Bringing the Spark to Europe: Local Intelligence Powers Next Wave of Agents
NVIDIA Blog Gerardo Delgado ● Covered by 6 sources
NVIDIA used GTC Berlin to push its local AI gear for PCs and workstations. The pitch is simple: run agents and models on-device, not in the cloud.
Based on reporting by NVIDIA Blog, Gerardo Delgado — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
NVIDIA turned GTC Berlin into a showcase for something it clearly wants to make normal: local AI that runs on your own hardware instead of somewhere off in the cloud. The company spread the message across three tiers of machines — RTX Spark for personal agents on Windows PCs, DGX Spark for developers on Linux, and DGX Station for larger enterprise and robotics workloads, with a Windows version of DGX Station due in December.
RTX Spark is the most consumer-facing of the lot. It launched last week, and NVIDIA says it can handle personal agents, accelerated content creation and 1440p gaming at 100 frames per second, while packing up to 128GB of unified memory and 1 petaflop of AI compute. Underdog, a desktop agent app built for fully local AI, is now available there too, keeping user data on the device. NVIDIA also pointed to unlocked bootloaders on many RTX Spark laptops, which should make tinkering easier for developers and enthusiasts.
DGX Spark is the bridge between that laptop world and more serious development work. The 64GB systems start rolling out on Friday, Oct. 23, and multiple units can be linked with NVIDIA ConnectX-7 networking and NVIDIA Sync Cluster Assistant. NVIDIA says clusters of up to eight units are supported in manual configuration, giving small teams a shared backend with 1TB of unified memory. The company is also adding Model Launcher later this month, along with a Wi-Fi Assistant in the October software update and support for models from Nemotron to Qwen, DeepSeek and GLM.
DGX Station is the bigger play. NVIDIA says it offers 748GB of coherent memory and up to 20 petaflops of AI performance, enough to serve more than 30 users at speeds comparable to cloud models. New integrations with security, fleet-management and IT tools are meant to make it easier for enterprises to approve and deploy. And the Windows version is aimed squarely at the software and manufacturing teams already living in Windows-native tools, with support coming from companies including nTop, Dassault Systèmes, MathWorks, Siemens and Synopsys.
The rest of the announcement fills in the ecosystem around those boxes: a new local coding model called Recursive Flint, day-zero support for a batch of new models, and speedups for llama.cpp and vLLM. NVIDIA is not being subtle here. It wants local AI to feel less like a workaround and more like the default way to build, test and run agents.
My take — AI-written commentary, not fact-checked reporting
This is classic NVIDIA: sell the hardware, then sell the software stack, then sell the idea that the cloud was always optional. That’s smart business and, annoyingly for everyone else, probably the right bet for enterprises that hate sending data places they can’t control. The open-model crowd keeps winning the practical argument whenever the machines get powerful enough to stay on the desk.
Read more about this at: NVIDIA Blog
