Jensen Huang's Biggest AI Announcements at NVIDIA GTC 2026 (Supercut)

Key Concepts

• Vera Rubin Platform: Nvidia’s next-generation AI supercomputing architecture.
• NVLink (6th Gen): High-speed, liquid-cooled interconnect technology for GPU communication.
• Disaggregated Inference: A methodology separating "prefill" (math-heavy) and "decode" (bandwidth-heavy) tasks across different processors.
• CPO (Co-Packaged Optics): Integrating optical interfaces directly onto silicon chips to translate electrons to photons.
• Open Claw: An open-source agentic AI operating system/framework for managing LLMs, tools, and scheduling.
• AI Factories: Large-scale, optimized data centers designed for high-throughput, low-latency AI workloads.
• Omniverse DSX: A virtual simulation platform for designing and optimizing data center infrastructure.

1. The Vera Rubin Platform & Hardware Evolution

Nvidia has transitioned from a chip manufacturer to an "AI factory" architect. The Vera Rubin system represents a massive leap in compute density, offering 40 million times more compute power than systems from a decade ago.

• Design Efficiency: The system is 100% liquid-cooled, utilizing 45°C hot water to eliminate traditional data center cooling costs. Installation time has been reduced from two days to two hours.
• Vera Rubin System: Features eight Groq chips (LP30) and the world’s first CPO Spectrum X switch.
• Reuben Ultra & Kyber: A new compute node architecture that slides vertically into the "Kyber" rack, enabling 144 GPUs in a single NVLink domain, bypassing the distance limitations of copper cabling.

2. Performance Scaling and Throughput

Jensen Huang emphasized the trade-off between throughput (high-volume processing) and latency (real-time interactivity).

• Grace Blackwell vs. Vera Rubin: Vera Rubin provides a 5x revenue generation increase over Blackwell.
• The "Enemies" Problem: High throughput and low latency are traditionally conflicting goals. Nvidia solves this by integrating Groq chips for low-latency "decode" tasks while using Vera Rubin for high-throughput "prefill" tasks.
• Disaggregated Inference: Using the "Dynamo" operating system, Nvidia offloads the feed-forward network (token generation) to Groq chips, while keeping the attention-heavy math on Vera Rubin. This results in a 35x performance increase for the most valuable AI workloads.

3. Roadmap: Oberon, Fineman, and Beyond

Nvidia maintains a yearly cadence for new architectures:

• Oberon: Supports copper scale-up and optical scale-out, allowing for NVLink 576 configurations.
• Fineman (Next Gen): Will feature the LP40 LPU, the "Rosa" CPU, and Bluefield 5. It will support both copper and co-packaged optics (CPO) scale-up simultaneously.
• LP35 Chip: Will introduce Nvidia’s MVFP4 computing structure for further speed gains.

4. Agentic AI and Open Claw

The presentation highlighted the shift toward "Personal Agents" via Open Claw.

• Definition: Open Claw is an operating system for agentic computers that manages resources, tools, file systems, and scheduling. It decomposes complex prompts into step-by-step tasks.
• Enterprise Security: To address the risks of agents accessing sensitive corporate data, Nvidia introduced Open Neoclaw, which integrates "Open Shell" technology to ensure enterprise-grade security and privacy.
• Strategic Imperative: Huang argued that every company now requires an "Open Claw strategy," similar to the historical necessity of having Linux or HTML strategies.

5. Ecosystem and Simulation

• Omniverse DSX: A platform for virtualizing the design of "gigawatt AI factories." It simulates mechanical, thermal, and electrical systems to ensure no power is squandered.
• Max Q: A dynamic power management system that adjusts cooling and electrical loads in real-time to maximize token throughput.
• Space Initiatives: Nvidia is developing "Vera Rubin Space One" for satellite-based data centers, overcoming the lack of convection/conduction in space by utilizing radiation-based cooling designs.

6. Synthesis and Conclusion

The core takeaway is Nvidia’s move toward vertical integration and horizontal openness. By combining extreme hardware co-design (NVLink, CPO, liquid cooling) with software-defined AI factories (Dynamo, Open Claw), Nvidia is enabling a 350x increase in token generation rates over two years. The strategy is clear: provide the hardware for high-throughput AI factories while creating the software ecosystem (Open Claw) that allows businesses to deploy secure, agentic AI at scale. As Huang noted, "We are at the frontier on every single domain of AI models," and the company intends to maintain this lead through continuous, yearly architectural innovation.