Novatech AI server in a SilverStone RM600 6U rackmount chassis Hand-built · burn-in tested
NVIDIA RTX PRO 6000 Blackwell Workstation Edition 96 GB GPU memory · 4,000 AI TOPS
AI SERVERIn stock · ships in 7–10 business days

Threadripper PRO 9995WX AI Server with RTX PRO 6000

96 GB of GPU memory in your own rack. Run 70B-class models and gpt-oss-120b for your whole team.

Best forModels up to 120B70B models at 8-bitFine-tuning larger models
GraphicsRTX PRO 6000 · 96 GB
ProcessorRyzen Threadripper PRO 9995WX
Memory512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)
Storage4 TB PCIe 5.0 NVMe SSD (9100)
$73,999Full specs
Buying for a business? Get a quote, PO or net terms
  • 48-hour burn-in test
  • 3-year limited warranty
  • Free shipping
  • Lifetime tech support
RTX PRO 6000$73,999
RTX PRO 6000 server$73,999

Overview

Private AI for your whole team, in your own rack

The Threadripper PRO 9995WX AI Server with RTX PRO 6000 pairs NVIDIA's 96 GB Blackwell workstation GPU with AMD's 96-core Zen 5 Threadripper PRO 9995WX and 512 GB of ECC memory, in a 6U rackmount chassis. That's enough GPU memory for 70B-class language models at 8-bit and OpenAI's gpt-oss-120b, entirely on the GPU, served to your whole team from your own rack. Inside a SilverStone RM600 you get an ASRock WRX90 WS EVO board with dual 10 GbE and IPMI, a 4 TB PCIe 5.0 NVMe SSD, a SilverStone XE360-TR5 liquid cooler and a 2500 W SilverStone HELA power supply.

96 GBGPU memorySets which AI models fit on the RTX PRO 6000.
4,000AI TOPSNVIDIA's rating for AI throughput.
96 coresThreadripper PRO 9995WXFor data pipelines, VMs and every user it serves.
512 GBSystem memory512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)

What's inside

What's inside the RTX PRO 6000 server

Every part is listed by name, so you know exactly what's going into your rack and what you can upgrade later.

1 2 3 4 5 6 7
  1. 1
    GraphicsNVIDIA RTX PRO 6000 Blackwell Workstation Edition96 GB GPU memory · 4,000 AI TOPS
  2. 2
    ProcessorAMD Ryzen Threadripper PRO 9995WX96 cores, 192 threads · up to 5.4 GHz
  3. 3
    CPU coolerSilverStone XE360-TR5 360 mm liquid cooler
  4. 4
    Memory512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)
  5. 5
    Storage4 TB PCIe 5.0 NVMe SSD (9100)
  6. 6
    Power supplySilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1
  7. 7
    CaseSilverStone RM600 6U rackmount

Use cases

What you can do with it

It lives in your rack or server room, so every user and every job runs on hardware you own: no per-hour bill, no usage caps, and nothing leaves the building.

Running locally

Serve AI to your whole team

Host open models like gpt-oss, Qwen and Llama behind your own endpoint with vLLM or Ollama. Prompts and data stay on your network.

contract.pdf Page 4

Search your company's data

Build RAG over contracts, tickets and code. The embeddings, the vector search and the model all run on one machine you control.

Training loss LoRA adapter

Train and fine-tune

Fine-tune on your own data while 96 CPU cores handle data prep, evaluation and everything else around the GPUs.

Generate

Run jobs around the clock

Image generation, transcription and batch inference as shared services that run day and night, with no per-hour bill.

Model fit

What AI models can the RTX PRO 6000 server run?

GPU memory (VRAM) decides which models fit. This build has 96 GB, so every bar that ends left of the line runs fully on the GPU.

Fits in GPU memory Needs more memory This build: 96 GB
27B models · 4-bitGemma 3 27B
~19 GB
Runs on this buildRuns
32B models · 4-bitQwen3 32B
~22 GB
Runs on this buildRuns
FLUX.1 [dev] · full precisionFLUX image model
~28 GB
Runs on this buildRuns
70B models · 4-bitLlama 3.3 70B
~44 GB
Runs on this buildRuns
gpt-oss-120bOpenAI open-weight model
~66 GB
Runs on this buildRuns
70B models · 8-bitLlama 3.3 70B
~80 GB
Runs on this buildRuns
235B models · 4-bitQwen3 235B-A22B
~130 GB
Step up to the 2× RTX PRO 6000 build2× RTX PRO 6000
70B models · full precisionLlama 3.3 70B
~150 GB
Step up to the 2× RTX PRO 6000 build2× RTX PRO 6000
405B models · 4-bitLlama 3.1 405B
~230 GB
Ask about larger buildsAsk us

Approximate memory for the model plus a working context. On the dual-GPU server the model is split across both cards. Larger models can run partly on the CPU and system memory, much more slowly.

Compare

Compare the AI server lineup

Same 6U chassis, Threadripper PRO platform and build process. The GPUs set which models you can run; the CPU and memory set how many users and jobs it handles at once.

RTX 5090Threadripper PRO 7995WX$34,999View build
RTX 5090Threadripper PRO 9995WX$52,999View build
YOU'RE VIEWINGRTX PRO 6000Threadripper PRO 9995WX$73,999
2× RTX PRO 6000Threadripper PRO 9995WX$92,998View build
GPU memory32 GB32 GB96 GB192 GB
AI TOPS3,3523,3524,0008,000
ProcessorThreadripper PRO 7995WXThreadripper PRO 9995WXThreadripper PRO 9995WXThreadripper PRO 9995WX
CPU cores96969696
System memory256 GB DDR5-4800 ECC RDIMM (8 × 32 GB)512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)
Storage2 TB PCIe 5.0 NVMe SSD (9100)4 TB PCIe 5.0 NVMe SSD (9100)4 TB PCIe 5.0 NVMe SSD (9100)4 TB PCIe 5.0 NVMe SSD (9100)
Power supplySilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1SilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1SilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1SilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1

Support

Built and backed by people you can call

Burn-in report Threadripper PRO 9995WX AI Server with RTX PRO 6000 GPU stressCPU stressMemory testStorage check PASS PASS PASS PASS GPU temperature Start End of test

48-hour burn-in

Every system is assembled and stress-tested in the USA for 48 hours before it ships.

Terminal $ nvidia-smi --query-gpu=name --format=csv,noheader NVIDIA RTX PRO 6000 Blackwell Workstation Edition $ python3 -c "import torch; print(torch.cuda.is_available())" True $ ollama run gpt-oss:20b >>> Send a message (/? for help)

Ready on first boot

Windows 11 Pro with current NVIDIA drivers, set up and tested before it ships.

3-year limited warranty

Hardware coverage handled by the team that built your system.

Lifetime technical support

Questions about drivers, models, networking or upgrades go straight to our build team.

Specifications

Full specifications

Threadripper PRO 9995WX AI Server with RTX PRO 6000

Compute

Processor
AMD Ryzen Threadripper PRO 9995WX96 cores, 192 threads · up to 5.4 GHz
Graphics
NVIDIA RTX PRO 6000 Blackwell Workstation Edition24,064 CUDA cores · 96 GB GDDR7 with ECC · 1,792 GB/s memory bandwidth · 4,000 AI TOPS
GPU memory
96 GB

Memory and storage

System memory
512 GB DDR5-5600 ECC RDIMM (4 × 128 GB)
Storage
4 TB PCIe 5.0 NVMe SSD (9100)

Platform

Motherboard
ASRock WRX90 WS EVO
Power supply
SilverStone HELA 2500Rz · 2500 W 80 PLUS Platinum · ATX 3.1
CPU cooler
SilverStone XE360-TR5 360 mm liquid cooler
Case
SilverStone RM600 6U rackmount

Software and connectivity

Operating system
Windows 11 Pro
Connectivity
2 × 10 Gb Ethernet (Intel X710) · IPMI (ASPEED AST2600) · 2 × USB4

Size, power and warranty

Dimensions and weight
6U rackmount · 17.2 × 10.4 × 20.6 in (W × H × D)
Warranty
3-year limited hardware warranty

FAQ

Questions before you buy

Can't find your answer? Contact us.

What can I run on it?

Check the model list above: anything marked “Runs on this build” fits fully in GPU memory. Working with something that isn't listed? Ask us before you buy.

Will it fit my rack?

It's built in a SilverStone RM600, a 6U rackmount chassis about 17.2 in wide, 10.4 in tall and 20.6 in deep, made for standard 19-inch racks. Tell us your rack model and we'll confirm the depth and rails before it ships.

What power does it need?

The single-GPU servers run from a standard 120 V outlet; give them their own circuit. The dual RTX PRO 6000 server can draw close to 1,700 W at full load, and its 2,500 W SilverStone HELA power supply only delivers full output on 200–240 V, so plan for a 200–240 V circuit, which most server rooms have.

How do I connect and manage it?

The ASRock WRX90 WS EVO board has two 10 Gigabit Ethernet ports (Intel X710) for data, plus a dedicated IPMI management port (ASPEED AST2600) for monitoring and remote power control over the network.

Which operating system does it come with?

Windows 11 Pro, set up with current NVIDIA drivers and tested before it ships.

Can I upgrade it later?

Yes. The board has seven PCIe 5.0 x16 slots and eight ECC memory slots (up to 2 TB), and the 512 GB builds leave four memory slots free. Contact us first: hardware installed after shipment isn't covered by the warranty, and we can confirm which parts are compatible.

How long until it ships?

Each system is built to order, burn-in tested for 48 hours and ships in 7–10 business days. Shipping is free.

Do you sell to businesses?

Yes. We accept purchase orders and GSA, offer net terms, and quote multi-unit orders. Request a quote.

Ready to order?

$73,999

Ships in 7–10 business days

Need a different configuration?

Multi-GPU builds, more memory, rack-mount servers, or several units for a team. Tell us the workload and we'll send a quote.

Request a quote