The case for local-first AI
The argument is usually made about privacy. Privacy is the weakest version of it.
The argument for running AI on your own device is almost always made about privacy, and privacy is the least interesting reason.
It is not a bad reason. It is just the one that loses. Privacy arguments have a poor record against convenience arguments, and they lose in a specific way: people agree with them, express concern, and then continue using the convenient thing. Everybody knows their data is being harvested. Almost nobody switches. A technology that requires the public to finally start caring about privacy after twenty years of not doing so is a technology waiting for a bus that is not coming.
There are three better arguments.
It works when nothing else does
A cloud model has a dependency graph. It requires your device, your local network, your ISP, the backbone, the provider’s edge, the provider’s inference cluster, and the provider’s continued willingness to serve you. Every link is a place where the answer stops arriving.
A local model requires your device.
This sounds like a marginal difference until you enumerate when the graph breaks. On a plane. On a train through a tunnel. In a hospital basement. In rural areas where coverage is theoretical. During the outage that takes down a third of the internet because one provider’s certificate expired. In countries where the state periodically switches off the network — one accounting found nearly two hundred social media shutdowns across forty-one African countries between 2016 and 2024, each affecting at least ten million people, and rising sharply since.
For a large share of the world, “the AI works only when the network does” is a severe constraint rather than a theoretical one. The provider does not have to be malicious for the service to be unavailable. It merely has to be far away.
It has no marginal cost
Every cloud inference call costs someone money. This is not a detail; it is the shape of the entire industry.
It means AI features must be metered, and metering means rate limits, tiers, and the persistent low-grade calculation of whether this question is worth a credit. It means free tiers exist to convert, and therefore exist to be insufficient. It means that a product built on cloud inference has a variable cost that scales with success, which forces monetization strategies that would be unnecessary if compute were free.
Local inference has a fixed cost — the device, which the user already bought — and a marginal cost of zero. You can call the model in a loop. You can run it on every keystroke. You can build features that would be economically absurd at a tenth of a cent per call, because the economics do not apply.
This is a larger change than it sounds. Consider what happened to software when storage stopped being metered: applications began to assume they could keep everything. Whole categories of product became possible because nobody had to ask whether a file was worth the disk. Compute is on the same trajectory, on the edge, and the products that assume free inference have not been built yet.
Nobody can take it away
A cloud model is a service, and every service is a relationship you do not control.
The model you built your product on will be deprecated. The pricing will change. The terms will be revised, and one of the revisions will restrict a use case you depend on. The provider will be acquired, or will decide your category competes with theirs, or will be compelled by a government to filter something you needed. Your account will be flagged by a classifier and you will discover there is no human to appeal to.
None of this is speculative. All of it has already happened to somebody, repeatedly, and the pattern will not stop because the incentives producing it have not changed.
A model that runs on your device runs whether or not the company that made it still exists. This is the property that open weights and local execution have in combination, and neither has it alone. Open weights on somebody else’s server is a service. Closed weights on your device is a lease. Open weights on your device is possession.
Tether’s own phrasing is that intelligence should not be a service one rents. Strip the grandiosity and the claim is precise: rental and ownership differ in who can end the arrangement.
The honest costs
Local inference is slower for large models. A phone GPU is not a rack of H100s, and pretending otherwise is how this argument gets discredited.
It burns battery. It requires the developer to care about the difference between an A18 and a five-year-old mid-range Android, because both are users and only one will have a good time. It shifts responsibility for optimization and correctness from a provider with a dedicated infrastructure team onto whoever ships the app.
And the frontier is not local. It is not going to be local. The largest models will always live where the largest amount of power and cooling lives, and there will always be questions for which those models are meaningfully better.
Where the line falls
The useful question is not whether local beats cloud. It is what fraction of what people actually ask an AI to do requires the frontier.
Watch what people use assistants for. Rephrase this. Summarize that. What does this error mean. Draft a reply. Explain this concept. Find the note where I wrote about the thing. Turn this photograph of a receipt into a number.
Very little of that needs a frontier model. Most of it needs a competent one, close to the data, with no latency and no meter running. And the tasks touching the most sensitive material — your health record, your finances, your notes, your messages — are exactly the ones where the case for keeping the computation local is strongest on every axis at once, privacy included.
That is the shape of it. Not local instead of cloud. Local for the large and unglamorous middle of what people do, where the frontier is overkill, the network is a liability, and the data is nobody else’s business. The frontier stays where it is, for the things that need it.
The interesting question about the next decade is not which side wins. It is how much of the middle turns out to be the middle.