Luisa Crawford
Sep 15, 2026 19:21
NVIDIA unveiled AI factory breakthroughs at the AI Infra Summit, highlighting Vera Rubin’s role in optimizing energy efficiency and token output.
NVIDIA unveiled significant advancements in AI infrastructure at the AI Infra Summit, held at the Santa Clara Convention Center on September 15, 2026. With over 8,000 attendees—more than double last year’s event—the company highlighted the Vera Rubin platform and DSX MaxLPS technology as cornerstones of energy-efficient AI factory operations.
Ian Buck, NVIDIA’s VP of hyperscale and high-performance computing, emphasized the growing importance of optimizing “tokens per watt,” a new metric for evaluating AI infrastructure efficiency. As agentic AI workloads push demand for scalability and performance, NVIDIA’s full-stack approach—from silicon to software—aims to make AI factories more productive and economically viable.
Key Announcements: Power Optimization and Performance Gains
Among the announcements was the NVIDIA DSX MaxLPS, a power optimization system that dynamically reallocates energy across GPUs and racks. Lambda, an AI cloud provider, showcased results highlighting a 23% improvement in performance per watt. By running 19 nodes within the energy footprint of 16, Lambda achieved a 24% increase in token throughput, demonstrating the system’s ability to boost efficiency within existing power constraints.
The Vera Rubin platform itself is designed for agentic AI, where models perform complex tasks like multi-step reasoning and tool calls. NVIDIA’s Groq 3 LPX technology complements this by offering deterministic low-latency inference, particularly effective for long-context models like Qwen 3.8. These innovations reportedly push token throughput per megawatt 35 times higher than prior-generation systems, reducing energy costs for large-scale inference workloads.
AI Factories as Grid-Responsive Systems
An intriguing development involved Emerald AI’s collaboration with Silicon Valley Power. Using NVIDIA’s DSX Flex software, Emerald demonstrated grid-responsive AI loads that autonomously reduce power usage during peak demand without compromising critical workloads. This flexible-load program could enable AI factories to act as dynamic grid resources, supporting energy stability while maintaining productivity.
Why It Matters
NVIDIA’s Vera Rubin platform is a response to the rising operational costs of AI factories, where electricity is often the largest expense. By treating data centers as unified compute units, integrating technologies like NVLink interconnects and Spectrum-X networking, the platform promises significant efficiency gains. According to NVIDIA, Vera Rubin systems can deliver 30x higher throughput per megawatt for agentic workloads compared to older architectures, translating to lower token production costs and higher profitability for operators.
The economic implications are evident. With AI market capitalization exceeding $5 trillion as of September 15, 2026, and NVIDIA securing long-term partnerships—such as its June deal with Safe Superintelligence Inc.—the company is cementing its position as a leader in AI infrastructure. For investors, the focus on energy efficiency and scalability may strengthen NVIDIA’s competitive moat in a rapidly expanding industry.
Looking Ahead
As Vera Rubin systems ramp into full production, NVIDIA aims to accelerate adoption with tools like its DSX AI Factory reference designs and Omniverse DSX digital twin simulations. These resources are designed to help operators optimize power, cooling, and networking configurations, shortening the time to deploy AI factories.
With the shift toward energy-efficient AI infrastructure, NVIDIA appears well-positioned to capitalize on the next wave of demand for agentic AI workloads. Investors and industry stakeholders alike will be watching for further performance metrics and adoption rates as these technologies scale into broader commercial use.
Image source: Shutterstock
Credit: Source link
