Sovereign AI, at home
Own your intelligence.
Ark One is a sovereign AI appliance that sits beside your router. It runs frontier-class models locally, learns from you privately, and gets better — and cheaper — the more you use it.
- Runs 240B-class models locally
- Post-quantum encrypted
- Works fully offline
- No subscription required
The problem
Renting intelligence is a trap.
Your data trains your competitors.
Every prompt you send a cloud AI is a disclosure. Enterprises are paying for tokens while handing over their proprietary edge.
Access can vanish overnight.
Repriced, throttled, deplatformed, or cut off by a severed cable — a rented capability is not infrastructure, it's a subscription.
You don't own your algorithms.
The models, the rankings, the feeds — all opaque, all optimized for someone else.
AI today is in its mainframe era: chat windows are terminals, data centers are mainframes, and the meter runs by the token. Ark One is the personal computer moment.
The solution
The whole loop, in your home.
It runs the model locally.
Frontier-class open models on efficient AMD silicon — full quality, no cloud required.
It remembers like you do.
Your documents and interactions become a private knowledge graph you can inspect and edit — not a black box.
It learns overnight.
While you sleep, Ark One distills the day into private model improvements. Nothing ever leaves the box. The more you use it, the better and cheaper it gets.
It runs itself.
Self-healing OS, hardware kill-switch for total air-gap (paranoid mode), and a physical recovery button that restores a known-good state. No IT team required.
Product tiers
Three appliances. One philosophy.
All figures projected · final specs at launch
Ark One Lite
Automates your daily loops.
Target pricing
Target pricing at launch
- Memory
- 64 GB
- Models
- Runs a 27B model at ~70 tokens/s
The near-silent ambient agent for the home. Enough capability for daily assistance, drafting and household knowledge.
Ark One Pro
BeachheadFor professionals whose data can't leave the office.
Target pricing
~$2,600
- Memory
- 128 GB
- Models
- Up to 120B-class models
Purpose-built for regulated professions. Runs mid-frontier models with headroom for RAG over sensitive documents.
Ark One Max
Frontier-class, owned outright.
Target pricing
~$5,200
- Memory
- 256 GB pooled (dual-node)
- Models
- 240B-class MoE at 40+ tokens/s (projected)
Built for coding and agentic workflows. Two nodes pool memory for the largest open models available.
All figures projected from the published architecture study; final specs at launch.
The economics
It pays for itself.
At typical professional usage, Ark One Max breaks even against cloud AI spend in a projected 9–18 months. After that, every month is yours. And when it's idle, ArkNet lets it earn: lease compute, cache content, serve your neighborhood — the appliance becomes an asset, not an expense.
9–18 mo
Projected break-even vs. cloud AI spend
>90%
Projected cut in residual API spend
$0
Mandatory subscription
Illustrative; break-even depends on usage profile.
ArkNet
Stronger together, private apart.
Every Ark can stand alone, join a private cluster (your clinic, your firm), or opt into ArkNet — a decentralized, post-quantum-encrypted mesh with no central operator. Distributed inference, agent swarms, edge caching, and collective learning that shares model improvements, never data.
Who it's for
Professionals whose data can't leave the room.
A law firm
…that can't send case files to the cloud. Contract review and discovery, on-premises, under privilege.
A medical practice
…where transcription AI is banned but used anyway. Ark One brings it inside the compliance boundary.
A developer
…who wants a frontier coding agent that's actually theirs. Long-context, private repo, no rate limits.
Professionals tell us nothing works below $100K. Ark One changes that by an order of magnitude.
Roadmap
Four waves.
- Wave 1
Regulated professions
Law, medicine, accounting — the first 2,000 units.
- Wave 2
Prosumer & homelab
Open reference program for builders and enthusiasts.
- Wave 3
Neighborhood utility
ArkNet market: lease idle compute; earn credits.
- Wave 4
Open ecosystem
Third-party appliances, models, and agents on ArkNet.
First 2,000-unit production run in preparation; silicon platform access from September 2026.
Team
Built by operators, not tourists.
A founding team spanning edge-AI hardware manufacturing, two decades of data-center liquid-cooling and OCP leadership, kernel-level OS engineering, post-quantum cryptography, and consumer brand building.
- Edge-AI hardware & supply — Melbourne / Shenzhen
- Data-center liquid cooling & OCP leadership · patent portfolio
- Kernel-level OS engineering
- Post-quantum cryptography
- Consumer brand building
Named bios published as we close the round.
Contact
Get in touch.
Questions about the product, partnerships, or press — send us a note and we'll get back to you personally.
Stay in the loop
Get launch updates.
Short, infrequent, and only when it's worth reading.
FAQ
Questions we get.
Entry-tier workload validated on existing hardware; flagship tiers are specified on next-gen AMD silicon (access September 2026). The full architecture study is published.