Software Stack

From building systems to model hub · Orchestration and digital twins · Flagship: Nvidia AI Factory Research Center, Manassas

The software that runs an AI data center reaches from building controls to the model hub, and Nvidia now owns a leading product at most layers. Since 2024 it has bought Run:ai and SchedMD, the company behind Slurm, launched Mission Control and Dynamo, and agreed to buy Hugging Face; building systems stay with Schneider, Siemens and Vertiv.

The data center software stack from building systems to model hub, showing which layers Nvidia now owns or controls through Omniverse DSX, Slurm, Run:ai, Mission Control, Dynamo, CUDA and, pending review, Hugging Face, and which stay with Schneider, Siemens, Vertiv, the open-source Kubernetes and PyTorch projects, and the hyperscalers THE SOFTWARE STACK, AND WHO OWNS EACH LAYER Model hubHugging FaceNvidia deal pending FrameworksPyTorch, JAXOpen source Accelerator softwareCUDA; ROCm; Huawei CANNNvidia Inference servingDynamo, vLLM, SGLangNvidia + open source Cluster operationsMission ControlNvidia GPU orchestrationRun:ai; KubernetesNvidia; CNCF Batch schedulingSlurmNvidia (SchedMD) Grid-aware powerDSX Flex; Emerald ConductorNvidia; Emerald AI Facility digital twinOmniverse DSX; Cadence; SiemensNvidia + partners DCIM, BMS, EPMSEcoStruxure, DesigoSchneider, Vertiv, Siemens Purple: Nvidia owns or controls a leading product at this layer. Gray: building-systems vendors and open-source projects.
DCIM market, 2024
~$2.2B
Kubernetes for AI
66% of hosts
Hugging Face deal
~$12.9B, pending
Slurm owner
Nvidia
Hub
Stack

What it is

Software From the Breaker to the Model

The data center software stack runs from building systems that watch breakers and chillers up to the schedulers, serving engines and frameworks that run AI jobs.

  • Facility: building management (BMS), electrical power monitoring (EPMS) and data center infrastructure management (DCIM) come mainly from Schneider Electric, Vertiv and Siemens. The DCIM software market was about $2.2 billion in 2024.
  • Digital twins: simulated copies of a building used to plan power, cooling and racks before they are built, from Nvidia's Omniverse DSX, Cadence and Siemens.
  • Cluster: Slurm schedules batch jobs on most of the top supercomputers; Kubernetes runs containers, and 66% of organizations hosting generative AI use it for inference (CNCF, 2026).
  • Grid: software that slows AI jobs when the grid is stressed. Emerald AI cut a Phoenix cluster's power 25% for three hours with Oracle, Nvidia and Salt River Project.

Consolidation

Nvidia Is Buying the Stack

  • Scheduling: Nvidia bought SchedMD, the developer of Slurm, in December 2025 and pledged to keep it open source and vendor-neutral.
  • Orchestration: Run:ai, about $700 million, closed December 2024; Mission Control, launched March 2025, runs Slurm and Kubernetes and recovers failed jobs automatically.
  • Inference: Dynamo 1.0 entered production in March 2026, used by AWS, Azure, Google Cloud, CoreWeave and, in China, ByteDance and Alibaba Cloud. Nvidia also licensed Groq's technology for about $20 billion.
  • Model hub: Nvidia agreed to buy Hugging Face for about $12.9 billion in September 2026, pending US and EU review.
  • Limits: building systems stay with Schneider, Siemens, Vertiv and Eaton, which plug into Nvidia's DSX as partners; Google and Meta run their own stacks; China is building its own on Huawei Ascend.

Flagship

Nvidia's AI Factory Research Center, Manassas, Virginia

  • Site: a nearly 100 MW Vera Rubin research facility built by Digital Realty in Manassas, where Nvidia validated its Omniverse DSX blueprint.
  • Grid: the first site built to a flexible-power standard led by Nvidia with Emerald AI, EPRI and PJM, running Emerald's Conductor to cut load on grid signals; tested with Dominion Energy.
  • Status: slated to come online later in 2026; the first-half 2026 target slipped.

China

China Builds a Parallel Stack

Huawei pledged in September 2025 to open-source its CANN accelerator software, its Mind toolchains and its openPangu models, as an alternative to Nvidia's CUDA on Ascend chips. Huawei's Volcano scheduler is a CNCF project, and its NetEco DCIM and iCooling software run its modular data centers. AMD's ROCm 7, released in September 2025, is the main open alternative to CUDA in the US.

Layers

Software Layers and Owners

Data center software layers
LayerLeading productsOwnerLicensing
BMS, EPMS and DCIMSchneider EcoStruxure and AVEVA, Vertiv Environet, Siemens Desigo, Sunbird, Nlyte; Huawei NetEcoBuilding-systems vendorsProprietary
Facility digital twinNvidia Omniverse DSX; Cadence Reality; Siemens; Vertiv SmartRunNvidia and partnersProprietary on OpenUSD
Grid-aware power controlNvidia DSX Flex; Emerald AI Conductor; PhaidraNvidia, Emerald AIProprietary
Batch schedulingSlurmNvidia, via SchedMDOpen source
Container orchestrationKubernetes; VolcanoCNCFOpen source
GPU orchestration and cluster operationsRun:ai, Mission Control, Base Command ManagerNvidiaProprietary; Run:ai to be opened
Inference servingDynamo; vLLM, SGLang, llm-dNvidia; communityOpen source
Accelerator softwareCUDA; AMD ROCm; Huawei CANNNvidia; AMD; HuaweiCUDA proprietary; ROCm open
Frameworks and model hubPyTorch, JAX, MindSpore; Hugging FaceLinux Foundation, Google, Huawei; Hugging FaceOpen source
Hyperscaler internal stacksGoogle Borg; Meta MAST and TwineGoogle, MetaInternal only

Record

Dataset Record

Dataset
Data center software layers, owners and Nvidia's acquisitions
Sources
Nvidia, Schneider Electric, Siemens, Vertiv, Cadence, Emerald AI, CNCF, AMD, Huawei, Google Research, USENIX, DatacenterDynamics, Semafor, ResearchAndMarkets
Last verified
October 2026