For the past seven days, I've been digging through every trace of Meta's rumored 'Muse' AI agent โ and what I found isn't a product story. It's a trust story. The headlines say the same thing in slightly different words: strong early numbers, rapid adoption, investors demand proof. Three claims, zero numbers. No MAU. No DAU. No retention curve. No paid conversion rate. No ARPU. In the language of evidence, 'strong early numbers' without a denominator is not data โ it's atmosphere.
I've seen this atmosphere before. In 2017, I sat across from 120 first-time investors in Copenhagen coffee shops, listening to them describe why they'd put their savings into a project with a whitepaper and a Telegram group. They all said some version of the same thing: the numbers are strong, the community is growing, everyone is talking about it. What none of them could answer was simpler: what happens when the growth stops and the bill arrives? That question โ the bill arriving โ is exactly what's happening to Meta right now with Muse, and it's the most instructive thing in this entire story.
Here's what I can establish from the reporting and from Meta's own financial posture. Meta is a company where roughly 98% of revenue comes from advertising. That's not a footnote โ it's the gravitational field that shapes every product decision the company makes. A 'productivity agent' inside that field faces a structural question it cannot avoid: does it monetize directly, through subscription or usage pricing, or does it exist to feed the advertising flywheel by keeping users inside the ecosystem? Those are two completely different products with two completely different success metrics. And the coverage so far hasn't distinguished between them.
This is where my background in economics becomes useful, and where I want to push past what the headlines are telling you. When investors demand proof, they aren't asking whether people use the product. They're asking whether the unit economics close. In plain terms: does the revenue from a single user exceed the cost of serving that user โ including the substantial inference costs of running an agent that, unlike a chatbot, may execute multiple tool calls and planning steps per interaction? Agent products consume dramatically more compute per session than pure conversational models. If Muse's 'rapid adoption' is built on free or low-friction access, then high usage isn't a win. It's a bill that scales linearly with engagement while revenue stays flat.
I audited this pattern in DeFi Summer 2020. We were watching liquidity pools grow, TVL charts going parabolic, and everyone calling it adoption. But when my collaborators and I dug into the gas fee data, we found something different: the users most active in these pools were the ones least able to absorb fee volatility. Growth and value were diverging, and the dashboard was hiding it. The same divergence risk applies to AI agents today. A free-tier user exploring Muse's capabilities generates inference costs and produces zero revenue. Investors who understand this are not being unreasonable when they demand evidence. They're doing the math that the press release skipped.
The Distribution Advantage That Cuts Both Ways
Meta's strongest card in this entire hand isn't the model. It's the distribution. Facebook, Instagram, WhatsApp โ a user base measured in billions. If Muse gets embedded into that surface, the adoption curve becomes trivial to manufacture. You don't need a growth hack when you have the world's largest social graph to plug into.
But this is precisely why the proof demand is so sharp. Distribution can generate usage at near-zero marginal cost. It cannot generate revenue at near-zero marginal cost. If Meta pushes Muse to three billion users and only charges a fraction of a fraction of them, the company has created the most expensive free product in history โ and the market will notice.
There's a second tension that the coverage has missed entirely, and it's the one I'd be watching most closely if I were analyzing Meta's position honestly. Meta built its AI reputation on Llama โ an open-weight model family that gave it credibility with developers and researchers. Agent products, however, tend to commercialize on closed, controlled deployments. The tension between open foundation and closed application is not a minor strategic wrinkle. It's the difference between being the platform everyone builds on and being the vendor everyone competes with. Meta hasn't clarified which side of that line Muse falls on, and until it does, the strategic picture remains genuinely ambiguous.
The Verification Problem Nobody Wants to Name
I need to be direct about something, because it's the kind of thing that gets buried in breathless product coverage. I cannot independently confirm that Meta has shipped a product called Muse at all. The name collides with a Microsoft Research world model released in 2025, and the sourcing here traces back to a crypto vertical covering a non-crypto company โ a mismatch that should raise every reader's guard. The entire story may be secondhand. It may be AI-generated filler. I don't know, and neither, apparently, does anyone reporting on it.
This matters because the pattern of unverified claims with confident framing is itself the lesson. We spent 2022 watching exchange after exchange publish proof of reserves that proved considerably less than advertised. Sparse Merkle trees showing assets but not liabilities. Point-in-time snapshots presented as continuous audits. The form of verification without its substance. A headline that says 'investors demand proof' while providing no proof of its own is participating in the same theater it claims to critique.
So let me separate what I actually believe from what I can support. I believe the AI agent category is entering a monetization reckoning, and that the reckoning is healthy. I believe Meta's compute spending has reached a scale where the market will no longer accept narrative as collateral. I believe agent inference economics are structurally harder than chatbot economics, and that most current adoption metrics are measuring the wrong thing on purpose or by accident. What I cannot believe โ because I cannot verify โ is anything specific about Muse's numbers, roadmap, or even existence.
If Muse is real and Meta follows its playbook, watch for the product to appear inside WhatsApp or Workplace as a bundled capability rather than a standalone subscription. That's the path of least resistance for a company with an advertising center of gravity and a distribution monopoly. And watch the next earnings call, where the questions will be sharper than the product announcements. The analysts who cover Meta aren't asking 'is this cool?' anymore. They're asking 'what does it earn?'
The years I spent helping Nordic banks understand blockchain taught me one durable lesson: institutions adopt technology when the economics are legible, not when the philosophy is beautiful. Trust no one, verify everyone โ that's the discipline this story demands, whether the subject is a DeFi protocol or a trillion-dollar social platform. The winter of easy AI narratives is here. What gets planted in the spring depends entirely on whether the numbers, when they finally arrive, mean what the headlines promised.
Behind every adoption chart, a unit economic waiting to be read. Behind every early number, a question someone is hoping you won't ask. The ledger of AI agents is being written right now, and it will remember exactly which claims came with receipts.