Garnet · sovereign systems
01 — The complete sovereign AI system
A complete sovereign AI system — the hardware and the entire software stack — designed, built, and privately delivered to you. It runs itself, on machines you own, with nothing leaving your walls. Turnkey. Exclusive. Handed over, not rented. Three complete builds, scaling to the largest sovereign AI on earth.
An autonomous multi-agent AI workforce that operates across your business — planning, producing, and following through without you in the loop.
Every inference runs locally. Your prompts, your data, and your work never leave your walls — nothing is sent out, nothing is used for training.
It doesn't just answer questions — it works. A system built to originate and pursue real outcomes, around the clock.
When it needs a new capability, it builds it. The system grows into your problem instead of waiting for the next vendor release.
No cloud, no vendor lock-in, no per-token meter. You own the hardware and the entire system outright.
Hardware up front, electricity after. No subscription, no usage billing, no invoice that scales with how much you use it.
Three complete builds of the same sovereign system. Each is the whole thing — hardware and full stack. They scale by memory and capacity: larger models, more concurrent agents, higher throughput. One system, one owner, per build.
Tier I
$2,500,000
The complete sovereign AI system
The complete system — the full autonomous AI company, built and delivered on hardware you own. Everything above scales up from here.
Tier II
$5,000,000
1 TB memory build
more capacity
More of everything. A 1 TB high-capacity build for larger models, more concurrent agents, and higher sustained throughput.
Tier III
$10,000,000
2 TB memory build
Maximum capacity. A 2 TB build — the largest models, the most concurrent agents, the highest throughput. The ultimate sovereign AI.
Every build is a single, complete offering — not a subscription, not a rental. Sized and delivered for one owner.
Before you commission
$149
The written specification behind the builds: how to size a machine to the models you will actually run, why serving is memory-bound before it is compute-bound, the all-local software stack, the sovereignty and airgap posture, and the build-verify-handover sequence. Vendor-neutral, and a technical guide rather than a parts list — parts and prices change, so a build is sized to the workload at the time.
Read it, build it yourself, or use it to scope a commissioned build with us.
Request the Blueprint →A complete system, specced and assembled for the build, delivered ready to run.
The entire autonomous system, installed and configured — the whole thing, not a demo.
All models run on your own machines. No external provider is ever in the path.
Yours outright. No cloud dependency, no vendor lock-in, no per-token meter.
Fully operable disconnected. Network is optional, never required.
Delivered, set up, and handed over directly. The details are shared under NDA.
Sovereign means the models run on your hardware. Here is the plain arithmetic of what a given card can hold — the same trade-off each build is sized against. Bigger builds hold bigger models and more of them at once.
Fit calculator · will it run on your card
Largest model that fits: 34B at Q4 (4-bit), 8k context.
| Model | Weights | KV | Total | Fit |
|---|---|---|---|---|
| 7B | 3.5 GB | 0.5 GB | 5.2 GB | yes |
| 14B | 7.0 GB | 0.8 GB | 9.0 GB | yes |
| 20B | 10.0 GB | 1.0 GB | 12.2 GB | yes |
| 34B | 17.0 GB | 1.4 GB | 19.6 GB | yes |
| 70B | 35.0 GB | 2.4 GB | 38.6 GB | no |
| 120B | 60.0 GB | 3.6 GB | 64.8 GB | no |
| 235B | 117.5 GB | 5.6 GB | 124.3 GB | no |
Estimates, not benchmarks. Weights are bytes-per-weight × parameters; KV cache is a conservative per-class figure; 1.2 GB is reserved for runtime overhead. Real throughput depends on your card, and offload lets bigger models run slower rather than not at all.
A complete sovereign system is scoped, built, and handed over in person — never bought from a cart. Tell us who you are, which build fits, and what you need. Serious inquiries only; everything beyond this page is shared privately, under NDA.