Curated systems
Five tiers. Different capabilities.
These are compatibility-checked templates, not a dynamic parts lottery. Price bands include the complete system and preserve condition risk.
Entry
Useful small-model systems and controlled offload.
Compact used RTX A2000 workstation
A quiet, low-power entry point for 7B coding models and remote always-on use in very limited space.
Used RTX 3060 value tower
The cheapest serviceable DIY host with CUDA maturity, enough system RAM for controlled offload, and standard replaceable parts.
Value
Memory capacity per dollar with explicit software-stack tradeoffs.
Efficient RTX 5060 Ti 16 GB tower
Low operating cost and mature CUDA support for responsive small models, with 16 GB allowing constrained 24–30B Q4 use.
Radeon AI PRO R9700 32 GB workstation
Thirty-two gigabytes of ECC-capable VRAM at a lower acquisition cost than current NVIDIA 32 GB options, for users willing to validate ROCm and model support.
Performance
High bandwidth or high unified-memory capacity.
Used RTX 3090 24 GB tower
High bandwidth and 24 GB VRAM make a used 3090 a strong price/performance step for 24–32B Q4 models at moderate context.
Mac Studio M3 Ultra 96 GB
Quiet compact unified memory enables 70B-class Q4 models without multi-GPU topology, with excellent idle efficiency.
RTX 5090 32 GB coding flagship
The fastest practical consumer single-GPU CUDA option, with 32 GB VRAM for strong coding models and very high prompt/decode headroom.
Multi-agent
Topology sized for parallel request pools and supported distribution.
No-compromise
ECC capacity, professional reliability, or rack-scale operations.
RTX PRO 6000 96 GB workstation
A single 96 GB ECC GPU avoids consumer multi-GPU pooling pitfalls and supports 70B-class models with substantial long-context and concurrency headroom.
4× RTX PRO 6000 rack inference node
A rack-scale capacity tier for isolated agent pools or validated tensor parallelism where ECC, remote operations, and serviceability matter more than acoustics or residential power.