All machines

Configure your RTX PRO 6000

Blackwell serving power — 96 GB per card, in a 4U rack or a workstation tower.

Form factor
GPU quantity1–4 × RTX PRO 6000
1× 96 GB GDDR7 = 96 GB total VRAM
MemoryRAM is for your workloads — earnings are set by the GPU
256 GB
default
StorageNVMe · ~$1,200 steps
Default tier
Processor & boardone option per platform
AMD
CPU
Threadripper PRO 9965WX (24c)
Board vendor
Motherboard
WRX90 workstation board

Custom CPU, board, or beyond-max memory? Talk to sales →

Price

our landed cost — component prices updated Jun 2026
1× NVIDIA RTX PRO 6000 Blackwell 96GB· 1 GPU$11,500
Threadripper PRO 9965WX (24c)· processor · fixed$3,500
WRX90 workstation board· motherboard / barebones · fixed$1,100
256 GB DDR5 RDIMM· 4× 64 GB$8,000
NVMe· storage$1,200
Chassis· case + cooling$900
Networking & misc· NIC, cabling, sundries$300
Hand-build labor — 5 h × $22· assembled in Oregon$110
B3-IQ Build & AI-Guarantee — 5%$1,331
Total$27,941

Hardware at cost. We make 5% on the build. While your machine pays itself off we keep 70% of what it earns — once it's yours, you keep 70%.

Hosted at Andromeda by default · $262/mo B3-IQ Managed Hosting.

What it runs

70B-class inferenceIndependent 70B replicas (4×)LoRA fine-tuningDiffusion image / video · rendering
GPU
1–4× RTX PRO 6000

Blackwell silicon — a whole 70B fits on one 96 GB card, the serving sweet spot.

VRAM
96 GB GDDR7 each · up to 384 GB

1.8 TB/s bandwidth — ample for large-model inference and render scenes.

Motherboard
4U4G rack or WRX90 tower

Rack barebones (EPYC 9355P) or a Threadripper PRO workstation — pick your form factor.

CPU
EPYC 9355P / Threadripper PRO

Rack runs an EPYC 9355P; the tower auto-selects a Threadripper PRO by GPU count.

Cooling
Tuned rack or tower airflow

Blower-style cards in a tuned airflow path, monitored 24/7.

Power
~1.2 kW (1×) to ~1.9 kW (4×)

Sized for sustained multi-GPU serving load.

The RTX PRO 6000 pairs 96 GB of GDDR7 with Blackwell's render and AI cores — a whole Llama 70B fits on one card. Great for inference and serving (no NVLink, so not for large multi-GPU training — that's the H200's job). B3-IQ builds it as a 4U rack node or a WRX90 workstation tower; the 4× rack is rent-to-own.

Hand-built & hosted in Eugene, Oregon · ~90% carbon-free hydro · monitored 24/7 · 1–2 weeks lead time.

Search B3IQ

Search pages and machines.