash absorption is real; it does not establish that advertising has broken. Litigation this year and a tax charge last year distort EPS growth. Excluding those respective items, EPS rose from $7.25 to $7.40.[1,5] Our assessment is mixed: stronger evidence of advertising competitiveness, weaker visibility into shareholder cash returns. The next earnings report must connect the 2027 budget, depreciation and new AI revenue. Supplier benefits require their own order and revenue evidence.
Listen to the audio summary
Audio transcript
Meta’s third-quarter advertising grew about twenty-six percent, but cash returns face a harder test. Operating cash flow was about thirty-two point eight billion dollars, below capex of thirty-four point nine billion including finance lease principal. Free cash flow was negative two point one billion. These are consolidated figures, not AI-only attribution. Litigation this year and a tax charge last year distort earnings growth. Excluding those respective items, EPS increased only from seven dollars twenty-five to seven dollars forty. AI already improves the ad engine, while agent monetization and capital returns need more evidence. Management has not provided numerical 2027 capex guidance. Class C expands funding options; it does not mean cash has already been raised. ANET, compute, networking and power suppliers require independent order and revenue validation. The next report must connect budgets, depreciation, cash coverage and new AI unit economics.
Known facts and open questions
- Disclosed facts
- Financial tables and the SEC earnings exhibit support reported revenue, cash flows and segment results.[1–2]
- Management guidance
- 2026 capex of $135–145bn; aggressive investment continues in 2027, without a numerical range.[1,3]
- Analytical inference
- Cash absorption, adjusted-margin compression and conditional supply-chain transmission are ACIS calculations and judgments.
- Scenario assumptions
- Valuation and cash-flow sensitivities are neither company guidance, consensus forecasts nor price targets.
- Evidence limits
- The SEC earnings exhibit and 8-K were verified. The Q3 10-Q was not independently verified; asset lives, commitment schedules and AI-specific returns remain unresolved. Media-reported consensus lacks an original estimate timestamp.
Operating cash generation
$32.786bn, up 9.3% year over year.[1–2; calculation]
Cash equipment purchases
$33.942bn versus $18.829bn a year earlier.[1–2]
Finance lease principal
A further $0.964bn; do not omit it or double-count it as operating expense.[1–2]
Residual cash
FCF of −$2.120bn before $1.361bn of quarterly dividends.[1–2]
| Metric and basis | Q3 2026 | Q3 2025 or benchmark | Interpretation |
|---|---|---|---|
| Total revenue | $64.601bn | $51.242bn; +26% | About 2.1% above media-reported Visible Alpha consensus of $63.3bn.[1–2,7; calculation] |
| GAAP operating income / margin | $12.988bn / 20.1% | $20.535bn / 40.1% | Includes $9.89bn litigation charge; not all AI expense.[1–2; calculation] |
| Ex-litigation operating income / margin | $22.878bn / 35.4% | $20.535bn / 40.1% | Income +11.4%, but margin down about 4.7 percentage points.[1–2; calculation] |
| GAAP net income / diluted EPS | $9.860bn / $3.83 | $2.709bn / $1.05 | Distorted comparison; see earnings-quality bridge.[1–2,5] |
| Operating cash flow / capex | $32.786bn / $34.906bn | $29.999bn / $19.374bn | Cash flow +9.3%; capex +80.2%, including lease principal.[1–2; calculation] |
| Depreciation and amortization | $7.154bn | $4.963bn; +44.1% | Consolidated D&A, not verified AI-server depreciation.[1–2; calculation] |
| Item | Bridge or range | Research treatment |
|---|---|---|
| Q3 2026 EPS | GAAP $3.83 + $3.57 after-tax litigation impact = $7.40 | Company-provided single-item adjustment, not fully recurring EPS.[1] |
| Q3 2025 EPS | GAAP $1.05 + $6.20 one-time tax charge = $7.25 | Like adjustment comparison grows about 2.1%, versus GAAP growth of 265%.[5; calculation] |
| Tax quality | 24% reported; 17% ex-litigation; 14% prior year ex-tax charge | Tax rates still differ; stock compensation remains an economic expense.[1,5] |
| Shares and non-operating items | 2.573bn diluted shares vs 2.572bn; interest/other income $0.013bn vs $1.128bn | Shares broadly stable; non-operating income fell. No invented fully clean EPS.[1–2] |
| Q4 revenue guidance | $73–76bn; midpoint $74.5bn | About 1.0% above media-reported $73.75bn Visible Alpha consensus; not annual guidance.[1,7; calculation] |
| 2026 capex guidance | $135–145bn, previously $130–145bn | Lower bound +$5bn; midpoint rises from $137.5bn to $140bn.[1; calculation] |
| Assumption | Low multiple or spending | Middle | High multiple or spending |
|---|---|---|---|
| Normalized annual EPS $28 × P/E 20 / 25 / 30 | $560 | $700 | $840 |
| Normalized annual EPS $32 × P/E 20 / 25 / 30 | $640 | $800 | $960 |
| Normalized annual EPS $36 × P/E 20 / 25 / 30 | $720 | $900 | $1,080 |
| Annual operating cash flow $140bn less capex $140 / $170 / $200bn | FCF $0 | FCF −$30bn | FCF −$60bn |
02 · FACTS → IMPACT → VIEW
Why does this change matter?
Business evidence and stock evidence diverge
- What is confirmed
- Advertising grew about 26%, operating cash flow about 9%, and capex about 80%.[1–2; calculations]
- Why it matters
- Revenue growth did not automatically produce free cash flow. Ex-litigation operating margin still declined.
- ACIS view
- The business view is constructive; capital allocation remains conditional. Revenue estimates have upward support, but EPS and FCF require cost and 2027 budget evidence rather than an automatic parallel upgrade.
03 · EVIDENCE & ANALYSIS
Evidence and analysis
01｜Advertising strength and the limits of AI attribution
Advertising rose from $50.082bn to $63.061bn and represented 97.6% of revenue. Impressions increased 10% and average price 14%. Their product implies about 25.4% growth; rounding and mix explain why this is not an exact bridge. Adding the two rates is incorrect.[1–2; calculations] Management attributes gains to AI-driven engagement, recommendations and monetization. It reported nearly 2% higher conversions from generative retrieval, plus 1.7% higher offsite conversions and 9.6% higher app-ad conversions from GEM expansion.[3, p6] These are specific product or experiment metrics, not additive contributions to consolidated revenue. A dollar bridge for incremental AI advertising revenue was not disclosed.
There is operating evidence that AI improves the existing ad engine. New agents, model services and enterprise offerings remain a different earnings proposition. The follow-up call clarified that Muse currently monetizes through subscriptions; transaction monetization requires greater user and partner scale.[4, pp2,5] Core-ad efficiency and new-AI revenue must be modeled separately to avoid double counting. A credible future bridge needs adoption, paid conversion, retention, realized revenue and serving costs on consistent cohorts. User growth alone cannot establish profitable scale. Management targets more than tenfold compute-cost efficiency improvement over the next year; the follow-up stresses that further technological advances are needed. This is a target, not realized savings, and lower unit cost does not ensure lower total spending.[3, p8; 4, p6]
02｜Three gates for cash flow depreciation and returns
Quarterly FCF was negative, but nine-month FCF remained positive at $11.049bn. One quarter does not establish a negative full year.[1–2] Capex consumes cash before commissioned assets create depreciation expense. Net PP&E rose from $176.400bn at the end of 2025 to $253.222bn, expanding the asset base whose utilization and future costs must be tested.[1–2] Procurement, construction, commissioning and utilization occur at different times. Consolidated D&A is not AI-only depreciation and cannot establish server lives.
Return on invested capital, or ROIC, compares after-tax operating profit with invested capital. AI-specific profit, capital and utilization are not fully disclosed; an exact incremental AI ROIC would be false precision. First test incremental ad profit, agent gross profit and serving-cost efficiency, then compare them with cumulative investment and the cost of capital. Litigation is isolated for earnings analysis, but the agreement provides for payments up to $18.59bn and remains economically consequential.[1] It cannot be permanently excluded from cash projections. Working-capital movements and non-cash add-backs also affect operating cash flow, so it is not the same as unrestricted shareholder cash.
03｜Reality Labs and the 2027 funding structure
Reality Labs generated $0.443bn of revenue and a $4.129bn operating loss, compared with $0.470bn and $4.432bn a year earlier.[1–2] A smaller loss is not profitability. Management guides to $18–19bn of 2026 losses and around $15bn in 2027 while reallocating investment primarily toward AI.[1,3] Those are expectations, not realized savings; accounting loss reductions do not translate dollar for dollar into cash available for AI.
Cash and marketable securities were $82.449bn, long-term debt $80.922bn, and operating lease liabilities another $31.173bn. Nine-month net debt proceeds were $24.910bn, with none in Q3; repurchases were zero both quarterly and year to date.[1–2] Funding is already broader than retained cash. Gross liquidity alone does not establish unlimited capacity, and operating leases should not be mislabeled as finance leases. Management says third-party cloud supplements slower owned infrastructure but has not specified the mix.[4, p3] The Q4 call is the stated window for quantitative 2027 budget detail; aggressive-investment language is not a numerical capex forecast.[3, p15]
04｜Class C is a funding instrument rather than cash already raised
The SEC 8-K confirms non-voting Class C, expected charter effectiveness on 19 January 2027, and a planned one-for-one dividend to A/B holders. Record and payment dates require subsequent board action.[6] A pro-rata distribution neither doubles wealth nor itself raises capital or dilutes proportional economic ownership. Later issuance to outsiders can dilute economic ownership while preserving founder voting control. After the distribution, valuation must adjust the share denominator and consider combined A/C value and differing rights.
On the main call, management had no specific financing plan to announce. In the follow-up, Susan Li characterized the referenced hundreds-of-billions figure as a theoretical founder-control financing threshold, not present equity financing needs or plans.[3, p14; 4, pp2–3] Long-range planning, theoretical capacity, authorization and a completed transaction are distinct evidence categories. This earnings report complements the existing capital-structure article. The next task is to reconcile information-statement tables, commitment payment schedules and actual issuance documents, not equate future leases with current debt.
05｜The call debate and four scenarios
Eric Sheridan of Goldman Sachs questioned Susan Li on capital structure; Doug Anmuth of JPMorgan pressed on 2027 expenses and capex; Justin Post of BofA asked about advertising runway; Ross Sandler of Barclays challenged compute economics. Management described financing flexibility, ad improvements and efficiency opportunities without a numerical 2027 budget.[3, pp12–19] In the follow-up, Ron Josey of Citi questioned financing and transaction monetization, while Ken Gawrelski of Wells Fargo asked about lease-versus-own capacity. Susan Li’s answers limit extrapolations about committed financing and scaled new-AI revenue.[4, pp2–3]
Bull: ad improvements persist, agents reach commercial scale, unit serving costs fall faster than usage grows, and capex growth moderates. The next print must show better operating-cash-flow coverage. Base: healthy ads fund early AI businesses while depreciation and investment absorb most incremental cash; a capital-intensive valuation framework replaces historical asset-light assumptions. Bear: ads slow, depreciation rises and monetization slips, making funding and dilution more important; simultaneous growth and cash-conversion deterioration weakens the operating thesis. Tail risk: regulation or litigation, power and construction delays, rapid obsolescence and tighter funding coincide, damaging utilization, cash flow and valuation. No invented probabilities or company-endorsed forecasts are assigned.
06｜Valuation implications and the next earnings test
The sensitivity grid isolates earnings and multiple risk. EPS can grow while capital intensity, governance or funding risk compresses the multiple. The discipline is to wait for comparable earnings and cash validation rather than turn 26% ad growth into an automatic purchase or position increase. Current price and a same-source forward estimate series were not verified, so there is no precise expected return, target price or historical price chart.
The Q4 report is management’s stated window for the 2027 budget; its exact scheduled date was not verified.[3, p15] Check ad volume, pricing and conversion; capex and lease definitions; operating cash-flow coverage; depreciation and asset lives; Reality Labs losses; Muse paid retention and unit economics; and any Class C issuance or new commitments. If efficiency gains fail to offset serving costs and depreciation, valuation support weakens. If incremental AI gross profit and cash recovery consistently exceed incremental investment, the return assessment improves. Supplier claims need their own company reporting and delivery cadence rather than another repetition of Meta’s spending headline.
04 · INVESTMENT IMPLICATIONS
Industry and asset implications
Compute and memory
Analytical inference: capacity expansion can support accelerators, servers and memory. Training versus inference, internal versus purchased chips, cloud versus owned capacity, and procurement timing determine allocation. No supplier-share or pricing assumption is supplied, so Meta capex is not apportioned into supplier revenue.
Networking and ANET
Disclosed product evidence: Arista Networks (ANET) named Meta among partners for its open Ethernet rack-scale architecture announced on 7 October. Its 9 June 1.6T platform release also included a Meta infrastructure executive’s collaboration statement.[9–10] This provides a specific technical transmission path: dense compute requires both within-rack scale-up and cross-rack scale-out connectivity, linking switching, optics, cooling and power design. Analytical inference: ANET’s opportunity depends on Meta capacity additions, network spending intensity, supplier share and delivery/acceptance timing, rather than Meta’s aggregate capex growth. Neither announcement discloses the incremental Meta order value or corresponding revenue. No numerical ANET earnings upgrade follows. Its next results should verify AI-network revenue, 1.6T deliveries, customer concentration and product margin.
Power and cooling
Analytical inference: rising loads may increase electrical and cooling needs, but contracted megawatts are not energized capacity, and energized capacity is not high utilization. Grid access, permits, construction and cost allocation determine realization.
Data centers and financing
Analytical inference: land, shells, servers, leases and funding have different cash cycles. Cloud purchases may enter operating expense rather than Meta capex. Supplier revenue depends on contracts, acceptance and its own accounting; $34.906bn of Meta capex is not industry revenue.[1–2]
Industrialization stack: compute → networking and optical → memory → available compute capacity → power and cooling → capital and financing → application revenue and shareholder cash. These are conditional transmission links, not confirmed earnings at every layer.
05 · VALIDATION & RISKS
What to watch next
Q4 earnings: advertising
Volume, pricing and conversions support revenue and comparable profitability.
What would weaken the view: Traffic and pricing soften while efficiency fails to cover costs.
Q4 earnings: capital efficiency
2027 budget, depreciation, commissioning and operating-cash-flow coverage are disclosed.
What would weaken the view: Investment keeps outrunning cash generation or lease obligations remain opaque.
Product and financing disclosures
Muse paid retention, revenue and unit costs improve; any Class C issuance states size, price and use.
What would weaken the view: Usage or authorization rises without commercial profit or completed financing.
What would change our view?
Key risks include the ad cycle, regulation and litigation payments, delayed depreciation, underutilized compute, agent monetization and efficiency shortfalls, funding costs and equity dilution. Supplier upside is conditional. This research does not authorize a trade.
06 · FAQ
Key questions
Is AI capex already consuming cash flow?
Consolidated capex exceeded quarterly operating cash flow, leaving negative $2.120bn FCF. The spending is not fully separated into AI-only investment, so the entire gap cannot be attributed to AI.[1–2]
Why is 265% EPS growth misleading?
Last year included a one-time tax charge and this year a litigation charge. Excluding those respective items gives $7.25 and $7.40, about 2.1% growth, still without a common tax rate or full recurring normalization.[1,5; calculation]
Is there a firm 2027 capex number?
No numerical range was verified. Management was developing the budget and pointed to the Q4 call for quantitative detail.[3, p15]
Has Class C already raised hundreds of billions?
No. A pro-rata distribution differs from future financing; the follow-up explicitly separated theoretical capacity from present financing plans.[4, pp2–3; 6]
Can Meta capex growth directly raise ANET revenue estimates?
No. Network allocation, architecture, delivery and acceptance need verification in supplier disclosures. Meta’s total capex covers many assets and finance lease principal.
07 · TERMS & SOURCES
Terms, sources and related research
Key terms
- Free cash flow or FCF
- Company-defined non-GAAP operating cash flow less cash PP&E purchases and finance lease principal.
- Capital expenditures or capex
- Here, company-defined investment including finance lease principal, not all AI expense.
- Return on invested capital or ROIC
- After-tax operating profit relative to invested capital; AI-specific inputs are incomplete.
- Stock-based compensation or SBC
- Employee equity compensation whose cash-flow add-back does not remove economic cost.
- Class C
- Planned non-voting shares; distinguish the pro-rata dividend from later external financing.
[1] Meta IR | Q3 2026 results and financial tables ↗
[2] SEC | Q3 2026 earnings Exhibit 99.1 ↗
[3] Meta IR | Q3 2026 main earnings call transcript ↗
[4] Meta IR | Q3 2026 follow-up call transcript ↗
[5] Meta IR | Q3 2025 tax-charge reconciliation ↗
[6] SEC | Class C recapitalization and founder agreement 8-K ↗
[7] Fortune | Visible Alpha revenue consensus comparison ↗
[8] Barron’s | EPS market-expectation comparison ↗
[9] Arista | Rack-scale AI networking architectures, 7 October 2026 ↗
[10] Arista IR | 1.6T platforms and Meta collaboration, 9 June 2026 ↗
Research dated 23 October 2026; quarter ended 30 September. Financial data [1–2], prior tax reconciliation [5], guidance and call narrative [3–4], and capital arrangements [6]. Full Q3 10-Q and information-statement schedules were not independently verified. Consensus is a sourced media comparison with no original cutoff provided. Sensitivities are assumptions, not targets. Existing articles and ratings are preserved.