The first sovereign AI workstation with a dedicated ARM AI processor and a full-power x86 Windows engine — running simultaneously in a compact Micro-ATX chassis.
Two independent processors. Two operating systems. One chassis. Each does what it does best — simultaneously.
Without a dedicated GPU, this machine already outperforms most workstations built for the same tasks.
The Radeon 8060S delivers discrete-GPU performance out of the box. Add a dedicated card to the open PCIe slot and it becomes a full gaming rig.
The Radeon 8060S with 40 compute units and 256 GB/s memory bandwidth competes with RTX 4070 laptop class at 1080p and 1440p.
RDNA 3.5 · 40 CU · ~RTX 4070 laptop class
Drop any full-size PCIe GPU into the open expansion slot. x16 physical slot — bandwidth scales with your Jetson tier. The 16-core CPU doesn't bottleneck even the latest cards.
x16 physical slot · PCIe bandwidth scales with Jetson tier · No CPU bottleneck
Every PCIe slot on every board variant uses the same x16 physical connector. PCIe is backward compatible by design — the slot auto-negotiates to the available lane count at link training. Same connector footprint. Same x86 companion card. Same GPU. Different electrical bandwidth depending on which Jetson module is installed.
Standard runs x4 Gen3 in that x16 slot. Pro runs x8 Gen4 — 4× the bandwidth, zero GPU bottleneck. Ultra runs x8 Gen5 — 8× Standard bandwidth, absolute maximum throughput. Slot 2 is open on every tier for any full-size card.
Three AI cores. One x86 engine. One chassis. Reference specifications — subject to change.
|
Standard
Orin NX 16GB
Entry sovereign AI
|
Pro
AGX Orin 64GB
Professional AI workstation
|
Ultra
Thor T5000
Physical AI supercomputer
|
|
|---|---|---|---|
| NVIDIA Jetson AI Core | |||
| AI Performance | 157 TOPSINT8 | 275 TOPSINT8 | 2,070 TFLOPSFP4 · 1,035 FP8 |
| GPU Architecture | NVIDIA Ampere 1,024 CUDA · 32 Tensor |
NVIDIA Ampere 2,048 CUDA · 64 Tensor |
NVIDIA Blackwell 2,560 CUDA · 96 Tensor · MIG |
| On-Module Memory | 16 GB LPDDR5 | 64 GB LPDDR5 | 128 GB LPDDR5X |
| Memory Bandwidth | 102 GB/s | 204 GB/s | 273 GB/s |
| CPU | 8× Cortex-A78AE up to 2.0 GHz |
12× Cortex-A78AE up to 2.2 GHz |
14× Neoverse V3AE Arm v9 · 2.6× faster/core |
| Module Connector | 260-pin SO-DIMM | 699-pin Molex Mezz | 699-pin Thor Mezz |
| Module TDP | 10 – 25 W | 15 – 75 W | 40 – 130 W |
| Operating System | JetPack 6 · Ubuntu 22 | JetPack 6 · Ubuntu 22 | JetPack 7 · Ubuntu 24 |
| Combined TOPS | 207+50 x86 NPU | 325+50 x86 NPU | 2,120+FP4 + x86 NPU |
| x86 Windows Engine — AMD Ryzen AI Max+ 395 (same across all tiers) | |||
| x86 NPU | 50 TOPS · XDNA 2 · Copilot+ PC qualified (>40 TOPS) | ||
| x86 Memory | 64 GB LPDDR5x-8000 · 256 GB/s unified · shared CPU + GPU pool | ||
| CPU | 16× Zen 5 cores · up to 5.1 GHz boost · 32 threads · 4nm TSMC | ||
| Integrated GPU | Radeon 8060S · 40 CU RDNA 3.5 · ~RTX 4070 laptop class | ||
| Operating System | Windows 11 · Full x86_64 compatibility · Native driver support | ||
| Card Interface | PCIe x16 physical card · Custom KI x86 companion module · Supplemental 12V power · Negotiates x4/x8/x8 Gen5 by tier | ||
| Storage (same across all tiers) | |||
| Primary NVMe | M.2 Key M · PCIe 3.0 x4 · 2280 · Customer-specified | ||
| Secondary M.2 | M.2 Key M · PCIe 3.0 x4 · Open expansion | ||
| SATA | 8× SATA III · Dual ASM1064 controllers · RAID 0/1/5/6/10 via mdadm · 8 internal drive bays · Customer-configurable | ||
| Connectivity | |||
| Network — Jetson | 1 GbE RGMII | 1 GbE RGMII · MGBE 10 GbE | 4× 25 GbE · 1× 5 GbE |
| Network — Board | 2.5 GbE management port · 10 GbE SFP+ Switch Module uplink | ||
| SD Card | Full-size SD slot · Rear panel edge-mount · SDIO from Jetson · OS recovery · AI model loading · field service | ||
| USB — Rear Panel | 4× USB 3.2 Gen 2 Type-A · 10 Gbps · standard ATX rear I/O position · USB-C recovery port | ||
| USB — Internal Headers | 2× 19-pin USB 3.0 headers · mount port clusters anywhere in the case | 1× 9-pin USB 2.0 front-panel header | 1× 9-pin silkscreen-labeled CARD READER — any case-mount multi-format reader plugs in here | ||
| USB Hub | On-board VL817 USB hub IC · multiplies Jetson USB budget to feed all ports and headers simultaneously · transparent to OS | ||
| Display | DP 1.2 + HDMI 2.0 | Multi-DP + HDMI 2.0 | 4× independent displays HDMI 2.1 + 3× DP 1.4a |
| Wireless | M.2 Key E · WiFi 6E + Bluetooth 5.3 | ||
| PCIe Expansion | |||
| x86 Slot — Physical | PCIe x16 gold-finger — identical on all three boards. Same connector footprint. Same x86 companion card. PCIe auto-negotiates to available lanes at link training. | ||
| x86 Slot — Electrical | x4 Gen3 ~16 GB/s · via PCIe switch |
x8 Gen4 ~64 GB/s · native · no switch |
x8 Gen5 ~128 GB/s · native · double Gen4 |
| GPU Bottleneck? | Minimal 1080p/1440p ~10% at 4K ultra |
None | Zero |
| Open Slot 2 | Open · PCIe x16 physical · any GPU, NIC, AI accelerator, or expansion card | ||
| Slot 3 | PCIe x1 · Open expansion | ||
| Form Factor (same chassis across all tiers) | |||
| Board Standard | Micro-ATX · 244 × 244mm · Standard mounting · ATX 24-pin + 8-pin power | ||
| Enclosure | Custom KI chassis · Multi-bay · Active cooling · PCIe slot access panel | ||
| Recommended PSU No GPU · SSD only |
450 W | 500 W | 600 W |
| Recommended PSU No GPU · 8× HDD NAS build |
550 W | 650 W | 750 W |
| Recommended PSU + Mid-range GPU (~200W) |
750 W | 850 W | 950 W |
| Recommended PSU + High-end GPU (~450W) |
1000 W | 1000 W | 1200 W |
This is a concept product in active development. Leave your email and we'll reach out before anything goes public.