Meta AI / FAIR

Refresh delta (since 2026-06-17)

What changed

1. The margin broke. Q2-2026 revenue $60.80B (+28% YoY) but income from operations fell 8% to $18.78B; operating margin 31% vs 43% a year ago. Family of Apps — the engine, not the moonshot — saw operating income fall 6% to $23.39B on +28% revenue, margin 39% vs 53%. The CFO's own normalization: excluding $2.4B of legal charges and $1.18B of severance, operating income would have risen 9%. Either way, the prior dossier's central claim — "the ad engine is absorbing the AI investment so far" — no longer holds.

2. Free cash flow effectively went to zero. Company-reported Q2 FCFFree cash flowCash left after paying to run and maintain the business. Unlike profit, it is hard to flatter with accounting choices. $784 million on $31.08B of quarterly capexCapital expenditureMoney spent on long-lived things — buildings, machines, servers — rather than on running costs., against a prior-year comparable of roughly $8.4B. Meta repurchased zero shares in all of H1-2026 (vs $22.92B in H1-2025) with $25.03B still authorized and untouched, and issued $24.91B of new notes in May 2026, taking long-term debt to $83.66B from $58.74B. The buyback was the shock absorber and it has been spent.

3. Meta started selling the AI. This is the genuinely new leg. Muse Spark 1.1 shipped with a public, OpenAI-compatible API — listed on OpenRouter at $1.25 per million input tokens and $4.25 per million output tokens, roughly a quarter of frontier pricing, with a 1M-token context window. Alongside it: Meta One (a consumer subscription), the Meta Business Agent Platform with 1M+ businesses using agents weekly, 9 million SMBs on AI ad-creative tools (up from 4M+ at the boundary), and Advantage+ at a $75B annual run-rate (up from $60B). Management now openly discusses "monetizing compute directly given outsized market demand" and says it is "getting a lot of offers for compute at a significant premium over what we paid for it."

4. Off-balance-sheet and committed capital exploded. Non-cancelable contractual commitments went $237.67B → $349.31B in one quarter; leases not yet commenced went $182.88B → $278.99B, plus a further ~$68B of data-center leases signed in July 2026. A second unconsolidated VIE landed: the El Paso venture with BlackRock — 1GW, ~$14B development cost, BlackRock 80% / Meta 20% (Meta keeps operational control), Meta contributing ~$2.3B of land and construction-in-progress and taking a ~$1B distribution, with ~$13B of residual-value guarantees. Louisiana's max exposure is now $46.03B. And $10.80B of cash was reclassified to restricted, escrowed against multi-year infrastructure purchase agreements and locked until 2028-2030.

5. ~8,000 layoffs — the first big cut since 2023. Meta reduced headcount by approximately 8,000 on May 20, 2026 (~10% of the workforce), booking $1.18B of severance in Q2; headcount ended at 75,472, −1% YoY and −3% QoQ, with more cuts reported as planned for H2. Cutting people while raising capex is the "Year of Efficiency" reflex pointed at opex to protect the buildout.

6. Legal moved from contingent to cash — and the near-term AI-copyright catalyst vanished. $2.4B of legal charges hit Q2 G&A (+111% YoY). The EU issued preliminary DSA findings on 2026-07-10 that Facebook and Instagram breach the DSA through addictive design (infinite scroll, autoplay, push notifications), exposing Meta to a fine capped at 6% of worldwide turnover — >$12B on FY2025 revenue. Meanwhile the Kadrey AI-training distribution claim — a headline catalyst dated 2026-07-16 in the prior dossier — has slipped to a summary-judgment hearing on 2026-02-25 (2027), with the first AI-copyright trial (Entrepreneur Media) set for 2027-05-24.

What held

• The ad flywheel still turns. Ad impressions +14% and average price per ad +12% YoY; DAP 3.60B (+3%); Instagram passed 2B daily actives, Threads 500M monthly. Meta quantified the AI lift for the first time: the GEM ads-ranking work plus sequence learning drove an 8.3% increase in ad clicks and 15.7% uplift in conversions on Facebook. FoA other revenue crossed $1B/quarter, +73%, on WhatsApp paid messaging and subscriptions. The distribution moat is intact and is now, finally, being pointed at something other than ads.
• The cheap multiple persists. ~$592/share, ~$1.42T market capMarket capitalisationThe share price multiplied by the number of shares. What the market says the equity is worth., forward P/E ~17.3-18.7x. Down from ~$600 and ~$1.52-1.59T at the boundary.
• The structural thesis is intact but its polarity flipped. The prior dossier's #1 forensic flag — a depreciation wave deferred by the 5.5-year server-life extension — is no longer a warning. It is arriving: PP&E depreciation +40% YoY ($6.00B in Q2), servers-and-network depreciation +48% ($4.62B), net PP&E $176.4B → $225.7B in six months, construction-in-progress $50.5B → $80.3B. The bear mechanism was correctly identified; the timing was 12 months too generous.
• Open-weight is still dead, and that is now the less interesting fact — because closed-weight turned out to have a price list. Founder voting control, the single-revenue-vehicle concentration, and the capital moat are all unchanged.

Phase A — Understand the business

Company Overview — REFRESHED (materially touched)

At the boundary, Meta's AI franchise had exactly one monetization path: better ad ranking inside Family of Apps. As of Q2-2026 it has five, and management named them on the call:

Revenue mix is still overwhelmingly advertising — $59.36B of $60.80B in Q2 — so none of the four new paths is yet financially material. But the prior dossier's bear point #2 ("the incremental superintelligence spend has no incremental revenue line yet") is now partially falsified: there is a line, it has a price, and it is in market.

Management's own framing of the monetization endgame is worth recording verbatim in structure: subscriptions, volume pricing, then "evolve more of these products to be like our ad systems where businesses only pay us when we achieve results… let us run an efficient auction over our compute". That is Meta proposing to price inference on outcomes rather than tokens — the same auction mechanic that made the ad business, aimed at compute.

Product/leadership items new since the boundary: Muse Spark 1.1 and Muse Image shipped; Meta AI daily interaction count +60% since the Muse Spark rebuild; Meta Glasses with EssilorLuxottica (including a Kylie Jenner collaboration) shipping with Muse Spark on-device, "early sales… exceeding our expectations"; standalone Forum (Groups) and Seller (Marketplace) apps; Kunal Shah hired to run WhatsApp; Connect conference 2026-09-23.

Supply Chain — CARRIED

Unchanged in structure; see the previous dossier. One update belongs here rather than a re-run: the binding constraint the prior dossier identified as having migrated from chips to power + capital has now migrated one step further — to third-party capital. Meta is no longer funding the chain from operating cash flow alone (see Lens 3 and Lens 10).

Competitive Advantages — REFRESHED (capital moat changed shape)

• Distribution moat — strengthened and, for the first time, aimed at AI. 3.60B DAP, Instagram 2B DAU, Threads 500M MAU, WhatsApp peaking at 30M messages/second during the World Cup Final. Muse Spark rides this for free where OpenAI must buy it. Management leaned on exactly this asymmetry: "our distribution advantages will give us the opportunity to serve AI products that are valuable for everyone… This should be true regardless of whether our models are on the frontier" — a candid admission that Meta is hedging its own frontier position.
• Data + feedback moat — deepened, and now measurable. Every public Reels and Feed post on Instagram is processed through an LLM; over half of recommended Instagram Feed content is now <1 day old, more than double a year ago; the largest single-release Reels ranking change drove +15bps sessions.
• Capital moat — WEAKENED, and this is the most important change in this lens. At the boundary the moat was "$115.8B of operating cash flow self-funds the buildout." That is no longer true. In H1-2026 Meta funded capex with OCF plus $24.91B of new debt plus a halted buyback plus an 80/20 partnership with BlackRock (in which BlackRock's share is itself partly funded by $12.5B of debt financing). The CFO's language is explicit: "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." Self-funding has become co-funding. That is still a moat almost nobody else has — but it is a narrower one, and it introduces counterparties.
• New moat candidate: price. A frontier-adjacent model at ~25% of OpenAI/Anthropic list pricing, subsidized by an ad business, is a commoditization play executed with a price list instead of open weights. Whether it is a moat or a margin leak depends entirely on Meta's cost per token, which is not disclosed. Not sourced — do not assume it is profitable.

Segments — REFRESHED (the segment story is the story)

Q2-2026 vs Q2-2025 ($M):

Read: at the boundary, FoA operating income was growing +24% YoY (Q1-2026) and the prior dossier called the engine "accelerating, not decelerating." One quarter later FoA operating income is shrinking. A 14-point margin compression in Family of Apps in four quarters is not a rounding item — it is the AI bill arriving inside the profit centre, not beside it.

RL is now the quiet problem: the loss barely widened, revenue grew on AI glasses (offsetting lower Quest), and at $4.6B/quarter it is no longer the biggest drag on the P&L. The AI buildout is.

Geography (Q2-2026 YoY): US & Canada +32%, Europe +24%, Asia-Pacific +19%, Rest of World +36%. Note the shape: growth is fastest where ARPP is highest (US/Canada) and where it is lowest (RoW), and impression growth is concentrated in lower-monetizing surfaces and regions — which management flags as the offset holding price-per-ad growth down.

---

Phase B — Measure performance

Earnings Result (Q2-2026) — REFRESHED

Reported 2026-07-29; 10-QForm 10-QThe quarterly version of the annual report. Lighter, and not audited. filed 2026-07-30.

Beat/miss: revenue beat; EPS missed consensus by roughly 14%. The miss is mostly, but not entirely, explained: $2.40B of legal-proceeding charges and $1.18B of severance. Strip both and operating income is +9% rather than −8%. That normalization is fair as far as it goes — but +9% operating growth on +28% revenue is still 19 points of negative operating leverage, and the legal charge is the second consecutive quarter in which a non-operating item dominated the headline (Q1's $8.03B CAMT benefit ran the other way).

Cost drivers, as disclosed: R&D +67% on employee compensation (SBC and severance), data-center and third-party-cloud infrastructure, and third-party AI token costs. Cost of revenue +33% on the same infrastructure line. Marketing & sales +15% on third-party AI token costs and severance. Worth sitting with: Meta is now a meaningful buyer of other people's inference, disclosed by name in three separate expense lines. The company that spends $130B+ on its own compute is renting someone else's.

Guidance issued with the print:

• Q3-2026 revenue $61-64B — against Q3-2025 actual $51.242B, that is +19% to +25%, a deceleration from +28%, with FX guided as a ~1% headwind.
• FY2026 total expenses $165-169B (low end raised to absorb the $2.4B legal charge).
• FY2026 capex $130-145B, which management characterized as "narrowed from our prior outlook of $125-145B" — accurate as to the range, but the floor moved up $5B.
• FY2026 tax rate 15-17% for remaining quarters, raised from 13-16%.
• FY2026 operating income still guided above FY2025's $83.28B — note how thin that is: roughly +3% operating income on roughly +26% revenue.
• Explicit new risk language: youth-related trials this year "may ultimately result in a material loss."

Balance-sheet flags (all ``):

• $10.80B of cash reclassified to restricted, escrowed against multi-year infrastructure purchase agreements, released 2028-2030. Cash + marketable securities of $90.26B overstates usable liquidity by that amount.
• Long-term debt $83.66B (total principal $84.0B); net cash position has thinned to roughly $6.6B from $22.4B at Q1.
• Accounts payable $8.89B → $15.89B; accrued expenses $30.73B → $38.07B; $19.50B of property and equipment sitting in payables/accruals (non-cash investing) — a large and growing float against the capex program.
• Goodwill: $1.27B reclassified to held-for-sale within Family of Apps, no impairment recognized. Unexplained in the filing; worth a question.
• Unrecognized RSU compensation $79.79B over ~3 years — roughly $26B/yr of SBC still to come, against a buyback currently running at zero.

Market reaction: shares fell ~7.5% (reported as −9.6% in the after-hours session) on the print. That is the second consecutive punishing reaction to a revenue beat. The prior dossier called the Q1 reaction "a regime change in how META trades." Q2 confirms it: the market is marking capex and FCF, not revenue.

Earnings Calls — REFRESHED, and now primary-sourced

The prior dossier's Lens 6 was -inferred because no transcript was on the shelf. It is now .

Sentiment arc:

• 2023 — "Year of Efficiency": austerity rewarded.
• 2024-25: pivot to aggressive AI investment; capex framed as a bet better over- than under-made.
• Q3-2025 → Q1-2026 — conviction under fire: defend the capex, lean on adoption proof points.
• Q2-2026 — the tone changed again, in a specific direction: from conviction to optionality.

The Q2 call is not defensive. It is a systematic argument that the compute has multiple exits. The tell is how often management offers a fallback:

• "we expect to grow a large business serving large customers as well… we're getting a lot of offers for compute at a significant premium over what we paid for it."
• "Those enterprise offerings have the potential to take multiple forms – agentic tools, our API, or monetizing compute directly given outsized market demand."
• "remaining nimble about these opportunities will help us fund our buildout more efficiently… and provide us multiple pathways to generate returns on invested capital."
• "This should be true regardless of whether our models are on the frontier."
• On long-lived assets: "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."

What they started saying: "ROIC" (unprompted, by the CFO), "monetizing compute directly," "Meta Compute," "multiple pathways," "supply chain leverage" (on custom silicon), and the industry-underbuild argument — "The industry has underbuilt historically for the wave of AI adoption, making existing capacity, including our own, extremely valuable."

What they stopped saying: the unqualified frontier-leadership framing. Compare the boundary-era mission language ("AI that surpasses human intelligence") with Q2's carefully hedged "regardless of whether our models are on the frontier." Zuckerberg's own positioning also shifted — from building superintelligence to distributing it: "We're the only major company building AI with the primary goal of putting superintelligence directly into people's hands. Rather than centralizing superintelligence, we're focused on distributing it widely." That is a repositioning of the same spend toward a story the ad business can carry.

Read: an executive team that spends this much of a call enumerating alternative uses for its compute is an executive team that has been asked, hard, what happens if the primary use disappoints. The answers are good ones. The fact that they were prepared is the signal.

Tension carried forward: the Zuckerberg-Wang friction reported at the boundary now has a structural expression — a March 2026 reorganization created a parallel Applied AI Engineering unit under Maher Saba, reported as diluting Wang's autonomy. This is secondary-source commentary, not a filing fact, and is labelled accordingly. Yann LeCun's exit (Nov 2025) predates the boundary.

Comps — PARTIAL REFRESH

META's own figures are refreshed; peers were again not returned with sourced, dated multiples this pass. Fabricating them would be worse than the gap.

Read: the multiple barely moved (~17-19x → ~17.3-18.7x) while the forward earnings path deteriorated. That is the whole valuation problem in one line — the stock got cheaper-looking without getting cheaper, because the denominator is falling toward the price. Market cap is down ~7-11% from the boundary; the FY2027 base EPS estimate below is down ~22%.

Conflict flagged: one source reports META "modestly undervalued" against a GF Value of $833.92. That figure is a model output on a trailing-earnings base and is not reconciled with the Q2 margin break; it is recorded, not adopted.

Stock-Price Catalysts — REFRESHED

Additions since the boundary:

• 2026-07-29/30 — approximately −7.5% (−9.6% after hours) on the Q2 print: revenue beat, EPS miss ~14%, FCF $784M, capex floor raised, Q3 revenue guide below expectations.
• Year-to-date context: META was down ~10% for 2026 heading into the July print. So the drawdown is not a single-print event — it is a year-long de-rating of the capex story.
• 2026-07-10 — EU DSA preliminary findings on addictive design; up to 6% of global turnover (>$12B).
• 2026-07-28 — BlackRock/El Paso venture announced — received as a financing innovation rather than a spending shock; it did not arrest the post-earnings decline.
• 2026-08-07 — New Mexico youth-safety abatement order (see Lens 10 for the provenance conflict).

Updated pattern. The prior dossier concluded META reacts to (1) user/engagement trend, (2) margin direction, and (3) capex-vs-ROI perception, with the market having re-weighted toward (3). Q2 refines that: the market is now marking (2) and free cash flow specifically. Revenue at +28% bought nothing. The catalyst that would actually re-rate this stock is not a revenue beat — it is the first quarter where operating income growth exceeds revenue growth again, or a credible, sized disclosure of API/agent/compute revenue.

Catalyst that disappeared: the Kadrey distribution-claim summary judgment, dated 2026-07-16 in the prior dossier's catalyst list, has slipped to 2027-02-25. Anyone holding this name for a near-term copyright resolution should know that clock stopped.

---

Phase C — Judge people & books

Management — REFRESHED (four new capital-allocation facts)

The archetype is unchanged — founder-operator with absolute voting control, betting the franchise's free cash flow on superintelligence. Four things changed underneath it:

1. The buyback was switched off. Zero repurchases in H1-2026 against $22.92B in H1-2025, with $25.03B authorized and unused. Meta did not announce a suspension; it simply stopped. With ~$26B/yr of SBC still to recognize and no offsetting buyback, dilution is now unhedged — diluted shares are 2,566M, and 60 million shares were excluded from Q2 diluted EPS as anti-dilutive (vs ~1 million a year ago), which is itself a marker of how far the stock has fallen relative to strike prices.
2. 8,000 people cut while capex rose. May 2026, ~10% of the workforce, $1.18B severance, more reported for H2. Read charitably: opex discipline to protect the buildout. Read skeptically: the company is funding capex out of headcount because it will not fund it out of the buildout.
3. A moonshot option grant. In H1-2026 Meta issued options on 20 million Class A shares to executives and employees at a weighted-average exercise price of $2,788 per share, five-year term. Against RSUs granted in the same period at a $598.76 average fair value, the strike is roughly 4.7× the current share price. Read both ways honestly: it pays nothing unless shareholders make ~5×, which is shareholder-aligned; and it tells you what internal ambition is being underwritten, which is not a modest number. It is also a ~20M-share overhang if it ever lands. Only $471M of expense is recognized against it — the accounting cost is trivial relative to the payoff geometry.
4. Third-party capital entered the structure. BlackRock at 80% of El Paso, funded partly by $12.5B of debt, with Meta at 20% and full operational control. This is a good deal on its face — Meta gets a 1GW campus for ~$2.3B of contributed assets and a $1B distribution — but it is the second such structure in nine months, and the pattern is now a policy, not an experiment.

Governance note carried forward and slightly sharpened: at the 2026 annual meeting (2026-05-27), shareholder proposals on AI data-usage oversight (503.7M for / 4,446.9M against), dual-class capital structure, and a data-protection impact assessment on generative AI chatbots all failed. All twelve directors were re-elected. The prior dossier's central governance fact — the bet is one founder's conviction with no board check — was tested at the ballot box this quarter and confirmed.

Forensic Red Flags — REFRESHED

• The depreciation wave is no longer a forecast. PP&E depreciation $6.00B in Q2-2026 vs $4.28B (+40%); servers-and-network depreciation $4.62B vs $3.12B (+48%). Gross PP&E $292.87B, of which construction-in-progress $80.35B is not yet depreciating, against accumulated depreciation of only $67.15B. Servers and network assets alone are $119.68B gross; on the stretched 5.5-year life that is ~$21.8B/yr at full run-rate, and the depreciable base will roughly double as the $130-145B 2026 program and the $80B of CIP enter service. Depreciation plausibly reaches $45-55B/yr by 2027-28. The 5.5-year life extension that flattered FY2025 EPS by $1.00 is now deferring a visibly larger wave. Watch for a life shortening — it remains the single accounting decision that would most change this story, and management has not been asked about it on the record this quarter.
• Legal charges are now recurring, not exceptional. $2.40B booked in Q2, with the CFO guiding total expenses up specifically to absorb it and warning of possible "material loss" from youth trials. When a company raises its expense guide to make room for litigation, the litigation has stopped being a footnote.
• Off-balance-sheet capital footprint, updated: Louisiana Venture max exposure $46.03B (unconsolidated; Meta not the primary beneficiary); other unconsolidated VIEs $6.41B; El Paso adding ~$13B of RVG on close; $349.31B non-cancelable commitments ($53.52B due 2026, $81.65B due 2027); $278.99B of leases not yet commenced plus ~$68B signed in July. The true committed capital footprint is now roughly $690B+ against a $450B balance sheet.
• Restricted cash is a new and underrated flag. $13.55B total restricted, including $10.80B escrowed against infrastructure purchase agreements and unavailable until 2028-2030. Headline "cash and marketable securities of $90.26B" is roughly $79B of actually-usable liquidity. Suppliers are now demanding escrow from Meta — that is a supplier-power signal worth more than its size.
• Cash flow vs earnings. H1-2026 OCF $64.09B vs net income $42.62B — still healthy on the operating line. But H1 capex of $49.11B (PP&E) leaves H1 FCF of ~$15.0B, and the Q2 quarter alone converted $31.9B of OCF into $784M of FCF. Capex is no longer consuming FCF growth; it is consuming FCF.
• SBC: $7.60B in Q2 (+57% YoY), $13.63B H1; unrecognized $79.79B over ~3 years. Unhedged by buybacks for the first time in years.
• Held-for-sale: $1.27B of FoA goodwill reclassified to held-for-sale with no impairment, plus $2.3B of data-center assets destined for the El Paso venture. The goodwill item is not explained in the filing.

Regulatory findings (required sub-section):

• SEC LR / AAER: None. Re-run 2026-08-10 against EDGAR EFTS for 2021-2026 — zero LR, zero AAER. Unchanged from the boundary.
• EU DSA — NEW and material. On 2026-07-10 the European Commission issued preliminary findings that Facebook and Instagram breach the DSA through addictive design (infinite scroll, autoplay, push notifications, personalized recommendation). Exposure is capped at 6% of worldwide annual turnover — >$12B on FY2025 revenue of $200.97B. Preliminary findings do not prejudge the outcome; Meta may respond and appeal. This is separate from the DSA preliminary findings of 2025-10-24 on notice-and-action, appeals, and researcher data access described in the 10-Q.
• AI-training copyright — the near-term clock stopped. Kadrey (consolidated) — fair use was won on the named plaintiffs (2025-06-25); the surviving distribution claim now has summary judgment scheduled 2027-02-25, not 2026-07-16. Nine further cases filed from Nov 2025 (Entrepreneur Media, Carreyrou, TED, Chicken Soup, Cognella, Elsevier, Hobbs, Sullivan, Beaulier); SJ for four of them also 2027-02-25; first trial (Entrepreneur Media) 2027-05-24. Per-work statutory damages remain the tail risk and the filing still says so explicitly.
• Youth safety — PROVENANCE CONFLICT, surfaced not resolved. The 10-Q filed 2026-07-30 states the New Mexico Attorney General's case "has expanded to include various claims related to content moderation issues" and "is scheduled to begin on September 8, 2026," with the AG seeking up to $62.85B in penalties. Web reporting describes a $375M jury verdict in March 2026 and a $567M abatement order on 2026-08-07, totalling ~$942M, with an injunction covering youth push notifications, a 90-hour monthly usage cap for minors, and a required under-13 prediction model. These may be different proceedings or phases (a concluded liability phase plus a September trial on the expanded content-moderation claims), but the filing does not describe a verdict and the web reporting does not reconcile to the September date. Do not treat either as settled. The one thing both agree on: the youth-safety exposure moved from contingent to ordered-and-payable during this window, and the CFO now says on the record it "may ultimately result in a material loss."
• IRS transfer pricing — unchanged. 2010 Tax Court opinion valuing transferred IP at $7.79B ($1.48B above Meta's figure); the 2017-2019 Notice asserting an additional $15.89B plus interest and penalties, petitioned in Dec 2025. Long-term income taxes on the balance sheet fell to $18.33B from $21.01B.

Net: still clean on SEC accounting enforcement. The change since the boundary is that the AI-specific and youth-safety legal risk has converted from "reasonably possible" language into $2.4B of booked charges, an ordered abatement, and a raised expense guide — while the copyright tail extended by seven months.

---

Phase D — Project & stress-test

Forward Projection (FY2026 / FY2027 / FY2028 EPS) — REFRESHED

Model integrity — read this before any number below. A workbook was rebuilt this pass (models/ via our model, v2.0) and read back. It has no computed values — 51 formulas, no cached results — so value-per-share, WACC, EVEnterprise 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. and balance-check outputs are all blank and nothing in this lens is cited from it. Its seed diagnostics also disclose: only 2 quarters of financials on the shelf (so LTM figures are partial and revenue growth defaulted to a placeholder 8%), a partly-sourced opening balance sheet (inventory missing), and no share price set. The workbook is a scaffold, not a valuation. Every figure below is `` with arithmetic shown, anchored on filed actuals and management guidance.

Anchors (all `` unless noted): H1-2026 revenue $117.111B, H1 diluted EPS $16.62, H1 net income $42.621B [filings/10-q-2026-q2.md]; Q3-2026 revenue guide $61-64B; FY2026 total expenses guide $165-169B; FY2026 capex $130-145B; tax 15-17% for remaining quarters; FY2026 operating income guided above FY2025's $83.276B [transcripts/2026-q2.md]; Q3-2025 actual revenue $51.242B and Q4-2025 actual $59.893B (FY2025 $200.966B less 9M $141.073B) [filings/10-q-2025-q3.md, filings/10-k-2025-q4.md]; diluted shares 2,565M.

FY2026 build:

• Revenue: H1 $117.11B + Q3 ~$62.5B (guide midpoint) + Q4 ~$73B (Q4-2025 $59.89B at +22%) ≈ $252.6B (+25.7%).
• Operating income: $252.6B − $167B (expense guide midpoint) ≈ $85.6B — a ~3% increase on a ~26% revenue increase, consistent with management's "above 2025."
• Pre-tax: less ~$2.5B net interest/other ≈ $83.1B. H2 pre-tax ≈ $42.6B, taxed at 16% → H2 net ≈ $35.8B; plus H1 actual net $42.62B ≈ $78.4B.
• FY2026 EPS ≈ $78.4B ÷ 2,565M ≈ $30.6. Note this still includes the Q1 one-off $8.03B CAMT benefit (~$3.13/sh) — normalized FY2026 EPS ≈ $27.5.

The shape is the finding. The prior dossier's base case was FY2026 ~$32 → FY2027 ~$37 → FY2028 ~$43 — a smooth +16%/+16% compounding. This refresh's base is ~$30.5 → ~$29 → ~$36: FY2027 EPS goes sideways-to-down before recovering. That is a ~22% cut to the FY2027 base. The cause is not revenue (still guided to compound at high-teens) — it is that ~$130-145B of 2026 capex plus $80B of construction-in-progress lands as depreciation in 2027 while the new revenue lines are still too small to offset it.

Street consensus clustered at $30-34 for FY2026 at the boundary; my FY2026 base sits inside that. I have not sourced a post-Q2 consensus for FY2027 this pass — n/a, not sourced, and I am not going to infer one.

(Per unattended-run rules: no our model Brier forecast logged.)

Bull vs Bear — REFRESHED

Bull case. Meta is the only company that can put a competitive model in front of 3.6 billion people and undercut the frontier on price, because ads pay for the compute. Muse Spark 1.1 ships at $1.25/$4.25 per million tokens — roughly a quarter of OpenAI/Anthropic list — with a 1M-token context, on OpenRouter, with enterprise rollout queued. Underneath it, the ad engine is not broken but compounding: impressions +14%, price +12%, and for the first time a quantified AI contribution (+8.3% ad clicks, +15.7% conversions on Facebook), Advantage+ at a $75B run-rate, 9M SMBs on AI creative tools, 1M+ businesses running agents weekly, and FoA other revenue crossing $1B/quarter at +73%. Management is guiding FY2026 operating income above FY2025 even after absorbing $2.4B of legal charges, and has three independent ways to monetize the compute if the fourth disappoints — including reselling it at "a significant premium over what we paid for it." At ~17-19x forward you are buying the second-cheapest form of AI exposure in megacap, at the moment the spend peaks and just before the assets start earning.

Bear case. The thing the prior dossier said would break the thesis has begun breaking it, on schedule but a year early. Operating income is falling while revenue grows 28%. FCF is $784M. The buyback — the mechanism that offset $26B/yr of SBC — is off. Debt is up $25B in a quarter, $10.8B of cash is escrowed to suppliers until 2028, and committed capital now runs to several hundred billion against a $450B balance sheet, with two off-balance-sheet ventures and third-party financiers inside the structure. The depreciation wave is real and quantified: +40% YoY already, with $80B of construction-in-progress not yet depreciating and a 5.5-year server life that is generous and reversible. Legal has converted to cash, with an EU decision carrying up to $12B of exposure and a copyright tail that just extended to 2027. And the new revenue lines, however exciting, are not sized in any disclosure — "1 million businesses" and "$75B run-rate" are adoption metrics, not revenue.

Pre-mortem (18 months out, thesis broke). FY2027 depreciation lands at $50B+; ad revenue decelerates to low-teens as impression growth concentrates in low-monetizing surfaces; the Muse Spark API proves gross-margin-negative at $1.25/$4.25 and is quietly repriced or capped; the EU levies a multi-billion DSA fine and the US youth-safety judgments compound; server lives are shortened toward 4 years, producing a one-time depreciation step-up; FCF stays near zero for two more years with no buyback; and the market re-rates a no-FCF, flat-EPS compounder to 14-15x → ~$380-430. Note the pre-mortem now needs fewer assumptions than it did at the boundary — three of its six clauses have already partly occurred.

Are multiples too high? Still no — they are low. But the prior dossier's conclusion that "the risk is earnings, not the multiple" has been vindicated, and the earnings risk has now shown up in the actuals. A low multiple on a falling estimate is not a margin of safety.

Devil's Advocate — REFRESHED

Dismantling the bull case as it stands today:

• "The AI is finally monetizing" is the weakest strong-sounding claim on the call. Every new line was disclosed as an adoption number, never a revenue number: 1M+ businesses using agents, 9M SMBs on creative tools, "$75B run-rate" for a suite whose incrementality is unmeasurable because it is bundled into ad spend Meta would largely have won anyway. Meta discloses revenue by segment and geography every quarter. It did not disclose one dollar of API, subscription, agent, or compute revenue. Companies disclose numbers they like.
• The API price is a margin claim nobody has audited. $1.25/$4.25 per million tokens is ~25% of frontier pricing. Meta has not disclosed cost per token, and in the same quarter it disclosed paying for third-party AI tokens in three separate expense lines. A company buying inference from others while selling its own at a quarter of market price is either the lowest-cost producer on earth or subsidizing a land grab with ad profits. Nothing in the filings distinguishes these.
• The BlackRock structure is not free money. Meta gets a 1GW campus off balance sheet — and also a partner owning 80% of the economics, ~$13B of residual-value guarantees, and a lease it must keep paying whether or not AI demand materializes. RVG payments are deemed "not probable" so no liability is recorded. Two such ventures in nine months, from a company with $90B of cash, is a statement about what the balance sheet can no longer absorb alone.
• The buyback halt is the loudest undisclosed decision of the year. $25.03B authorized, zero used, no announcement. Management explained the capex, the layoffs, the legal charge and the tax rate on the call. It did not explain this. Meanwhile SBC is $7.6B/quarter and rising, and 60 million shares were anti-dilutive this quarter versus ~1 million a year ago.
• The most dangerous competitor is still Google, and the gap widened in one specific way. Google runs its own TPUs and pays no NVIDIA tax; Meta's MTIA is real but management describes custom silicon in future tense — "will provide long-term strategic flexibility and supply chain leverage." Meta is buying third-party tokens; Google is selling them. On the one axis that determines who can price inference to zero, Meta is behind.
• What must hold for $592 to be a floor: ad revenue growth stays ≥18%, FY2026 operating income does clear FY2025 as guided, depreciation lives are not shortened, the EU DSA decision lands well under the 6% cap, and the youth-safety judgments stay in the hundreds of millions rather than the tens of billions the New Mexico AG has demanded. That is five conditions, and the filing's own language on the last one is "may ultimately result in a material loss."
• Single scenario that permanently impairs. Unchanged in kind, sharpened in mechanism: an assistant-layer shift that reduces feed time-spent, arriving while $349B of non-cancelable commitments keep running. The prior dossier rated this "moderate" plausibility on a sticky 3.5B DAP base. DAP is still growing (+3%) — but the 10-Q now discloses that Q1-2026 saw a QoQ decline in total Family DAP (Iran internet disruptions plus a WhatsApp restriction in Russia). Exogenous, not competitive — but it is the first disclosed DAP decline in this cycle, and the fixed-cost base underneath it is now enormous.

Management Questions — CARRIED

All fifteen questions in the previous dossier remain live and unanswered on the record. Q2-2026 partially answered #1 (compute allocation — "a substantial amount… towards training models," no split given), #3 (Business AI monetization — subscriptions, volume pricing, outcome-based auction, no revenue disclosed), and #4 (open-weight — answered decisively by shipping a paid closed API). Question #2 (under what conditions would you shorten server useful lives, and what is the step-up?) is now the single highest-value unanswered question on the list and was not put to management this quarter.

Four to add: 16. What is the gross margin on the Muse Spark API at $1.25/$4.25 per million tokens, and at what utilization does it break even? 17. Why did you repurchase zero shares in H1-2026 with $25.03B authorized — is the buyback suspended, and what would restart it? 18. What are API, subscription, business-agent, and compute-resale revenues today, and when will you disclose them as a line? 19. Why was $1.27B of Family of Apps goodwill reclassified to held-for-sale, and what is being divested?

---