Six cases, each with what works today and what does not yet. The labels mean:
Available
works today
Pilot
being built with a first user or partner
Coming
next in line, not released
Planned
an idea, not available yet
Smart home & cameras
Available
Home Assistant · Frigate · MQTT
Camera snapshots and doorbell clips go to a cloud API to be tagged, with a bill per camera, and stop working when the internet does.
How it works with the gateway
Home Assistant, Frigate or any script sends the snapshot to the hub with its own scoped key.
The hub routes it to the node whose chip runs the model best, for example a TV NPU running EfficientNet-Lite4 int8 in 22 ms.
Labels or an embedding come back; the image never leaves your network and every call is in the audit log.
Today and next
Available
Native and OpenAI-compatible HTTP API, callable from Home Assistant automations and scripts today
Coming
MQTT bridge (an example integration): a motion, doorbell or new-photo event triggers a model and the result is published back
Collector & scanner apps
Pilot
e.g. a trading-card recogniser
A card scanner needs a detector and an embedding model on a GPU server: the app maker pays for every scan, and it fails offline at a tournament.
How it works with the gateway
The pipeline splits into two models: a rotated-box detector finds each card, and an embedding model plus a nearest-neighbour search over about 14k card arts names it.
Both install as one package on the user’s own phone or TV; the small card index ships with it and updates with each card set.
The app calls the user’s hub instead of its own server, so photos stay on the user’s devices.
Today and next
Available
Multi-model nodes, image embeddings and classification through the hub
Pilot
Packaging the card detector and embedder as a gateway package (export to TFLite, accuracy checked against the server version)
Coming
Signed, versioned packages distributed through the marketplace
Privacy-first consumer apps
Available
Android client library · same-device IPC
A photo app either uploads photos to embed them, or ships its own 582 MB model and loads it on every phone.
How it works with the gateway
The app asks the client library for an embedding. With the gateway on the same phone, pixels go over Binder IPC in shared memory: no network, no JPEG loss.
Away from the gateway, the library uses the home hub, then the app’s own bundled model as a fallback.
Every path uses the same model and preprocessing, so the vectors stay comparable.
Today and next
Available
Same-device IPC with a signature permission and a genuine-gateway check, plus hub and app fallbacks
Available
Identical results through the hub (cosine 1.000000); 0.996 against the app’s own on-device embedder
Available
CLIP ViT-L/14-336 at ~1.45 s per image on a Pixel 7 Pro GPU
Model makers cannot test a model on hundreds of real phones, TVs and boxes with different chips, firmware and NPUs. Owners of those devices get nothing for their idle hardware.
How it works with the gateway
Opt in, and your idle devices run real-device benchmarks for model makers: “test my model on 500 real devices”, like a device farm, but for AI.
Only while the device is idle (never while you watch), within the power and data budget you set; jobs are verified with spot checks.
Planned, not available yet: whoever uses your devices could pay you, with a small platform fee. Not built, not priced, no dates; today sharing is free and only within trusted groups.
Today and next
Available
The pipeline it reuses: nodes, benchmarks and the registry
Available
Trusted groups (free sharing with people you know), owner protections (idle-only, charging, thermal), end-to-end sealing
Planned
Payment between users, metering, result verification
What a shared device is worth
Honestly: not much per job. A TV is on a few hours a day and suspends shortly after it is switched off, and a datacenter GPU does bulk inference very cheaply. Sharing makes sense where a datacenter cannot help: work that must stay on real, local devices, and people who would rather use a friend's idle NPU than a cloud.
What a datacenter cannot offer is real-device diversity: thousands of different TVs, phones and boxes with different chips, firmware and NPUs. Testing a model on them could be useful to model makers, app developers and chip and TV makers. Planned, not available yet.