The sharpest positioning statement we have read this month comes from a non-profit that has not shipped anything yet. We took a close look at Kavai, a pre-launch initiative in Oslo building free AI tools that run entirely on the user's own device, and its three-line promise ("No subscriptions. No coding. No new hardware.") says more about who AI currently serves than most funded startup decks manage. Inside: why on-device is a wedge rather than a privacy checkbox, who the ICP is when nobody pays, and the wrong number the entire local-AI category keeps optimising for.
Picture the person this is for. There is a rental contract in the downloads folder, a laptop bought in 2018, and a quiet suspicion of chatbots that want the document uploaded somewhere. Cloud AI's answer to this person has been consistent: pay a monthly fee, accept the data terms, buy a newer machine, or sit this one out. The local-AI scene's answer has been no better in practice: download a runtime, pick a model file, follow a README written for people who already followed it once. Both roads lead to the same place. Kavai frames the gap bluntly: cloud AI sends your request away, and the people who most need help with their documents are the least able to pay subscriptions, learn tooling, or surrender personal information to a third party. Whether that framing becomes a product is unproven. As a description of a hole in the market, it is exact.
The wedge: three refusals, and patience as the moat
Most privacy-first AI projects fight one battle at a time. Kavai's wedge is a stack of refusals: no subscriptions, no coding, no new hardware, nothing leaving the device. Anyone can claim two of those three. The third, explicit older-hardware optimisation, is where the economics break competitors, because every hour a commercial vendor spends making a model run well on a 2015 laptop is an hour not spent on customers who upgrade. A non-profit can serve people whose lifetime value is zero by design. The site is unusually honest about where the difficulty sits: "The first trials will be deliberately hands-on because the difficult part is not merely running a model. It is making the entire experience clear enough that a non-technical person can say, 'That just worked.'" That is grunt work, not research, and grunt work is hard to copy because nobody wants to do it. The founder, Kristoffer Rolf Deinoff, has focused on running AI models locally since 2023, and the plan leans on guided deployments and public trials rather than a download-and-hope launch.
The ICP: two of them, and only one uses the product
Since end users pay nothing, the ICP splits in two. The beneficiary is non-technical, Norwegian, on ordinary or ageing hardware, and wants help with contract summaries, writing, photos and voice notes without a subscription or a data handover. Notice how specific that is. Not "privacy-conscious power users", which is who local AI actually gets built for, but renters, pensioners and anyone whose documents are sensitive partly because their finances are tight. The buyer is the funder: grant bodies, donors, sponsors and corporate CSR teams, plus the volunteers with grant-application experience Kavai is openly recruiting to its board. Recruiting grant-writing competence before shipping a product is a telling move. It names funders as a first-class audience rather than an afterthought. What can a funder evaluate today? Only the founder. Deinoff is a software developer, architect, speaker and teacher who has built systems for companies, organisations and the public sector, and in 2022 he co-founded Paracrew, which is still operating as Paracrew Humanitarian Aid, delivering medical supplies into Ukraine. That is a verifiable record of finishing things in a domain harder than software. No user numbers exist, and nobody should pretend otherwise. But a named founder with demonstrated follow-through is more than most pre-launch initiatives put on the table.
What the category still gets wrong: optimising for people who already have it
Local AI is built by and for the already-equipped. Watch any release cycle and the numbers that get celebrated are parameters, context length and tokens per second on hardware nobody's grandmother owns. The DIY stack of runtimes, model files and forum threads assumes the reader knows what quantisation means. Cloud vendors optimise a different wrong number: seat expansion among people with corporate cards and a tolerance for SaaS. Both maps leave blank the same territory, the people with sensitive documents, thin budgets and zero patience for setup. Kavai's implicit metric is different and worth writing down: how many people can get real help on the machine they already own, without reading a README. The privacy framing also inverts the usual pitch. Most of the industry sells privacy as a premium feature for people who care enough to pay extra. Here privacy is an access condition: a tenancy dispute or a benefit claim letter is more sensitive than the average corporate memo, and precisely because of that, its owner is least able to buy protection. Treat privacy as an upsell and you get the demand curve backwards.
For operators, the thing to watch is not the models, which will keep improving for everyone, but the funding model and the UX bar it sets. If grants and CSR partnerships can sustain genuinely free, on-device AI for non-technical users, then AI access starts to resemble a public good in the way libraries and public broadcasting did, and that changes who the competition is. If a one-person initiative in Oslo ever cracks "That just worked" on a seven-year-old laptop, the convenience argument for cloud subscriptions weakens for every vendor. The lesson generalises past Norway: the segments nobody can monetise are often the ones that reveal whether your category's ease-of-use story was ever true.