NVIDIA argues that the Vera Rubin platform and related tech (NVLink, Spectrum-X) can deliver up to 30x higher throughput per megawatt for agentic AI workloads, addressing rising energy costs in AI factories.

At its DSX AI factory platform launch, NVIDIA unveiled DSX as an integrated stack—combining libraries, APIs, reference designs, computing platforms, and partner technology to design, deploy, and operate AI factories, with components including DSX Reference Design, DSX Sim, DSX OS, DSX MaxLPS, DSX Flex, and DSX Exchange.

Lambda validated DSX MaxLPS on NVIDIA HGX B200 servers, achieving 5 million tokens per second cluster-wide and a 23% gain in performance per watt by reallocating headroom across nodes to maximize compute density within the same power budget.

DSX MaxLPS is a dynamic power-optimization system that reallocates energy across GPUs and racks, delivering a 23% improvement in performance per watt and a 24% increase in token throughput within a 16-node energy footprint.

NVIDIA highlighted Vera Rubin and DSX MaxLPS as core technologies for energy-efficient AI factories during the AI Infra Summit, drawing more than 8,000 attendees.

Emerald AI and Silicon Valley Power demonstrated grid-responsive DSX Flex workloads that autonomously cut power during peak demand while preserving critical AI workloads, enabling AI factories to function as dynamic grid resources.

Emerald AI’s Conductor has been deployed at Eos under Silicon Valley Power’s Flexible Load Interconnect Program, treating AI factories as dispatchable resources with over 200 demand signals received to date.

A notable demand-response milestone occurred at NVIDIA’s Eos AI factory, where SVP’s Conductor trimmed power from four to three megawatts by prioritizing higher-priority workloads, marking the first commercial DSX Flex-powered pilot.

NVIDIA plans a 96-megawatt Vera Rubin AI factory at its Manassas Center as the first dedicated DSX Flex commercial deployment, in partnership with Digital Realty, EPRI, and PJM Interconnection, expected later in 2026.

Looking forward, NVIDIA aims to broaden DSX adoption with AI Factory reference designs and Omniverse DSX digital twins to optimize power, cooling, and networking and accelerate deployment.

The AI market environment supports NVIDIA’s leadership, with a market capitalization exceeding $5 trillion as of mid-September 2026 and strategic partnerships, including a June deal with Safe Superintelligence Inc.

Early production results for DSX released on September 15, 2026, include Lambda’s 24% higher token throughput within a fixed power budget and a DSX MaxLPS deployment across Lambda’s five-rack, 19-node cluster showcased at the AI Infra Summit.