Meta

The company that already answered the question

Mark Zuckerberg is quietly building an AI to do part of his own job. He’s developing a personal “CEO agent” that retrieves answers he’d otherwise pull through layers of people, and it’s the headline node in a company-wide obsession: an internal chief-of-staff bot called Second Brain, local agents named My Claw that talk to each other across 78,000 employees, output per engineer up ~30%, and ~8,000 roles being cut as the coordination work compresses (The Next Web). His framing on the January earnings call was the quiet part said out loud — projects “that used to require big teams” can now be done “by a single, very talented person.” The most AI-native company on Earth is automating itself from the CEO’s desk outward.

Hold that scene, because it inverts the question every other lab is still asking. Can AI make money? Meta answered that years ago, at quarterly scale, on audited statements: its feed ranking, ad targeting, Reels recommendations, and automated campaign tools are all AI, and they’re why its advertising business is projected to pass Google’s in 2026 to become the largest digital-ad engine in the world.

So its real question is the inverse: what happens when the billion people using Meta AI for free become a revenue line?

Here it’s real money

Across this series, the headline number is usually a run-rate or a projection running ahead of money actually booked. Meta is the exception: most of its story is audited GAAP. One number to correct before it misleads you — the $164.5B figure that still circulates is FY2024; FY2025 was $200.97B total, $196.18B advertising, confirmed by the $56.3B Meta booked in Q1 2026 alone (+31% YoY). The only forward-looking figure is the 2026 ad crossover — and it’s clearly a projection. Read the badges.

The advertising trajectory, what’s banked, and where the projection points. Read the first two bars as audited fact and the last as a forecast: the firm ground is the $196B Meta actually earned.

Feed the flywheel, then flip the switch

Meta runs AI at a scale no one else touches, 3.56 billion daily users, served continuously by AI ranking and ad auctions. Its strategy has two moves. The first is already paying: AI as an ad multiplier. Advantage+ automates campaigns, Andromeda ranks billions of auctions a day, and the result is advertisers spending more because the AI makes their dollars work harder. It’s the tightest capex-to-revenue loop of any company in this series.

The second move is patience. Meta AI’s 700M–1B users generate $0 directly, on purpose. It’s the WhatsApp playbook: acquire in 2014, stay free for a decade, monetize from a standing start. If even 5% of 700M users eventually pay $10/month, that’s ~$4.2B a year of new revenue conjured from nothing, added to a $196B ad base, not replacing it. No Wall Street model prices an asset that doesn’t exist yet, which is exactly why it’s underpriced.

The pivot that gives nothing away

Now the move that complicates the winner’s story, and it’s the most significant strategic change in Meta AI’s history. For a decade, every Meta model — Llama 1 through 4 — was open-weight, downloaded 75M+ times, the default the regulated world ran on its own servers. In April 2026, the first model out of the new Meta Superintelligence Labs, Muse Spark, shipped closed: proprietary, available only through a private API preview, with paid access “planned for later” (Bloomberg, NYT). The company that taught the industry to give AI away no longer trusts open weights to protect a frontier edge.

The pivot has scaffolding behind it. Meta paid $14.3 billion for a ~49% stake in Scale AI and brought in Alexandr Wang to run MSL, internalizing the data and RLHF pipeline the whole industry otherwise rents — the difference, in Meta’s telling, between iterating every six months and every six weeks (Reuters). The next model, codenamed “Watermelon,” is already in training. So Meta now runs the same dual track Google does — Llama as the open tier, Muse as the closed frontier tier — and the honest weakness sits exactly where it hurts in a coding-dominated era: Muse Spark still lags rivals on coding (NYT).

Three moves tell the strategy — one buys the pipeline, one closes the model, one keeps the ecosystem.

And the quietest, most damning evidence is back in the opening scene. Second Brain — the AI chief of staff Meta built for itself — runs on Anthropic’s Claude, not on Muse Spark, not on Llama (The Next Web). When the company that already won at AI automates its own executives, it reaches for a competitor’s model — the coding-and-reasoning gap made concrete inside Meta’s own walls, the same hedge Microsoft made when it piped Claude into Copilot.

The one wave that can’t touch it

There’s a structural advantage worth naming against the rest of this series, because it’s the cleanest expression of the inversion: Meta never sold seats. The per-seat-software collapse bearing down on Microsoft and Salesforce — agents doing the work a license used to bill for — simply can’t reach a company whose meter is ad impressions and whose assistant is free. Meta’s exposure is the opposite kind: it depends almost entirely on attention, and on AI making that attention more valuable.

The silicon bet underneath is built for exactly that. An MTIA chip roadmap — 300 in production, 450 and 500 inference-first parts arriving 2027, Broadcom committed through 2029 — is designed for the one workload no rival runs at Meta’s scale: AI inference across 3.56 billion daily users, continuously. When those inference-first parts land, even a 20% cost cut on that volume is billions flowing straight to operating income.

Muse Spark 1.1 — and the landlord option

Meta’s July release, Muse Spark 1.1, is an incremental update to the consumer-facing generation stack rather than a bid for the frontier. That is consistent with everything above: Meta monetizes attention, so a model that makes feeds and ads better is worth more to it than a model that tops a coding benchmark.

The more interesting July item is unconfirmed. Reporting on July 17 said Meta is in early talks to lease up to $10B of AI compute to Anthropic over two years (NYT, CNBC). Neither company has commented; talks are not a contract, and this may never happen.

But treat it as a signal about what Meta’s capex is for. If Meta will rent capacity to a company it competes with, then the data centers are not purely a weapon in the model race — they are an asset that can be monetized either way. That is the same structural position Microsoft, Google and Amazon already hold: get paid by the labs whether or not your own model wins. It also lands the irony of the current market squarely: the company that lost the talent war loudest may end up collecting rent from the company winning it.

What argues against it

The coding gap is real, and in a world where coding is the dominant power-user case, Muse Spark trailing there matters more than Meta would like. Revenue is ~98% advertising, so anything that pulls people off-platform — Google’s AI Mode, a new discovery surface — is an existential exposure, not a side risk. EU and FTC youth-safety litigation is a standing liability Meta itself flags as potentially material. And the closed-model pivot risks the developer goodwill that Llama’s openness built — the 75M downloads were a moat made of trust, and Muse Spark spends some of it. Markets are already twitchy: the stock fell ~7% after a Q1 beat, purely because capex was raised again.

My read — the realest AI profit, hiding in plain sight

This is the strongest AI thesis in the consumer segment, and the most misread, because the proof is filed under “advertising.” Meta is the only company in this set where AI demonstrably prints returns at scale already — $196B of FY2025 advertising, AI-enabled — so the $125–145B capex funds a flywheel that has proven its economics rather than a bet on a future that may not arrive. The 700M–1B unmonetized Meta AI users are, to me, the single largest unrealized monetization asset in consumer technology. And the closed-model pivot is the signal that Meta is finally building toward charging rather than only giving away.

The one question I’d build the 2026–27 watch list around: when does Meta put a price on Meta AI?

When it does, it’ll be the most consequential monetization event since ChatGPT Plus, a multi-billion-dollar line from a standing start. If I’m building: Llama is my open-weight default for products that must run cheaply or privately, and I’m watching the Muse Spark API the moment it opens, Meta’s data edge shows up most in personalization and shopping, where its behavioral history is unmatched.

Sources: FY2025 and Q1 2026 financials, Meta IR and CNBC (audited); the 2026 ad-crossover projection, eMarketer via LinkedIn; the Scale AI investment and Alexandr Wang, Reuters; Muse Spark (closed model, private preview, coding gap) and the “Watermelon” successor, Bloomberg and NYT; the MTIA roadmap and Broadcom extension, Meta Blog and Tech Wire Asia; Meta AI usage and the $0 strategy, ValueAddVC and CNBC; Zuckerberg’s CEO agent, the Second Brain chief-of-staff bot (built on Anthropic’s Claude), My Claw, and the “single very talented person” / ~30% output framing, The Next Web and Cybernews. The honest frame: Meta’s ad revenue and profit are audited and real; it reports no “AI revenue” line, and the $243B is a forecast — the $196B is what’s banked.

This deep dive is part of “Deep Research on Frontier Companies.” Crafted with intent by Ravi Teja Palanki · June 2026.