NVIDIA has introduced DSX MaxLPS, a full-stack suite of chip, thermal, system, and software technologies designed to maximize AI factory throughput within a fixed power budget. The platform targets the three constraints that define an AI factory — land, utility power, and the physical shell holding power, cooling, networking, and compute — by combining dynamic power allocation, workload-aware performance-per-watt profiles, and 45°C liquid-cooling site design.

Traditional data center power planning reserves enough capacity as though every rack could draw its specified maximum simultaneously. That protects the facility against peak demand but treats each rack as an isolated power island. In one representative 100 MW AI factory model examined by NVIDIA, only about 60% of delivered site power reaches the AI load; the rest is consumed by facility overhead, rack losses, and operational inefficiencies during failures, restarts, and checkpointing. Static provisioning strands additional headroom inside the rack allocation itself — power reserved for one rack's peak may sit unused while a neighbor could turn it into tokens.

What's new

At the center of MaxLPS is Dynamic Power Software (DPS), currently in Developer Preview. DPS models the data center topology from the utility level down to individual GPUs. Operators define resource groups, power budgets, and policies; DPS then runs a continuous control loop that collects GPU-, rack-, and group-level telemetry, identifies unused capacity, and reallocates it within policy — all while keeping the site power envelope unchanged. When grid, maintenance, or emergency events alter available power, DPS adapts operating limits across the fleet without manual re-planning.

DSX Exchange, an open-source event bus also in Developer Preview, connects DPS and other DSX services to building management systems, electrical monitoring, cooling infrastructure, grid interfaces, and compute schedulers. MaxLPS does not require DSX Exchange to function, but the integration exposes signals such as seasonal cooling headroom and facility power events that DPS can act on.

Beyond rack-level steering, MaxLPS includes Workload Profile Power Solutions (WPPS) for common operating modes — inference, training, memory-bound, and compute-bound. The Application Performance and Power Manager (APPM) applies the selected profile to participating GPUs, while software such as NVIDIA Dynamo further optimizes inter-rack performance and power behavior for inference services.

Why it matters

On representative inference workloads, MaxLPS reduces provisioned rack power from 125 kW to 90 kW on GB200 NVL72 and from 136 kW to 101 kW on Vera Rubin NVL72. That enables 39% and 35% more racks respectively within the same power envelope while preserving workload throughput. Performance per watt improves approximately 1.5x on GB200 NVL72 and 1.3–1.4x on Vera Rubin NVL72. For Vera Rubin NVL72 AI factories, NVIDIA projects that MaxLPS combined with data center power planning can enable up to 40% more Rubin GPU capacity within the same power budget.

MaxLPS sizing starts with a fixed gross facility power envelope and works inward to determine the maximum number of GPU rack positions the site can support at the MaxLPS operating point. Day-one deployment can be lower than this infrastructure limit; as the workload mix shifts toward inference over the hardware lifecycle, average rack power can decline toward the MaxLPS average operating point, creating headroom to populate additional rack positions without a later facility retrofit.

Our take

MaxLPS reframes the AI factory constraint from "how many GPUs fit" to "how much intelligence per megawatt." The 40% capacity uplift is real but contingent on early site-level engagement for 45°C liquid cooling, optimized infrastructure design, and workload profiling — factors that favor greenfield builds over brownfield retrofits. The software layer (DPS) is the differentiator; the thermal target is the enabler.

Sources