Guided Path: Small Startup
Goal: run one capable open-source model as cheaply as honestly possible. No overprovisioning, no cluster — just enough hardware to clear the model's memory bar.
1. The model
Apache 2.0
Mixtral 8x22B — Mistral AI
Of our 3 open-license 100B+ models, this one has the smallest memory footprint, which makes it the cheapest to self-host.
Total parameters141B
Minimum memory169.2 GB
Math: 141B × 1 GB + 20% = 169.2 GB.
2. The build
The cheapest catalog hardware that clears 169.2GB is two RTX PRO 6000 Blackwell cards (96GB each = 192GB) in one workstation. A single one isn't enough on its own.
NVIDIA press image
2x NVIDIA RTX PRO 6000 Blackwell + workstation host
GPU / systemNVIDIA RTX PRO 6000 Blackwell (Workstation Edition) × 2
Combined memory192 GB
GPU cost$17,130
Host system (est.)$4,000
Total price$21,130
⚡ 1,450 W sustained draw
🏠 = 1.21x an average home (~1,200W continuous)
🔋 = drains a 90 kWh EV battery in 62.07 hrs
3. Why not cheaper?
Two RTX 5090s only total 64GB — nowhere near the 169.2GB this model needs, so it simply won't load. Desktop GPUs with less memory are a false economy here: the model has to fit, in full, before it runs at all.