The man predicting that every household will soon run its own AI assistant builds the largest computer chips ever manufactured, objects you could not fit behind a television, let alone inside a smart speaker. And in the same breath, he says those assistants will run partly on local devices and partly in the cloud. Both things are true at once, and the gap between them is where all the interesting engineering lives.
TechCrunch reported that Cerebras CEO Andrew Feldman believes we’re nearing a future where everyone has a personal AI assistant at home, one that becomes as commonplace as a household appliance. The split between local silicon and remote compute is the part of the prediction worth taking seriously, because it’s an architectural claim, not a marketing one. The sources don’t put a date on it or spell out what the device looks like. Fine. The shape of the problem is visible without those details.
Appliance is a stronger word than it sounds
Calling something an appliance is a commitment to a set of properties most AI systems currently fail. A refrigerator does not ask you to retry. A dishwasher does not degrade when your ISP has a bad afternoon. Appliances have bounded failure modes, predictable power draw, decade-long service lives, and interfaces your relatives can operate without instruction. They also do not quietly change behavior overnight because a vendor shipped a new model weight.
That last one is the hard part. An assistant that lives half in your house and half in a datacenter inherits the update cadence of the datacenter. Your local component is frozen hardware with a fixed memory budget; the remote component evolves weekly. Keeping those two halves behaviorally consistent is a versioning problem that the agent research community has mostly waved at rather than solved. If the local planner was tuned against one remote model’s tool-calling quirks and the remote model changes, the agent’s reliability shifts without anyone touching the device.
Where the split actually falls
The useful question is not whether inference happens locally but which parts of an agent loop belong where. A reasonable decomposition looks something like this:
- Local, always: wake-word detection, speech transcription, intent classification, personal memory storage, retrieval over your own documents, and the policy layer that decides what is allowed to leave the house.
- Local, preferably: short-horizon planning, routine tool calls against devices on your own network, and any action with physical consequences.
- Remote: long-horizon reasoning, anything needing a frontier-scale model, and tasks that already involve external services.
Put that way, the local side is less about raw model size and more about latency, state, and trust boundaries. Your personal memory is the asset that cannot reasonably be centralized, both because it’s yours and because the retrieval loop needs to be fast and offline-tolerant. The frontier model is the asset that cannot reasonably be localized, because its weights and its serving economics belong to someone else’s building.
The routing layer is the product
If that division holds, then the thing being sold into homes is not a small language model in a box. It’s a router with opinions, a component that decides, per request, whether to answer locally, escalate, or refuse. That decision has to account for privacy policy, cost, latency budget, network state, and confidence. Getting it wrong in one direction leaks your calendar to a third party. Getting it wrong in the other direction gives you a confidently bad answer from an undersized local model.
We don’t have good evaluation methodology for this. Benchmarks measure model quality on fixed inputs; they don’t measure whether a hybrid system made the right escalation choice under degraded conditions. An honest appliance-grade spec would include things like: behavior with the network down, worst-case response time, what data crossed the boundary this month, and whether the same question asked twice in different conditions yields compatible answers. None of those are standard today.
Why a wafer-scale company says this
There’s no contradiction in a company that builds enormous chips arguing for assistants in every home. Hybrid architectures increase total inference demand rather than replacing it. Every local assistant that escalates is a client for fast remote serving, and the cheaper and faster that escalation becomes, the more the local half can afford to defer. A fast cloud tier makes a modest local tier viable. That’s a coherent commercial position and also, as it happens, a defensible technical one.
What I’d want from the next version of this prediction is specificity about the contract between halves. Not the device, not the date. The protocol: what state lives where, who owns it, how versions are pinned, and what the system promises when the connection drops. Household appliances earned their trust through boring guarantees. Agents will have to earn it the same way, and the guarantees are the engineering work nobody has finished yet.
🕒 Published: