# Hardware for running AI on your own ground

> OnPrem World · https://onpremworld.com/en/shop · Last updated: 2026-10-16

Every machine we sell starts with what it can run. Specs and fit estimates come from our open dataset; everything is quote-based, so add what you like to your quote list and ask once.

We sell the hardware on this page. Specs, fit estimates and shelf prices are compiled independently of sales and are the same on every page of this site.

## Desktop AI supercomputers

### [DGX Spark](https://onpremworld.com/en/hardware/nvidia-dgx-spark.md)

CUDA on your desk with 128 GB of unified memory: the easiest way to run 120B-class MoE models in the office.

- 64–128 GB · 273 GB/s
- Runs: Qwen3.5-122B-A10B, gpt-oss-120b (Q4_K_M, 8K)
- Taiwan shelf price: NT$169,000–NT$199,500 (checked 2026-10-15)
- Quote on request
- Versions we can source: NVIDIA DGX Spark (Founders Edition), ASUS Ascent GX10, Acer Veriton GN100, Dell Pro Max with GB10, HP ZGX Nano, Lenovo ThinkStation PGX, MSI EdgeXpert, GIGABYTE AI TOP ATOM

### [DGX Station (GB300)](https://onpremworld.com/en/hardware/nvidia-dgx-station-gb300.md)

A data-center GPU beside your desk: 252 GB of HBM3e for department-scale inference and fine-tuning.

- 252 GB · 7100 GB/s
- Runs: Qwen3.8-Flash-Next, Qwen3.5-122B-A10B (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Project quote · long lead time
- Versions we can source: ASUS ExpertCenter Pro ET900N G3, Dell Pro Max with GB300, GIGABYTE W775-V10-L01, HP ZGX Fury AI Station, MSI XpertStation WS300, Supermicro Super AI Station GB300

## Workstation GPUs

### [RTX PRO 6000 Blackwell Max-Q Workstation Edition](https://onpremworld.com/en/hardware/nvidia-rtx-pro-6000-blackwell-max-q.md)

96 GB at 1,792 GB/s in a 300 W blower card: the building block for 2- to 4-GPU team workstations.

- 96 GB · 1792 GB/s
- Runs: Qwen3.5-122B-A10B, gpt-oss-120b (Q4_K_M, 8K)
- Taiwan shelf price: NT$575,000 (checked 2026-10-15)
- Quote on request
- Versions we can source: Leadtek RTX PRO 6000 Blackwell Max-Q

### [RTX PRO 6000 Blackwell Workstation Edition](https://onpremworld.com/en/hardware/nvidia-rtx-pro-6000-blackwell.md)

The 600 W flow-through version of the RTX PRO 6000, for single-GPU workstations.

- 96 GB · 1792 GB/s
- Runs: Qwen3.5-122B-A10B, gpt-oss-120b (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Quote on request
- Versions we can source: Leadtek RTX PRO 6000 Blackwell Workstation Edition

### [RTX PRO 5000 72GB Blackwell](https://onpremworld.com/en/hardware/nvidia-rtx-pro-5000-blackwell-72gb.md)

72 GB on one 300 W card: gpt-oss-120b at 4-bit, just.

- 72 GB · 1344 GB/s
- Runs: gpt-oss-120b, Llama 3.3 70B Instruct (Q4_K_M, 8K)
- Taiwan shelf price: NT$455,000 (checked 2026-10-15)
- Quote on request
- Versions we can source: Leadtek RTX PRO 5000 Blackwell 72GB

### [RTX PRO 5000 48GB Blackwell](https://onpremworld.com/en/hardware/nvidia-rtx-pro-5000-blackwell-48gb.md)

48 GB at 1,344 GB/s: a 70B dense model at 4-bit on one card.

- 48 GB · 1344 GB/s
- Runs: Llama 3.3 70B Instruct, Qwen3-32B (Q4_K_M, 8K)
- Taiwan shelf price: NT$325,000 (checked 2026-10-15)
- Quote on request
- Versions we can source: Leadtek RTX PRO 5000 Blackwell 48GB

### [RTX PRO 4500 Blackwell](https://onpremworld.com/en/hardware/nvidia-rtx-pro-4500-blackwell.md)

32 GB at 200 W: 27-32B models at high-quality quantizations, with ECC memory.

- 32 GB · 896 GB/s
- Runs: Qwen3-32B, Gemma 4 31B IT (Q4_K_M, 8K)
- Taiwan shelf price: NT$189,000 (checked 2026-10-15)
- Quote on request
- Versions we can source: Leadtek RTX PRO 4500 Blackwell

### [RTX PRO 4000 Blackwell](https://onpremworld.com/en/hardware/nvidia-rtx-pro-4000-blackwell.md)

A single-slot 24 GB card: the most affordable way into CUDA workstation cards.

- 24 GB · 672 GB/s
- Runs: Qwen3-32B, Gemma 4 31B IT (Q4_K_M, 8K)
- Taiwan shelf price: NT$124,000 (checked 2026-10-15)
- Quote on request
- Versions we can source: Leadtek RTX PRO 4000 Blackwell, Leadtek RTX PRO 4000 Blackwell SFF

### [Radeon AI PRO R9700](https://onpremworld.com/en/hardware/amd-radeon-ai-pro-r9700.md)

32 GB for much less per GB than NVIDIA cards, on AMD ROCm and Vulkan instead of CUDA.

- 32 GB · 640 GB/s
- Runs: Qwen3-32B, Gemma 4 31B IT (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Quote on request

### [Arc Pro B70](https://onpremworld.com/en/hardware/intel-arc-pro-b70.md)

32 GB on Intel's oneAPI and Vulkan stack: a budget option for small RAG and chat servers.

- 32 GB · 608 GB/s
- Runs: Qwen3-32B, Gemma 4 31B IT (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Quote on request

## Consumer GPUs

### [GeForce RTX 5090](https://onpremworld.com/en/hardware/rtx-5090.md)

32 GB at 1,792 GB/s. Off Taiwan retail shelves since Oct 2026; we quote it inside complete systems.

- 32 GB · 1792 GB/s
- Runs: Qwen3-32B, Gemma 4 31B IT (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Only in complete systems

### [GeForce RTX 5080](https://onpremworld.com/en/hardware/rtx-5080.md)

16 GB at 960 GB/s for 8-20B models; many Taiwan shops sell it only with a system build.

- 16 GB · 960 GB/s
- Runs: gpt-oss-20b, Qwen3-8B (Q4_K_M, 8K)
- Taiwan shelf price: NT$53,888 (checked 2026-10-14)
- Quote on request

## Mini PCs

### [Ryzen AI Max+ 395 (Strix Halo) systems](https://onpremworld.com/en/hardware/amd-ryzen-ai-max-plus-395.md)

Up to 128 GB of unified memory in a mini PC: big models on a budget, slower than CUDA boxes.

- 32–128 GB · 256 GB/s
- Runs: Qwen3.5-122B-A10B, gpt-oss-120b (Q4_K_M, 8K)
- No Taiwan shelf price recorded
- Quote on request

Request a quote: https://onpremworld.com/en/quote · alexma@painpoint-ai.com