Nvidia DGX Station GB300 Makes the $100,000 AI Tower Feel Real

As of 31 August 2026, the Nvidia DGX Station GB300 is no longer just a high-end AI workstation announced on a stage and hidden behind enterprise quote forms. On 23 August 2026, Tom’s Hardware reported that Exxact’s Valence VWS-158270643 configuration based on Nvidia’s DGX Station platform had appeared online with a starting price of $94,930. That does not make it a consumer desktop, but it does make the machine easier to understand for labs, AI startups, research teams, and enterprise developers trying to budget for local AI hardware. (tomshardware.com)

The core of the system is the Nvidia GB300 Grace Blackwell Ultra Desktop Superchip. Nvidia’s own DGX Station documentation lists a server-class Blackwell Ultra GPU with fifth-generation Tensor Cores, paired with a 72-core Grace Arm CPU. The GPU carries 252GB of HBM3e with up to 7.1TB/s of bandwidth, while the CPU side adds 496GB of LPDDR5X with up to 396GB/s of bandwidth. Together, the CPU and GPU memory form up to a 748GB coherent memory pool, linked through NVLink-C2C rather than a conventional PCIe-only workstation layout. Nvidia also lists up to 20 petaFLOPS of sparse FP4 AI compute, which places this tower closer to data-center development hardware than a graphics workstation. (docs.nvidia.com)

Explore NVIDIA DGX Spark Personal AI Desktop Supercomputer as a more compact AI development option

The interesting part is not just that the DGX Station GB300 is powerful; it is that this class of hardware is being packaged into a full-size deskside tower. The Exxact Valence model reported by Tom’s Hardware uses a 1,600W 80 Plus Titanium power supply, liquid cooling for the Grace CPU and Blackwell Ultra GPU, and a direct-to-chip coldplate design. It is meant to sit beside a desk or in a lab, but it still behaves more like infrastructure hardware than a normal workstation. Nvidia’s software documentation describes a Linux-based environment using Ubuntu 24.04 with Nvidia AI Developer Tools, along with CUDA, cuDNN, TensorRT, the Nvidia Container Toolkit, Data Center GPU Manager, DOCA-OFED, and Nvidia’s optimized kernel stack. (docs.nvidia.com)

See ASUS GeForce RTX 5090 price for a high-end GPU workstation build

For AI development, the headline spec is the 748GB coherent memory pool. Standard multi-GPU workstations can offer large aggregate VRAM, but splitting large models across several cards can add software complexity. DGX Station GB300 is designed around a tighter CPU-GPU memory model, which helps developers experiment with large models, fine-tuning, inference, and agentic AI workflows locally before moving work to larger cloud or data-center systems. Nvidia says DGX Station can run large LLMs locally and supports workflows around frameworks and tools such as PyTorch, TensorFlow, Hugging Face Transformers, vLLM, SGLang, LLaMA-Factory, Unsloth, Omniverse, Isaac, and RAPIDS. That makes it less of a general-purpose PC and more of a local AI development node. (docs.nvidia.com)

View Lenovo Legion Pro 7i laptop with Core Ultra 9 and NVIDIA RTX 5090 24GB on partner website

The tower is also built to scale beyond one box. Nvidia lists ConnectX-8 networking for DGX Station, with 800Gbps connectivity when paired with a second DGX Station. The software guide also describes wired networking that includes 1x 10GbE RJ45, 2x 400GbE QSFP through the ConnectX-8 NIC, and a dedicated 1GbE RJ45 BMC management interface. Display output is handled differently from a normal desktop too: Nvidia states that host OS display output comes from a PCIe add-in GPU, while the BMC Mini DisplayPort is reserved for management rather than the primary desktop display. (docs.nvidia.com)

Buy Crucial T705 13 GB/s NVMe SSD here for fast local AI project storage

The DGX Station GB300 is still enterprise hardware with enterprise economics. The nearly $95,000 Exxact listing simply turns a vague quote-only category into something more concrete. It gives teams a clearer comparison point against cloud GPU spending, shared cluster access, and larger rack-scale DGX deployments. For individual developers, the price is obviously extreme. For organizations that need local access to Blackwell Ultra architecture, large coherent memory, and Nvidia’s AI software stack, the DGX Station GB300 marks an important shift: rack-class AI development hardware is now available in a tower that can be ordered, configured, and placed beside the people building the models.

See MacBook Pro M5 price as a portable developer workstation alternative