NVIDIA JETSON · RUGGED ON-DEVICE INFERENCE

Physical AI that never
leaves the field.

From a fanless Orin box to an IP66 Jetson AGX Thor running frontier models.
Your cameras, your data, your silicon — sealed against dust, water, and the cloud.

See the hardware Which model fits →
128 GBunified memory · Jetson AGX Thor
2,070FP4 TFLOPS · AGX Thor
IP66sealed · -20 to 60°C
$0per-token. Ever.

The hardware.
Sealed. Fanless. Certified.

WCO-6000-THR-A IP66 Physical AI Computer with NVIDIA® Jetson AGX Thor™, Holoscan-ready

IP66 Physical AI · Jetson AGX Thor

WCO-6000-THR-A

Jetson AGX Thorup to 2,070 FP4 TFLOPS128 GB

Typical uses

  • Holoscan real-time sensor AI
  • Outdoor autonomy & robotics
  • 8-camera GMSL machine vision
View specs →
JCO-6000-ORN High-Performance AI Edge Computer

AI Edge Inference · Jetson AGX Orin

JCO-6000-ORN

Jetson AGX Orinup to 275 INT8 TOPS64 GB

Typical uses

  • Multi-camera vision inspection
  • AMR / AGV navigation
  • On-box VLM & LLM endpoints
View specs →
WCO-3000-ORN-A IP66 AI Edge Computer with NVIDIA® Jetson Orin™ NX Super/Nano Super

IP66 Edge AI · Jetson Orin NX / Nano

WCO-3000-ORN-A

Jetson Orin NX Superup to 157 INT8 TOPS16 GB

Typical uses

  • Sealed outdoor edge AI
  • Surveillance & NVR analytics
  • Compact robotics payloads
View specs →

Four Jetson modules.
From 8 gigabytes to 128 gigabytes.

Jetson AGX Thor · FLAGSHIP

2,070 FP4 TFLOPS

128 GB LPDDR5X

GPU
Blackwell
CPU
14-core Arm Neoverse
Memory
128 GB
Power
40–130 W

Premio’s WCO-6000-THR brings Thor-class generative and physical-AI compute to an IP66 sealed chassis.

Jetson AGX Orin

275 INT8 TOPS

64 GB

CUDA
2,048
Tensor
64
Memory
64 GB
Power
15–60 W

The workhorse for multi-stream vision and on-device LLMs — the JCO-6000-ORN.

Jetson Orin NX

157 INT8 TOPS

16 GB

CUDA
1,024
Tensor
32
Memory
16 GB
Power
10–40 W

Compact, sealed, and rugged in the WCO-3000-ORN for outdoor autonomy.

Jetson Orin Nano

67 INT8 TOPS

8 GB

CUDA
1,024
Tensor
32
Memory
8 GB
Power
7–25 W

Entry-class edge AI for single-stream vision and small language models.

THE SILICON

GPU, CPU and memory
on one sealed board.

Every inference this hardware makes happens at the edge — behind your firewall, on your network, at your site. Jetson’s system-on-module puts a datacenter-class GPU inside a fanless, ruggedized enclosure.

Pick your model.
We sized them for Jetson.

Qwen2.5-VL 7BVision-Language

~6 GB4-bit

Machine vision QA

Runs on:Orin NX and up

Llama 3.1 8BGeneral

~6 GB4-bit

Local RAG & chat

Runs on:Orin NX and up

Qwen2.5 14BReasoning

~10 GB4-bit

Agentic workflows

Runs on:AGX Orin

Llama 3.3 70BFrontier · MoE-class

~40 GB4-bit

On-prem frontier LLM

Runs on:AGX Thor (128 GB)

SOFTWARE

NVIDIA’s stack.
Preflighted for the edge.

SDK

JetPack

The full Jetson software stack — CUDA, cuDNN, and the Linux BSP — validated on every Premio Jetson platform.

INFERENCE

TensorRT / TensorRT-LLM

Compile and quantize models to run at the lowest latency the silicon allows, INT8 to FP4.

VISION

DeepStream · Metropolis

Multi-stream video analytics pipelines for inspection, surveillance, and traffic at the edge.

SENSOR AI

Holoscan

Real-time sensor processing for medical, scientific, and industrial AI — WCO-6000-THR is Holoscan-ready.

ROBOTICS

Isaac ROS

GPU-accelerated ROS 2 packages for perception, navigation, and manipulation on AMRs and AGVs.

SPEECH

Riva

On-device speech recognition and TTS for hands-free operator interfaces on the plant floor.

What people
actually run.

Manufacturing

Vision QA that explains itself

Multi-camera defect detection with a VLM that reads gauges, labels, and codes — logged on-box, no cloud.

JCO-6000-ORN · Qwen2.5-VL · DeepStream
Robotics

Autonomy at the payload

Isaac ROS perception and navigation running sealed and fanless on a moving platform.

WCO-3000-ORN · Isaac ROS · TensorRT
Physical AI

Sensor fusion in the field

Thor-class Holoscan pipelines fusing eight GMSL cameras in an IP66 enclosure, -20 to 60°C.

WCO-6000-THR · Holoscan · TensorRT-LLM

Questions.

Can I run a local LLM on a Premio Jetson computer?

Yes. Jetson’s unified memory lets a Premio edge system serve models on-device via an OpenAI-compatible endpoint. An AGX Orin (64 GB) runs 14B-class models; an AGX Thor (128 GB) reaches 70B-class — all behind your firewall with no per-token cost.

Which Jetson module should I choose?

Orin Nano/NX (WCO-3000-ORN) for single- and multi-stream vision and models to ~8B. AGX Orin (JCO-6000-ORN) for heavy multi-camera vision and 14B-class agents. AGX Thor (WCO-6000-THR) for Holoscan sensor AI and frontier-class models.

How much memory do I need for a model?

At 4-bit: a 3B model needs ~3 GB, an 8B ~6 GB, a 14B ~10 GB, and a 70B ~40 GB, plus context headroom. Jetson shares memory between CPU and GPU, so the module’s total memory is your budget.

Are these systems rugged enough for the field?

The WCO series is IP66-sealed against dust and water with a wide -20 to 60–65°C operating range, M12 connectors, and power-ignition management. The JCO series adds modular EDGEBoost I/O for 10GbE, 5G, and GMSL cameras.

What about outdoor cameras and sensors?

Premio Jetson platforms expose waterproof Fakra-Z GMSL2 connectors — four on the WCO-3000, eight on the WCO-6000-THR — for synchronized multi-camera capture feeding DeepStream or Holoscan.

Do these need an internet connection?

No. Models and pipelines run on the device. For regulated or air-gapped sites that is the point: imagery, telemetry, and inference stay local.

Ready when you are.

Tell us the workload — cameras, models, environment. We’ll spec the Jetson platform.

Request a quote Browse AI edge computers