I keep getting a version of the same question from customers and partners: what does it cost to take an Omniverse project into production? The answer changed in May: it’s free now, and a lot of the people asking me haven’t caught up yet. It’s documented plainly on the Omniverse licensing page, but the change showed up in the license wording first and only got a short announcement on the Developer Forums weeks later, on July 1, without much noise. That gap is part of why it’s easy to miss. I wrote this down to separate what changed from what didn’t, and to work out why the new model fits where the platform is going.
What the page says now
The wording is direct: “As of May 2026, Omniverse is freely available for both development and production use with no NVIDIA AI Enterprise subscription required.” Software built with Omniverse can also be redistributed under the same terms.
“Free” does not mean the licensing went away. Omniverse is still governed by a license agreement: you accept the NVIDIA Software License Agreement and Product-Specific Terms when you download and use it, exactly as before. What changed is the commercial side of that license: production used to require a paid NVIDIA AI Enterprise subscription, and now it doesn’t. The license still applies; the commercial requirement attached to production is what lifted. Free to develop, free to deploy, free to redistribute, under a license you still agree to.
What it was before, and why that made sense too
Before this, development was already free, and production deployment went through an NVIDIA AI Enterprise subscription. NVIDIA’s published list pricing puts that at $4,500 per GPU per year, with perpetual ($22,500 per GPU) and pay-as-you-go ($1 per GPU-hour on cloud marketplaces) options as well. That’s a completely standard way to license an enterprise platform: try it for free, and when you deploy at scale you’re in a supported, subscription-backed relationship. It fit where Omniverse was at the time, and plenty of serious infrastructure is still sold exactly that way. The change isn’t a correction of something that was wrong. It’s a move to a different model that fits where the platform is going.
What actually changed for a team
The new model keeps a support path; it just makes it opt-in rather than required. You can now build, deploy, and redistribute without a commercial licensing conversation. Separately, if you want Enterprise Support with an SLA, NVIDIA AI Enterprise is still available through an NPN reseller or a cloud marketplace, and partners who embed Omniverse in their own products can use the embedded licensing model: the partner handles front-line support for their users, NVIDIA provides back-line support behind them. Without a subscription, support runs through the community channels: the Developer Forums and Discord.
So the practical read is two separate decisions where there used to be one. First: can I put this into production? Now yes, no commercial-licensing step stands in the way. Second, and independently: do I want a support contract on top? That depends on what you’re building. A prototype or an internal tool is right at home on community support; a customer-facing production system might well want the SLA. Both are legitimate, and now you get to make those calls separately instead of clearing the licensing one before you can ship anything at all.
Why this is a natural move
When a team asks me about this, I separate the cost of production into hardware, software, and support. Hardware cost scales with adoption. You buy more of it as the thing actually gets used, so it tracks real business success. Software licensing, in the old model, was a fixed cost you committed to up front, before any of that success had shown up. The teams I see hit this hardest are the early ones: startups and partners building something genuinely new, still chasing market fit, not yet at the scale where a per-GPU subscription is a rounding error. Committing to a fixed software fee before the product has proven itself is the wrong shape of cost for where they are. Making the software free lines it up with the hardware, so you spend as you succeed rather than ahead of it.
That also fits what NVIDIA is after. Omniverse isn’t really a standalone product to be sold. It’s the foundation under digital twins, robotics simulation, autonomous-vehicle simulation, and AI-factory design, all of which NVIDIA wants broadly adopted and all of which run on the GPUs. Free-to-adopt software with an optional support tier makes it the natural place to start that work, and the hardware scales with it as adoption grows. For anyone who was weighing whether to start a project in Omniverse, the calculus just got simpler.
Libraries changed the shape of adoption
The library direction changed how I read the price change. NVIDIA has been breaking Omniverse into modular libraries you pull into your own application: take the piece you need and build it in, rather than adopting the whole suite. Several are already available in early access:
- ovrtx: RTX rendering and sensor simulation; you can render frames and generate synthetic data in roughly ten lines of Python, with zero-copy handoff to PyTorch, NumPy, and Warp.
- ovphysx: USD-native, GPU-accelerated PhysX simulation that runs headless, with no UI dependency.
- ovstream: a single
pip installfor streaming a live session over WebRTC, RTSP, or shared memory.
More is on the way, and NVIDIA’s own write-up on the libraries is the best current overview. These are early-access builds today, with production releases and long-term API stability planned later this year.
A free license fits that library model. When the platform is something you consume as a library inside your own app (and increasingly something an agent reaches for, calling the specific capability it needs at the moment it needs it) the old rhythm of a long build, a test phase, another test phase, and only then a commercial licensing conversation before production doesn’t match how the work actually happens. Agentic development goes from dev to prod quickly and keeps iterating; you want to stay agile while the thing stabilizes underneath you, rather than freeze the design to fit a procurement step. One licensing model can’t be the right shape for every use case in that world. Free to adopt, with an optional support tier, can: you build with the library that matters, ship when you’re ready, and add enterprise support if and when the workload calls for it.
Why I keep bringing it up
The change takes a step out of the middle of building. You can move an idea from development into production without stopping to sort out commercial licensing first; that conversation only becomes necessary once you reach a workload that genuinely needs an SLA or enterprise support behind it. For me, the useful change is not simply the lower cost. It is the smaller gap between trying something and running it for real. If you ever set an Omniverse idea aside because the path to production looked like it came with a commercial licensing negotiation attached, that reason is gone.
Disclaimer: The views and opinions expressed in this account are those of my own and do not represent those of my employer, NVIDIA.