What Is NVIDIA DSX OS?
NVIDIA DSX OS is the software half of the NVIDIA DSX platform. Where a DSX reference design specifies how compute, networking, power, and cooling are integrated into an AI factory, DSX OS is the open source, modular software that operates it once it is built.
It is not a single product or a monolithic distribution. It is a set of components, most of them independently released and Apache 2.0 licensed, that operators can adopt individually or together. An operator already running a scheduler can take only the health and remediation pieces. An operator starting from bare metal can take the provisioning layer too.
Components
| Component | Role |
|---|---|
| NVIDIA Infra Controller (NICo) | API-driven bare-metal lifecycle management, with tenant isolation enforced in BlueField DPUs |
| NVIDIA AI Cluster Runtime (AICR) | Version-locked runtime recipes that hold driver and library versions consistent across a fleet |
| KAI Scheduler | GPU-aware workload placement, fractional GPU allocation, hierarchical quotas |
| NVIDIA Dynamo | Distributed inference serving with disaggregated prefill and decode |
| NVIDIA Grove | Kubernetes orchestration for multi-component, multi-node inference workloads |
| NVSentinel | GPU fault detection with automated cordon, drain, and remediation |
| NVIDIA Fleet Intelligence | Fleet-wide health, utilization, and integrity visibility |
| DSX Exchange | MQTT-based hub bridging IT and OT signals (grid events, thermal data, power anomalies) |
| DSX MaxLPS | Power policy enforcement at the GPU, rack, cooling, and workload level |
| DSX Flex | Grid integration libraries for demand response and load shedding |
NVIDIA reports that MaxLPS lets a fixed power budget support up to 40% more GPUs at peak energy efficiency, with minimal impact on inference workload performance. That claim reflects the underlying economics of large AI factories, where the binding constraint is usually megawatts rather than rack space.
What DSX OS Covers, and What It Does Not
DSX OS covers the operational layer: getting nodes provisioned, keeping runtime versions consistent, placing workloads on GPUs, serving models, and detecting and remediating hardware faults without a human in the loop.
What it deliberately leaves out is the commercial layer. Per-token metering, per-tenant quotas tied to a price list, chargeback, and billing are not part of DSX OS. For an operator selling inference capacity rather than consuming it internally, that gap is the difference between infrastructure that runs and infrastructure that can be sold.
Where Saturn Cloud Fits
Saturn Cloud runs on DSX OS components and adds the layer above them. Dynamo and Grove serve the models, the KAI Scheduler places the work, NVSentinel and Fleet Intelligence keep the fleet healthy, and Saturn Cloud provides per-token metering, per-tenant quotas, and billing on top, along with fine-tuning and single-tenant or multi-tenant serving on the same control plane.
For neocloud operators, telcos, and sovereign AI clouds, that combination turns a DSX-based buildout into a token factory: capacity that is metered, priced, and sold per token rather than rented by the GPU hour.
