AI $30K Workstation
A community post laid out a roughly $30,000 workstation built for local AI work with explicit parts and cooling choices. The build lists an AMD Threadripper PRO 7965WX, ASUS Pro WS WRX90E‑SAGE SE motherboard, 256GB DDR5 ECC, two NVIDIA RTX PRO 6000 Blackwell GPUs (192GB VRAM total), a Corsair 9000D case, MSI 1600W PSU, and a Samsung 9100 PRO 8TB PCIe 5.0 SSD — builders noted liquid‑cooling tweaks and community optimizations for heavy GPU utilization (x.com) (x.com) (x.com).
A roughly $30,000 desktop build is circulating as a template for running large artificial intelligence models locally, without renting cloud servers. (nvidia.com) Local AI work depends heavily on memory: the bigger the model, the more data has to sit close to the chips instead of being swapped in and out from storage. This build centers on two NVIDIA RTX PRO 6000 Blackwell workstation cards, each with 96 gigabytes of error-correcting GDDR7 memory, for 192 gigabytes of video memory in total. (nvidia.com ) The processor and motherboard are workstation parts built for expansion rather than gaming. AMD’s Threadripper PRO 7965WX is a 24-core, 48-thread chip, and ASUS says its Pro WS WRX90E-SAGE SE board provides seven PCI Express 5.0 x16 slots plus dual 10-gigabit Ethernet. (techpowerup.com) (asus.com) That matters because local AI rigs are increasingly constrained by video memory and power delivery, not just by raw processor speed. NVIDIA markets the RTX PRO 6000 Blackwell series for “agentic AI,” scientific computing, rendering, and real-time video work, a sign that workstation hardware is being pitched directly at model developers as well as designers and engineers. (nvidia.com) The rest of the parts list is built around feeding and cooling those cards. Corsair says its 9000D Airflow super-tower is designed for very large systems with extensive radiator support, while MSI’s MEG Ai1600T PCIE5 power supply is rated for 1,600 watts and includes two 600-watt PCI Express 5.1 power connectors. (corsair.com) (msi.com) Storage is fast, but it is not the main bottleneck once a model is loaded. Samsung says its 9100 PRO solid-state drive reaches sequential read speeds up to 14,800 megabytes per second and is sold in capacities up to 8 terabytes, which helps with moving large model files and datasets on and off the machine. (samsung.com) Builders in the discussion focused on liquid-cooling changes and airflow tuning because two top-end workstation graphics cards can turn a desktop into a heat problem as much as a compute problem. Third-party listings for the RTX PRO 6000 Blackwell put the card at 96 gigabytes of ECC memory and show workstation availability around $9,989 per card, which helps explain how the total build approaches $30,000 before extras. (bhphotovideo.com) The appeal of a machine like this is control: models, data, and inference stay on hardware you own, with no hourly cloud bill attached to every experiment. The tradeoff is that local AI at this scale now looks less like a hobby PC and more like a small on-premises server built into a tower case. (amd.com)
