A slow directory of computers that run real models on your desk. We list what it is, who it is for, what it actually runs, and the catch.
Prices are USD snapshots.
Start here
Cheap door · not ChatGPT
For students, robots, and anyone who needs CUDA at the edge for $249.
The cheap door into JetPack and physical AI. Not a 70B desk box.
Default desk box
For people who want a repairable Strix Halo box they can live with for five years.
The grown-up midrange pick. Not CUDA-native.
70B CUDA brick
For CUDA people who want 128GB of unified memory in a paperback-sized brick.
Buy this if you already live in CUDA and need capacity, not the fastest tokens per dollar.
All boxes
Same 495 class
For people deciding between the two 192GB Gorgon Halo cubes.
Same job. Pick ports, size, and when it arrives. Bandwidth stays Halo-class.
For people who want 192GB Gorgon Halo in a 2L cube with OCuLink, and will wait on a pre-order.
Same 495 class as the EVO-X5 Pro. Slightly cheaper ticket, weaker network, not shipping yet.
For people who want a shipping Intel Panther Lake mini with 64GB, dual 10GbE and OCuLink.
The site’s first Intel box. Quiet 64GB x86 with fast networking. Not a 70B machine.
For people who want 192GB Gorgon Halo in a sealed GMKtec cube with dual 10GbE.
192GB Max+ PRO 495. Halo-class bandwidth, not a new architecture. Buy for capacity over the X2.
For people who want Strix Halo 128GB in an RGB gaming cube with three M.2 slots.
Same Strix Halo silicon and ~256 GB/s as Framework/EVO-X2. Buy for the cube and three M.2; skip if you want quiet or 10GbE.
For CUDA people who want GB10 128GB through Acer Veriton channel.
Acer’s GB10 cube. Same silicon and 273 GB/s as Spark. Buy for channel/support, not new silicon.
For macOS users who want 128GB–512GB unified capacity for local models in a quiet deskside box.
Apple’s capacity deskside. Pre-order open; first units 22 Sep. Buy the RAM you need.
For CUDA people who want the GB10 128GB capacity tier through Dell channel and support options.
Dell’s GB10 cube. Same silicon and 273 GB/s as Spark. Buy for channel/support, not new silicon.
For CUDA people who want GB10 128GB through HP Z / Z Toolkit.
HP’s GB10 nano. Same silicon class as Spark/Dell/Lenovo PGX. Buy for HP channel, not new silicon.
For CUDA people who want GB10 128GB through ThinkStation channel.
Lenovo’s GB10 nano. Same silicon class as Spark/Dell/HP ZGX. Buy for ThinkStation channel, not new silicon.
For robotics and edge teams who need CUDA + JetPack + 64GB between Nano Super and Thor.
The Orin workhorse kit. JetPack and sensors first. Not a 128GB desk cube.
For people who want an upgradeable Gorgon Point mini with OCuLink for 7B–14B local models and a compact expandable desk box.
Quiet upgradeable HX 470 cube with OCuLink. Fine for smaller models. Not a 24B–70B capacity box.
For people who want an upgradeable Gorgon Point mini with 10GbE for 7B–14B local models and a home lab.
Quiet upgradeable HX 470 cube. Fine for smaller models. Not a 24B–70B capacity box.
For people who want a compact CUDA gaming mini that also runs 8B–14B local models well.
3L CUDA mini with laptop RTX 50. Fast while the model fits in 16GB. Not a 70B capacity box.
For developers who want AMD’s official Strix Halo reference box with dual-OS ROCm playbooks and will pay the Micro Center list.
AMD’s first-party Strix Halo developer platform. Same ~256 GB/s ceiling as the cheaper minis. Buy for the official stack, not the silicon.
For people who want prebuilt Strix Halo under a known brand with dual M.2 and a two-year warranty.
The brand-name prebuilt Strix Halo mini. Same ~256 GB/s ceiling. Buy 128GB only if the ticket still makes sense.
For CUDA people who want max tokens on models that fit in 32GB, in a small case.
The used CUDA bandwidth play. Fast while it fits. Not a 70B capacity box.
For people who want Strix Halo in a certified, mountable mini workstation with HP support options.
The enterprise Strix Halo mini. Same bandwidth ceiling as the sealed cubes. Buy 128GB.
For robotics and physical-AI teams who need CUDA + JetPack + 128GB under JetPack.
High-end Jetson for physical AI. Same memory class as Spark, Arm + sensor I/O.
For people who want Strix Halo with dual 10GbE in a sealed Mac Studio–shaped cube.
The dual-10GbE Strix Halo cube. Same bandwidth ceiling as Framework and EVO-X2. Confirm NIC revision.
For people who want Strix Halo with dual 10GbE and a limited PCIe slot in one chassis.
The networked Strix Halo mini. Buy 128GB. Same bandwidth ceiling as Framework and EVO-X2.
For macOS users who want a silent 14B–27B (or 64GB M5 Pro) box without leaving Apple’s stack.
Quiet Apple desk box for smaller models. Not a capacity cube.
For people who want a sealed Strix Halo mini with dual 10GbE and a lower ticket than Framework.
The cheaper sealed Strix Halo mini. Same ~256 GB/s ceiling. Buy 128GB. Fan noise is real.
For buyers who want the DGX Spark recipe without the Founders Edition.
Same class as Spark, often cheaper. Read the SKU.
For people who want Strix Halo in a 5-bay NAS with dual 10GbE and hybrid storage up to the 200TB class.
Halo silicon in a real 5-bay NAS. Buy 128GB. Same ~256 GB/s ceiling. Not a desk cube.