إنفيديا تكشف في GTC تايبيه 2026 عن منظومة متكاملة لوكلاء الذكاء الاصطناعي المستقلين في المؤسسات

Complementing the orchestration layer is OpenShell, a secure, isolated runtime environment that enforces privacy controls for hosting and executing agents on any dedicated platform—whether an RTX workstation, a cloud VM, or the new DGX Station . Together, NemoClaw and OpenShell create a sandboxed operating environment for digital coworkers that can run indefinitely without compromising sensitive data.

Nemotron 3 Ultra: The Reasoning Engine for Agents

Fueling the intelligence of those agents is the Nemotron 3 Ultra, a 550-billion-parameter open-source model that Nvidia claims is the most powerful open model ever released by a U.S. company . It sits at the pinnacle of the Nemotron 3 family, above the previously released 4B-parameter Nano and 120B-parameter Super models .

According to Nvidia’s official documentation and white paper, the model uses a hybrid Mamba-Transformer mixture-of-experts (MoE) architecture . This design enables up to ~55 billion active parameters per token while keeping inference computationally efficient . Nvidia reports that this translates to 5x faster inference and up to 30% lower cost for complex agentic workflows compared to its predecessors .

Crucially, Nemotron 3 Ultra features a native 1-million-token context window, which Nvidia’s developer blog states is key for sustained reasoning across vast codebases, deep multi-document research, and long-running agent memory without the typical fragmentation caused by chunking heuristics . The model is set for an open release on June 4, 2026 .

DGX Station for Windows: The Trillion-Parameter Desktop

Nvidia completed the stack from silicon to software with the unveiling of the DGX Station for Windows. Billed as the world’s most powerful deskside AI supercomputer, the system is engineered to bring frontier-scale AI development directly to enterprise desks, running Windows natively .

Powered by the new NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip, the DGX Station pairs a Blackwell Ultra GPU with a 72-core Grace CPU through a high-speed NVLink-C2C interconnect, creating a unified and coherent memory pool of up to 748 GB . The result is a system that delivers up to 20 petaflops of FP4 AI compute performance and can run frontier AI models of up to 1 trillion parameters entirely locally, or support hundreds of parallel AI agents simultaneously .

The move is a direct play to shift enterprise AI development from shared cloud instances to dedicated, always-on local hardware. Nvidia announced that the DGX Station for Windows will be available in Q4 2026 through an ecosystem of OEM system builders .

Enterprise Adoption: The Partnership Landscape

Nvidia didn’t just launch products; it launched them with a pre-committed coalition of enterprise partners ready to embed the stack into their core products .

Engineering and EDA software leaders are among the first movers. Cadence, Dassault Systèmes, Siemens, and Synopsys announced plans to use the NemoClaw framework to build autonomous AI engineers. These digital coworkers are designed to execute complex simulation and verification workflows, a task Nvidia claims can compress weeks of engineering work into hours .

In the cybersecurity and data analytics space, CrowdStrike and Palantir are integrating the agent platform to run long-lasting autonomous agents powered by the Nemotron open models, aiming to let security and operations teams analyze data dramatically faster .

On the hardware and platform side, Microsoft is collaborating to deliver a native Windows experience for personal and enterprise agents that seamlessly connects to Windows applications . System builders Dell, HP, and ASUS are all named as OEM partners for the DGX Station deskside form factor, ensuring the supercomputer will be widely available through standard enterprise IT procurement channels .

With these moves at GTC Taipei, Nvidia has positioned itself not just as a component supplier for the agentic AI era, but as the architect of a complete, open, and enterprise-ready stack—from the model weights to the workstation on the office floor.