One-Click Digital Twins: Deploying the NVIDIA Omniverse DSX Blueprint using Rafay

AI factories are no longer just racks of GPUs in a data hall — they're power plants, cooling loops, electrical topologies, and compute fleets that all have to be designed, simulated, and operated together. NVIDIA's answer to that problem is the Omniverse DSX Blueprint for AI Factories: a digital twin reference application, built on OpenUSD and the Omniverse Kit SDK, that lets teams reconfigure a 50-acre AI factory design and watch the power and thermal simulation respond in real time, inside an interactive 3D viewport.

It's a genuinely impressive piece of engineering. It's also a lot of infrastructure to stand up before anyone gets to see a digital twin.

What does it take to stand up DSX today?

If you read through the NVIDIA Omniverse DSX Blueprint, the prerequisites alone tell the story:

• A workstation-class GPU (NVIDIA RTX Pro 6000 Blackwell), a pinned driver version, 64GB RAM, and fast local NVMe
• Git LFS, Docker, and the NVIDIA Container Toolkit configured for non-root GPU access
• Node.js 20+ for the React web portal, plus a full C++ toolchain (VS2019 with the v142 toolset on Windows, build-essentials on Linux) to compile the kit-cae submodule on first run
• A separate multi-gigabyte USD content pack pulled from NGC and unpacked to a local path that then has to be wired into dsx.kit
• Two long-running processes — a Kit streaming server and a web frontend — that have to find each other over WebRTC
• A decision, at deployment time, between packaging for NVIDIA Cloud Functions (NVCF) or standing up your own GPU-enabled Kubernetes cluster with the provided Helm charts

None of this is a criticism of the blueprint itself. The README is explicit that it's a starting point for developers to extend, not a turn-key production app.

If you are a neocloud, sovereign cloud operator, or enterprise platform team that wants to offer digital-twin design sessions to CFD engineers, design engineers, and AI engineers across your organization, repeating that setup by hand for every user, every session, and every GPU generation just will not scale!!

Turning a Blueprint into a 1-Click SKU

Rafay exists to take exactly this kind of clone it, build it, configure the GPU, wire up streaming, remember to tear it down workflow and turn it into a governed, self-service catalog entry. Applied to the DSX blueprint, that looks like:

Package the Blueprint

The dsx_nvcf.kit container build, the kit-cae and kit-usd-agents submodules, the pinned USD content pack, and the web portal are baked into a versioned SKU with the Helm chart from helm/dsx as the deployment artifact. Every subsequent user pulls the same known-good image — no one is running ./repo.sh build on their laptop and waiting out a multi-minute shader compile.

GPU Sizing and Partitioning

Rafay's turnkey support for NVIDIA GPU Operator and DRA integration exposes MIG or time-slicing modes so that a single NVIDIA RTX Pro 6000 Blackwell or an L40S/L4 fleet, for teams standardizing on a shared GPU pool for streaming rather than dedicating a GPU per session. This allows you to host multiple concurrent DSX streaming sessions with real isolation between them.

End Users never touches Kubernetes

They login into Rafay's self-service portal, pick the "DSX Digital Twin" SKU, choose a session size and TTL, and click launch. Rafay provisions the cluster (bare metal or VM, depending on the tenant's existing footprint), deploys the Helm chart, and hands back a streaming URL. That's the entire "getting started" guide, compressed into one click.

Benefits

Sessions Expire

DSX streaming sessions are exactly the kind of workload that gets left running. An user opens a viewport, gets pulled into a meeting, and forgets about it.

Rafay's TTL-based expiry reclaims the GPU automatically so that idle digital-twin sessions don't quietly burn through GPU-hours the way an unmanaged deployment might.

Multi-tenancy is Native

DSX's own target audience spans five different personas i.e. end user, CFD engineer, design engineer, application developer, AI engineer. They are often from different teams or even different partner organizations (OEMs, CAD vendors) collaborating on the same facility design.

Rafay's built in multi-tenancy keeps each persona's session, and each partner's data, walled off on shared infrastructure.

Summary

For Rafay customers running GPU capacity for neoclouds, sovereign clouds, or as NVIDIA Cloud Partners, DSX is a good example of a broader pattern: NVIDIA ships increasingly sophisticated Omniverse and NIM-based blueprints, and the value an infrastructure provider adds isn't in re-implementing NVIDIA's application logic. Instead, it is in making that application consumable.

A one-click DSX SKU means:

• OEMs and CAD partners can validate SimReady assets against a live digital twin without ever requesting a bare-metal GPU allocation
• Internal facilities and CFD teams get self-service access to power and thermal simulation scenarios, metered in GPU-session-minutes for straightforward chargeback
• Platform teams get a 1-click deployment path they can point at any GPU generation in their fleet, self-managed streaming infrastructure or NVCF-style patterns included, without maintaining bespoke runbooks per team

The DSX Blueprint proves out what an AI factory digital twin can do. Rafay is what makes it something you can actually hand to fifty engineers across three companies without a support ticket queue forming behind it.