Sovereign AI Box
Local, unmetered AI that never leaves your building. Run the whole suite on hardware you control, with the models and data staying home. Pick the box that fits your team.
5 to 15 users · 30 GB VRAM. 24 to 32B dense at Q4 or a ~30B MoE. One 32 GB GPU, RTX 5090 class. Throughput-first single SKU.
15 to 30 users · 50 GB VRAM. 32B high precision, 64k+ context. Two 24 GB GPUs, tensor-parallel. 70B Q4 fits tight.
30 to 60 users · 75 GB VRAM. 70B Q4 comfortably, long context. 96 to 128 GB unified or two 48 GB pro cards. Roughly 100B MoE.
60 to 200 users · capacity-first. 100 to 140B MoE at Q4, airgapped. 192 GB unified memory. Best capacity per dollar.
60 to 200 users · throughput-first. Frontier models, batched throughput. Four 48 GB pro cards. Book a call to configure.