THE APEX TIMES
NVIDIA’s GTC Berlin pitch: local “Spark” systems bring AI agents closer to where work happens
At GTC Berlin, NVIDIA highlighted a new lineup of local AI hardware and software, positioning “local intelligence” as the next platform for personal agents, developer workflows, and enterprise deployments that keep data on device.
NVIDIA used its GTC Berlin keynote to argue that the next wave of artificial intelligence agents will run closer to users and developers, not only in the cloud. The company showcased a platform of “local AI” systems it says is designed to accelerate agentic workloads on different hardware tiers, ranging from Windows laptops to Linux-based enterprise servers.
The centerpiece for personal computing is RTX Spark, which NVIDIA says launched on October 16. RTX Spark is positioned as a PC platform for running personal agents, accelerated content creation, and gaming at 1440p with up to 100 frames per second. NVIDIA also claimed RTX Spark configurations can include up to 128GB of unified memory and deliver “1 petaflop of AI compute,” framing it as a machine built for on-device inference rather than round-tripping user data to remote servers.
NVIDIA also tied RTX Spark to a companion desktop agent application, Underdog, described as a fully local and private AI app. NVIDIA said Underdog runs models using its Husky model-specific inference engine, keeping model execution on the device so that “user data [does not have to] ever leaving the PC.” The company said Underdog is available for download now and added that many RTX Spark laptops include an unlocked bootloader to help developers and enthusiasts build and customize local agents.
For developers building and scaling local AI applications, NVIDIA pointed to DGX Spark. The company said new DGX Spark 64GB systems begin rolling out from partners on Friday, October 23. NVIDIA said these systems can run models including Nemotron 3.5 Bolt 74B (24GB), Qwen3.8 27B, and Gemma 4 26B-A4B, and that when more capacity is required, developers can connect multiple DGX Spark systems using NVIDIA ConnectX-7 networking plus NVIDIA Sync Cluster Assistant.
NVIDIA described the DGX Spark clustering workflow as enabling larger models, longer contexts, concurrent agents, and distributed development. It said DGX Spark clusters can be configured up to eight units in a manual setup, and it claimed such configurations deliver 1TB of unified memory. NVIDIA framed the result as a backend suitable for small teams where multiple developers can run concurrent sessions against the same model.
The company’s enterprise tier, DGX Station, was positioned as a “frontier workload” system running on Linux, with a Windows version expected to begin shipping in December. NVIDIA said DGX Station offers 748GB of coherent memory and up to 20 petaflops of AI performance, and it claimed it can serve 30-plus users at speeds “comparable to” cloud models. It also said NVIDIA Sync now supports connections between two Linux-based DGX Station systems.
On the software and operations side, NVIDIA said DGX Station integrates with a list of enterprise management and security tools, including Absolute Security, Splunk by Cisco, Fleet Device Management, Microsoft Intune, Perforce Puppet, Progress Software Chef, and Tanium. The company said these integrations are aimed at helping IT teams manage DGX Station systems through enterprise setup, management, and conformity processes, with the goal of accelerating approvals and reducing deployment risk.
NVIDIA also used the Berlin stage to connect the platform to agent workflows beyond generic coding. During the keynote, CEO Jensen Huang demonstrated AI agents running on RTX Spark used to train Hugging Face’s Microduck robot using NVIDIA Isaac Lab, Hermes Agent, and Nemotron 3.5 Bolt 74B. NVIDIA further described additional local agent and workflow tooling for enterprise, including DGX Station reference workflows that it said bring GPT-6 Astra, Isaac Lab, Isaac Lab-Arena, and Cosmos to DGX Station.
On what is still unclear, NVIDIA’s post emphasizes performance targets and system availability, but it does not provide pricing, detailed benchmark methodology, or specific availability timelines for every configuration across partners. It also does not break out how much of the claimed agent capability depends on particular software components, model sizes, or customer-specific deployments. As a result, the practical gap between “local” demos and day-to-day production reliability and manageability will likely hinge on integration work that the company says it is enabling through its management tools and orchestration features.
Looking ahead, NVIDIA is effectively asking developers and enterprises to adopt local inference and local agent workflows as a parallel path to cloud. The next concrete milestones to watch are partner rollouts of DGX Spark 64GB systems beginning October 23, shipping cadence for DGX Station systems, and the arrival of the Windows-based DGX Station later this year alongside the company’s software updates aimed at simplifying model access and operations in local environments.
Why It Matters
- If NVIDIA’s pitch lands, local AI agent platforms could reduce latency and data transfer constraints for enterprise workflows while still supporting multi-user development and inference.
- The push across tiers, from RTX Spark to DGX Spark clusters and DGX Station, suggests NVIDIA wants developers to standardize agent tooling across device classes rather than treating local AI as a one-off experiment.
- Enterprise integrations and device-management support point to a key adoption hurdle for on-prem systems: fitting into existing IT governance and security processes.
- NVIDIA’s emphasis on “model access” and workflow guidance inside its software stack indicates a bid to lower the operational overhead of running large models locally.
Sources
Key Facts
- NVIDIA showcased local AI “Spark” hardware and software at GTC Berlin, including RTX Spark for Windows PCs, DGX Spark for developers on Linux, and DGX Station for enterprise workloads on Linux with a Windows version expected in December.
- RTX Spark launched October 16 and NVIDIA described it as supporting personal agents, accelerated content creation, and up to 100 frames per second gaming at 1440p, with configurations up to 128GB unified memory and “1 petaflop” of AI compute.
- Underdog is described as a desktop agent app for fully local, private AI, powered by a Husky model-specific inference engine that keeps model execution on-device; Underdog is available for download now.
- NVIDIA said new DGX Spark 64GB systems begin rolling out from partners October 23, and it described scaling via NVIDIA ConnectX-7 networking and NVIDIA Sync Cluster Assistant for clusters up to eight units with 1TB unified memory.
- DGX Station was described as offering 748GB of coherent memory and up to 20 petaflops of AI performance on Linux, plus enterprise management integrations and connections supported by NVIDIA Sync.
- NVIDIA said DGX Station for Windows is expected to begin shipping in December and will target workflows that use professional applications with access to AI models on NVIDIA RTX PRO GPUs and agents backed by NVIDIA GB300 systems.
Technology Related
Netflix posts a profit gain, but investors want proof on hits and ads ahead of January
A stronger bottom line is not yet convincing Wall Street. The market is pressing Netflix to show it can deliver a true global breakout and to fix what it views as shortcomings in its ad-supported rollout, while also demonstrating that revenue momentum is stabilizing.
Armada says it will build “superintelligence” factories using NVIDIA DSX POD, positioning deployments as a months-long rollout
At NVIDIA GTC, Armada announced support for the NVIDIA DSX POD reference design, aiming to help partners assemble standardized, factory-integrated compute systems faster than traditional large-scale deployments.
Tanium and NVIDIA team up to bring enterprise endpoint management to local AI systems
The companies say the integration will extend Tanium’s enterprise-grade device management capabilities to NVIDIA’s local AI infrastructure, aiming to make deployments easier to control at scale.
TrendAI adds NVIDIA Nemotron open models to its enterprise security platform for sovereign and private cloud deployments
The Trend Micro unit says the move is aimed at letting regulated organizations run AI-driven security controls with model and data staying within customer-managed environments.
Nvidia debuts DSX POD, a shipping-container-sized data center concept at GTC Berlin
The company introduced DSX POD at its GTC Berlin event, indicating continued push to make AI compute deployments more standardized and faster to install, even as details of configurations and pricing remain unclear.
Nvidia unveils AI agent safety push and new humanoid robot design elements at GTC Berlin
The company used its Berlin developer conference appearance to highlight software and robotics themes aimed at limiting unsafe behavior in AI agents and advancing humanoid robot concepts.
Volkswagen Group and partners build a shared “physical AI” data flywheel for autonomous-vehicle development, with Microsoft among integrators
Porsche is set to lead the initial implementation of a joint vehicle-data platform, combining computer-vision and AI tooling with cloud and platform integrations aimed at accelerating AV training and validation.
Yahoo Finance column pins hopes on Apple’s next leg after Oct. 28
A market-focused prediction from Yahoo Finance says Apple’s stock could benefit after Oct. 28, citing investor enthusiasm around the company’s innovation push. Apple did not provide details in the cited post beyond general business framing.
Crusoe backs NVIDIA DSX POD, aiming to plug its modular AI factory approach into NVIDIA’s data center stack
The cloud infrastructure provider says its Crusoe Spark platform, built from years of running modular AI factories, will be integrated with NVIDIA’s DSX POD offering, expanding a modular ecosystem for enterprises that want faster deployment of AI capacity.
Nvidia CEO Jensen Huang pitches “AI factories” that can be deployed faster, with new DGX Pod aimed at financing hurdles
In remarks at Nvidia’s GTC Berlin, Huang said the company wants AI factory deployments to take weeks, not months, and introduced DGX Pod as a way to package compute and simplify the path to implementation.