Home / Services / Hardware

Your models. Your building.

We spec, build and tune AI hardware for the work you actually do, then hand it over with the software already on it. No per-token bill, and no data leaving the building.

Four tiers, sized by what they run

People supported are conservative planning estimates. We confirm them on our own bench before quoting.

Desk

A private AI on one quiet desktop box

1–5 people

Runs ~120B mixture-of-experts models at about 45 tokens/s, and 30B models fast. Sits on a desk and draws less than a lamp-heavy office.

Quote

Specs
Box
128 GB AMD Ryzen AI Max+ 395 mini PC (Windows or Linux). Option: NVIDIA DGX Spark for CUDA and long documents
Model memory
96–128 GB unified
Power
Under 250 W, desk-quiet
Studio

One serious GPU for a team

5–25 people

70B dense models (FP8) or ~120B MoE models, served to many people at once with batching. About 6× the memory bandwidth of Desk.

Quote

Specs
Box
Tower workstation, 1× NVIDIA RTX PRO 6000 Blackwell 96 GB, 256 GB RAM, 4 TB NVMe
Model memory
96 GB
Power
~1 kW, normal outlet, office-quiet
Team

Several models at once, company-wide

25–100 people

A fast model, a smart model and a coding model side by side. Agents and pipelines for the whole company, plus light fine-tuning.

Quote

Specs
Box
4U server, 4× RTX PRO 6000 Server Edition, EPYC, 512 GB RAM
Model memory
384 GB
Power
~3.5 kW, 240 V circuit, server closet
Rack

Frontier-size open models and training

100+ people

The largest open models, training runs, and many departments on one platform. Starts with a room plan.

Quote

Specs
Box
8× RTX PRO 6000 server, HGX B200 class, or multi-node
Model memory
768 GB+
Power
8–9 kW+, needs a room plan

Every tier is quoted, not list-priced. Hardware prices checked Nov 2026. Memory prices are moving; your quote locks the number.

Hardware prices checked Nov 2026. Memory prices are moving; your quote locks the number.

Own vs rent

Enter what you spend on AI each month. We compare it with owning a box. This is a rough first look: it assumes the box can take over that whole workload, which depends on what you run. We'll check that on a call.

Start with a free exploration call.

Tell us what you're trying to do and what's in the way. If AI fits, we'll show you where. If it doesn't, we'll say that too.