Power Bottleneck Solved! NVIDIA Unveils DSX MaxLPS, Boosting Token Output by Up to 40% at Same Power Level

Recently, AI chip giant NVIDIA unveiled DSX MaxLPS, a full-factory-level power optimization system, at the 2026 AI Infra Summit. According to Ian Buck, Vice President of Large-Scale and High-Performance Computing at the company, DSX MaxLPS can increase the number of verifiable agent tokens produced per megawatt by up to 1.4 times without adding new power infrastructure.

Buck pointed out that AI infrastructure evaluation has shifted from peak computing power to 'tokens per megawatt,' and AI factories must engage in co-design from chip to grid.

The core of MaxLPS is dynamic reallocation of GPU and rack power, distributing the total power budget across more nodes. Instead of running a few nodes at full load, it operates more nodes at moderate load—keeping total power consumption unchanged while increasing throughput.

AI cloud provider Lambda conducted real-world testing on a 19-node HGX B200 cluster. By running all 19 nodes at 85% power instead of 16 nodes at full load, token throughput increased from approximately 4 million to 5 million per second—a 24% improvement—with per-watt performance rising 23%.

Dave Ward, President of Lambda's Cloud Business, stated that in environments constrained by power and cooling, this 'increase density without adding power' scheduling is more critical than simply stacking more hardware.

MaxLPS is part of the DSX platform, integrating with Dynamo inference software, NeMo, NVLink, Spectrum-X, ConnectX SuperNIC, and BlueField DPU to deliver end-to-end optimization.

NVIDIA claims that in the Vera Rubin NVL72 architecture, MaxLPS can increase GPU capacity per rack by up to 40% and token throughput by up to 35%, all without requiring new power cabling.

NVIDIA also demonstrated DSX Flex in a flexible load project with Emerald AI and Silicon Valley Power. During high grid pressure, the AI factory's power draw automatically drops from 4MW to 3MW—keeping critical workloads running while pausing low-priority tasks—and automatically resumes when grid conditions improve.

For hyperscale customers, MaxLPS transforms the challenge from 'buying more power' to 'reallocating existing power,' potentially easing the tension between data center power connectivity, cooling, and compute delivery. However, broader applicability to heterogeneous clusters and different inference models still requires further third-party benchmark validation.