NVIDIA presented new deployment results for its DSX AI infrastructure platform at the AI Infra Summit, including Lambda’s validation of higher token throughput under a fixed power budget and an Emerald AI demand-response deployment with Silicon Valley Power. Lambda reported 24% more cluster-wide token throughput and 23% better performance per watt using DSX MaxLPS on NVIDIA HGX B200 GPU servers.
Lambda tested the software on a five-rack, 19-node cluster, running 19 nodes within the same power budget as 16 nodes at full power. Token throughput increased from roughly 4 million to 5 million tokens per second. DSX MaxLPS monitors GPU- and rack-level consumption and reallocates available power across nodes according to workload type, with the stated aim of recovering capacity left unused by static provisioning.
NVIDIA also described a commercial-scale grid-flexibility deployment at its Eos AI factory in Santa Clara. Emerald AI’s Conductor platform receives signals from Silicon Valley Power and changes a predefined workload hierarchy: lower-priority work is slowed or rescheduled while high-priority inference continues. During an August event, power consumption fell automatically from four megawatts to three, and NVIDIA said Silicon Valley Power has since sent more than 200 demand signals to the facility, with Conductor responding successfully each time.
The Santa Clara deployment is not a DSX Flex installation, according to NVIDIA, but an earlier demonstration of the operating model that DSX Flex is intended to support. Eos participates in Silicon Valley Power’s Flexible Load Interconnect Program, which NVIDIA described as the first commercial utility program designed to treat AI factories as dispatchable resources. Conductor responds to utility signals in under a minute and is expected to integrate with DSX Flex as the platform develops.
NVIDIA said the first dedicated commercial DSX Flex deployment will be a 96-megawatt Vera Rubin AI factory at its AI Factory Research Center in Manassas, Virginia, building on five demonstrations across two continents. The broader DSX suite also includes DSX OS for lifecycle management and resiliency, DSX Sim for predeployment modeling, and generation-specific reference designs covering compute, networking, storage and facilities.
The company is also incorporating an 800V DC power architecture into DSX reference designs to reduce conversion complexity and support denser accelerated-computing racks. NVIDIA projects a 3%-5% end-to-end efficiency gain compared with 54V distribution when the architecture becomes available with Vera Rubin NVL72 in 2027. It further projects that DSX MaxLPS, combined with data center power planning, could enable up to 40% more GPU capacity for suitable next-generation Vera Rubin NVL72 deployments within the same megawatt budget; those figures are projections, not reported deployment results.