One Card. No Compromises.
The top of the family: an NVIDIA RTX PRO 6000 Blackwell carrying 96 GB of ECC GDDR7 at its full 600 W GPU power budget, fed by a 1300 W 80 PLUS Platinum supply and a liquid-cooled 16-core EPYC — in a mid-tower that still fits under a desk. OpenClaw ships pre-installed on Ubuntu.
AI Silicon Spotlight
What This Silicon Does for AI
96 GB Makes Model Size Someone Else's Problem
- 96 GB of GDDR7 on one card loads the largest open-weight models this family can run — with context to spare — and skips multi-GPU sharding entirely.
- ECC video memory guards weights on services that run for weeks; a stray bit-flip gets corrected, not served.
- The full 600 W GPU budget (a power-tuned MaxQ sibling exists on this listing) buys sustained tokens-per-second, not just peak numbers.
- Universal MIG partitions the flagship into isolated instances — one card, several researchers or services, safely.
- Blackwell 5th-gen Tensor Cores with FP4 run modern low-precision inference at architecture-native speed.
The Quiet Coworker That Keeps a 600 W Card Fed
- 16 server-class cores run tokenization, dataloaders, and OpenClaw agent orchestration so the GPU never idles waiting on prep work.
- Up to 5.7 GHz max turbo keeps single-thread latency low — agent decisions and eval scripts stay snappy.
- 64 MB of L3 holds hot token pipelines close to the cores.
- EPYC ECC platform + liquid cooling: 24/7 duty cycle, clocks that hold, office-quiet acoustics under a full-power flagship.
What It Unlocks
Three Doors Only 96 GB Opens
The Biggest Models, Kept In-House
Models that need serious VRAM run on one card with long context — private, fixed-cost, behind your firewall. The optional pre-engineered LLM/RAG/automation environment (by quote) makes it turnkey.
MIG Turns the Flagship Into a Fleet
Universal MIG splits the GPU into isolated instances with dedicated resources — researchers, services, and experiments share one card without stepping on each other, served over 10 Gb/s LAN.
Sustained Throughput You Can Sell
Full 600 W GPU power, ECC end to end, a 1300 W 80 PLUS Platinum supply, and OpenClaw agents pre-installed for the operational glue — production inference on hardware you own.
Scene photos are AI-generated illustrations of typical deployments, not photographs of this exact product.
Full Config
The Supporting Cast
| ZRE4565P-RP6000 · As Configured | |
|---|---|
| GPU | NVIDIA RTX PRO 6000 Blackwell — 600 W GPU power budget, 96 GB GDDR7 with ECC |
| CPU | AMD EPYC 4565P — 16-core, liquid-cooled, 4.3 GHz base / 5.7 GHz max turbo, 64 MB L3, Socket AM5 |
| Memory | 128 GB DDR5 ECC UDIMM — expandable to 256 GB across 4 DIMM slots |
| Storage | 2× 2 TB M.2 NVMe (4 TB total) |
| Network | 1× 10 Gb/s + 2.5 Gb/s LAN + dedicated management port |
| Expansion | PCIe 5.0 x16 + PCIe 4.0 x16 · ASUS Pro WS B850M-ACE SE |
| Power | 1300 W 80 PLUS Platinum |
| OS | Ubuntu pre-installed, OpenClaw included (Windows available by quote) |
Know Your Rung
Top of the Ladder
RTX PRO 4500 tier — LLM apps, RAG, and agents.
RTX PRO 5000 tier — mid-size model serving.
RTX PRO 6000 Blackwell at full 600 W GPU power — MaxQ variant and an AMD Ryzen 9950X3D CPU alternative also on this listing.
FAQ
Quick Answers Before You Buy
What does 96 GB on a single card change?
Model size stops being the first constraint: the largest open-weight models this family can run fit on one card with room for context — no multi-GPU sharding, no cross-card complexity.
This vs. the MaxQ variant?
Both carry 96 GB. This one runs the full 600 W GPU power budget for maximum sustained throughput; the MaxQ sibling on the same listing is power-tuned. Pick this when tokens-per-second is the metric.
Can several people share it?
Yes — Universal MIG partitions the GPU into isolated instances with dedicated resources, served to the team over 1× 10 Gb/s + 2.5 Gb/s LAN.
What does "OpenClaw pre-installed" mean?
The OpenClaw enterprise AI agent framework comes installed and ready on Ubuntu — agentic workflows on day one, no setup. Windows is available by quote.
Government / public-sector eligible?
TAA compliant and assembled in the USA, with compliance documents available for purchase review.
How is it supported?
1-year parts and labor limited warranty backed by ABS manufacturer support. Build to Order — custom built and stress-tested after purchase.
