NVIDIA DSX AI Factory Platform
NVIDIA DSX is a platform, not a single product: a set of reference designs, simulation tools, open-source operations software, power management and data-exchange schemas for designing, building and running AI factories, with each part usable by different partners in the build.1
Also known as NVIDIA DSX, DSX, NVIDIA Air (now NVIDIA DSX Air)
At a glance
- What is it?
- DSX is NVIDIA's platform for AI factories at facility scale. It brings together reference designs, APIs, software libraries and technologies covering design, simulation, construction and operation. NVIDIA's FAQ names the technologies as DSX Reference Design, DSX Sim, DSX OS, DSX MaxLPS, DSX Flex and DSX Exchange. DSX Sim includes NVIDIA DSX Air (formerly NVIDIA Air), the Omniverse DSX Blueprint for AI factory digital twins, and SimReady asset pipeline samples. Around these sit qualification programs such as DSX Ready and an AI factory marketplace of validated products.123
- What does it do?
- DSX gives the many parties in an AI factory build a shared design standard. Reference designs specify generation-specific architectures for compute, networking, storage and facilities. DSX Sim lets builders simulate the network and the physical site before hardware arrives. NVIDIA describes DSX OS as open-source, modular software for running multi-tenant AI factory infrastructure, from schedulers to bare-metal provisioning and GPU health; some listed components, such as Run:ai and Fleet Intelligence, are commercial or managed services. MaxLPS manages power at GPU, rack and workload level within a fixed site budget. DSX Flex models grid requests such as demand response, and DSX Exchange carries operational signals between IT, power, cooling and facility systems.12
- Who needs it?
- NVIDIA lists a broad ecosystem: NVIDIA Cloud Partners, sovereign clouds and land, power and shell providers; makers of power and cooling equipment; ISVs for design, simulation and facilities operations; OEMs and system integrators; architecture, engineering and construction firms; and infrastructure ISVs for IaaS and PaaS. In practice it matters most to organizations planning or operating multi-megawatt AI facilities, and to vendors who want their equipment or software to fit those builds.2
- What does it need?145
- NVOnline access for the DSX reference designs (through an NVIDIA representative)
- Kubernetes clusters for DSX OS components and the DSX Exchange event bus (Helm-based deployment)
- A building management system able to publish facility data over MQTT 3.1.1 for Exchange integration
- Omniverse skills and OpenUSD assets for the DSX Blueprint digital twin
- A DSX Air account for network simulation
- Approved early-preview engagement for Dynamic Power Software
- What it is not
- DSX is not one product you buy and install. It is a collection of reference designs, open-source software, schemas, simulation tools and partner programs, and NVIDIA's own Flex documentation says the families are shown together to explain responsibilities, not as one required stack. Parts differ in access and maturity: some reference designs require NVOnline access, Dynamic Power Software is limited to approved early-preview engagements, and the MaxLPS capacity figures are illustrative scenarios for Vera Rubin where testing is still in progress. DSX Flex does not take over control of the factory; each participant keeps its local controls. In our view, DSX does not replace facility engineering by qualified firms.167
Availability and licensing. The Omniverse DSX Blueprint is generally available. NVIDIA describes DSX OS and DSX Exchange as open source, although some components listed under DSX OS, such as NVIDIA Run:ai and the Fleet Intelligence managed service, are commercial products or managed services. DSX Air is cloud-hosted and NVIDIA offers a free trial. Some reference designs require NVOnline access, and Dynamic Power Software is limited to approved early-preview engagements.128
The problem it solves
An AI factory is built by many parties: grid and power providers, building and cooling contractors, hardware makers, network and software vendors and the operator. Each tends to work from its own specifications, so problems surface late, when equipment is already installed.
Power is the other constraint. Sites are sized for peak GPU power, which leaves capacity stranded when workloads draw less, and grid operators increasingly ask large loads to respond to demand events. DSX addresses both with common reference designs and simulation before construction, and with shared schemas and software that let compute, power and facility systems coordinate during operation.26
How it works
DSX is organized as layers that partners can adopt separately.
- DSX Reference Design: validated, generation-specific architectures such as the Vera Rubin NVL72 reference design and the Vera Rubin facilities infrastructure reference design.
- DSX Sim: DSX Air for logical simulation of GPUs, SuperNICs, DPUs and switches; the Omniverse DSX Blueprint for physically accurate digital twins.
- DSX OS: platform software (KAI Scheduler, Run:ai, NVIDIA Cloud Functions, Grove, Dynamo) and infrastructure software (GPU Operator, Network Operator, Topograph, NVSentinel, Infra Controller, DOCA Platform Framework, Rack Management Service and others).
- DSX MaxLPS: facility design for 45 degree Celsius cooling, Dynamic Power Software and performance-per-watt techniques within a fixed power envelope.
- DSX Flex and DSX Exchange: a grid-facing schema for load targets and power states, carried over the DSX Event Bus (NATS with MQTT 3.1.1), plus an Agent Gateway that routes Model Context Protocol requests to authorized servers.147
Diagram as a list
Applications & solutions
- DSX Sim (DSX Air, Omniverse DSX Blueprint)Logical and physical simulation before and after deploymentConnects to DSX Reference Design
Inference & runtime software
- Dynamo and Grove (DSX OS platform software)Optimized inference serving on the factoryConnects to NVIDIA compute, networking and storage
Operations & orchestration
- DSX OSOpen-source scheduling, provisioning, health and multi-tenant operationsConnects to NVIDIA compute, networking and storage
- DSX Exchange (Event Bus, Agent Gateway)Carries operational signals between IT, OT and agentsConnects to DSX MaxLPS and Dynamic Power Software, DSX OS
Accelerated computing
- NVIDIA compute, networking and storageThe token-producing infrastructure
Networking, power & facilities
- DSX Reference DesignValidated architecture from chip to gridConnects to DSX Sim (DSX Air, Omniverse DSX Blueprint), Facilities: land, power, shell and cooling
- Facilities: land, power, shell and coolingPhysical site built to the reference designConnects to DSX Exchange (Event Bus, Agent Gateway)
- DSX MaxLPS and Dynamic Power SoftwarePower allocation at GPU, rack and workload levelConnects to NVIDIA compute, networking and storage
- DSX FlexGrid load targets, power state and exceptionsConnects to DSX Exchange (Event Bus, Agent Gateway)
Programs & resources
- DSX Ready and AI factory marketplaceQualification of partner productsConnects to DSX Reference Design
Capabilities
DSX Reference Design12
Generation-specific, validated AI factory architectures covering compute, networking, storage, facilities infrastructure and cluster design.
Why it matters: Gives every contractor and vendor the same target design.
Limits: Several documents require NVOnline access through an NVIDIA representative.
DSX Sim (Air and Omniverse DSX Blueprint)1
DSX Air simulates NVIDIA hardware and software infrastructure logically; the Omniverse DSX Blueprint builds physically accurate digital twins of the site.
Why it matters: Bottlenecks and configuration errors can be found before hardware ships.
Limits: We recommend checking the models and assets fed into a simulation, because its results depend on them.
DSX OS19
Open-source, modular software for lifecycle management, scheduling, resiliency, monitoring, provisioning and multi-tenant operations.
Why it matters: Operators can assemble an operations stack from documented components.
Limits: Components are separate projects that must be integrated; use of NVIDIA software is optional per the NCP guide.
DSX MaxLPS2
Power and efficiency software that manages power at GPU, rack and workload level to fit more useful compute into a fixed power budget.
Why it matters: Addresses stranded power in sites sized for peak draw.
Limits: NVIDIA's capacity figures are vendor claims and illustrative scenarios; Dynamic Power Software is in early preview.
DSX Flex26
A schema for receiving grid signals such as load shedding, demand response and pricing events, and for reporting the factory's power state and exceptions.
Why it matters: Lets an AI factory take part in utility or energy programs.
Limits: Does not transfer control of the factory response. The Flex overview leaves onsite generation and storage control interfaces outside the current schema, while the DSX product page and docs landing page describe Flex as orchestrating power across utility, on-site renewables and storage.
DSX Exchange4
Open-source platform with a NATS-based event bus (MQTT 3.1.1), AsyncAPI schemas and an Agent Gateway for Model Context Protocol routing.
Why it matters: Connects IT, power, cooling and facility systems through one message layer.
Limits: Uses MQTT 3.1.1 rather than MQTT 5 for compatibility with building systems.
DSX Ready qualification2
A designation for qualified and validated solutions aligned with DSX requirements, plus an AI Cloud-Ready validation initiative for infrastructure software.
Why it matters: Helps buyers shortlist compatible equipment and software.
Limits: In our reading, qualification shows alignment with DSX requirements; we recommend still checking performance at your own site.
Practical use cases
An AI factory builder needs confidence in network, power and cooling design before committing capital to construction.
- Approach
- Build a digital twin with the Omniverse DSX Blueprint and simulate the network in DSX Air, against a DSX reference design.
- Role of NVIDIA DSX AI Factory Platform
- DSX Sim and the reference designs supply the models and the target architecture.
- Data, infrastructure and skills
- CAD and equipment data, Omniverse and simulation skills, NVOnline access.
- Type of benefit
- Risk reduction before construction
- Caveats
- Simulation results depend on model fidelity and do not replace engineering sign-off.
- First step
- Explore the DSX Blueprint repository and the facilities reference design overview.
Sources 2
A utility asks a large AI site to cut its grid draw during peak events.
- Approach
- Receive time-bound load targets through DSX Flex, carry them over DSX Exchange, and let factory-side power management such as MaxLPS decide how to meet them.
- Role of NVIDIA DSX AI Factory Platform
- DSX Flex defines the request and reporting; Exchange transports it; MaxLPS is one option for the response.
- Data, infrastructure and skills
- A grid ISV participant, factory power management, BMS integration.
- Type of benefit
- Grid flexibility and program participation
- Caveats
- The Flex overview places onsite generation and battery controls outside the current schema, although NVIDIA's product page describes Flex orchestrating utility, on-site renewable and storage power.
- First step
- Read the DSX Flex overview and the Power Management Schema.
A cloud partner wants to sell multi-tenant AI services on its own NVIDIA hardware.
- Approach
- Use the NCP Software Reference Guide and DSX OS components to build IaaS, CaaS and AI PaaS layers.
- Role of NVIDIA DSX AI Factory Platform
- DSX OS supplies optional NVIDIA components mapped to each layer.
- Data, infrastructure and skills
- Kubernetes operations team, tenant isolation design, partner ISVs.
- Type of benefit
- A documented reference architecture
- Caveats
- The guide is a reference, not an implementation manual.
- First step
- Use the guide's capability checklist to find gaps in your current stack.
Sources 9
Works with
Optional integration
- NVIDIA OmniverseThe Omniverse DSX Blueprint is part of DSX Sim.
- NVIDIA Run:aiRun:ai is a workload management component of DSX OS.
- NVIDIA DynamoDynamo and Grove provide optimized inference in DSX OS.
Complementary tools
- NVIDIA BlueFieldThe DOCA Platform Framework in DSX OS operates BlueField-accelerated infrastructure services.
- NVIDIA OmniverseThe DSX Blueprint for AI factories is listed among Omniverse blueprints.
- NVIDIA Run:aiListed as a workload management component of DSX OS.
- NVIDIA Spectrum-X EthernetDSX Air simulates Spectrum-X Ethernet fabrics.
Optional integration for
- NVIDIA Spectrum-X EthernetDSX Air simulates Spectrum-X fabrics before deployment.
Relationship labels follow NVIDIA's documentation. "Alternative approaches" does not mean one is better: each profile says when it fits.
Getting started
Map the DSX components to your role
Start on the DSX documentation landing page and identify which of Sim, OS, MaxLPS, Flex, Exchange and the reference designs apply to your part of the build.
Check: A short list of DSX components you will evaluate, with an owner for each.
Request reference design access
Ask your NVIDIA representative for NVOnline access to the generation-specific reference designs.
Check: You can open the Vera Rubin reference design documents.
Simulate the network
Open DSX Air at dsx-air.nvidia.com and run a prebuilt demo or a custom topology.
Check: A simulation of your planned fabric boots and passes your configuration scripts.
Try the digital twin blueprint
Clone the Omniverse DSX Blueprint repository and run its sample AI factory digital twin.
Check: The sample twin loads with SimReady assets and the web UI.
Stand up DSX Exchange in a lab
Deploy the event bus with its Helm chart and publish test facility data following the BMS Companion Guide.
Check: A test subscriber receives BMS topics from the event bus.
Official resources
Could this technology help you?
Describe your project to the Solution Architect. It starts with NVIDIA DSX AI Factory Platform as context but recommends independently, including when you do not need it.
Sources
Each statement above links to the source it comes from. Labels say who reported it.
- NVIDIA DSX documentation (opens in a new tab)
- NVIDIA DSX AI Factory Platform (opens in a new tab)
- NVIDIA DSX Air Platform (opens in a new tab)
- NVIDIA DSX Exchange architecture (opens in a new tab)
- NVIDIA DSX Air User Guide (opens in a new tab)
- NVIDIA DSX Flex overview (opens in a new tab)
- NVIDIA DSX MaxLPS overview (opens in a new tab)
- NVIDIA Spectrum-X Ethernet Networking Platform (opens in a new tab)
- NVIDIA Cloud Partner Software Reference Guide: introduction (opens in a new tab)
- NVIDIA newsroom: Vera Rubin ramps into full production (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.