Central to the Vera Rubin system is a move away from conventional air cooling toward 100% liquid cooling. By utilizing hot water at 45°C to regulate internal temperatures, the platform removes the immense energy burden typically placed on data center cooling systems. This shift allows that saved energy to be redirected into the computational process itself. Furthermore, the physical architecture has been streamlined; what once required two days of complex cable installation can now be completed in just two hours. This reduction in cycle time is a critical step toward making gigawatt-scale AI factories a operational reality rather than a theoretical goal. [01:13]
 |
| The Sixth Generation Interconnect: NVLink and the Power of Photons |
The Sixth Generation Interconnect: NVLink and the Power of Photons
A defining feature of this new era is the sixth-generation NVLink, a scale-up switching system that moves beyond the limitations of standard Ethernet or Infiniband. This technology allows for the tight coupling of 144 GPUs within a single NVLink domain, essentially turning a massive rack into one giant, unified computer. The engineering behind this is incredibly rigorous, involving a transition from electrons to photons through co-packaged optics (CPO).
In this system, optical interfaces are integrated directly onto the silicon. Electrons are translated into photons at the chip level, allowing data to move at the speed of light with minimal latency. This process, developed in collaboration with TSMC, represents a frontier in manufacturing technology. By bypassing the physical limits of copper cabling in high-demand zones, the Vera Rubin system achieves a throughput density that was previously unthinkable. This "optical scale-up" is the secret to maintaining the high bandwidth necessary for the next generation of massive language models and complex agentic systems. [02:42]
Disaggregated Inference: The Synergy of Throughput and Latency
One of the most innovative breakthroughs in the Vera Rubin platform is the concept of disaggregated inference. In traditional systems, high throughput and low latency are often at odds; optimizing for one typically sacrifices the other. However, by integrating specialized processors like Grock chips alongside the Vera Rubin GPUs, the system can split the workload. The "prefill" and complex attention mechanisms - which require massive memory and mathematical operations - are handled by the Vera Rubin chips. Meanwhile, the "decode" and token generation - which are bandwidth-limited and demand low latency - are offloaded to specialized inference engines.
This vertical integration allows for a 35-fold increase in throughput for premium-tier AI services. It enables a "Dynamo" operating system to manage the pipeline, ensuring that every token is generated at the optimal rate. For a data center operator, this means the ability to distribute power dynamically across different tiers of service, from free consumer-facing models to high-value engineering agents. The result is a system capable of scaling from two million to 700 million tokens per second - a 350-fold increase in output rate within a single gigawatt factory. [15:46]
 |
| Agentic Operating Systems: The Open Claw Revolution |
Agentic Operating Systems: The Open Claw Revolution
Beyond the hardware, the way we interact with AI is being redefined by Open Claw, an open-source system that effectively serves as the operating system for personal agents. Much like Windows enabled the era of personal computing, Open Claw provides the framework for agentic computers that can access tools, manage file systems, and decompose complex prompts into executable steps. It is a multi-modal system that understands voice, gestures, and text, allowing for a seamless integration of AI into daily workflows.
However, the rise of agentic systems brings significant security challenges. These agents often require access to sensitive corporate data, the ability to execute code, and the capacity to communicate externally. To address this, the development of enterprise-ready versions of these systems has become a priority. By integrating security toolkits and private "shells," organizations can deploy custom models - such as Neotron for language, Cosmos for world foundation, and Groot for robotics - within a secure, governed environment. This horizontal openness, combined with vertical technical integration, ensures that the AI revolution is accessible to everyone while maintaining the integrity of the data that powers it. [26:47]
The Vera Rubin era is not just about faster tokens; it is about building a scalable foundation for a world where intelligence is a utility. By optimizing across the entire stack - from the physics of light to the logic of agentic software - we are witnessing the birth of a new industrial paradigm. [29:59]
 |
| The Architecture of Infinite Intelligence: Unveiling the Vera Rubin Era |