NVIDIA DSX MaxLPS is a combined chip, thermal, system and software approach intended to maximize AI throughput within a fixed site power envelope. MaxLPS — short for Maximum Land Power Shell — refers to the site-level constraints that define an AI factory: the land, utility power, and the physical shell that houses power distribution, cooling, networking and compute infrastructure.
Why performance-per-watt matters for AI factories
At large AI deployments the central question is no longer how many GPUs can fit in a data center, but how much AI output can be produced per available megawatt. In practice, a sizable portion of incoming site power does not reach compute: facility overhead, power distribution losses, cooling, networking, storage and backup consume part of the budget. NVIDIA cites a representative example where about 60% of delivered site power is available for compute that produces AI output.
Static rack provisioning worsens this: traditional planning reserves each rack’s maximum power as if every rack could draw peak simultaneously, which treats each rack as an isolated power island and strands unused headroom.
For example, in an illustrative 100 MW AI factory power-budget waterfall, 20 MW goes to facility overhead, 10 MW to rack losses and another 10 MW is unavailable due to operational inefficiencies during failures, restarts and checkpointing — leaving 60 MW for AI load.
How MaxLPS reclaims stranded power with dynamic allocation
DSX MaxLPS replaces static provisioning with dynamic power allocation using Dynamic Power Software (DPS), currently in Developer Preview. DPS models the data center topology from the utility level down to racks, nodes and GPUs. Operators define resource groups, power budgets and policies; within those boundaries DPS continuously compares allocated power to actual consumption.
When GPUs or racks operate below their reserved level, DPS makes that headroom available to others within the managed group while the site-level power envelope remains unchanged. DPS runs a continuous control loop: it collects GPU-, rack- and group-level power telemetry, identifies unused capacity, reallocates power subject to policy, validates compliance with group budgets and reacts to power events and emergency policies.
The benefit derives from coordinating many power decisions across the fleet rather than treating each rack as an independent island. When site budgets change or grid/maintenance events alter available power, DPS can adapt operating limits across the data center without manual re-planning of every rack.
DSX Exchange, an open-source event bus (Developer Preview), can connect DPS and other DSX services to building management, electrical monitoring, cooling infrastructure, grid interfaces and compute schedulers. DSX Exchange is optional for MaxLPS, but exposing additional facility and seasonal signals can enable further optimizations.
Software techniques to raise performance per watt
Beyond rack-level steering, MaxLPS includes software features that improve how each GPU uses its allocated power. The suite provides workload profile power solutions (WPPS) for common modes — inference, training, memory-bound and compute-bound — so operators apply validated profiles rather than tuning node behavior per job. The Application Performance and Power Manager (APPM) applies these configurations, and tools like NVIDIA Dynamo can further tune inter-rack performance and power for inference services.
For inference workloads, maximizing tokens per second per watt directly translates into measurable throughput gains. NVIDIA reports that MaxLPS improves performance-per-watt roughly 1.5× on GB200 NVL72 and about 1.3–1.4× on Vera Rubin NVL72 in representative tests.
Measured rack and capacity impacts
NVIDIA validation showed provisioned rack power reductions: GB200 NVL72 decreased from 125 kW to 90 kW, and Vera Rubin NVL72 from 136 kW to 101 kW. Within the same 540 kW site envelope these reductions enabled approximately 39% and 35% more racks respectively, while preserving workload throughput. NVIDIA projects that MaxLPS combined with data center power planning can enable up to 40% more Rubin GPU capacity within the same power budget for NVIDIA Vera Rubin NVL72 sites.
Sizing an AI factory for MaxLPS
MaxLPS site design starts with a fixed gross facility power envelope and derives the maximum number of GPU rack positions supportable at the MaxLPS operating point. Day 1 deployment can be smaller — a site that begins with training-heavy workloads may run at higher average rack power and therefore populate fewer racks initially. As workload mix shifts toward inference over the hardware lifecycle, average rack power can decline toward the MaxLPS operating point and free capacity to populate additional rack positions without increasing the facility envelope.
Sizing decisions include choosing the GPU product family, setting a PUE target for 45 °C DLC inlet operation, sizing east–west network capacity for the MaxLPS GPU count, and deriving the total deployable rack-position count. With space, power distribution, cooling and network capacity configured up front for the full MaxLPS capacity target, operators can add racks incrementally as workloads evolve.
Why 45 °C liquid cooling matters
MaxLPS’s third pillar is thermal and facility design: Vera Rubin NVL72 racks are designed for 45 °C liquid-cooling inlet operation. Higher coolant temperatures can sound counterintuitive, but they enable more frequent use of free cooling (outside air or water) and reduce demand on energy-intensive chillers or water-hungry adiabatic coolers. Chillers remain necessary under hot conditions and for resilience, but using them less often lowers average annual PUE and frees electrical power for compute.
The technology cooling system (TCS) loop moves heat from GPU cold plates to the coolant distribution unit (CDU), which hands it to the facility cooling loop for rejection by chillers, dry coolers or towers. Efficient operation balances matching coolant flow to GPU thermal demand, staying within inlet-temperature and reliability limits, and responding to workload-driven thermal events without instability.
Many facilities use PID control for CDU behavior, which is reactive and often forces conservative, colder operation to preserve margin. That conservatism wastes cooling power. Predictive or learned control (agentic control) can anticipate thermal behavior and reduce buffer needs, although rack hardware and facility design to support 45 °C inlet operation are primary enablers.
Getting started and validation
Physical infrastructure is costly to change later, so NVIDIA recommends designing or retrofitting sites for DSX MaxLPS at the site level even if not every rack is populated on Day 1. Validate power topology, redundancy, cooling behavior, telemetry availability, policy requirements, network capacity and rack-position optionality before fixing infrastructure decisions.
For software evaluation, the NVIDIA Dynamic Power Software documentation and the NVL72 inference power pilot provide starting points. A scoped validation compares an unmanaged static baseline against a MaxLPS-managed run using representative workloads and tracks throughput, latency, error rate, power draw, utilization and policy compliance.
Facilities teams should engage NVIDIA early to validate 45 °C thermal design, MaxLPS rack-position optionality and power distribution flexibility. Software testing can proceed in parallel or afterward.
Conclusion
Static rack provisioning can strand usable power that real AI workloads could convert into additional compute output. NVIDIA DSX MaxLPS addresses this with dynamic power-allocation software that unlocks stranded rack capacity, performance-per-watt optimizations that raise output per watt, and 45 °C thermal and facility design that reduces cooling overhead. For operators planning Vera Rubin NVL72 deployments, MaxLPS is presented as a path to prepare sites to support up to about 40% more Rubin GPU capacity within fixed power budgets.
