NVIDIA DGX Station Review: 748 GB, 1.6 kW and a $160,195 Dell List Price

Dell lists its Pro Precision with GB300, a DGX Station tower with 748 GB of coherent memory and a 1,600 W power budget, at $160,195 on Dell.com in the US, the price shown on Nov. 1. It is the first US list price for NVIDIA's deskside GB300 platform that Voltgrid has found from any of its system makers.

As Voltgrid reported on Oct. 24, NVIDIA says DGX Station ships from seven partners, and two labs had found a multi-user inference node rather than a personal PC. Since then, Dell has posted its price and HP has opened sales in Japan. Four measured write-ups from ServeTheHome and StorageReview now cover units from Gigabyte, MSI, ASUS and HP. This review sets NVIDIA's claims against those measurements. Voltgrid did not test the machine.

What's new

The chip is the change. DGX Station uses one GB300 Grace Blackwell Ultra Desktop Superchip: a Blackwell Ultra GPU with 252 GB of HBM3e at 7.1 TB/s, joined over a 900 GB/s NVLink-C2C link to a 72-core Grace CPU with 496 GB of LPDDR5X at 396 GB/s, according to NVIDIA's spec page. NVIDIA adds the two pools together and calls it 748 GB of coherent memory. The DGX Spark below it has 128 GB at 273 GB/s.

The compute figure needs its label. NVIDIA rates FP4 at 20 PFLOPS with sparsity and 15 PFLOPS without, and FP8 at 10 PFLOPS, which is also a sparse figure under NVIDIA's footnote. The 20 PFLOPS headline is the sparse number. NVIDIA says the machine supports models of up to 1 trillion parameters.

Networking is new for a desk machine. A ConnectX-8 SuperNIC brings two 400 Gb/s QSFP112 ports, and the GPU can be split into up to seven Multi-Instance GPU partitions, NVIDIA says.

The price point is new as of this week. Dell's default configuration, with an RTX PRO 2000 Blackwell display card and 16 TB of Gen5 SSDs, shows $160,195. When the page was read, Dell's Add to Cart button returned a notice that the action was unavailable because of a technical issue, and no ship date was shown. HP's ZGX Fury G1n AI Station went on sale in Japan on Oct. 27 from 28,490,000 yen including tax, HP Japan said. MSI publishes no price; ServeTheHome describes its WS300 as a $100K+ system.

What it solves

DGX Station removes the cluster booking for a team that wants to serve open models privately. A 120B-class model such as GPT-OSS-120B loads entirely into the GPU's HBM; StorageReview notes that even the 96 GB RTX PRO 6000 can hold it. In StorageReview's vLLM runs on an MSI XpertStation WS300, GPT-OSS-120B in its native MXFP4 peaked at 9,294 output tokens per second with 512-token prompts and 512-token answers, in a sweep of 1 to 128 concurrent streams.

Larger models load at all because of the 748 GB, but speed depends on where the weights sit. StorageReview found GLM-5.2 in NVFP4 is about 433 GB, so roughly 215 GB of expert weights spill into Grace memory on one Station. Every token that touches them crosses to LPDDR5X at 396 GB/s instead of HBM at 7.1 TB/s. The capacity sets which models run; the HBM sets how fast.

The third problem it solves is growth. NVIDIA's networking links two Stations as a two-node cluster, and StorageReview showed that this removes much of the spill penalty.

Pros and Cons

The strengths are throughput and memory. StorageReview's GPT-OSS-120B result came from one lab running the same vLLM test on three machines: the MSI tower, an RTX PRO 6000 in a Dell Pro Max Tower T2 at 3,384, and a DGX Spark, represented by an Acer Veriton GN100, at 500. On the smaller GPT-OSS-20B it measured 22,161 output tokens per second at 128 streams against 9,000 and 1,469. ServeTheHome, on a Gigabyte W775-V10-L01, measured well over 26,000 tokens per second on Qwen3.6-35B-A3B NVFP4 with short 24- to 30-token prompts, 1,024-token answers and 512 concurrent requests. It did not name the serving tool. StorageReview's thermals follow-up reported a 60C die at a flat 2.07 GHz over a 2.8-hour run on the MSI unit, with a cooling loop MSI rates at 1,400 W; the review itself logged a 71C peak at 1,292 W GPU draw. ServeTheHome said the MSI unit performed flawlessly.

The weaknesses are price, power and single-user speed. ServeTheHome measured the Blackwell Ultra GPU at about 148 W at idle and the Grace module at about 78 W, and said the whole platform never fell below 250 W. It measured 37 to 40 dBA at light load. In StorageReview's runs on a single ASUS ExpertCenter Pro ET900N G3, GLM-5.3 in NVFP4 began at 44 output tokens per second for one stream and reached 188 at 32 streams, using NVIDIA's vLLM 26.08 container (vLLM 0.27.1) with multi-token prediction. GLM-5.2 began at 35. Storage specifications also conflict across sources: NVIDIA lists four M.2 Gen5 slots, StorageReview lists four PCIe 5.0 x4, ASUS lists PCIe 6.0 x4 and ServeTheHome's MSI table lists two Gen6 and two Gen5. No source has explained the gap.

Who it's for

It fits a team that serves agents or internal users around the clock on 30B to 120B-class open models, where aggregate throughput in the thousands of tokens per second matters more than any one user's speed. It fits a group building for GB300 racks that wants the same GPU architecture under the desk. It fits a lab that plans to start with one Station and add a second for 400 GB-plus models, the case StorageReview measured.

It does not fit one developer who wants the fastest single chat. At 35 to 44 tokens per second for one stream on the largest MoE models StorageReview ran, and at $160,195 in Dell's configuration, a $6,950 DGX Spark or a Mac Studio costs a small fraction as much for one user on smaller models. It also does not fit an office without a dedicated 20 A circuit.

Why the second Station matters

StorageReview linked the ASUS tower to an HP ZGX Fury with two 400G direct-attach copper cables and NVIDIA's own setup guide. NCCL collectives reached 98 GB/s against the 100 GB/s the pair can carry. Splitting GLM-5.3 across both GPUs in tensor parallel kept all of it in HBM, and output rose from 188 tokens per second at 32 streams on one tower to 5,018 at 256 streams on two. Those are different concurrency points, which StorageReview states; at matched 32 streams, GLM-5.2 rose 11x, from 139 to 1,525. With prefill on one Station and decode on the other, DeepSeek V4.1 Flash in FP8 reached 5,248 tokens per second at 256 streams.

How it compares

On StorageReview's single test, DGX Station's 9,294 tokens per second on GPT-OSS-120B is about 2.7 times the RTX PRO 6000 and nearly 19 times the DGX Spark in that test; StorageReview's earlier Acer GN100 review reached 713 tok/s aggregate on the same model under a different setup. Those are the only rivals measured on the same model, quantization, tool and workload. The table under this story lists each product's memory, bandwidth, price and one independent result.

The DGX Spark has 128 GB at 273 GB/s and lists at $6,950 for the 128 GB, 4 TB model on NVIDIA's marketplace (a copy cached Oct. 31, stock unverified). It runs the same NVIDIA software stack, so it is the cheaper place to develop for a Station. The RTX PRO 6000 Blackwell Workstation Edition has 96 GB of GDDR7 at 1,792 GB/s and a 600 W board power, and lists at $16,000 on NVIDIA's US marketplace, where it was listed out of stock on Aug. 10. Its 96 GB cannot hold the 433 GB models that two Stations serve.

The Mac Studio with M5 Ultra offers up to 512 GB of unified memory at 1.2 TB/s, at $22,299 minimum for the 512 GB model, according to MacRumors. That is more memory than the Station's HBM at about a seventh of Dell's price, but at about a sixth of the HBM bandwidth and on a different software stack. No lab has published a head-to-head run of the two on the same model and tool, so their speeds cannot be compared directly.

What to watch

Partners are due to ship DGX Station for Windows in December 2026, NVIDIA said on Oct. 21, and HP Japan says Windows support for the ZGX Fury is planned from December. The 512 GB Mac Studio, the closest large-memory alternative, has first orders due Dec. 3, according to Digital Trends. Still missing are US list prices from ASUS, HP, MSI and Gigabyte, a Dell ship date, a measured full-load wall-power figure from any outlet, and a single-user llama.cpp result on a dense 70B-class model.