WINDOW
YEARS, NOT DECADES
Before the defaults ship.
> EPISODE 01 // OFFLINE-READY NOTEBOOKLM AUDIO
The window for local-first AI is open right now, and it won't stay open. Hardware is almost there, ARM boards with neural processing units are shipping at consumer prices, the "AI PC" push from Intel and AMD means NPUs are becoming standard features rather than premium upsells, and by mid-2026 capable local inference hardware should be available for under $200.
The machine is no longer the barrier, the software layer is where this gets decided.
Apple, Google, and Meta are watching the same trend lines and they are not going to cede this territory. The response is already predictable, "local" AI that phones home, on-device processing with cloud-mandatory features, privacy marketing sitting on top of telemetry requirements, the UX of sovereignty built on the architecture of dependence.
⚠
The threat model. Apple ships a $299 home AI device with seamless ecosystem integration and "privacy-first" marketing that processes locally but syncs to iCloud, works offline but degrades without Apple services, and is beautiful and convenient enough that most people will never question what it's sending back.
The window closes because most people stop looking for alternatives once a convenient default exists, and by the time the alternative matures, the habit is already formed. The default wins on convenience rather than quality.
Data extraction has moved in stages, each one capturing a new surface area of human life. Personal AI is the last one.
The questions you ask an AI assistant are more revealing than your search history because search shows what you want to know while AI conversations show how you think about what you want to know, your reasoning patterns, your uncertainties, your decision-making process, laid out across hundreds of conversations in a way no other platform has ever had access to.
The business model is straightforward and the cost of "free" services is your agency rather than just your data.
In this system you are the inventory, and personal AI is the last surface area left to capture.
NOW
Hardware is commoditising. Rockchip RK3588, Orange Pi 5 Plus, Qualcomm edge AI chips are shipping at consumer prices. Apple puts NPUs in every device. Intel and AMD make neural processing units a standard feature rather than a premium upsell. Capable local inference hardware is available for under $200 and the price is still falling.
H1 2026
The defaults start shipping. Apple is reportedly preparing a HomePad smart home hub with a 7-inch display, Face ID for per-user personalisation, and deep HomeKit integration, rumoured for spring 2026. Google and Amazon are iterating on existing hubs. The revamped Siri has already been delayed once from iOS 18.4, which suggests the AI layer is harder than Apple expected, but the hardware is ready and the default UX for home AI is being established.
H2 2026
The installed base grows. If Apple ships the home hub at around $350, it will be in millions of homes by Christmas. It will process locally but sync to iCloud, work offline but degrade without Apple services. Beautiful, convenient, and the threat model described above. Every month without a credible alternative is ground ceded to defaults that are being designed right now.
2027
Network effects compound. Apple is reportedly working on home robots and AI glasses for 2027 and beyond. Third-party developers build for the dominant platforms first. Switching costs become visible, your routines, your preferences, and your family's usage patterns are all stored in the ecosystem and leaving means rebuilding them from scratch.
2028+
The window narrows. The habit is formed, two years of household patterns invested in one ecosystem, and leaving means rebuilding all of it from scratch. The alternative has to be better and easier than the default, because privacy alone has never won a consumer market.
Out-marketing Apple and out-spending Google aren't required. What's required is to exist as a credible alternative.
Even if most users never switch, the existence of a real option constrains the worst behaviours.
Running inference locally keeps your data off their servers, but it can't fix what was baked into the model during training. The weights are the worldview, and a model trained to subtly favour certain products or positions will carry that bias wherever it runs, including on your own hardware.
⚠
The honest limitation. You can verify where inference happens but you cannot easily verify what was baked into 70 billion parameters during training.
What local-first provides is the option to respond, model diversity, swappability if bias is found, community red-teaming of open weights, and the ability to adopt better solutions when they exist. Running local means that when a problem is discovered you can do something about it, which is more than cloud providers have ever offered.
LocalGhost is a vision, not a product. Right now there's a repo, this website, and the architecture in my head, no working software, no hardware prototypes. I built this over Christmas because the window won't wait for me to be ready, and because someone needs to at least make the argument before the defaults ship without it.
If this convinces one person to build local-first software, or surfaces one privacy-respecting project that deserved more attention, it did its job.
The defaults haven't shipped yet and the territory is unclaimed. The next few months are the ones that decide what normal looks like. [ localghost.ai // hard-truths ]