NGARi+ · Sovereign AI appliances
NGARi+ is the hardware line. A NVIDIA Jetson at the centre, the sovereign stack on top, and nothing leaving the box. The model runs here. The agents run here. Your data stays here — because there is nowhere else for it to go.
Ships pre-flashed and configured · typically 7–14 business days · 2-year hardware warranty · 30-day returns after delivery · installment plans available
Three machines, one principle: the intelligence is yours. Pick by the largest model you need to run and how many agents you need working at once.
The one that fits in a drawer and replaces a subscription.
Installments: 2 × $499.50 or 4 × $249.75 — 0% interest; the payments total the same $999.00. How plans work
The one that runs the big model and the whole C-suite.
Installments: 2 × $2,499.50 or 4 × $1,249.75 — 0% interest; the payments total the same $4,999.00. How plans work
The one that trains, sees, and drives machines.
Installments: 2 × $4,499.50 or 4 × $2,249.75 — 0% interest; the payments total the same $8,999.00. How plans work
This is not a faster chatbot. It is a machine that removes three bills, one dependency, and one ceiling at the same time.
Six autonomous executives — CEO, CRO, Engineer, Worker, CMO, Chief of Staff — running continuously on hardware you own, with hash-chained audit records of everything they do.
Ingest contracts, records, research and client files, then search across them and answer from them with zero outbound traffic. The answer never leaves the room.
Train the model on what makes your business yours. No data-licensing clause, no upload, no one else's terms of service. The weights live on your NVMe.
Denied, degraded, intermittent, low-bandwidth. Ships, clinics, farms, field teams, basements, borders. Inference does not stop when the link does.
Your 10,000th question costs the same as your first: nothing. Usage cannot be metered against you because usage is not observable by anyone but you.
Standardise how your team uses AI — one box, one stack, one governance model — then certify people on it through AI-Corps.
A rented model is a subscription that never ends and never becomes an asset. At a blended frontier-cloud price of roughly $15 per 1M tokens, a team using 50M tokens a month spends $750 every month — $27,000 over three years. NGARi+ Orin is $4,999, once, and it is still yours in year four.
At that volume, NGARi+ Nano pays for itself in about six weeks and Orin in about seven months — and unlike the cloud bill, the asset never resets to zero.
Illustrative. The frontier-cloud figure assumes a blended $15 per 1M tokens at 50M tokens per month; the NGARi API figure uses our published $2.00 per 1M tokens at the same volume. Third-party pricing varies widely by model and provider and is often higher. NGARi+ carries no recurring fee.
Comfortable quantized model range. Smaller models always run faster; larger models fit on Thor with reduced context. Vision and multimodal run on Orin and Thor.
| Specification | NGARi+ Nano | NGARi+ Orin | NGARi+ Thor |
|---|---|---|---|
| Price | $999 one-time | $4,999 one-time | $8,999 one-time |
| Accelerator | NVIDIA Jetson Orin Nano class | NVIDIA Jetson AGX Orin 64 GB | NVIDIA Jetson AGX Thor 128 GB |
| AI performance | 67 TOPS (INT8, sparse) | 275 TOPS (INT8, sparse) | Up to 2,070 TFLOPS (FP4, sparse) |
| GPU | Ampere, 1,024 CUDA cores, 32 Tensor cores | Ampere, 2,048 CUDA cores, 64 Tensor cores | Blackwell-class, 2,560 CUDA cores, 96 Tensor cores |
| CPU | 6-core Arm Cortex-A78AE | 12-core Arm Cortex-A78AE | 14-core Arm Neoverse-V3AE |
| Memory | 8 GB LPDDR5 · 102 GB/s | 64 GB LPDDR5 · 204.8 GB/s | 128 GB LPDDR5X · 273 GB/s |
| Storage (NGARi+ config) | 512 GB NVMe SSD | 1 TB NVMe SSD | 2 TB NVMe SSD |
| Model range | 7B–14B quantized | Up to 32B quantized (incl. qwen3.8) | 70B–100B+ quantized, multimodal |
| Agents | 6 included · 1–2 concurrent | 6 included · all concurrent, plus custom | 6 included · unlimited concurrent |
| Typical power | ~25 W | 15–60 W | 40–130 W |
| Networking | GbE · optional Wi-Fi/BT module | Dual GbE · optional 10 GbE | Dual 10 GbE · high-speed fabric options |
| Form factor | Palm-sized, desk | Desk, small team | Desk or shelf, robotics/lab |
| Enclosure | NGARi-designed enclosure | NVIDIA AGX Orin developer-kit enclosure | NVIDIA AGX Thor enclosure |
| Operating range | 0–45 °C ambient · fan-cooled · designed for continuous operation | ||
| Warranty | 2 years hardware (see Hardware warranty) | ||
Accelerator, TOPS, GPU, CPU, memory and power figures are the published capabilities of the corresponding NVIDIA Jetson modules that NGARi+ is built on. Storage and the pre-installed sovereign stack are NGARi+ configuration, as are the agent limits. Enclosures are not: NGARi+ Nano ships in a NGARi-designed enclosure, while NGARi+ Orin and NGARi+ Thor ship in NVIDIA’s own developer-kit enclosures.
No subscription is required for the device to be useful. Support plans and training are optional add-ons.
We chose NVIDIA Jetson because it is the only edge platform that gives a small appliance data-centre class AI and a long, inspectable lifecycle. Arm64 Linux, unified memory, CUDA and TensorRT, and a module roadmap that keeps getting faster without changing the software contract.
The GPU and CPU share one memory pool, so a 27B model loads without shuttling weights across a bus — and there is no separate VRAM ceiling to hit.
CUDA, TensorRT and the Jetson software stack are production-grade on aarch64. That is why our fine-tuning and distillation pipeline runs on the same silicon we sell.
Jetson modules have multi-year supply and software support windows. Buy once, and keep it useful — which is the whole point of ownership.
NGARi is an NVIDIA Inception member. NVIDIA and Jetson are trademarks of NVIDIA Corporation. NGARi+ is an independent product built on NVIDIA Jetson modules and is not endorsed or sponsored by NVIDIA.
The Jetson photographs show NVIDIA developer kits. Source and licence record: assets-src/jetson/CREDITS.md.
NGARi+ Thor is not only a desk appliance. The NVIDIA Jetson AGX Thor module at its centre is the reference compute for humanoid robotics: NVIDIA built its open Isaac GR00T Reference Humanoid on this silicon, and says AGX Thor is powering the next generation of humanoid and robotic systems.
That matters because every commercial humanoid today is cloud-tethered by default — it is how fleets receive model updates and how vendors collect training data. A robot whose brain never connects is the only kind that can be deployed inside a hospital, a bank, a defence plant or a national lab. That is the appliance we already sell.
Up to 2,070 TFLOPS and 128 GB of unified memory in a 40–130 W envelope — enough to run vision, language, planning and control on one module, on a moving machine, with no network in the loop.
No telemetry, no fleet learning, no phone-home. Your facility, your floor plan and your people are never training data for anyone else. Sovereign by architecture, not by policy.
The hash-chained audit log and signed attestation that receipt our data-processing work extend to physical work: what was read, what was produced, which model ran and when — signed so it verifies without us.
We will not sell a capability we have not built, so here is the honest line.
| Layer | Status |
|---|---|
| AGX Thor compute, 128 GB unified memory, 2 TB NVMe, 40–130 W | Shipping — NGARi+ Thor, $8,999 |
| Sovereign stack, air-gap mode, hash-chained audit, signed attestation | Shipping |
| Physical action receipt — motion, force, document and byte provenance | Specified, not built |
| ROS 2 and Isaac GR00T integration onto the sovereign stack | Engineering engagement |
| Local teleoperation — operator outside the room, no internet path | Engineering engagement |
Building robots? The brain is available as an integration engagement, and we will tell you plainly which parts are shipping and which are not. If you need an air-gapped brain for a regulated deployment, that is worth a conversation now rather than after your customer’s security review.
NVIDIA, Jetson and Isaac GR00T are trademarks of NVIDIA Corporation. NGARi+ is an independent product built on NVIDIA Jetson modules and is not endorsed or sponsored by NVIDIA.
Hardware is only half of confidence. The other half is knowing what to do with it on day one. You have two honest paths.
You know what you want to run. The box arrives flashed, you set an admin account, pick a model and start. Best for teams that already have an AI plan.
AI-Corps is our sovereign venture and workforce program. You get the same class of hardware pre-flashed, plus the curriculum, the clinics, and a credential that is anchored to the machine that earned it.
If you are deploying sovereign AI inside a team for the first time, AI-Corps is usually the better purchase: the failure mode of ownership is not the hardware, it is not knowing which workflow to move onto it first. If you already know, buy direct and keep the margin.
No. NGARi ONE is the palm-sized sovereign handheld. NGARi+ is the desk and team line built on NVIDIA Jetson modules — it runs bigger models, hosts all six agents at once, and is designed to sit in an office, lab or vehicle rather than a pocket.
Yes. The NS-BOS kernel is free and Apache 2.0 at ngari.ai/kernel — it is the same runtime this appliance ships with. It needs Linux with Python 3.10+, and it does not include a model, so you point it at your own runtime and weights. What NGARi+ adds is the flashed, verified, thermally-enclosed machine that removes that setup entirely.
It stays on the device. Nothing is transmitted for inference, telemetry is off by default, and the sovereignty meter can show outbound bytes from the network interface as proof. You can run the whole stack with the network cable unplugged.
No. The device is complete without one. Subscriptions exist for hosted agents, managed support and training, and for the API — none of them are required to use your hardware.
Yes. The model manager supports open weights — Qwen, Llama, Mistral, Gemma, Phi and others — and you can import your own fine-tunes. The open kernel is Apache 2.0.
The hosted preview and API currently serve qwen3.8-orin — a 27.3B model (Apache 2.0) — on our edge hardware. The same model is what NGARi+ Orin and Thor run locally, with no network in the loop. NGARi's own distilled 1.5B fine-tunes remain published and are used in the distillation and evaluation pipeline, but they are no longer served on the public path.
Yes — Orin and Thor run fine-tuning and distillation locally, which is how our own distilled models were produced. Nano is intended for inference and small adapters.
Yes. Institutional licences cover on-prem clusters, white-label platform and curricula, and a dedicated team. Email [email protected] or start with Corps.
Pick by model size. 7B–14B and one or two workflows: Nano. A 27B–32B model and the full C-suite: Orin — that is the machine we run in production ourselves. Training, vision or robotics: Thor. If you want help deciding, join an AI-Corps clinic first.
Every question you ask a rented model is a payment and a disclosure. NGARi+ makes it neither.
Extended support $149/mo · add it at purchase · questions: [email protected]
Installment plans are part of the offer on all three NGARi+ appliances, and on every Corps Edition tier. Each appliance can be reserved over 2 or 4 monthly payments. A plan changes when you pay, never how much: the payments add up to the list price exactly, with no interest and no fee. Your node is flashed, configured and shipped once the plan is complete — or immediately if you pay in full.
| Device | 2 payments | 4 payments | Total |
|---|---|---|---|
| NGARi+ Nano | 2 × $499.50 | 4 × $249.75 | $999 |
| NGARi+ Orin | 2 × $2,499.50 | 4 × $1,249.75 | $4,999 |
| NGARi+ Thor | 2 × $4,499.50 | 4 × $2,249.75 | $8,999 |
Every plan above is live: the buttons on each appliance and the amounts in this table check out at exactly the figures shown — you are never charged an amount you did not agree to. Prices are USD; each plan's payments sum to the one-time price.