NVIDIA's new capacity math: 60 of every 100 megawatts actually reaches the AI load

NVIDIA has published a design pitch for what it calls DSX MaxLPS, and the framing matters more than the product: the company argues the capacity question is no longer how many GPUs fit in a data center but how much AI output each available megawatt delivers ^[1]. Buried in the post is the number operators should be modelling against, which is that in one representative power-budget view examined by NVIDIA, roughly 60% of delivered site power is allocated to compute for AI output ^[4].

The arithmetic is laid out explicitly for a 100 MW site. Of 100 MW of grid input, 20 MW goes to facility overhead, 10 MW to rack losses, and 10 MW is unavailable to the AI load because of operational inefficiency during failures, restarts and checkpointing, leaving 60 MW ^[5]. That is 40 MW of every 100 MW that never becomes tokens ^[6]. Read as a procurement rule, adding 10 MW of AI load at those same ratios means buying about 16.7 MW of grid input ^[7]. Facility overhead alone is twice the size of either the rack-loss or the restart-inefficiency slice ^[18].

Power distribution, cooling, networking, storage, backup and facility systems all take their share before electricity reaches a GPU ^[3]. On top of that, NVIDIA singles out static rack provisioning: the practice of allocating each rack its maximum possible draw to cover worst-case peaks, with further capacity reserved for failures, operational flexibility and expansion ^[8]. The consequence is that each rack becomes an isolated power island, and a rack sitting on unused reserve cannot lend it to a neighbour that would convert it into output ^[9]. That is not a rounding error, because training, post-training and inference cycle through compute bursts, memory-bound execution, synchronization, checkpointing, prefill, decode, idle gaps and network-bound communication, each drawing power differently, so a rack provisioned for its peak runs below that level for meaningful periods ^[10].

MaxLPS, which NVIDIA expands as Maximum Land Power Shell, is described as a suite of chip, thermal, system and software technologies aimed at maximising throughput inside a fixed power budget ^[11]. It works on three layers: dynamic power allocation, software performance-per-watt techniques, and 45C warm-water liquid cooling with site design changes intended to improve PUE ^[12]. The allocation piece is Dynamic Power Software, currently in Developer Preview, which models the site topology from utility down to racks, nodes and GPUs while operators set resource groups, budgets and policies ^[13]. It compares allocated against actual consumption and hands headroom from underused GPUs and racks to others in the same managed group, leaving the site envelope unchanged ^[14], running a continuous loop of telemetry collection, reallocation within policy, and validation against the approved group budget ^[15].

Note where that acts. NVIDIA presents the stranded rack headroom as a separate problem from the facility overhead, rack losses and restart inefficiency that shrink the AI load in the first place ^[16], which means dynamic allocation is redistributing inside the 60 MW rather than clawing back the 40 MW ^[19]. Only the thermal layer is described as attacking cooling overhead directly ^[12]. And the post does not publish a recovered-megawatt or percentage throughput figure for any of the three layers ^[20].

For inference specifically, NVIDIA nominates application-level performance per watt as the efficiency metric ^[2]. That is the number to demand in any vendor comparison, because it is the only one that survives contact with a fixed shell.