Right now, artificial intelligence works like a mainframe. A handful of enormous models live in a handful of enormous buildings, and everyone queues for a turn. We think that arrangement is a stage, not a destination — and that the thing it becomes is an AI that belongs to you.
This is a position, not a prediction we can prove. We build things on the assumption that it is right, which is a reason to read it critically. At the end we have set out what would show us to be wrong, because a position that cannot be falsified is just an advertisement.
You have a conversation with an assistant that has read a substantial fraction of everything ever written. It is genuinely remarkable. And it does not know who you are.
It does not know the client you have been arguing with for three weeks, or how you decided the last time this came up, or that you already tried the obvious approach in March and it failed for a reason you would rather not repeat. Every conversation starts from nothing. You spend the first several minutes explaining your own situation to something that has forgotten you since yesterday.
Meanwhile, the material that would make it actually useful — your files, your correspondence, your accounts, your history — sits on your machine, and you are reluctant to send it anywhere. Reasonably so.
So the most capable tool most people have ever had access to is kept deliberately ignorant of the thing it would be most useful for. That is the shape of the current arrangement, and it is worth noticing how odd it is rather than accepting it as the natural order.
In the 1960s, computing meant a machine in a temperature-controlled room, owned by an institution, accessed by appointment. The idea that everyone would eventually have a computer on their desk was not obvious. It was a niche opinion, and the people holding it were told the economics would never work.
The centralisation was real, and it was also a consequence of cost. When the machine cost as much as a building, sharing it was the only sane arrangement. When the cost fell, the arrangement changed — not because anyone won an argument, but because the constraint that produced it went away.
We are early in the same curve. Training a frontier model costs enormous sums, which is why the organisations that can do it are few. But training a model and running one are different problems with different costs, and the second has been falling fast enough to change what is possible on ordinary hardware.
A laptop with 16GB of memory now runs a model that would have been close to state of the art a couple of years ago. Not the best available — but good enough to summarise a contract, draft a reply, sort a hundred documents, or answer a question about your own files. That is not a marginal capability. For most of what most people need, it is the capability.
Open models keep arriving, and keep closing the distance. The pattern of the last few years has been repeated: a frontier capability appears, and within a year or two an openly available model reaches roughly that level. The gap at the very top has not disappeared, and may not. But the gap at the level most work requires has narrowed to the point where, for a great many tasks, it stopped mattering.
Hardware is moving toward this. Unified memory architectures, neural accelerators in ordinary consumer machines, phones with processors designed for inference. The industry is building hardware that assumes models will run on the device. That investment is a bet by people with better information than ours, and it is a bet on local.
And the third one is the argument that actually convinces us.
An AI's value to you scales with how much of your world it can see. An assistant that has read your last three years of correspondence, your contracts, your notes and your accounts can do things that no general model can do, however capable — because the limiting factor stopped being intelligence and became context.
But you will only give that access to something you trust completely. And "trust completely" is a very high bar for a service operated by a company you have no relationship with, subject to terms that can change, in a jurisdiction that may not be yours, whose incentives are not identical to yours.
This is why we think the direction is not merely "some AI runs locally for compliance reasons". It is that the most useful configuration — an assistant that genuinely knows your work — is only reachable when the thing runs somewhere you control. The privacy and the capability are not in tension. They are the same requirement seen from two sides.
We work between Nouméa and Auckland, which gives this argument a texture it does not have when written from a city with reliable fibre and a data centre down the road.
In much of the Pacific, connectivity is a single cable and a set of assumptions. Outages are not hypothetical. A business dependency that requires a working connection to a server on another continent is a dependency with a failure mode that people here have actually experienced. Software that keeps working when the link goes down is not a preference; it is a specification.
There is a second dimension, less technical and more consequential. New Caledonia has twenty-eight Kanak languages. We built a dictionary that brings nine of them together with nearly twenty thousand words and recordings of native speakers — and the question of where that material lives, and who controls it, is not an infrastructure decision. Language is held by the people who speak it. Whether a community's own words end up as training data on servers belonging to a company that will never be accountable to that community is a question about power, and it does not become less of one because the technology is impressive.
Digital sovereignty reads as an abstraction in places that have always had it. It reads differently here, and that difference is part of why we build the way we do.
Not a chatbot with your name on it. Something with four properties:
On your machine, or on hardware you control. Which is what makes everything below safe enough to actually do.
Your documents, your correspondence, your history — not as training data absorbed into a shared model, but as material it can search and cite when you ask.
Context that persists. It gets more useful the longer you have it, in the way a colleague does and a stranger never can.
Not access you rent. Nobody can deprecate it, reprice it, or change what it is allowed to do for you.
That last property is the one people underestimate until it costs them something. Every business that has built a process around a service has eventually watched the terms change underneath it. An AI you own does not do that.
That the cloud is going away. It is not, and it should not. For the hardest reasoning, the largest models remain better, and it would be dishonest to pretend the gap has closed. Frontier capability will keep living in large data centres for the foreseeable future, because training is where the enormous costs are and those costs are not falling the way inference costs are.
What we expect is a split that is already visible in how careful businesses work: the personal, continuous, context-heavy work moves close to the person, and the occasional hard problem goes to the big model. Not one replacing the other — each doing what it is shaped for.
We are also not claiming this is easy today. Running a model locally still means caring about memory, quantisation and hardware. Building an assistant that has genuinely read your work is a project, not a download. Most of the tooling assumes a technical user. That is a real gap, and it is a large part of what we spend our time on.
A position worth holding is one you could abandon. These would do it:
We watch all four. None of them is currently happening, which is why we still hold the position — and why we build systems where the model can be swapped rather than wired in, because being right about direction is not the same as being right about timing.
It is why the outreach system we sell runs on your own machine and offers a local model as a first-class option rather than an afterthought — your prospect list is your business, and it should not have to leave. It is why the private systems we build for law firms, accountants and clinics run inside their walls. It is why we would rather sell something once than rent it forever: the same conviction applied to the commercial arrangement.
It is also why we spend time on tooling that makes local AI reachable by people who are not engineers. The technology is closer to ready than the experience of using it, and that gap is ordinary, unglamorous work rather than a breakthrough.
If you are thinking about where your own systems should run, we are glad to talk about it — including in the cases where the honest answer is that the cloud is the right choice for you.
We build private AI systems in environments you control — your data never trains public models, and engagements are available under NDA. Start with a free audit of where AI would actually help.