NVIDIA Says Smarter AI Factories Can Produce More Tokens From the Same Power
NVIDIA is pushing a new approach to the biggest constraint facing large-scale artificial intelligence infrastructure: electricity. Instead of simply adding more power to AI data centers, the company is developing software and infrastructure that can squeeze more useful compute from the electricity already available.
At NVIDIA’s AI Infra Summit on September 15, the company highlighted results from its DSX platform, including a Lambda deployment that produced 24% more cluster-wide token throughput within the same power budget. NVIDIA also showcased work with Emerald AI and Silicon Valley Power demonstrating how an AI factory can automatically reduce electricity consumption when the local grid needs relief.
The developments point toward a broader shift in how AI infrastructure is measured: not simply by how many GPUs a facility contains, but by how much useful AI work it can produce for every megawatt it consumes.
Lambda Gets 24% More Token Throughput
One of the clearest demonstrations came from Lambda, a GPU cloud provider that tested NVIDIA DSX MaxLPS on a five-rack cluster containing 19 NVIDIA HGX B200 systems.
According to NVIDIA, the software allowed Lambda to operate 19 nodes within the same power budget previously used for 16 nodes at full power. The result was approximately 24% higher cluster-wide token throughput and a 23% improvement in performance per watt.
For an inference workload, NVIDIA reported that the baseline configuration produced about 4.04 million tokens per second. With 19 nodes operating under an 85% DSX MaxLPS power policy, throughput reached roughly 5 million tokens per second.
The idea is to avoid leaving power capacity stranded.
AI workloads do not consume electricity in perfectly uniform patterns. Training can create bursts of power demand, while inference can have different and steadier requirements. DSX MaxLPS dynamically manages available power at the GPU, rack and workload levels to redistribute that headroom.
That means a facility can potentially run additional compute without increasing its overall power allocation.
AI Data Centers Can Also Respond to the Power Grid
NVIDIA’s strategy goes beyond optimizing power inside the data center.
The company is also working with Emerald AI and Silicon Valley Power on systems that allow AI factories to respond automatically to grid conditions.
In the Santa Clara demonstration, Emerald AI’s Conductor platform receives signals from the utility and adjusts flexible workloads according to predefined priorities. Lower-priority jobs can be reduced or delayed while higher-priority AI workloads continue operating.
NVIDIA said the system can respond to grid signals in under a minute. The technology is designed to handle events such as demand-response requests, load shedding and pricing signals.
This approach effectively turns some AI computing capacity into a controllable resource for the electricity grid.
When electricity demand rises, an AI factory could temporarily reduce flexible computing rather than requiring the grid to supply the facility’s maximum possible load.
NVIDIA and Emerald AI have already demonstrated similar flexibility in multiple commercial data-center trials. NVIDIA’s case study says Emerald AI’s Conductor successfully met more than 200 simulated grid power-reduction targets, with reductions of up to 40% in power demand in under a minute while maintaining critical workloads.
What NVIDIA DSX Actually Does
DSX is broader than a single power-management product.
NVIDIA describes it as a platform for designing, simulating, operating and optimizing AI factories.
Its components include DSX MaxLPS, which focuses on maximizing performance per watt; DSX Flex, which connects AI factories with grid services; DSX OS, which provides operational software; DSX Sim, which helps model infrastructure before construction; and DSX reference designs for validated hardware and facility architectures.
The common objective is to optimize the entire AI factory rather than treating GPUs, networking, cooling and electricity as separate problems.
That matters because adding more GPUs does not automatically translate into proportionally more useful AI output if power, cooling or networking becomes the bottleneck.
NVIDIA Is Preparing for Even Denser AI Racks
NVIDIA is also working on the electrical architecture that will support future generations of increasingly dense AI systems.
The company has proposed an 800V DC power architecture for next-generation AI factories, replacing traditional lower-voltage distribution approaches as rack power requirements move toward the megawatt range.
NVIDIA says the architecture is designed to reduce conversion complexity and improve power-delivery efficiency. Its technical material projects up to a 5% improvement in end-to-end power efficiency and says full-scale 800V DC data-center production is targeted to coincide with Kyber rack-scale systems in 2027.
The significance is straightforward: as AI racks become more power-hungry, the electrical infrastructure itself becomes a major part of computing performance and cost.
The Bigger AI Infrastructure Problem
The race to build AI infrastructure is increasingly constrained by access to electricity.
NVIDIA has argued that improving the productivity of existing power capacity can complement the much slower process of building new generation and transmission infrastructure. Its DSX reference architecture is designed around that concept, with software and hardware working together to improve the amount of productive AI compute obtained from available energy.
That does not eliminate the need for new power plants, transmission lines or grid upgrades.
Instead, it creates another way to increase AI capacity: use existing electricity more intelligently.
For cloud providers, the economic incentive is also direct. More tokens produced from the same power budget can mean more usable computing capacity without proportionally increasing electricity consumption.
What Happens Next
NVIDIA says the next phase will move these technologies from demonstrations and early deployments toward broader commercial use.
The company has identified a planned DSX Flex deployment at its AI Factory Research Center in Manassas, Virginia. NVIDIA says the facility is planned as a 96-megawatt Vera Rubin AI factory and will build on earlier demonstrations of power-flexible AI infrastructure.
The broader direction is clear: AI factories are increasingly being designed as energy-aware computing systems.
The question for the industry is no longer only how many GPUs can fit inside a data center. It is increasingly about how many useful tokens those GPUs can produce from every megawatt — and whether the facility can adjust its demand when the surrounding power grid needs flexibility.
NVIDIA’s DSX platform is an attempt to make that optimization part of the AI factory itself.
