NVIDIA RTX PRO 6000 Blackwell

NVIDIA announced the RTX PRO 6000 Blackwell at GTC in March 2025 — 96 GB of GDDR7 with ECC on a 512-bit bus, 24,064 CUDA cores, 752 fifth-gen Tensor Cores, and the same ~1.8 TB/s of memory bandwidth as the RTX 5090 — with three times the VRAM. It is the workstation flagship of the Blackwell generation: the largest memory pool you can put on a single PCIe card that boots in a normal tower.

Three variants share the silicon, and the name matters when you order:

• Workstation Edition — 600 W, dual-slot active cooler (5.4″ × 12″), 1× 16-pin connector. The fastest single card NVIDIA sells for a desktop.
• Max-Q Workstation Edition — the same 96 GB and dies, capped at 300 W in a smaller dual-slot body (4.4″ × 10.5″). Built for multi-GPU boxes and sane power budgets — you give up peak clocks, not memory.
• Server Edition — passively cooled, 400–600 W configurable, meant for rack chassis airflow. Not what you put in a tower.

All three carry 4× DisplayPort 2.1 (Workstation/Max-Q) and PCIe Gen 5 ×16, plus the professional driver stack and ECC that consumer Blackwell doesn’t get.

Honest math: 96 GB changes the model list, not the speed

Memory bandwidth is essentially identical to the 5090, so on any model that fits both cards, token speed is similar — you are not paying for tok/s. What the PRO 6000 buys is capacity in one slot:

• Dense 70B at Q8 (~75 GB) fits with room for real context — the 5090 needs Q4 and a short leash.
• gpt-oss-120b (~61 GB in native MXFP4) runs on one card with tens of gigabytes left for KV-cache.
• Qwen3.5 122B-A10B — 81 GB as Ollama ships it (qwen3.5:122b) — just fits: a frontier-class MoE on a single card, with the MoE’s 10B active parameters keeping generation fast.
• The 235B-total class still doesn’t fit (Qwen3 235B wants ~142 GB at Q4) — that remains multi-GPU territory, or a different trade entirely via a unified-memory appliance.

Fine-tuning is the quieter argument: 96 GB with ECC means LoRA work on 30B-class models at real precision, on the CUDA stack every production cluster runs.

Pricing, honestly

As of July 2026: the Workstation Edition launched around $8,565 USD in March 2025; the same GDDR7/DRAM shortage that inflated the 5090 hit a 96 GB card three times as hard. NVIDIA’s own marketplace now lists it at $13,250 USD, partner boards run ~$11,400–12,100 at major US retailers, and refurbished cards trade around $9,500–11,000. That is roughly three to four RTX 5090s, or nearly three DGX Sparks. You are paying for VRAM density in one slot — not FLOPS per dollar, and we won’t pretend otherwise.

Who should buy it — and who shouldn’t

Buy it if: you need the biggest open-weight models a single PCIe card can hold at full bandwidth, you fine-tune mid-size models for a living, or you’re building a multi-GPU Max-Q box where 300 W per card and 96 GB each is exactly the right shape. On power: the 600 W Workstation Edition plus the rest of the box wants a dedicated 120V/20A or 240V circuit — miners will recognize the math. The Max-Q is the sane 120V/15A pick.

Skip it if: your models fit in 24–32 GB (an RTX 5090 — or two — delivers the same bandwidth for a fraction of the price), or you want maximum gigabytes per dollar and can live with slower generation (the DGX Spark holds 128 GB for about a third of the money, at a seventh of the bandwidth). The full decision table by model size and budget lives in our local AI hardware guide.

The sovereignty frame, as always: this is the card for plebs and small shops who want frontier-class open weights running on hardware they own — no per-token meter, no retention policy but your own, the argument of the Sovereign AI Manifesto. Credit where due: the PRO 6000 stands on three decades of Quadro workstation lineage and the same GB202 silicon the gaming side funds — professional AI hardware exists because millions of gamers paid for the fab runs first.