Hook
Three numbers. HomeKit: roughly 1,000 certified devices. Amazon Alexa: 140,000. Google Home: 85,000.
I spent the first week of this bear market reading MCP server manifests — the JSON files that describe what a device can actually do — instead of candles. That was the wrong instinct and the right call. The chart whispers before the market screams, and this chart is whispering something crypto hasn't priced.
Two orders of magnitude is not a product gap. It's a liquidity gap. An agent's usefulness scales non-linearly with the number of actions it can execute. A household assistant with 1,000 callable endpoints isn't 140 times worse than one with 140,000 — after the third task it stops being an agent and goes back to being a light switch with a voice.
That number should be on a terminal right now. It isn't.
Context
For a decade the smart home was a protocol war — Matter, Zigbee, Thread, and the certification badges that came with them. Apple fought there and fought well. That war is over, it ended in a draw, and the draw is worthless, because value migrated up the stack to the orchestration layer: model reasoning, tool invocation, device API execution.
Amazon is charging $19.99 a month for Alexa+. Google is selling a $99.99 Gemini for Home speaker with tiered subscriptions behind it. Samsung went self-contained — a Bespoke AI Family Hub running on-device AI Vision to track food expiry, which is a multimodal model executing on a refrigerator controller, not a keynote slide.
The connective tissue beneath all of it is MCP, the Model Context Protocol, which Amazon opened to Bosch, Whirlpool and iRobot. Read that as infrastructure, not as a press release. MCP flattens a device's capabilities into a tool set a model can call. It's the standardized socket between reasoning and the physical world. Whoever extends that socket decides who pays rent on it.
Meanwhile Apple's HomePad is reportedly hardware-complete and software-unready. In 2017 I built a Python scanner that aggregated 150-plus ICO whitepapers overnight and published alerts before the crowd woke up. The lesson from that year holds: hardware that ships without its software layer is inventory, not product.
Core
Three crypto reads, ordered by how tradable they actually are.
First: the oracle problem just got rebuilt for actuators. Chainlink's founding insight was that contracts can't see the world, so you need a feed. MCP is the same primitive one layer down — not "what is the price" but "what can this machine do, and may I do it." DePIN has been selling decentralized physical infrastructure for four years. The physical interop standard actually being adopted is a centralized one from Amazon. That is not a DePIN win. That is a DePIN rug-pull executed by adoption. DePIN tokens that route around this stack need a reason to exist that isn't the word "decentralized." So far, none has shipped one.
Second: the demand curve bends toward inference. Multi-step planning plus tool calls plus multimodal perception multiplies tokens per interaction by roughly an order of magnitude versus a single wake-word command. That's a direct tailwind to cloud inference bills and, by extension, to decentralized compute networks. But there's a kill switch, and it's physics. Household interaction budgets live under a second; multi-step cloud reasoning already runs into seconds. No decentralized compute marketplace wins a sub-second household SLA until edge nodes physically sit in the last mile. Renders of idle GPUs don't move fiber.
Third, and the one I keep returning to: the sequencer problem, restated. My position on Layer 2s has been boringly consistent — "decentralized sequencing" has been a PowerPoint for two years while a single operator orders your transactions, and the marketing improved faster than the architecture. Cloud-native home agents are the identical pattern with a nicer voice. One operator sequences every command in your house, holds the index of your cameras, and can reorder, censor or monetize your request stream. Google's searchable video history makes it concrete: household footage indexed, stored, retrievable through a sentence. That's not a feature. That's a centralized sequencer with a lens.
Apple's roughly 95% local processing is architecturally real, not a slogan. But right now it protects a capability that doesn't exist yet. Privacy that costs you the feature isn't a moat — it's a tariff.
Which raises the question crypto keeps refusing to ask: who accrues the fee? A $99.99 speaker is almost certainly sold at or below cost, so the entire model rests on subscription lifetime value. Bosch and Whirlpool are already learning the answer — once their capabilities are flattened into someone else's tool set, they stop being brands and become endpoints. Device makers are being demoted from products to API calls, and most of them signed the paperwork inside a press release.
Contrarian
The consensus read will be simple: Apple lost, Amazon and Google won, buy compute. That's the trade everyone can see, which usually means it's already priced.
The unreported angle is that the compliance asset may outperform the capability asset. Indexing every household camera into a natural-language-searchable corpus is the most regulation-bait feature shipped in consumer AI this decade. Under GDPR and China's PIPL it looks less like a product and more like high-risk processing with elevated obligations attached. Regulators move slowly, and when they move they move on the largest indexed corpus, never the most private one. The jurisdiction that writes the household-AI data rule first doesn't just regulate the market — it becomes the venue for it, in the same way licensing regimes don't attract builders so much as they redirect capital toward whoever secured the license first. Note the pattern: the same jurisdictional competition that turned licensing into venue-selection sport now applies to household data residency.
So the mispriced assets aren't the frontier AI names. They're zero-knowledge proving, fully homomorphic encryption, and edge inference silicon. The code is cold, but the hype is hot — and the regulatory file is where this trade actually prints.
And the second blind spot, stated plainly: do not put execution on-chain. Running agent orchestration on a base layer is the same category error as minting tokens on Bitcoin — expensive, degraded, and insulting to both sides. Chains should settle identity, payment and attestation. Everything else belongs off-chain, where it can be fast.
Takeaway
Three signals to watch. One, whether Amazon or Google discloses household agent subscription conversion and ARPU — a subscription without disclosed conversion is a pricing theory, not a revenue line. Two, whether Apple's next developer cycle ships a rebuilt Siri instead of a reskinned one. Three, whether Brussels or Beijing enforces on indexed household footage before 2027.
Bear market rule holds: liquidity is the only truth that bleeds. Right now the liquidity in agent execution sits on three cloud balance sheets. Hedge that. Don't hold it.