Three numbers, one column, none reconciling. Amazon's Alexa: 140,000 certified devices. Google Home: 85,000. SmartThings: 2,500. Apple HomeKit: roughly 1,000.
The spread between the first and last is 140x. That is not a market. It is a measurement artifact. SmartThings' 2,500 almost certainly counts certified brand integrations; Alexa's 140,000 counts SKUs. You cannot subtract a brand from a SKU and call the remainder "market share." Two of those four figures are unusable for any conclusion about competitive position.
I flag this before anything else, because the entire smart home narrative โ Apple is losing, Amazon and Google are winning โ sits on numbers that were never normalized to the same standard. When I built a SQL ledger tracking 400,000 CryptoPunk transactions in 2021, step one was normalizing for gas-fee epochs. Same discipline applies here. The dataset is broken before the thesis starts, and a broken dataset produces a broken conclusion.
The frontier moved. In 2026, smart home competition left the connection layer โ Matter, Zigbee, Thread, HomeKit's classic turf โ and dropped into the agent orchestration layer: LLM reasoning, tool calling, device-API execution. Amazon opened Model Context Protocol endpoints to Bosch, Whirlpool, and iRobot. Google shipped Gemini for Home at $99.99 hardware plus tiered subscription. Samsung embedded on-device AI Vision in refrigerators.
MCP is the event worth circling. If you audit protocols, you recognize it: a standardized tool-calling and context protocol that abstracts physical device capability into a callable tool set for a language model. It is the socket standard between agent and appliance. Whoever implements and extends MCP controls the plug.
That is value migrating from connection to execution โ the same migration that moved on-chain value from consensus to sequencer. And it is exactly where decentralization claims quietly die.
I will state my bias as a variable, not a conclusion. In 2017 I audited LendingBot's time-lock contracts and found a reentrancy flaw in withdrawal logic before mainnet; I shipped the patch. The lesson held: source over whitepaper, every time. My 2020 arbitrage bot across Uniswap V2 and Curve ran 150 trades a day at 99.8% accuracy because the interaction was deterministic. I trust execution layers that can be read. I distrust the ones that cannot.
Apply that lens to the agentic home.
Claim one: "decentralized sequencing" is a PowerPoint. The agentic core โ a persistent runtime with state memory, multi-step planning, cross-device orchestration โ is functionally an operating system. There is no working model in which a resident, stateful execution layer is decentralized at the consumer tier. It runs on one node. Call it a hub or a sequencer; the topology is identical. Two years of rollup specs promised otherwise. The sequencer still runs on one box. The smart home just made the L2 sequencer problem legible to people who never read a spec.
Claim two: the subscription is a subsidy wearing a business model. Alexa+ lists at $19.99/month. Gemini for Home pairs $99.99 hardware with tiers. Hardware at that price sits at or below bill-of-materials. The entire logic depends on subscription lifetime value. I have seen this structure before, and the judgment holds: too good to be true. My 2022 LUNA forensics taught me the tell is never the headline yield โ it is the wallet clusters funding it. Here, the tell is the missing datum: subscriber conversion and ARPU. Neither Amazon nor Google has disclosed it. Absent that number, "the subscription era of home AI" is a slide, not a scoreboard.
Claim three: the platform takes margin, the manufacturer takes the tool call. Bosch, Whirlpool and iRobot traded their agent layer for ecosystem access. Their devices become callable tools inside someone else's runtime. This is structurally identical to exchange monetization decay: Binance Launchpad returns compressed from 100x to 10x because the venue captured the upside and the project absorbed the risk. Intermediated user relationships degrade brands into endpoints. Value accrues to the socket, not the plug.
Claim four: the liability surface is undocumented. Gemini for Home offers searchable video history and daily summaries โ household camera footage indexed, stored, and retrievable by natural language. GDPR and comparable data-protection regimes treat household surveillance processing with elevated obligations. The source material omits this dimension entirely.
Then the number that matters most. Apple's privacy architecture โ roughly 95% local processing โ is architectural, not marketing. But "95%" is a disclosure convention, scoped almost certainly to functions like on-device photo semantics, not multi-step agentic dialogue. A scoped percentage presented unscoped is a soft claim. Verify before you value it.
If agent execution is where margin lives, the on-chain question is whether it settles permissionlessly and verifiably. Tornado Cash already answered the liability half: writing code was treated as proximate to the crime. If a resident agent executes a command that causes harm, the code author becomes a defendant. That is the overhang on every runtime โ and a reason the execution layer centralizes by default. Centralization is legal insulation.
The consensus read is clean: Apple is in a positioning vacuum, so Amazon and Google win. Correlation is not causation. Apple's weakness across software, ecosystem density and openness is real and simultaneous. But "cloud wins" carries the same too-good-to-be-true signature.
Cloud-native agents inherit a seconds-long latency budget and a compliance burden that compounds with every jurisdiction. Searchable video is a regulatory bomb, not a moat. And the winners are running the identical platform-tax play that compressed exchange returns. The durable beneficiary is not the assistant. It is the substrate: on-device NPU, ISP, multimodal sensor, and the settlement rail underneath โ the same place value migrated on-chain.
Watch, not device counts, but the number of MCP endpoints that accept on-chain settlement โ and whether any consumer agent runtime exposes verifiable execution. If the agent layer settles permissionlessly, the platform tax breaks. If it does not, the smart home becomes a walled subscription zone with a regulatory alarm attached. Track the ARPU disclosure. Track the video-retention policy. Everything else is a dashboard of unaudited claims.