Meta Platforms

Refresh delta (since 2026-07-22)

A · What changed — six material moves

1. The P&L cracked. Operating margin went 43% → 31% and operating income fell 8% year-over-year on 28% revenue growth. Q2-2026 revenue $60,801M (+28.0%), income from operations $18,775M vs $20,441M. Diluted EPS $6.18 vs a ~$7.22 consensus — a 14.4% miss [primary: 10-Q Q2-2026 income statement; consensus web: TradingKey, 2026-07-30]. Revenue beat ($60.80B vs ~$60.29B expected). This is the first year-over-year decline in operating income since the 2022 efficiency trough.

2. Free cash flow went to $784 million. Q2 operating cash flow $31,862M against capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. of $31,078M [primary: 10-Q Q2-2026 cash-flow statement; capex figure disclosed verbatim in MD&A and repeated by the CFO on the call]. Down 90.8% from $8,549M in Q2-2025 ``. CapexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs. consumed 97.5% of operating cash flow in the quarter.

3. "Meta Compute" became primary. On 2026-07-28 Meta announced a strategic venture with BlackRock (with Global Infrastructure Partners and HPS) to develop a 1GW, ~$14B data-center campus in El Paso, Texas — BlackRock funds 80%, Meta 20% ``. Zuckerberg named it on the call: "as part of our Meta Compute effort" [primary: Q2-2026 transcript]. The prior dossier's Section 2 finding — "there is NO primary evidence… a press report plus a CEO's 'on the table'" — is superseded. See §2.

4. The buyback went to zero and the balance sheet flipped. No Class A repurchases in H1-2026, against $22,921M in H1-2025; $25.03B of authorisation sits untouched [primary: 10-Q Q2-2026, Note 10 + MD&A Capital Return Program]. Long-term debt $83,664M (from $58,748M). Net cash has collapsed from −$41.4B (Q1-2025) to −$6.6B `` — Meta is roughly one quarter from being net-debt.

5. Forward commitments grew ~55% in a single quarter. Leases not yet commenced $182.88B → $278.99B; non-cancelable contractual commitments $237.67B → $349.31B; plus ~$68B of additional data-center leases signed in July 2026 (post-quarter, commencing 2027–2028) [primary: 10-Q Q2-2026, Note 9]. Total disclosed forward obligation ~$711B vs ~$459B at the prior boundary ``.

6. Meta booked its first material litigation accrual — $2.40B — and the state-AG trial starts in two days. G&A rose 111% YoY on "$2.40 billion of charges related to legal proceedings" [primary: 10-Q Q2-2026 MD&A]. Jury selection in the MDL state-AG trial (four states + a federal disgorgement claim by all 29 suing AGs) begins 2026-08-12 in Oakland, openings 2026-08-18, expected to run over a month [primary: 10-Q Q2-2026 Legal Proceedings; web: Law360 / Newsweek, 2026-08].

B · Grading the prior dossier's nine thresholds

The 2026-07-22 refresh named specific numbers that would confirm or break its frame. Graded honestly:

Five of nine landed as framed, two came back negative (which is itself the finding), one was blown through, and one — capex — was answered by a guidance change rather than by the number. The frame held; the magnitude on free cash flow did not.

C · What held

The structural picture is intact and is carried: the two-segment structure, the ad-auction business model, dual-class control, the moat locus in distribution and data, and the supply chain. Nothing in the delta window touched them. Reality Labs remains a ~$19B/yr drag guided to stay that way [primary: 10-Q Q2-2026 MD&A: *"we expect our full-year 2026 RL operating losses to remain similar to 2025"*].

The prior dossier's central mechanical claim held and got sharper: the guided capex line is a narrower thing than "Meta's AI build," and the narrowing is widening. Two co-development ventures now sit outside it where there was one.

1 · The capex line, re-anchored to the print

A · The guidance moved again — the floor, not the ceiling

[primary: 10-Q Q2-2026 MD&A liquidity; Q2-2026 transcript — CFO Susan Li: *"in the range of $130-145 billion, narrowed from our prior outlook of $125-145 billion"*]

The definition is unchanged and still stated only in the press release and on the call: including principal payments on finance leases, a cash measure, calendar year. It reconciles: Q2-2026 purchases of P&E $30,116M + finance-lease principal $962M = $31,078M ``.

Note what the narrowing actually is. Meta did not raise the ceiling. It removed the bottom $5B — the outcome in which the buildout underspends. That is a commitment signal, not a cost signal, and it is the opposite of what the Q1 raise was (which Meta attributed first to component pricing).

B · The run-rate arithmetic, updated

Actuals [primary: 10-K FY2025; 10-Q Q1-2026; 10-Q Q2-2026; research-layer our figures]:

What the back half now requires ``:

The prior dossier's version of this table demanded 1.77–2.10× off Q1. Q2 nearly doubled Q1 and cut the required step-up roughly in half — but a 27% to 51% further acceleration off a record quarter, held for two consecutive quarters, is still the ask. FY2026 at the range implies +80.0% to +100.8% on FY2025 ``.

C · What the $130–145B still does not cover — now three structures, not two

[primary: 10-Q Q2-2026, Note 5 and Note 13 (Subsequent Event)]

Disclosed maximum exposure to loss today: $52.44B, going to ~$65.4B once El Paso closes ``.

The El Paso commercial terms, from the announcement ``: $14B total development, 1GW, BlackRock funds 80% / Meta 20%, BlackRock contributing **$4.9B cash at close partially funded by $12.5B of debt financing**, Meta sole occupant leasing the entire campus on four-year initial terms with four extensions (up to 20 years), online 2028. Meta provides construction, administrative and property management services.

Read the ownership split against the occupancy. Meta owns 20%, leases 100%, guarantees ~$13B of residual value, and runs the site. The economics are Meta's; the balance sheet is BlackRock's. That is the Louisiana template, repeated four months later with a bigger partner — which answers the prior dossier's own Lens 14 question ("does the Louisiana Venture structure become the template?") with yes, on the first opportunity to answer it.

One detail worth not skipping: Meta reclassified $1.266B of goodwill to held-for-sale in Q2 alongside the contributed assets, with no impairment recognised [primary: 10-Q Q2-2026, Note 7]. Goodwill moving into a data-center co-development vehicle implies the contribution carries a business, not just land and steel. That is unexplained in the filing and is a fair question for the Q3 call.

D · Commitments — the number that moved most

[primary: 10-Q Q1-2026 Note 8; 10-Q Q2-2026 Note 9]

Total disclosed forward obligation ≈ $711.0B ``, versus ~$459.3B on the prior dossier's post-April basis — +54.8% in one quarter.

The $10.80B of escrowed money-market funds, released between 2028 and 2030 as underlying purchase obligations are satisfied, is worth naming separately: it is cash on the balance sheet that cannot be spent, and it doubled in a quarter. Meta's reported "$90.26B cash and marketable securities" is $79.46B of usable liquidity once the escrow is netted ``.

---

2 · The compute-resale question — resolved at the primary level (partially)

The prior dossier's finding was categorical: "there is NO primary evidence that Meta sells or leases compute externally. None." That finding was correct on the record as it then stood. It is now out of date in three of four respects, and still correct in the fourth — which is the one that matters for the income statement.

And the counter-fact got louder, not quieter. The prior dossier's strongest evidence that Meta is structurally a compute buyer was $237.67B of commitments "mostly related to third-party cloud." That number is now $349.31B, and the Q2 10-Q adds a new named expense driver that did not exist in any prior filing:

"third-party AI token costs" — cited as a driver of the increase in research and development (+67%), marketing and sales (+15%), and total costs and expenses. [primary: 10-Q Q2-2026 MD&A, four separate occurrences]

Meta is now disclosing purchases of other providers' inference as an expense line material enough to name three times. Zuckerberg was asked about exactly this tension on the call and did not dodge it:

"there's just nowhere near enough compute for all the demand… it would be foolish to basically just sell all of the compute and take a short-term profit" — and, on the margin ranking: "we believe that there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly." [primary: Q2-2026 transcript]

Stated plainly: Meta Compute is real as an organisation, a partner, a 1GW asset and a financing structure. It is still zero as a revenue line, and Meta's own CEO ranks it below selling intelligence on margin. The correct read of the BlackRock venture is probably not "Meta is becoming a neocloud" — it is "Meta found a way to build 1GW while putting 80% of the capital on someone else's balance sheet." The first primary evidence of a resale business would be a Note 2 revenue line, a third reportable segment, or a named external compute customer. None exists at 2026-06-30.

Provenance note carried forward: the arXiv preprint (2607.07207) cited by the desk remains downstream of the 2026-07-01 Bloomberg report and is not independent corroboration. That has not changed.

---

Company overview — touched, brief

Carried from the previous dossier with three additions:

• Scale, per the call [primary: Q2-2026 transcript]: 3.6B daily actives across the family; Instagram reached 2B daily actives; Threads crossed 500M monthly actives; WhatsApp peaked at 30M messages/second during the World Cup final.
• A model family, not a model. Muse Image and Muse Video (2026-07-07), Muse Spark 1.1 on the paid Meta Model API (2026-07-09, $1.25/M input, $4.25/M output), Muse Spark 1.2 powering a Muse Code terminal agent (2026-08-06), and Muse Glimmer, an open-weight model (2026-08-10) ``. The open-weight release partially reverses the closed-weight turn the prior dossier flagged at Muse Spark's April launch.
• Meta Compute is now a named unit with a named strategy (see §2).

Moats — touched, brief

Unchanged in locus (distribution + ad auction + data). Two delta items:

• Reinforcing: the LLM-into-ranking work is producing measured engagement gains — Instagram global time spent up double digits YoY; Facebook video time spent +9% globally, +10% US/Canada [primary: Q2-2026 transcript]. Ad impressions +14%, average price per ad +12% [primary: 10-Q Q2-2026 MD&A]. Both legs of the auction are growing, which is the healthy configuration.
• Attacking: the European Commission opened an antitrust investigation in December 2025 into Meta's policy barring general-purpose AI providers from the WhatsApp Business API, and in June 2026 imposed an interim measure requiring WhatsApp to offer that API access for free to such providers. Meta intends to appeal. Similar investigations are open in Africa and Brazil [primary: 10-Q Q2-2026, Legal Proceedings]. This is new since the prior dossier and it is a direct regulatory attack on the distribution moat — it forces Meta to carry rival AI assistants into its highest-engagement surface at zero price.

Segments — refreshed

[primary: 10-Q Q2-2026, Note 12; research-layer our figures]

The headline is in the FoA line, not the RL line. Reality Labs did what it always does (−$4.62B, guided to ~$19B for the year, in line with 2025). Family of Apps — the profit engine — grew revenue 28% and shrank operating income 6%. FoA employee compensation rose 56% and FoA other costs and expenses rose 74% ``. That is the whole story of the quarter in two ratios.

Revenue by geography, Q2-2026 [primary: Note 2]: US & Canada $23,863M (+29.3%), Europe $14,009M (+25.9%), Asia-Pacific $16,073M (+25.0%), Rest of World $6,856M (+35.1%). No geography is decelerating disproportionately.

Latest reported print — Q2-2026, reported 2026-07-29

[primary: 10-Q Q2-2026; Q2-2026 transcript]

(Quarterly cash-flow lines derived as H1 cumulative less Q1 cumulative; the $31,078M capex and $784M FCF are both stated verbatim by the company.)

Vs guidance and consensus: revenue hit the upper half of the $58–61B guide and beat the ~$60.29B consensus; EPS missed by 14.4% ($6.18 vs ~$7.22) ``.

The earnings-quality flag has flipped sides. The prior dossier warned that Q1-2026's net income was inflated by an $8.03B CAMT benefit (effective tax rate −23%; 14% ex-benefit). Q2's tax rate is a clean 16% and there is no discrete item — so Q2's weak net income is real operating weakness, not a tax artefact, while Q1's strength was partly artefact. H1-2026's effective rate of (5)% is meaningless as a run-rate. Read the operating line, as before — and the operating line is down.

Management's own bridge, which is fair and should be stated:

"Excluding the Q2 legal charges and severance expenses, our second quarter operating income would have increased 9% year-over-year." [primary: Q2-2026 transcript, CFO Susan Li]

That is arithmetically right: $18,775M + $2,400M + $1,180M = $22,355M, or +9.4% on $20,441M ``. Margin ex-items ≈ 36.8% — still 6.2 points below Q2-2025. The adjustment rescues the direction, not the compression. And "legal charges" is a weak candidate for non-recurring when four trials are scheduled inside the next seven months.

Balance-sheet flags:

Two things to watch, neither yet a problem:

• Accounts payable +78.7% while revenue grew 28%. The cash-flow supplement explains most of it: "property and equipment in accounts payable and accrued expenses" was $19,502M at Jun-30 vs $10,618M a year prior [primary: 10-Q Q2-2026, supplemental cash-flow data]. This is capex sitting in payables, not deteriorating supplier terms — but it means ~$19.5B of the build is not yet in the cash capex number.
• The debt is marked below par and the gap is widening. Fair value of the Notes $79.75B against $83.66B carrying — a $3.91B discount, versus a $2.81B discount at 2026-03-31 [primary: 10-Q Q2-2026 Note 8; 10-Q Q1-2026 Note 7].

Market reaction: shares closed $585.61 on 2026-07-29 and fell 9.64% to $529.15 in after-hours trade . **By 2026-08-10 the stock is $599.73 — above the pre-print close** . The market sold the print and then bought it back within eight sessions. Sell-side cut targets into the recovery: Cantor Fitzgerald $770 → $680; DA Davidson $850 → $700; JPMorgan $875 → $800 . Consensus target is still **$756.95** .

Management narrative — the shift, with a transcript under it for the first time

The prior dossier's Lens 6 was web-grounded only because transcripts/ was empty. It now has a primary transcript. Comparing the Q2-2026 call against the prior three calls' documented posture:

What shifted, and it is a real shift:

1. From "efficiency/superintelligence" to an explicit enterprise revenue roadmap. Zuckerberg opened with a three-part frame — AI accelerating the core business, personal agents as "new revenue lines," and "a large enterprise opportunity to sell to businesses, including APIs, business agents, potentially selling compute directly." The prior dossier recorded a CEO who deflected ROI questions as "a very technical question." This call answered them for forty minutes.
2. The ROI deflection is gone, replaced by an explicit margin ranking. "there will continue to be a significantly higher margin on selling intelligence rather than selling compute directly." That is a falsifiable strategic statement, not a deflection.
3. A new vocabulary of financing. The CFO: "the strength of our balance sheet gives us the ability to attract capital from a wide range of markets to supplement the cash flow generated by our business. Our announcement with BlackRock yesterday is an example of the partnerships that we can structure." No prior Meta call framed infrastructure as a capital-markets exercise.
4. An honest refusal on 2027. Asked directly for 2027 capex, Susan Li: "we aren't providing a specific outlook for 2027 CapEx at this time. Infrastructure planning remains highly dynamic." The sell-side is nonetheless modelling $176B (2027) and $185B (2028) `` — a number with no company source behind it. Treat it as sell-side, not guidance.

What did NOT shift, and is the tell: the prior dossier asked whether the FY2026 expense range would hold a second time against a second capex move. It did not hold — but it broke for the legal charge, not for depreciation. The CFO was precise: "We are raising the lower-end of our expense outlook to incorporate the $2.4 billion charges related to legal proceedings." Depreciation is still not the binding constraint on the expense line. That is the single most load-bearing fact for 2027 modelling, and it is a timing reprieve, not a structural one.

One more unforced disclosure worth noting: the tax outlook for the remaining quarters rose from 13–16% to 15–17% [primary: Q2-2026 transcript]. Small, but it moves ~$1.5B of FY2026 net income and nobody flagged it.

Comps — NOT SOURCED

Peer multiples were not re-sourced this run and the June 2026 table remains stale. Do not reuse it.

META reference points only, all ``: trailing P/E 22.31, forward P/E 18.48, EPS (ttm) $26.54, revenue (ttm) $228.25B, market capMarket capitalisationThe share price multiplied by the number of shares. What the market says the equity is worth. $1.53T. Peer multiples: n/a. Any relative-value claim in this dossier would be invented; none is made.

Price catalysts — the delta

Net over the delta window: −6.8% ($643.81 → $599.73) ``, against a prior window that was +7%.

The pattern the prior two dossiers named — "this stock trades on the spend-vs-discipline narrative, not on ad fundamentals" — was tested hard and held, in both directions inside three weeks. Ad fundamentals were the best part of this print (revenue beat, impressions +14%, price per ad +12%, engagement up). The stock fell 9.6% anyway, on free cash flow and capex. Then it recovered 13% without a single new fundamental datapoint — on, as best as can be sourced, the digestion of the BlackRock structure and the ex-items margin bridge. Nothing about ads moved the stock in either direction.

Forward catalysts, dated:

• 2026-08-12 — jury selection, MDL state-AG trial, Oakland (openings 08-18, >1 month expected). Two days out.
• 2026-09-08 — New Mexico AG second matter; AG has indicated it will seek up to $62.85B.
• Q3-2026 — El Paso / BlackRock venture expected to close; ~$2.3B asset contribution, ~$1B distribution, ~$13B RVG onto the disclosure.
• 2026-10-28 — next user bellwether trials.
• Late Oct 2026 — Q3 print. Q3 revenue guided $61–64B (~1% FX headwind).
• February 2027 — next school-district bellwether.

Management & governance — touched, brief

Carried, with three additions:

• ~8,000 employees cut in May 2026, $1.18B severance in Q2, majority off the headcount by end of Q3 [primary: 10-Q Q2-2026, Note 12 footnote + MD&A]. Headcount 75,472, −1% YoY and −3% quarter-over-quarter [primary: transcript]. Meta is cutting heads while raising R&D 67% — the money is going to compute and to "technical hires… particularly AI talent."
• Kunal Shah joined as head of WhatsApp [primary: Q2-2026 transcript].
• Board: Hock E. Tan and Tracey T. Travis did not stand for re-election (notified 2026-04-08); twelve directors elected 2026-05-27 [primary: 8-K 2026-04-14; 8-K 2026-05-29] — carried from the prior dossier, unchanged.

Forensic — the delta

Accounting integrity: no change, and independently re-verified this run. regulatory/regulatory-findings.md re-fetched 2026-08-10: 0 SEC Litigation Releases, 0 AAERs across the five-year window. Note 1: no material changes to significant accounting policies. Critical Accounting Estimates: no material changes. The 5-to-5.5-year server life is intact and there is no arithmetic footprint of a quiet revision — see the depreciation check below.

Earnings are still cash-backed, but the cushion thinned: Q2 OCF $31,862M vs net income $15,848M (2.01×, flattered by $7,601M of SBC and $6,356M of D&A). H1 OCF $64,088M vs net income $42,621M (1.50×).

1 · The depreciation conveyor — loaded further, still not tipped

[primary: 10-Q Q2-2026, Note 6]

Depreciation expense [primary: Note 6]: total P&E $6.00B in Q2-2026 vs $4.28B (+40.2%); servers and network $4.62B vs $3.12B (+48.1%). H1: $11.67B and $9.01B respectively.

Straight-line consistency check — the quiet-change test, re-run. $119.683B of gross servers/network at 5.5 years implies $21.76B/yr if fully in service for a full year . Q2-2026 annualised came in at **$18.48B** — below, as it must be for a base that grew $13.70B during the quarter. The 5.5-year assumption is being applied consistently. No arithmetic footprint of a change.

Retirements narrowed. Accumulated depreciation rose $5,657M in Q2 against $6,000M of recognised P&E depreciation — an implied ~$343M of retirements/disposals ``, versus ~$1.51B in Q1. The Q1 gap was the held-for-sale reclass; it did not persist. Clean.

$80.35B of construction in progress is not depreciating at all — up from $61.02B. The barrel is 32% more loaded than one quarter ago.

Forward conveyor arithmetic `` Servers at 5.5 years, structures at 27.5 (midpoint of the 25–30 band):

The 2026 cohort alone carries ~$18–20B of eventual annual depreciation — roughly the entirety of FY2025's $18.00B total P&E depreciation, added on top. Nothing about that changed; the timing did not accelerate this quarter, which is precisely why the FY2026 expense range survived on operating grounds.

Useful-life sensitivity, on today's servers/network base only ``:

Every row is ~13% larger than the equivalent row one quarter ago, purely because the base grew $13.7B. Run it again on a post-2026 base and it roughly doubles.

2 · The financing has changed character — this is the quarter's real story

``

Meta's net cash position has fallen 84% in five quarters and is roughly one quarter from zero. H1-2026 repurchases were nil against $22.92B in H1-2025, with $25.03B of authorisation left untouched [primary: 10-Q Q2-2026 MD&A]. The dividend continues ($2.70B H1) but the buyback — historically the dominant return channel — has been switched off without being formally suspended.

The cost and shape of the debt [primary: 10-Q Q2-2026, Note 8]:

• Face $84.00B; carrying $83.66B; fair value $79.75B (marked $3.91B below carrying).
• May 2026 notes: six series, effective rates 4.60%–6.48%, maturities 2031–2066.
• Maturity ladder: $0 in the remainder of 2026, $2.75B in 2027, $1.50B in 2028, $1.00B in 2029, $5.00B in 2030, and $73.75B thereafter — 87.8% beyond 2030.
• Future interest obligations: $4.40B short-term + $84.98B long-term = $89.38B [primary: MD&A Long-term Debt]. Meta will pay more in interest over the life of these notes than it borrowed.
• Q2 interest expense $783M vs $241M (+225%).

The duration mismatch the prior dossier named is now larger and better documented. Meta is issuing money out to 2066 — forty years — against a capital program whose dominant disclosed asset class carries a five-to-5.5-year depreciable life, a ~7× to 8× tenor gap. The CFO gave the counter-argument on the call, and it deserves to be recorded because it is the strongest version available:

"our longer-term capacity strategy aims to give us the flexibility to continue growing compute in 2028 and beyond by laying down data center and network foundations to accommodate future server decisions. The long-lived nature of these assets inherently provides the flexibility that will make it possible to adjust our investment to the pace of AI adoption." [primary: Q2-2026 transcript]

That is a direct claim that the long-dated money is term-matched against land, shells, power and network — the 25-to-30-year assets — with servers deliberately deferred as a later, shorter decision. It is a coherent answer, and it remains unverifiable, because Meta does not disclose the capex split by asset class. The mix is still the unanswered question; the tenor is still the disclosed fact. This is now the single most valuable disclosure Meta could add.

3 · Litigation — the first real accrual, and the trial calendar arrives

[primary: 10-Q Q2-2026, Note 9 Legal Proceedings, unless noted]

The accrual: $2.40B of charges related to legal proceedings in Q2 G&A. The prior dossier's standing observation — "Meta records no aggregate youth reserve" — no longer holds. Meta has started paying.

Meta's own aggregate framing, verbatim and new: "the damages, disgorgement, or penalties that plaintiffs have indicated they could seek range widely in amount, including in certain cases up to more than a trillion dollars." Plus mass arbitration demands from counsel for over two hundred thousand individual claimants since November 2024.

New/updated non-US matters:

• EU — WhatsApp Business API: EC investigation opened Dec 2025; June 2026 interim measure requires free API access for general-purpose AI providers; Meta appealing. Parallel investigations in Africa and Brazil. New since the prior dossier and the most strategically dangerous item on this list — it is an access remedy, not a fine.
• Spain — AMI: judgment 2025-11-19, unfair competition, €542M damages, appealed. UTECA claim trial October 2026; a parallel 67-company claim in France.
• Carried unchanged: €798M Marketplace decision (appealed), €200M DMA decision (appealed), with Meta's own warning that further modifications "could result in a materially worse user experience for European users and a significant impact to our European business and revenue."

The asymmetry the prior dossier named has narrowed but not closed. Adjudicated amounts remain small ($375M + $6M + a settled bellwether). The accrual is now $2.40B — a 6× step onto the P&L. The claimed amounts run to $62.85B in one matter and "more than a trillion" in aggregate. What moved this quarter is not the size of the verdicts; it is that Meta's auditors and management concluded a loss was probable and estimable enough to book $2.4B.

4 · Tax — a $15.89B notice, largely uncovered in prior dossiers

[primary: 10-Q Q2-2026, Note 11]. Gross unrecognized tax benefits $18.74B (from $16.45B at Dec-31); accrued interest and penalties $2.97B. The IRS 2017–2019 Notice (September 2025) asserts an additional $15.89B in tax plus interest and penalties, on the same transfer-pricing transaction litigated for 2010; Meta petitioned the Tax Court in December 2025. The May 2025 Tax Court opinion valued the transferred IP at $7.79B, $1.48B above Meta's reported figure, with the final decision — and both parties' appeal rights — still pending. Meta states its accrual is adequate. This is a real, dated, quantified contingency that neither prior dossier surfaced.

---

Forward mechanics — three-year EPS path

Model integrity — read this before any number below

our model rebuilt the workbook this run on six quarters of newly-populated primary data (model.xlsx, version 2.0). our model returns:

"No computed values. The workbook has 51 formulas and no cached results."

Consequences, stated honestly:

1. Every valuation output is blank — value per share, enterprise valueEnterprise valueWhat it would cost to buy the whole company: its market value plus its debt, minus the cash you would get with it. Often very different from market cap., equity value, WACC, cost of equity, FY+1/FY+3 EPS, terminal growth. The workbook must be opened in Excel and saved before it computes. I am not citing a single model-derived number.
2. "Opening balance sheet sourced: partial (9/10 lines)" — Inventory is missing. This is a Meta reporting fact, not a data failure: Meta does not present inventory as a separate balance-sheet line (Reality Labs inventory sits inside prepaid expenses and other current assets). The 9 populated lines are all primary.
3. Share price and CAPM inputs are unset on the Assumptions sheet.
4. The balance check cannot be verified — it returns blank, so I cannot assert the pro-forma balances.

Everything below is hand-computed `` with the arithmetic shown, not model output. No forecast was logged (our model create deliberately not run — unattended).

FY2026 — mostly determined already

``

• H1-2026 actual: revenue $117.111B, operating income $41.647B, diluted EPS $16.62.
• Q3 guided $61–64B; take $62.5B. FY2025 Q4/Q3 seasonality was 1.169 → Q4-2026 ≈ $73.1B (+22.0% YoY, matching the Q3 guide's implied growth). FY2026 revenue ≈ $252B (+25.4%).
• Total expenses guided $165–169B; take $167B → operating income ≈ $85B, consistent with the company's "above 2025 operating income" ($83.28B) — and only just.
• Interest and other ≈ −$3.0B for the year (H1 −$1.139B, interest expense rising ~$0.85B/quarter).
• Tax: H1 was a $2.113B benefit; H2 pre-tax ≈ $41.5B at the guided 15–17% → ~$6.6B. FY tax ≈ $4.5B (~5.5% effective).
• FY2026 net income ≈ $77.5B → diluted EPS ≈ $30.2 on ~2,566M shares.
• Ex the $8.03B CAMT benefit: ≈ $27.1. Use the ex-CAMT figure for any 2027 comparison; the reported number will overstate the base.

FY2027 / FY2028 — base, bull, bear

``

Worked example (base FY2027): revenue $295B − expenses $204B = $91B operating; less $6B net interest = $85B pre-tax; × (1 − 0.17) = $70.6B; ÷ 2,570M diluted = $27.5. Shares held at ~2,566–2,570M in base and bear (buyback off, RSU dilution roughly offsetting); 2,500M in bull's FY2028 only.

The shape is the finding, not the digits. In the base case, EPS is flat for three years at roughly $27–30 — the buildout consumes the entire revenue growth, and the depreciation conveyor arrives in 2027 exactly as the expense line stops being able to absorb it. At $599.73 that is ~22× a flat 2027–28 EPS, i.e. the current price already requires the bull case to be at least partly right. The bear case is not a demand-collapse story — it needs only ad deceleration to low double digits plus the useful-life question going the wrong way plus litigation becoming recurring. None of those three is exotic.

Sell-side is modelling 2027 capex at $176B and 2028 at $185B ``. There is no company source for either. Susan Li declined to guide 2027. Any model built on those figures is built on an analyst consensus, not on Meta.

FY2026 free-cash-flow grid ``. H1-2026 OCF was $64.088B; FY2025 H1 was 42.8% of the full year; applying the same shape gives FY2026 OCF ≈ $150B:

Against FY2025's $43.59B, that is a 55% to 100% decline in free cash flow, inside Meta's own guidance range, with no operational disappointment required. The prior dossier called negative FY FCF "arithmetically reachable"; after Q2 it is reachable at the high end of capex without any OCF shortfall at all.

---

Bull vs bear, updated — with a pre-mortem

The bull case, stated at its strongest

1. The ad machine is not the problem and is accelerating. +28% revenue, impressions +14%, price per ad +12%, engagement up on every surface, no geography decelerating. Q3 guided to +22% at the midpoint. This is a business compounding at scale.
2. The margin compression is 60% explained by two identified items. Ex the $2.4B legal charge and $1.18B severance, operating income grew 9%. Both are one-quarter events on their face.
3. The depreciation conveyor has not arrived and Meta has bought time. Q2 P&E depreciation $6.00B, below the $6.5B trip-wire. The expense range broke for legal, not depreciation. Whatever 2027 holds, 2026's operating income will exceed 2025's.
4. Meta found a way to build 1GW with 20% of the capital. The BlackRock structure gets Meta full operational control, sole occupancy and 100% of the compute for 20% of the equity — with 80% of the balance sheet risk borne by someone whose cost of capital is lower. If this is the template, the "capex will crush Meta" thesis is measuring the wrong denominator.
5. The demand signal on compute is real and priced above cost. "we're getting a lot of offers for compute at a significant premium over what we paid for it." A firm with excess-demand pricing power on its fleet has an option the market is not paying for.
6. The balance sheet can fund it. $90B of liquidity, investment-grade access proven twice in seven months, $25B of unused buyback authorisation as a dry-powder switch, and net cash still positive.

The bear case, stated at its strongest

1. The profit engine shrank. Family of Apps grew revenue 28% and shrank operating income 6%. That is not a capex story or a Reality Labs story — it is the core business's own cost base outrunning the best revenue quarter in its history.
2. Free cash flow is $784M and the funding source has changed. The build is now paid for by debt, escrow, off-balance-sheet partners and a switched-off buyback. Every one of those is finite or reversible. Net cash has fallen 84% in five quarters.
3. The 2027 depreciation wall is arithmetic, not opinion. $80.35B of CIP plus a $130–145B 2026 cohort carries ~$29–31B of eventual annual depreciation on top of a base already running ~$24B annualised. There is no version of 2027 where that does not land.
4. "Legal charges" is a poor candidate for non-recurring. Four trials inside seven months, one starting 2026-08-12 with all 29 suing AGs on a federal disgorgement claim, one in September with a $62.85B demand, and Meta's own filing saying claimed amounts run "up to more than a trillion dollars." The $2.4B accrual is a first instalment, not an anomaly.
5. The EU can force the moat open without a fine. The June 2026 interim measure requiring free WhatsApp Business API access for general-purpose AI providers is an access remedy on Meta's highest-engagement surface. Fines are payable; access remedies are structural.
6. Meta is buying tokens. "Third-party AI token costs" now appears as a named driver of three separate expense lines. A company positioning to sell compute is simultaneously disclosing that it buys other people's inference at material scale. Both can be true; only one is in the P&L.
7. The valuation needs the bull case. ~22× a base-case EPS that is flat for three years, with a consensus target ($756.95) that has not been marked to the new expense and capex reality.

Pre-mortem — it is 2028 and the position lost 45%. What happened?

Most likely sequence, in order of probability:

1. 2027 was the depreciation year and management could not offset it. The 2026 cohort landed in service through H1-2027, total P&E depreciation stepped from ~$24B to ~$40B+, and the expense line — which absorbed 2026 by growing 40% against 25% revenue growth — could not absorb a second year without operating income falling outright. Consensus had modelled the capex and forgotten the depreciation, exactly as it did in 2022.
2. Litigation became a recurring line. The August 2026 state-AG trial returned an adverse verdict with a disgorgement remedy; the September NM matter followed; the $2.4B quarterly accrual became a $4–6B annual expense, and the injunctive terms (not the money) forced product changes that cost engagement in the US.
3. The compute-resale option expired unexercised. Offers "at a significant premium" never became a revenue line because Meta's own margin ranking said intelligence beats compute — so the market repriced the fleet from "optionality" back to "cost," and the July 2026 +9% narrative reframe unwound.
4. Ad growth decelerated to low double digits on a tougher comp while the cost base was already committed through $711B of contractual obligations that do not decelerate with revenue.
5. The financing tightened. Notes already marked $3.91B below par; a wider credit spread or a downgrade watch made the third $25B+ issuance expensive, forcing either a capex cut (which the market reads as demand weakness) or an equity-adjacent structure.

The single most important thing to be wrong about: whether the servers-vs-structures capex mix is closer to 70/30 (my assumption, which produces the wall above) or closer to 40/60 (which the CFO's "long-lived assets… laying down data center and network foundations" language implies, and which would push most of the depreciation out past 2030). Meta does not disclose it. That one undisclosed ratio moves FY2027 EPS by roughly $4–6. Everything in the bear case routes through it.

---

Devil's advocate — the short-seller's version, updated

Thesis: Meta is running a capital-intensive utility inside a software multiple, and has begun moving the utility off balance sheet to keep the multiple.

The five claims a short would press, each with the primary line under it:

1. "They stopped returning capital and nobody said so." Zero buybacks in H1-2026 against $22.92B a year earlier, with the authorisation untouched, disclosed in a single MD&A sentence and not addressed in the prepared remarks [primary: 10-Q Q2-2026]. A company that believes its stock is cheap and its cash flow is durable does not do that.
2. "The 'ex-items' bridge is a treadmill." Ex-items operating margin is still 6.2 points below last year ``, and the items themselves are structural — 8,000 severed employees and four scheduled trials.
3. "$711B of forward commitments against $90B of liquidity, $10.8B of which is escrowed." Commitments grew 55% in one quarter. The 2027 due-in-year line alone went $47.65B → $81.65B, +71%, before a single 2027 dollar of revenue is guided.
4. "The RVGs carry no liability." ~$28B (Louisiana) plus ~$13B (El Paso) of residual value guarantees, ~$41B in aggregate, with no recorded liability because payments are "not probable" [primary: Note 5, Note 13]. That judgment rests on the future fair value of purpose-built AI data centers in 2029+ — precisely the asset whose terminal value is the entire open question of this cycle. If AI data-center residuals disappoint, the RVGs are where it shows up first, and they are currently valued at zero.
5. "They're buying what they claim to be about to sell." "Third-party AI token costs" as a named driver of R&D, marketing and total expenses, alongside $349.31B of commitments "mostly related to third-party cloud capacity arrangements" and a contingent obligation whose release clause depends on Meta's vendor finding another tenant [primary: Note 9].

The strongest rebuttals, which a fair reading must carry:

• On (1): a buyback pause during peak capex with an investment-grade balance sheet is textbook capital allocation, not distress. The authorisation is intact and re-arms instantly.
• On (3): commitments are not debt. Leases not yet commenced carry no liability precisely because Meta has not taken the space, and a large share is cancellable-in-effect through renegotiation with vendors who are themselves capacity-constrained.
• On (4): "not probable" for an RVG on a 1GW campus with Meta as sole occupant and 20 years of renewal options is a defensible judgment — Meta would have to walk away from its own campus for the guarantee to bite.
• On (5): buying third-party tokens while building your own fleet is what a capacity-constrained firm does. Zuckerberg addressed it directly: "there's just nowhere near enough compute for all the demand."

Where the short is weakest: it has to argue that a business growing revenue 28%, with pricing power on both sides of its ad auction and 3.6B daily users, is a utility. It isn't — yet. Where the short is strongest, and where it cannot be answered from the record: the servers-vs-structures capex mix and the 2027 depreciation landing. Meta declined to guide 2027 capex and does not disclose the mix. The two facts a bear most needs are the two Meta has chosen not to provide.

---

Lenses carried forward unchanged

• Lens 2 · Supply chain — unchanged, see the previous dossier and the previous dossier. One reinforcement: Meta's position as a very large buyer of third-party compute strengthened materially ($237.67B → $349.31B of commitments, plus the new third-party AI token cost line).
• Lens 7 · Comps — not sourced this run. See Lens 7 above. Peer multiples are n/a.
• Lens 14 · Management questions — carried from the previous dossier. Question (b) — "does the Louisiana Venture structure become the template?" — was answered YES by the El Paso/BlackRock venture. Question (a) — the servers-vs-structures split inside the guided capex number — remains unanswered and is now the highest-value unanswered question in the file (see Lens 12 pre-mortem). Three to add: (i) why did $1.27B of goodwill move into the El Paso contribution, and what business went with it? (ii) on what basis are ~$41B of aggregate residual value guarantees carried at zero liability? (iii) what is the run-rate of third-party AI token costs, and against which providers?

---