At Hot Chips 2026, Nvidia highlighted its DSX MaxLPS site power management architecture for the Rubin GPU, demonstrating how it extracts additional compute density within fixed data center power envelopes.

While individual CPU, GPU, and chip performance often dominate discussions around rack-scale AI systems, the ultimate constraint remains the amount of power that can be delivered to a facility and distributed across its racks. Managing and allocating that power efficiently is now a primary concern for maximizing productivity in next-generation data center installations. During Nvidia’s Hot Chips 2026 presentation on the Rubin GPU, the company emphasized this hard capacity limit and highlighted how much compute Vera Rubin NVL72 systems can deliver within a fixed facility power budget of 100 MW.

According to Nvidia, combining Vera Rubin’s internal power management technologies with its DSX MaxLPS (Land, Power, Shell) suite of design and site-level dynamic power management tools will allow operators to provision approximately 40,000 next-generation GPUs—or roughly 40 Rubin DGX SuperPODs—within that 100 MW envelope. Nvidia expects this hardware to deliver up to 2 zettaFLOPS (ZFLOPS) for NVFP4 inference and up to 1.4 ZFLOPS for NVFP4 training. Based on publicly available Rubin specifications, these performance figures should be treated as estimates rather than measured benchmarks; real-world workloads will likely achieve lower FLOPS due to various operational factors. Nevertheless, the underlying premise holds: extracting maximum compute from constrained power budgets will require far more refined planning, monitoring, and facility management than simply applying coarse, estimated peak power draws to every electrical component.

Historically, data center operators have relied on static power provisioning, basing assumptions on fixed, worst-case power peaks per rack. This approach frequently inflates power budgets and strands allocated capacity in racks that rarely, if ever, reach their theoretical limits. Under a static scheme, power cannot be re-routed to address application load differentials between racks. Consequently, underutilized racks leave power idle while heavily loaded systems still hit conservative guard bands, capping cluster-wide performance.

The DSX MaxLPS approach addresses these limitations by implementing an intelligent, dynamic power allocation scheme that continuously monitors usage at the chip, rack, and group-of-racks levels. Nvidia’s Dynamic Power Software control loop identifies unused capacity resulting from workload characteristics or idle resources and redistributes it to systems requiring it most at any given moment. This maximizes both the number of installable systems and overall performance per watt over the facility’s lifecycle.

In a measured example using current GB300 racks, Nvidia demonstrated the tangible impact of this shift. Within a 540 kW power budget and under static provisioning, an operator could typically install four 135 kW systems by provisioning for peak draw rather than actual workload measurements. In practice, however, up to 170 kW of that budget often sits unused due to uneven rack utilization. With modern rack-scale architectures like Nvidia’s NVL72, that reclaimed capacity safely accommodates an additional system. Across five dynamically provisioned systems, the reserve power allocated for peak headroom can be significantly reduced.

At the rack level, DSX MaxLPS introduces further flexibility through workload-specific power profiles. Similar to quiet, balanced, and high-performance modes on client PCs, Nvidia has developed rack-level profiles tailored to specific tasks such as inference, training, and general memory-bound or compute-bound workloads. Although Nvidia did not publish measured Rubin power or performance-per-watt results, it validated the MaxLPS concept using prior-generation hardware. For a Grace Blackwell GB300 system running DeepSeek-R1, the traditional fixed-peak regime assumed a 1,400 W GPU TGP and an estimated rack power draw of 136 kW. Applying MaxLPS reduced the typical GPU TGP to 1,000 W and total rack power to 101 kW without impacting delivered performance. This tighter operating envelope directly translates to higher performance per watt, greater rack density within the same footprint, and increased token output for data center operators and tenants.

Designing facilities with MaxLPS from the outset also provides long-term operational flexibility. A site initially deployed as a training-focused facility with cutting-edge hardware will naturally demand a larger share of site power per rack. Operators may therefore choose not to populate every available floor space immediately. As training workloads migrate to newer hardware generations and older systems transition to inference roles, per-GPU and per-rack power demands decrease. Facilities equipped with dynamic power provisioning can then leverage freed capacity to deploy additional hardware within the same footprint, driving higher revenue-generating throughput.

Another critical component of MaxLPS for data centers deploying Vera Rubin hardware is the adoption of higher liquid coolant temperatures. The exclusively liquid-cooled Rubin NVL72 racks are engineered to operate with 45 °C inlet coolant temperatures, a significant increase over previous liquid-cooled generations. This “dry cooling” strategy matters because mechanical chillers used to reject waste heat in non-evaporative systems traditionally consume a substantial portion of a site’s power budget—up to 40% in past installations, according to Nvidia. Much like server power provisioning, chillers have historically been oversized for worst-case thermal scenarios, even when operating well below capacity for most of the year. Stranding this capacity prevents dynamic reallocation to compute workloads during favorable conditions. While chillers remain necessary during peak ambient temperatures, the higher coolant temperature generally enables more site power to be directed toward productive computation, improving the facility’s power usage effectiveness (PUE) under standard operating conditions.

Power availability for AI data centers—whether supplied by public utilities or generated behind the meter—will remain one of the most critical constraints for the foreseeable future, a concern echoed by multiple presenters at Hot Chips. Nvidia’s DSX MaxLPS framework delivers the essential building blocks for dynamically allocating this scarce resource, enabling operators to extract maximum performance per watt from Rubin-based facilities. As power consumption continues to scale, the interplay between facility power management and achievable compute performance will only grow in strategic importance.