Grid giants sign on to Nvidia’s DSX as power, not chips, gates AI data centers

Nvidia says power has become the single biggest bottleneck in AI infrastructure, pointing to more than $300 billion in backlogged equipment and over 200 gigawatts of projects waiting in U.S. interconnection queues, the waiting lists for new power projects and large loads to connect to the grid. Its answer is DSX, a platform meant to plan, simulate and operate the grid connection, the building and the racks as one system. Components named include DSX Sim for simulation, DSX Flex for syncing demand with grid conditions, and DSX MaxLPS and DSX Exchange for efficiency and data sharing.

GE Vernova, Hitachi Energy, Schneider Electric with AVEVA and ETAP, and Siemens Energy all announced work aligned with it, and Hitachi Energy’s battery storage qualified as ‘DSX Ready.’ It extends Nvidia’s push to make AI data centers arrive like manufactured goods (DSX POD) and to squeeze more output from stranded rack power (DSX MaxLPS).

The core idea is to treat an AI facility as a load that can flex with grid conditions rather than a fixed block of demand served at all times. That could bring capacity online without waiting years for new transmission, and simulating generators, cooling and switchgear together before construction is sound practice. There is early evidence: Emerald AI’s Conductor has taken Silicon Valley Power demand signals across thousands of Nvidia GPUs, with more than 200 events where power dropped on cue (megawatts to tokens).

The headline figures differ in quality. Phaidra reports about 10% more compute inside the same power envelope, a specific and meaningful claim. Emerald AI’s figure of up to 100 gigawatts unlocked on the existing U.S. grid is described as modeled, a ceiling from a model rather than delivered capacity. Google, Nvidia and Anthropic have backed Emerald AI’s demand-response coalition on that premise (Emerald AI coalition).

‘Tokens per watt’ is the data center’s metric. Utilities and ratepayers care about reliability and who pays for grid upgrades. If flexible AI facilities genuinely shed load during peaks, everyone benefits; if flexibility mainly helps projects move up interconnection queues while households absorb grid costs, the politics could turn quickly. Utilities will want proof that flexibility appears when the grid actually needs it, and regulators will face questions about whether to credit flexible load in interconnection decisions.

There is also a concentration point. When transformer and switchgear makers design to a chip company’s reference architecture, here tied to the Vera Rubin generation, one vendor’s roadmap begins to shape how substations and storage are specified. That can speed standardization, and Nvidia has opened its Vera Rubin rack designs through the Open Compute Project (Vera Rubin at OCP), but it deserves scrutiny. More in our Nvidia coverage and AI tools coverage.