Why this matters
For forty years computing got more personal. Mainframe, mini, desktop, laptop, phone.
Then intelligence arrived — and went straight back into somebody else's building.
You rent it by the token. It reads everything you give it. That is time-sharing with a subscription button.
Own 1 puts it back where it belongs. Intelligence should run where its owner decides, cost what the market says it costs, and remember only what its owner allows.
Local and cloud models available across the entire system as shared intelligence services.
Agents can pay for approved models, APIs and compute under system-level spending policy.
TPM2-protected keys, containerised execution, Btrfs snapshots and one-command rollback.
The category
Apple, NVIDIA and a handful of others are now building for the same desk. That is a good sign: it means the argument for running intelligence locally has been won. The question is no longer whether a machine can hold a large open model. It is what the operating system does once the model is loaded.
An AI-native computer. The intelligence is resident, and the OS is built around it.
Superb hardware. With enough unified memory it will hold a large open model — as a program you run, not a system you own.
Illustration, not a product photograph. Verify current Apple configurator pricing before publishing any figure.
| Capability | Own 1 | Mac Studio |
|---|---|---|
| Runs open models locally | Yes | Yes |
| Local model runtime built into the OS | Yes | Install it yourself |
| Two local LLMs at once, plus 8 agents | Yes | Manual orchestration |
| Agent identity as an OS primitive | Yes | None |
| Per-agent budgets and spending limits | Yes | None |
| Native wallet and machine payments | Yes | None |
| Atomic rollback to a previous system | Yes | Time Machine restore |
| Bursts to a priced compute market | Yes · 34 providers | None |
| App store for models, agents, skills, MCPs | Yes | General apps |
| Shared by 5–10 concurrent users | Yes | Single-seat |
| Open source, forkable, inspectable | Yes | Proprietary |
| User-replaceable storage | Yes | Soldered |
Comparison at broadly matched memory and storage. Apple leads on display pipeline, creative tooling and ecosystem maturity — if those are your constraints, buy the Mac.
Proof
No datacenter, no cloud, no outbound API calls — one fanless Own 1 with integrated graphics, 7 May to 30 June 2026.
Thirty-one days of optimisation on identical hardware — every gain came from software. The 104 failures stay in the published dataset because they mark exactly where a small machine stops and OwnCloud starts.
What it is
Own 1 runs open language, vision and voice models on your desk — with on-device privacy, AI-native security, native agents, a machine wallet and a direct path to OwnCloud when a workload genuinely needs more.
Models, agents and data stay on device by default. Not as a setting — as the design.
Silent thermals tuned for continuous inference. Agents that run all night, not until it throttles.
OwnCloud on tap when a workload needs global compute, under rules you set.
Rollback recovery, open software, no forced services, nothing soldered shut.
The object
A fanless aluminium body sized for a desk, not a rack. Silent enough to sit beside while eight agents work through the night.
USB-A, SD card reader, USB-C, 3.5 mm audio and a fingerprint reader — the things you reach for, on the side you face.
A single machined shell with a soft radius on every edge. No seams, no vents on the faces you see.
Dual 2.5 GbE, HDMI, DisplayPort, USB-C and the full-width thermal array that lets it run silent under sustained inference.
Smaller than a sheet of A5 paper, holding two language models, eight agents and your private memory.
The economics
A SuperAgent coordinates eight specialists, and between them they cover twelve job roles. They do not sleep, they do not context-switch, and they cost electricity.
Modelled at a blended $85,000 per role. Illustrative, not a promise of equivalence — an agent is not a person, and the roles above describe workloads, not job titles.
The operating system
A lightweight open operating system on a NixOS base, with TPM2-protected keys, Btrfs snapshots, atomic updates and one-command rollback. It boots to a desktop where intelligence is already resident.
Every agent inherits ten systems it cannot opt out of: identity, permissions, memory, data access, wallet, budget, payment policy, usage metering, audit history and compute routing. Nothing spends without a rule you wrote. Every action leaves a receipt.
OwnApp Store
Models, agents, skills, MCPs, data sets and dev tools — searchable, rated and versioned, the way an app store should be. Except what you install here is a mind, not an app.
Eight shelves, one hundred and ninety listings at launch, all installable without leaving the machine.
Open weights carry a Get button — they download and run on your hardware, offline, forever. Closed models can only ever be an API call to someone else's building. Same shelf, and the difference is impossible to miss.
OwnCloud connectivity
Own 1 decides what should stay personal — weighing privacy sensitivity, model size, local hardware, cloud price, latency, energy and budget before a task is allowed to leave the device at all.
Private files, memory, everyday reasoning, code, voice. Nothing transmitted, nothing metered.
Bursts to 34 providers across 482 regions when a job needs more than a desk can give. You choose the country.
A frontier API is never reached silently. It requires your explicit approval, every time.
Costs modelled on one open model run in four places, where only the location changes. A rented GPU for the same work runs $10.48 per 1,000 tasks.
Stablecoins & agents
The moment an agent needs a model, an API or an hour of GPU, it hits a checkout built for humans. Own 1 gives it a wallet instead — and a policy it cannot exceed.
A compromised agent cannot drain an account it was never given. It can spend what you allowed, on what you approved, until the date you set — and every attempt beyond that is refused and logged.
When one is not enough
Own 1 was designed to sit on top of itself. Four nodes occupy the footprint of one, share a 10 Gbps fabric, and present as a single pool to OwnOS — so a studio grows by adding a box rather than renting a rack.
The part that compounds
The price of a model falls every quarter. Five years of your memory, working style, documents, workflows and trusted relationships does not. Own 1 builds a Personal Intelligence Profile that lives locally — the more it learns, the more useful it becomes, without any of it accumulating in someone else's dataset.
That is the whole bet. Rent intelligence and you own nothing when the subscription lapses. Own the machine and the commodity underneath you keeps getting cheaper while what you have built keeps getting more valuable.
Specifications
| Feature | Specification |
|---|---|
| Processor | Intel Core Ultra 9 285H · 16 cores, up to 5.4 GHz |
| Graphics | Intel Arc 140T · PCIe x8 for external GPU |
| On-device AI | 99 TOPS (Int8, platform) |
| Unified memory | Up to 96 GB |
| Storage | Up to 8 TB NVMe · user replaceable |
| Local models | 2 compatible LLMs simultaneously |
| Agents | 8 concurrent, coordinated by a SuperAgent |
| Shared users | 5 concurrent · up to 10 seats |
| Networking | 10 Gbps Ethernet · Wi-Fi 7 |
| Operating system | OwnOS · NixOS base · TPM2 secure enclave |
| Security | TPM2 keys · full-disk encryption · Btrfs snapshots |
| Thermals | Fanless · tuned for continuous inference |
Preliminary and subject to change before production. Figures follow the current engineering configuration; earlier published material quotes different core counts and memory ceilings, and will be reconciled at launch.
Reserve
Around $4,000 once, and roughly $8 a month in electricity after that. A single light user should not buy this machine. A household, a studio or a small team should.
For AI power users, developers and builders.
Maximum capacity for heavier workloads and long-running agents.
Ten seats works out at roughly $450 per person, once. Payback against equivalent AI subscriptions lands around month eight, excluding electricity.
Questions
Two compatible local models simultaneously, alongside vector storage, private memory, voice services and up to eight specialised agents coordinated by a SuperAgent. That is the difference between running a model and running a team.
It weighs privacy sensitivity, model size, local hardware, cloud price, latency, energy and budget before a task is allowed to leave at all. Local when possible, OwnCloud when it is genuinely cheaper or faster, a hosted API only if you approve it.
The context, working style, tools, documents, workflows, agent history and trusted relationships the machine learns about you — held locally. It makes the device more useful over time without any of it accumulating in a central dataset.
Yes. The models are on the internal NVMe. The network is used only when you send a workload out, and the Control Centre shows you when that happens.
Up to five people reach it from their phones, and the machine supports up to ten seats. A household, a studio or a small team shares one device.
No. The operating system, the local models and the applications carry no recurring fee. Beyond electricity there are no API bills and no credits to run out of. OwnCloud is optional and billed only when a workload uses it.
OwnOS is built on NixOS, so every update is a complete system generation with a Btrfs snapshot behind it. If a change causes a problem, you reboot into the previous generation.
1,224 runs across 44 models and seven modalities, over 55 days, on one fanless machine with integrated graphics — no datacenter, no outbound API calls. The 104 failures stay in the published dataset because they mark exactly where a small machine stops and OwnCloud begins.
Own 1
The future of intelligence must not belong to four companies. It starts with one machine, on one desk, answering to one person.
