Help me choose
Two common situations, two specific recommendations. Pick the one that sounds like you.
"I want to run one big open model as cheaply as it can honestly be run."
We picked our smallest-footprint open model, Hy3 (295B parameters, Apache License 2.0), and sized the cheapest build that a real production setup would actually use.
Recommended build: 3x NVIDIA H200 NVL
Why this fits: NVIDIA H200 NVL is the cheapest hardware in our catalog that's actually built to link multiple GPUs into one real memory pool (NVLink bridging), not just a card in a box. A cluster of 4x NVIDIA RTX PRO 6000 Blackwell Workstation Edition would look cheaper on paper ($64,000), but those are isolated workstation cards with no high-speed link between them — see why that matters. For a startup that needs this to actually work in production, NVIDIA H200 NVL is the honest minimum.
"We expect more users and traffic — we need capacity and room to grow."
We sized this around our largest open model, DeepSeek-V4-Pro (1,600B parameters, MIT License) — if you can comfortably run the biggest model we carry, everything smaller fits with room to spare.
Recommended build: 1x NVIDIA GB200 NVL72
Why this fits: the model only needs 1,920 GB, but this rack has 13,400 GB — using only about 14% of it for the model itself. The rest is exactly the "room to grow" a mid-size company needs: headroom for many concurrent users, or to host additional models later, without buying new hardware. It also arrives as one already-integrated cluster — 72 GPUs wired together over a single NVLink switch fabric — instead of you having to network separate GPUs together yourself. See what makes that a real cluster.