Phase Space AI

Research

Meta Platforms [META]

Meta Platforms [META] — Research, re-underwritten 2026-07-29

Anchored on the Q2 2026 8-K filed today (2026-07-29, EX-99.1) plus the Q1 2026 10-Q (filed 2026-04-30) and nine quarters of earnings-call transcripts with speaker/title tags.

---

1. The analytical question: is the AI capex generating attributable revenue, or funding Reality Labs under a better name?

Answer: it is generating attributable revenue. Reality Labs is not the recipient — its budget is frozen. The entire margin cost of the AI build sits inside Family of Apps.

Segment decomposition, filed

(Q1 2026 FoA/RL derived: H1 2026 FoA operating income $50,294m − Q2 $23,394m = $26,900m; H1 RL loss $(8,647)m − Q2 $(4,619)m = $(4,028)m.)

The three facts that answer the question

(a) Reality Labs' loss is flat. $(4,530)m in Q2 2025 → $(4,619)m in Q2 2026: +2.0% yoy, while total revenue grew 28%. H1 2026 RL loss $(8,647)m is actually lower than H1 2025's $(8,739)m. The CFO guided on the Q4 2025 call that "we expect expense growth to be driven by the family of apps with Reality Labs operating losses remaining similar to 2025 levels." Reality Labs is a fixed, capped, ~$18bn/yr line item. It is not where the money is going.

(b) Family of Apps operating income is 5.1x the Reality Labs loss — $23,394m against $(4,619)m. FoA alone, at a 38.75% margin, is a larger and more profitable business than it was a year ago in absolute revenue and dollar terms; its margin fell because of R&D, not because of RL.

(c) The margin compression is inside FoA and is R&D. R&D went $12,942m → $21,656m, +67.4% yoy, adding $8,714m of cost against $13,285m of incremental revenue. Decomposing the 14.2pp fall in FoA margin: - $2,400m legal-proceedings charge (G&A, one-off) = 4.0pp - ~$1,180m May-2026 severance (~8,000 employees) = ~1.9pp - remainder ~8.3pp = genuine AI-investment margin cost (R&D compensation + AI infrastructure depreciation flowing through R&D and cost of revenue)

So of the 14.2pp, 5.9pp is one-off and 8.3pp is the real price of the AI build. Ex the one-offs FoA margin is ~44.5% and total operating margin is 36.8%.

Where the return shows up: impressions vs price per ad

This is the decomposition that separates a volume business from a pricing-power business.

The Q4 2025 print (+18% impressions / +6% price) was the one quarter consistent with "META is buying growth with inventory." It did not persist. The two most recent quarters show +14–19% impression growth and +12% price growth simultaneously — the strongest combination in the series. The company's stated cause is ad performance: "benefiting from increased advertiser demand, largely driven by improved ad performance" (Q3'25, Q4'25, Q1'26 CFO, prepared remarks) and, on Q1 2026, "broad-based growth as we benefited from ad performance improvements, better macro conditions... and currency tailwinds."

Management's own quantification, conceded in Q&A rather than volunteered (Q1 2025 CFO): "Year-over-year conversion growth remains strong, and in fact, we continue to see conversions grow at a faster rate than ad impressions." Conversions > impressions is the closest thing to a hard return-on-AI metric META discloses. It is not a dollar figure. META does not disclose an incremental-revenue attribution for Advantage+, AI ranking, or Meta AI — that absence is the honest limit of this analysis and is stated rather than filled with an estimate.

Capex and the guided trajectory

FY2025 purchases of P&E: $69.7bn. FY2026 guidance walk, raised at every call: $115–135bn (Q4'25) → $125–145bn (Q1'26) → $130–145bn (Q2'26, narrowed). At the $137.5bn midpoint that is +97% on FY2025, i.e. capex roughly doubling in one year.

Consequences already visible: free cash flow $8,549m → $784m (−91%); long-term debt $58,744m → $83,664m with $24,910m of net issuance in the quarter alone; buybacks stopped entirely ($10,167m in Q2 2025 → $0 in Q2 2026; $22,921m → $0 for the half). Only the $1,353m dividend remains. The capital-return programme has been converted into capex. That is the clearest single statement of how management is financing the build, and it is a balance-sheet fact, not a narrative.

---

2. Mention frequency over time — prepared remarks (P) / Q&A (Q)

Counts of regex-matched mentions per call, from AV EARNINGS_CALL_TRANSCRIPT with speaker/title. Q&A is defined as everything from the first utterance whose title contains "Analyst". The Q2 2026 call is being held this afternoon (1:30pm PT) — its transcript does not exist yet.

Readings: - capex is mentioned more in Q&A than in prepared remarks in seven of eight quarters. Analysts are pulling this out of management, not receiving it. That is the pattern the brief says to weight. - Llama went from 9/17 to 0/0. The open-weights narrative is dead as an investor-facing story; it has been replaced by "Meta Superintelligence Labs" (first named in the FY2026 capex guidance on the Q4 2025 call). - Advantage+ went from 5/1 to 0/0. The branded product name has been retired from the script even as the underlying AI-ranking claim (13/12 in Q1'26) stays central. This is a disclosure regression: the named, trackable product has been absorbed into an unnamed capability. It matters because Advantage+ was the one place a reader could follow adoption over time. - Reality Labs collapsed to 1/0. Consistent with a frozen budget and no longer the story. - The Q4 2025 spike to 11/25 on AI ranking is the fingerprint of the +18%/+6% impressions/price quarter: analysts pressed on whether the ranking gains were real when pricing decelerated.

---

3. Accounting quality — is the reported growth real?

Conclusion: the growth is organic, recurring and repeatable. There is no 10x-Genomics-style settlement, no Applied-Optoelectronics-style concentrated distributor, no AIOT-style acquired base. The accounting-quality risk on META is not revenue overstatement — it is cost understatement via useful-life and capitalisation choices, which is a P&L-timing question, not a revenue-reality question. The prior memo's work on the server useful-life extension (4–5 → 5.5 years effective Jan-2025, disclosed ~$2.9bn FY2025 depreciation benefit) and the unconsolidated Hyperion VIE (maximum exposure to loss $45.99bn at Mar-2026, $28bn residual-value-guarantee threshold with no liability recorded) stands and is not re-litigated here — it is carried forward as the principal downside amplifier and is the reason the Q2 2026 10-Q (expected ~2026-07-31) is a dated catalyst.

---

4. Share count — the priority defect

META is dual-class: Class A (1 vote) and Class B (10 votes), both $0.000006 par.

Reconciliation netIncome / reportedEPS: $15,848m ÷ $6.18 diluted = 2,564.4m, against the filed 2,566m — a 0.06% match. ✓ Basic: $15,848m ÷ $6.23 = 2,543.8m vs filed 2,543m. ✓

The cross-check does not catch split-basis error, so SPLITS was queried explicitly: META has NO recorded stock splits, ever (SPLITS returns data: []). There is no split-basis exposure on this name, and price history was pulled with adjustment=split regardless.

Float invariant (public_float ≤ market_cap): AV carries SharesFloat 2,192,690,000 against SharesOutstanding 2,196,045,000 — 99.85%. For a company where insiders hold 342m super-voting Class B shares, a 99.85% float ratio is impossible; float should sit meaningfully below total shares. The ratio is only near-unity because both fields are measured on Class A alone. The invariant correctly flags the dual-class tagging defect on META.

But run on the correct source, the invariant PASSES — and its residual validates the share count. The test belongs in dollars, against EDGAR's dei:EntityPublicFloat, not on AV's share-count fields, which are Class A in both numerator and denominator and therefore cannot fail:

The 13.5% shortfall of float against market cap reconciles exactly to the Class B share of total shares outstanding (342.4 / 2,538.4 = 13.5%). This is precisely the signature a genuine dual-class float should show, and it is independent confirmation that both the per-class share counts and the float tag are right — a stronger result than a bare pass, because the size of the gap is predicted rather than merely permitted.

Contrast the failure mode the invariant exists to catch: a record elsewhere in this corpus carried float $83.7bn against market cap $24.9bn — a 3.4x violation, sitting unexamined. META's 86.5% is what a clean dual-class record looks like.

Adopted: 2,566m diluted for EV; 2,538.4m outstanding for ownership questions. Note that the Q1 10-Q also discloses ~40m further shares excluded from diluted EPS as antidilutive, so true fully-diluted is higher still; the EV basis is if anything mildly understated.

---

5. Mechanism — named, evidenced drivers

1. Gemini-class model deployment across ad ranking and creative — evidenced not by a product name but by the disclosed metric pair: price per ad +12% on impressions +14%, with management attributing it to "improved ad performance," and conversions growing faster than impressions.
2. Threads / WhatsApp monetisation — "Other revenue" $583m → $1,007m, +72.7% yoy, the fastest- growing line in the P&L. Small ($1.0bn/qtr) but doubling.
3. Business AI / agentic messaging — mentions rose to 4/8 in Q1 2026 with the weight in Q&A. No revenue disclosed. Named but unquantified; carried as optionality, not modelled.
4. Capacity as the binding input — FY2026 capex $130–145bn, "Meta Superintelligence Labs" the named recipient. The constraint management describes is compute, not demand.

Counter-mechanism, equally named: the $2.40bn legal charge and the CFO's explicit warning of "youth-related trials scheduled for this year in the U.S., which may ultimately result in a material loss." This is disclosed, dated within FY2026, and unquantified.

---

6. Open items this run could not resolve

• Finance-lease liability is not on the condensed balance sheet. Principal payments $962m/quarter imply a material liability; it sits inside accrued expenses and other liabilities. Resolvable in the Q2 2026 10-Q (~2026-07-31), not before.
• Restricted cash rose from $1,662m to $13,107m (in other assets) year over year, an 8x increase with no explanation in the 8-K. Likely debt or JV collateral. Flagged, not assumed.
• No dollar attribution for AI-driven revenue. META discloses conversions-vs-impressions directionally and nothing more. Any specific "$X of revenue from AI" figure would be fabricated.
• EARNINGS_ESTIMATES returns an empty array for META ({"symbol":"META","estimates":[]}), so the 7/30/60/90-day revision history the brief describes was unavailable. Consensus is absent from this memo rather than estimated.