Michael

$META You’re thinking about the ROIC math with the wrong denominator and the wrong auction mechanics. A 2% “conversion lift” in Meta’s system does not show up as a 2% revenue lift or a 200 bps margin lift in a static way; in a target‑CPA/ROAS auction it largely capitalizes into higher clearing prices because advertisers raise bids until their cost‑per‑result drifts back to their threshold. That is why Susan Li emphasized that “value‑weighted conversions continue to grow faster than impressions,” and why price per ad rose 10% y/y in Q3 while impressions rose 14%, yielding 26% ad revenue growth. The 10% price move is the P&L proof that model quality improvements are monetizing as higher eCPM, not just as abstract “better performance.” Meta also quantified discrete lifts that explain the pricing power: rolling Lattice to app ads drove “nearly 3%” conversion gains for that objective, the new runtime ranker delivered “more than 2%” conversion lift on Instagram, and Andromeda improved “ads quality on Facebook surfaces” by 14%. Management’s point is they are stacking dozens of small, system‑wide lifts, not shipping a single 2% improvement. Translate that into dollars. Annualized FOA ad revenue is roughly $200B off the Q3 run rate. A uniform 2% revenue lift would be ~$4B. At a conservative 65% incremental margin and a 14% normalized tax rate, that is ~$2.2B of incremental NOPAT per year. If you attribute that to the portion of capex actually funding ads/recommendation training and serving—not the whole $70–$72B FY25 capex envelope that also covers Reality Labs, non‑AI infra, regulatory, and staged sites, plus off‑balance‑sheet JV spend and rising cloud opex—the year‑1 ROIC is already well into double digits. Even if you force the entire $20B‑type step‑up you might pencil for 2026 into the denominator, a 2% system‑wide lift is ~11% year‑1 ROIC; a more realistic stack of multiple 2%–3% lifts across major surfaces that cumulatively drive, say, a 6% price improvement yields ~$12B of incremental revenue and ~$6.7B of NOPAT, implying ~33% year‑1 ROIC on $20B. The call’s actual telemetry argues for the latter: price per ad was up 10%, weighted conversions outran impressions, and the end‑to‑end automated ads run‑rate is now $60B and still gaining share, which expands the portion of spend that auto‑reprices as model quality rises. You also need to account for the second channel of monetization that the capex is enabling: more impressions at constant or higher quality. Zuckerberg said AI recommendations drove “5% more time spent on Facebook in Q3 and 10% on Threads,” and “video time spent on Instagram up more than 30% since last year,” which is why ad impressions were up 14%. That engagement tailwind compounds with the auction effect; you are getting both more “I” and a higher “P” in the revenue identity at the same time. Management explicitly said they are “perennially… compute starved,” which is why they are front‑loading capacity: the core business can “profitably use much more compute than we’ve been able to throw at it.” A third channel shows up through new supply that is now ready to monetize on the same AI stack. Ads are “now running globally in feed on Threads,” WhatsApp Status ads will “complete the rollout next year,” and click‑to‑WhatsApp ads revenue grew 60% y/y. Business AI is live in Mexico and the Philippines with “millions of conversations” since July and a U.S. website embed starting, which boosts the effectiveness of click‑to‑message formats already scaling. These are incremental dollars that would not clear at attractive economics without the same ranking, retrieval, and creative systems the capex funds. Finally, your 2%–3% ROIC back‑of‑the‑envelope implicitly applies the full capex to a single quarter’s measured lift. That is not how this pays back. The GPU and data center spend supports a multi‑year series of model upgrades and product rollouts. Li said they “expect to make significant progress on our longer‑term ranking innovations in 2026,” to “significantly scale up the amount of data and compute we use to train our recommendation models in 2026,” and to “leverage LLMs to improve content understanding.” The company is also consolidating “approximately 100” smaller models already and plans to fold “another 200” into larger, general models, which both sustains performance gains and reduces model maintenance overhead. On financing, the Blue Owl JV moves some build costs out of capex and preserves option value, while third‑party cloud ramps give them elasticity to avoid overbuilding. The “real obvious use case” that shows up in the P&L is already there: 10% higher ad prices at scale, 14% more impressions, a $60B ARR automated ads channel absorbing more budgets, and accelerating time spent that keeps the system fed. The capex is what keeps those flywheels compounding; if you measure ROIC against the portion of that capex that actually powers ads and then roll the benefits over multiple model releases rather than a single 2% metric, the returns are materially above 2%–3% and comfortably above Meta’s cost of capital.