What NVIDIA shipped this week was a configuration that makes capacity predictable and a digital twin for designing the facility itself. The product is drifting from the GPU toward the building.
NVIDIA presented DSX POD, a configuration for securing AI capacity faster and more predictably, and published a workflow for building an AI factory digital twin with agent assistance. It also released Cosmos3-H-Dreams, which recreates clinical surgery scenes through generative simulation. Taken together, the announcements lean on how a facility is assembled and used up rather than on per-chip performance.
Raw GPU performance does not determine how much the buyer can actually use. Effective capacity shifts with how power, cooling and networking are arranged, so selling a standardised configuration as "capacity you can forecast" makes the investment case easier on the buying side. The digital twin pulls that forecast forward, before ground is broken. Cosmos3's surgical simulation sits on a different layer, placed here as a concrete answer to what the capacity is for. Note: DSX POD's effective performance rests on published figures, with no third-party verification yet.
Whether operators that adopt DSX POD publish measured utilisation. Whether anyone reports how much rework the digital twin removed at the design stage.