RTX PRO 6000 Server (4×–8×)

Where unified-memory boxes optimise for capacity, the RTX PRO 6000 Blackwell optimises for speed: 1.79 TB/s of memory bandwidth is 6–7× a DGX Spark. A single card serves gpt-oss-120b at ~150 tok/s for one user — or thousands of tokens per second aggregate under vLLM batching. This is the silicon for department-scale production serving on 120B-class models, and the pragmatic fine-tuning platform: QLoRA up to 120B fits on one card, with full CUDA, vLLM, SGLang and TensorRT-LLM support.

Two cards give 192 GB — notably, enough to serve DeepSeek V4 Flash's native FP4/FP8 checkpoint, the most capable open model that fits a workstation. Four give 384 GB; beyond that the passive Server Edition scales to 8 GPUs in a rack chassis for colocation. UK-built systems come from Scan's 3XS line with local warranty and support.

Buyers should know the market context: the GDDR7 shortage pushed the card from its $8,565 launch MSRP to ~$13,250 (+55%) by mid-2026 — about £11,300 inc VAT in the UK — and no successor exists or is announced. The Max-Q variant (300 W) sacrifices ~13% compute for half the power and heat, and since LLM decode is bandwidth-bound, its single-stream speed is nearly identical — it is the card of choice for dense multi-GPU builds.