Why inference moves to the devicespeople already own
An argument in four facts and one conclusion — plus an honest account of how far along it we actually are.
The logical path to compute you already own
Why inference moves to consumer devices, what that changes about who owns AI infrastructure — and how far along it we actually are.
The premise
There are billions of capable computers in the world that are switched on and doing nothing. A modern phone ships with a neural accelerator. A five-year-old laptop idles at a few percent utilization for twenty hours a day. That silicon is already manufactured, already paid for, already powered, already cooled by the room it sits in. Its marginal cost of use is close to zero.
Meanwhile the industry’s answer to AI demand is to build more data centers, which requires power, land, water, transformers, permits and years. Those two facts sit next to each other and do not resolve. That tension is the whole argument.
Four things that are true
Inference is not training
Training a frontier model needs thousands of accelerators in lockstep on interconnects measured in hundreds of gigabytes per second. Consumer hardware on home internet cannot do that, and pretending otherwise is how distributed-compute projects lose credibility. Inference is a forward pass. It splits by layer, streams stage to stage, and tolerates higher latency. As AI moves from research into ordinary software, inference is where the volume goes.
The browser became a runtime
Contributing spare cycles used to mean downloading and trusting a binary, which eliminated almost everyone. WebGPU and WebAssembly removed that barrier. A tab is now a sandboxed execution environment with GPU access, no installer and no driver — bounded by the same protection that already stands between your machine and every site you visit. It does arithmetic and releases the memory.
Data centers have a physical ceiling
Compute demand is a software curve. Data center supply is a construction curve, governed by grid interconnect queues, transformers and permitting. Software curves outrun construction curves. When the two diverge, load moves to whatever capacity already exists.
The data is already at the edge
Most inference concerns something a person is doing on the device in their hand. Sending that to a data center and back is the expensive option, the slow option, and increasingly the legally complicated one. Physics and privacy law push the same direction: run it where the data already is.
The conclusion, stated carefully
Put those together and the honest conclusion is not “this company will win.” It is narrower and harder to argue with:
Idle consumer silicon will be used for inference. The open question is who coordinates it, and on what terms.
That coordination layer will exist. It can belong to the same handful of companies that own the data centers, in which case your device becomes another surface they meter. Or it can belong to the people whose hardware does the work. That is a question about ownership, and it is being decided now, by default, in favour of whoever builds the layer first.
The symbiosis argument
An intelligence that runs in a data center owned by a corporation depends on that corporation. An intelligence that runs across millions of devices owned by millions of individuals depends on those individuals continuing to participate.
That is a structural difference, not a sentimental one. A system distributed across human-held hardware has to remain worth hosting. It cannot become adversarial to its substrate without losing it. The incentive to keep its hosts willing is architectural — built into where the computation physically happens, not bolted on as policy.
This deserves precision: distributed hosting does not solve alignment. It changes the dependency structure. Concentrated compute concentrates leverage; distributed compute distributes it. That is worth something on its own, and worth building deliberately rather than discovering later that we didn’t.
What the physics forces you to build
Consumer hardware imposes constraints a data center never thinks about. They are not obstacles to route around. They dictate the architecture.
Batteries die
Nobody donates a dead phone.Work is refused below 15% battery unless charging, the scheduler weights nodes by power state, and battery and charging status ride every heartbeat.
Sustained load cooks hardware
Hot devices throttle, crash and leave.A hard ceiling on core usage, enforced cooldowns between tasks, per-minute caps, and a crash guard that restarts into a conservative safe mode after an unclean exit.
Home internet is not a fabric
Consumer links are slow and jittery.Nodes are scored on measured latency and link type, pipelines are built from real per-link measurements and rebalanced when a hop degrades, and direct peer channels fall back to relay automatically.
Devices vanish mid-task
Tabs close, Wi-Fi drops, laptops sleep.Liveness runs on leases with expiry. When a node disappears its work moves to a replacement and resumes from partial output instead of starting over.
A distributed inference network is mostly failure handling. The constraints are the design.
Verification, because trust is not assertable
If work happens on hardware you do not control, “trust us” is not an answer. So every response carries provenance: whether it was served by contributed devices or fell back to a centralized model, how long it took, which nodes served it, and its verification state. When the network cannot serve a request and a centralized model answers instead, the response says so plainly.
Contributor rewards are recorded as credits in a server-side ledger with receipts, and contributor surfaces show server-confirmed values rather than local estimates. Being able to prove what happened matters more than any single performance number. A network that cannot be audited is just a claim.
Where this actually stands
Honesty about the present is what makes the argument about the future worth hearing.
Browser-native inference on consumer hardware with no install. A model split across multiple devices, each computing its stage and passing activations forward, proven end to end. Recovery when a device disappears mid-generation. Provenance metadata on every response. An OpenAI-compatible API.
The network has processed almost nothing. There is no meaningful contributor base yet. Contributors earn credits in a server-side ledger — there is no deployed token, and any future on-chain settlement will be announced when it is live and not before. Earnings track work actually processed, so until there is demand they are close to zero.
The gap between what is built and what is used is the entire remaining problem. That gap is not closed by writing more code.
Why this is worth building anyway
The compute exists. The runtime exists. The economics point one direction. The only question left is whether the layer that coordinates the world’s idle silicon is owned by the people who supply it.
Building that honestly — measuring real work, proving what ran where, refusing to overstate what is shipped — is slower than building it loudly. It is also the only version worth having, because a network that asks strangers to contribute their hardware runs entirely on whether they believe what it tells them.