Look at the billing line, not the product page. Five companies are racing to put an AI agent inside your house, and the only number that matters is not download count or wake-word latency — it is who collects rent when your hallway light turns on. Apple charges it in hardware. Google charges it in a $20 monthly protocol fee. Meta charges it in usage. Microsoft gives the software away to sell the appliance. Four business models, one shared blind spot: none of them fix the three apps, two subscriptions, and pile of incompatible hardware that started this war. They monetize the friction.
That should sound familiar to anyone who watched the Layer 2 sequencing wars.
I spent the last two weeks treating this five-player race the way I treat an unfamiliar token: I traced the value flows before I read the marketing. What I found is a structural replay of a fight crypto already fought — and, in most cases, already lost.
The article that seeded this analysis reads as a product survey. Apple ships a home hub and a $129 mini speaker. Google opens its agent layer to third-party assistants — Claude, ChatGPT, a "Hermes" — for a subscription. Meta's assistant pulls five million downloads in weeks but, by the piece's own admission, "lacks native device control," reaching Hue bulbs and a Tesla through connectors. Microsoft injects a real-time voice feature into an LG hub through a firmware update, free. And a model lab builds a $300–400 AI-native device that will not ship until, at the earliest, 2027.
I have to flag provenance before I analyze anything. The source carries no author, no outlet, no date. Three product names — the Apple hub, the Meta assistant, the Microsoft voice feature — diverge from every public naming convention I can verify, and it attributes to Google a protocol that another lab open-sourced in late 2024. That does not make the analysis wrong. It makes it a trend signal, not a fact set. Based on my audit experience, I have downgraded every confidence rating one notch, and so should you. Trace the wallet, ignore the tweet.
With that caveat anchored, the structure is legible. Read the four monetization models as a single motion, and they describe value migrating from the execution layer to the interface layer — exactly the migration that defined Ethereum's rollup era.
Apple is the hardware sequencer. It runs inference on-device, trading model capability for privacy and latency. In crypto terms, this is a rollup that keeps execution local and refuses to post state to a shared settlement layer. The moat is real and the TAM is narrow: you only get in if you already own the phone. Apple is selling the ledger, not the transaction.
Google is the protocol toll. It opens its agent layer to competitors and charges for access to the interface — the classic platform move of subsidizing the complement and taxing the chokepoint. This is the same posture as a sequencer that invites builders to deploy, then captures the ordering rent. The article's sharpest line lands here, and it lands hard: Google, in its telling, "turned the open network into a paid utility." An open protocol that charges admission is not open. It is a toll road wearing the flag of a public good. The code does not lie, only the narrative.
Meta is the distribution play without settlement. Five million downloads is a real signal — but a wallet that can talk and cannot sign is not a wallet. Without native device control, Meta's agent has no actuator layer, no ability to execute. It can narrate the light turning on. It cannot turn the light on. In a market where the entire value proposition is reliable tool-calling against heterogeneous device APIs, that is a structural disqualification, not a feature gap.
Microsoft is the side-channel. It gives the voice layer away and recovers through the cloud and the hardware partner. This is the airdrop strategy: buy the ecosystem position with free tokens, monetize the infra behind it. Cheap, fast, and dependent on someone else's distribution.
The model lab is the late sequencer. Building $300–400 hardware is a low-margin, capital-heavy path for a software-native company — the balance sheet does not justify it unless the real objective is controlling the interaction entry point, not selling the box. A 2027 timeline, in a category iterating quarterly, means arriving eighteen months after the map is drawn. Whales do not whisper; they shake the ledger — but a whale that shows up after the market closes is just a spectator with a big wallet.
Run the pricing matrix in prose and the arithmetic of extraction sharpens. Apple extracts $129 at the door. Google extracts $240 a year for protocol access. Microsoft extracts nothing visible and buries the cost in the cloud. The model lab asks $300–400 up front, once. Only Meta lets you in free, and only because it cannot yet charge for execution it does not control. This is the same shape I documented during the 2020 DeFi summer, when I tracked $2.4 billion in Uniswap liquidity and found that 40% of the top-yield pools were unsustainable structures dressed as innovation. The yield was the marketing. The fee was the product. The same inversion is happening in your living room: the experience is the marketing, and the rent is the product.
Two absences matter more than any presence. The article names five players and omits the one with the largest installed base in the category, treating it only as a benchmark to beat. That is not a rounding error; it is a framework that cannot close. And it omits an entire geography — the Chinese ecosystem that ships smart-home agents at a scale the West rarely measures. An analysis that leaves out the largest incumbent and the largest market is describing a race it has already scoped incorrectly.
Now the mechanism under all of it. The reason nobody wins is not that competition is fierce. It is that interoperability is unprofitable for every incumbent. A world where Apple devices obey a Google agent, and Meta's assistant signs commands on Microsoft's hub, is a world where every toll booth disappears. So each player optimizes for a fragmented market — and fragmentation, conveniently, gets rebranded as an unsolved problem worth new products. I have watched this exact maneuver in DeFi liquidity. "Fragmentation" was never a technical failure to be solved. It was a revenue model to be preserved.
That reframes the whole board. The competition is not over models. It is over who owns the settlement layer for the home. And the answer, per the evidence, is the layer that never has to fix anything because fixing it would cost it the fee.
The consensus reading of this race is that it is a technology contest — better models, lower latency, smoother voice. The data does not support that. Every differentiator in the piece is an engineering-integration choice, not an architectural breakthrough: edge versus cloud inference, open versus closed protocols, real-time full-duplex voice. None of these is a moat the way a settlement layer is a moat. They are features, and features converge. This is the same logic that decides OP Stack versus ZK Stack — the winner is not the cleaner architecture but the stack that convinces more projects to deploy on it first. Distribution eats engineering.
The deeper counter-intuitive claim is about causation. A reader sees five giants fighting and concludes "intense competition." I see five incumbents agreeing, silently and profitably, to stay incompatible. The correlation between "many players" and "a healthy market" is exactly the trap I warn against in every cycle. The right question is not who is winning. It is who benefits from the game never ending — and the answer is everyone currently charging rent.
There is also a safety ledger nobody is reading. A home agent that controls locks, ovens, and thermostats is not a chatbot that hallucinates a sentence. It executes against the physical world, and physical mistakes do not roll back. An open protocol that invites third-party connectors widens the attack surface at the exact point where the consequences are irreversible. The article never mentions this. Audits reveal the skeleton, not the soul — and this one has no audit at all.
Watch one signal next quarter, not the download charts: which player first publishes a real tool-calling success rate against third-party devices. That number — not the model, not the price — is the only honest measure of who can actually settle in the home. Until it appears, treat every "open protocol" announcement as a toll schedule, and every "fragmented" complaint as a business plan.
Pegs break, principles remain, portfolios vanish. So do ecosystems that confuse an open door with a free one.
