NVIDIA expects partner-built DGX Station for Windows systems to begin shipping in December. That is the current timetable in the company's October 21 announcement, rather than confirmation that Windows systems are already on sale. Separately, NVIDIA says Linux systems are shipping. The distinction matters for teams planning local AI infrastructure around a particular operating system.
A GenAI Works LinkedIn post, checked on October 23, draws attention to a demonstration mentioning 768 GB of memory. The current product page instead specifies up to 748 GB of coherent memory. We use the current specification for planning purposes; we have no confirmation that the figure in the clip describes a separately offered configuration.
Large capacity does not mean identical speed throughout
NVIDIA's development guide separates memory into 252 GB of GPU HBM3e and 496 GB of CPU LPDDR5X, with respective bandwidths of 7.1 TB/s and 396 GB/s. The sum does not mean uniformly fast video memory or a single resource delivering 8 TB/s throughout.
The practical question is where the data most frequently used by a particular workload resides. Fitting a model into the combined capacity does not establish how quickly it will respond under a real workload. A prospective buyer should request measurements with the chosen model and the settings the team will actually use, rather than compare memory capacity alone.
The up to 20 PFLOPS figure also comes with a condition: the product page associates it with sparse FP4 Tensor Core computation. It is not a measurement of generated tokens per second. This number alone cannot establish a chat application's speed or the number of simultaneous users it can serve.
What changes in the workflow
The Windows version was previewed at a Microsoft event on October 7. NVIDIA described local agents and access to Linux AI tools through WSL. This is a proposal to bring the working environment and AI computation into one machine, rather than evidence that every existing enterprise application is already compatible.
It is useful to distinguish development, model adaptation and inference with existing models. Our editorial conclusion is to begin with the task: an internal document assistant, a model experiment or a longer processing job. Each needs its own criteria for quality, completion time and resource use.
For example, a pilot involving internal documents could use a prepared set of questions with checked answers. The team should record incorrect answers, time to the first useful result and behaviour under several parallel requests. This is a proposed evaluation method, not a test SIExpo has conducted on DGX Station.
A local machine still needs operating rules
Local execution offers a different place to run computation, but privacy needs to be checked across the entire workflow. A pilot should identify external services, the agent's permissions and where requests and outputs are logged. Having the computer physically inside an office does not replace that review.
The same approach applies to costs: compare actual quotations, support arrangements and the expected workload. The sources reviewed did not establish a price or confirmed availability in Bulgaria. We also make no independent finding that the system will cost less than a cloud service.
The news is the expected expansion of DGX Station to Windows and a clearer shipping timetable. Deployment remains a decision requiring verified compatibility and a team's own pilot. The technical capabilities quoted are manufacturer specifications; SIExpo has not tested the machine.
*Cover: a conceptual AI illustration, not a photograph of a real product.*
*Sources checked: 2026-10-23T16:10:40.049366+00:00.*
