The Takeaway
- Nvidia’s Omniverse software for robotics shows slow returns despite big ambitions.
- Nvidia shuttered Omniverse Cloud service due to nearly nonexistent demand.
- Omniverse software is difficult to use and often breaks, frustrating developers.
In the past two years, Nvidia’s business of selling chips for AI has rocketed into the stratosphere, lifting the company’s revenue to nearly $148 billion in the nine months through October, up from $27.5 billion for the same period in 2023. But CEO Jensen Huang isn’t satisfied. Perpetually worried about threats to his business, Huang has been prodding his team to build a business selling chips and software to design robots and solve manufacturing problems—a venture that could be a future source of growth.
The software, which runs exclusively on Nvidia chips, helps people design and simulate digital twins of real-world objects, such as a car or piece of machinery. Huang has said that these tools, known as Omniverse, could allow Nvidia to capture a slice of the $50 trillion manufacturing and logistics industries. But after four years of effort, Nvidia’s attempts to turn the Omniverse software into a money-making business have made little progress, according to four current and former Nvidia employees familiar with the business.
Publicly, Huang continues to tout the business as a multitrillion‑dollar opportunity for Nvidia. He’s expected to do so this week at the annual Consumer Electronics Show in Las Vegas, where Huang is speaking and Nvidia has a packed schedule of events showcasing the role of Nvidia’s chips and software in manufacturing and robotics. Guests include the CEO of industrial engineering conglomerate Siemens and the CTO of Agility Robotics, a leading manufacturer of humanoid robots, along with executives from PTC and Synopsys, two of the world’s largest providers of design and simulation software.
Privately, however, Huang is frustrated that the Omniverse business hasn’t yet taken off, according to four people familiar with Huang’s thinking, although Nvidia’s mushrooming revenue means few people would notice the Omniverse shortfall.
Nvidia spent hundreds of millions of dollars renting thousands of graphics processing units, or GPUs, from Oracle, Google and Microsoft so it could offer what it calls an Omniverse Cloud service for large companies that wanted to run large-scale simulations on servers, according to two people with direct knowledge of the matter. But while Nvidia has promoted a long list of companies that use Omniverse software—ranging from BMW and Siemens to Foxconn and Boston Dynamics—few signed up to run large-scale simulations on the Omniverse Cloud servers, said the former Nvidia employees.
Demand for Omniverse Cloud was nearly nonexistent from its launch in 2022 through at least early 2025, leaving the company scrambling to find internal uses for the chips so they wouldn’t sit idle, they said. In August, Nvidia shuttered the service due to a lack of demand, a person with direct knowledge of the matter said.
Some Omniverse tools allow customers to build 3D scenes or objects and simulate how they behave in virtual environments—for example, running conveyor belts in a factory. Others such as DRIVE Sim allow customers to run large-scale simulations to help develop self-driving cars. One of the tools, Isaac Sim, lets people train robots in virtual environments.
Developers who have used Omniverse tools for building and simulating scenes and objects often say the software was hard to use, that it easily broke and that its features felt incomplete.
One reason, according to three former Nvidia employees, is that the Omniverse division’s engineers were often tied up building demonstrations of new features and products that worked in controlled settings but didn’t hold up in real-world use.
In meetings with subordinates, Huang has voiced frustration with the Omniverse team for wasting engineering time on demonstrations that don’t turn into real products, according to two people familiar with the business.
His vision is for every company to simulate its robots, vehicles and factories before building them in the real world, driving billions of dollars in purchases of Nvidia’s chips, much as companies do now for training and running AI models. Behind the scenes, though, he has grown impatient with his team over its slow adoption among large companies, according to five people familiar with the business.
Long-Term Bet
The Omniverse business is important enough to Huang that both his children have jobs in the group. His daughter Madison is a senior manager in Omniverse’s marketing team while his son Spencer is a product manager for robotics.
Nvidia is also investing in companies that make simulation tools, in hopes of driving Omniverse’s adoption by getting those companies to ensure their software works with Omniverse’s tools. In the latest example, Nvidia in December announced a deal to invest $2 billion in Synopsys, one of the world’s leading providers of software for designing semiconductors. Synopsys completed its $35 billion acquisition of Ansys in July, whose software is widely used to simulate real-world physical behavior.
In theory, demand for simulation could be enormous. One of the biggest potential opportunities is in robotics, where developers lack spatial and motion data on how humans move and use objects needed for the development of humanoid robots. Some robots, such as Tesla’s Optimus, have been relying on humans wearing motion-capture devices to gather such data. But that approach is costly and time-consuming, compared to using simulators.
But Nvidia is already facing fierce competition in this area. Unity Technologies, whose software has been used to develop popular videogames such as Pokémon GO, also offers tools for generating scenes and objects and lets customers design, test and train robotic systems in virtual environments. Another rival is Gazebo, an open-source robot simulator launched by nonprofit Open Robotics. And many companies are building their own simulation software in-house as there isn’t any standardized software that fits all of their needs.
To be sure, it’s still early in the field. Nvidia executives compare Omniverse to earlier long-term company bets on technologies like Cuda, Nvidia’s software for training and running AI models on its GPUs. Huang backed that effort aggressively despite objections from analysts and activist investors who argued that Nvidia was spending billions of dollars on a business that generated little immediate revenue.
“Just as we invested in Cuda for over a decade before it revolutionized deep learning, Nvidia has spent years building Omniverse as the foundational software to open the next massive market for accelerated computing: physical AI,” Rev Lebaredian, Nvidia vice president overseeing Omniverse, said in a written statement.
“That long-term vision is now paying off as major cloud service providers spin up dedicated simulation infrastructure, as robotics and auto companies use our frameworks to build and train the next generation of autonomous machines,” he added.
Half-Baked Software
Nvidia officially launched Omniverse in 2021, positioning it as a software and simulation tool that would let designers work together on 3D designs virtually, similar to a shared Google doc but for 3D models and environments.
BMW was one of the company’s first high-profile customers, using Nvidia’s Isaac robotics software to design and improve how materials flowed through its car factories.
In the past few years, Nvidia has announced a dizzying array of Omniverse customers with industrial manufacturers such as Dassault and Foxconn, autonomous vehicle developers like Mercedes-Benz and Jaguar Land Rover and robotics firms like Boston Dynamics and Figure AI.
The idea was that Omniverse software would both drive more demand for Nvidia’s GPUs and get people to pay as much as $4,500 a year per chip for enterprise software licenses and support. Alternatively, customers could sign up for Omniverse Cloud, which bundled access to GPU servers with those licenses.
But even as Nvidia’s Omniverse team was working on these partnerships, the company’s AI chip business was taking off. The late 2022 introduction of OpenAI’s ChatGPT kicked off a race among big tech companies to develop competitive AI technologies—with most companies using Nvidia’s AI chips for their development work.
Despite that success, Huang is known for his paranoia about losing ground to rivals and for pushing the company to find new sources of revenue.
His frustration about the lackluster performance of Omniverse has been particularly pointed. During a 2023 all-hands meeting, Huang chided Lebaredian about Omniverse, asking whether his group had made any money yet, according to a person who attended the meeting.
Around that same period, Huang became angry with the Omniverse team when he learned that a partnership Nvidia had announced with BMW was much smaller in scope than he initially was led to believe, one of the people familiar with the business said. BMW didn’t respond to emails and calls seeking comment. An Nvidia executive said in 2023 that BMW had over 20,000 factory planners worldwide using Omniverse software, though he didn’t elaborate on how that translated into sales.
In another example, Huang grew agitated during a conference call after a team in the Omniverse division asked for more staff to develop a new product, according to a person who was on the call. Huang berated them for the next hour, accusing the team of wasting their engineers’ time and lecturing them on how Nvidia builds products. The team already had more than enough people, he told them.
Throughout 2023 and 2024, the Omniverse team showed Huang numerous demonstrations of software ranging from tools that allow multiple people to collaborate on 3D models to concepts for streaming software and new simulation programs including one for cars. But the team hadn’t turned any into products. Huang told the team they were trying to do too many things at once and that they should focus on one thing that could become a product, the person said.
That prompted the team to focus more on software for building digital twins. At the same time, robotics became a bigger part of Omniverse, as more companies began using Isaac Sim to generate synthetic data.
But the robotics industry is still in its early days and major players such as Tesla are still building their own simulation software rather than turning to Nvidia’s offerings, according to people familiar with the industry.
Moreover, roboticists complain about shortcomings in Nvidia’s software. Isaac Sim, for example, isn’t useful for training a robot to manipulate complex objects, especially objects such as clothing, whose shape changes constantly, according to this cofounder and the cofounder of another robotics startup.
The Omniverse tool does work for teaching a robot how to walk or move around on its own, the startup cofounders said. But getting the robots to interact with other objects is critical, they said.
Getting users to pay for subscriptions to Omniverse Cloud has also been an uphill battle. Dan Cole, chief operating officer at Loupe, which designs automated manufacturing equipment for customers such as aerospace companies, said he bought a handful of computers running Nvidia chips for his developers to simulate machines in a factory. Although he is aware that companies can license Omniverse servers in the cloud for his employees, it didn’t make sense given the cost, he said.
It’s not just the robotics software that can cause problems. Several software developers said Omniverse’s scene creation tools are hard to use and can malfunction outside a set of narrowly defined scenarios. The instructions for how to use these tools are poor and outdated, making it hard to fix problems building virtual objects and worlds, they said.
Valentin Forager, a software developer, said Omniverse software worked well for creating virtual retail-store shelves, which his firm uses to develop sensors that track how shoppers look at products in real stores. But when he tried to simulate humans in the same virtual environment, his computer crashed.
“As soon as you try to do things that are a little bit outside the box, this thing is broken and almost never works,” Forager said.
Forager said that when he approached an Nvidia representative at a company-sponsored event in November in Seoul, the representative admitted that Omniverse wasn’t ready for his particular needs and suggested that he use Unity’s software.
Industry Demand
Nvidia may be ahead of its time when it comes to simulation. The company’s executives say they are trying to create a market that doesn’t yet exist and nobody knows they need. Many industrial manufacturers also don’t have the expertise and money to invest in it, according to former Nvidia employees and industry insiders.
It’s also hard for Nvidia to build simulation software that works broadly for users in many different types of businesses. Simulation software is different for robotics compared with cars and industrial equipment.
One solution Nvidia has come up with is investing more in startups that offer digital twins and simulation services so the company can better understand what customers want. This also encourages them to build their industry-specific services on top of Omniverse’s software tools.
One of those investments is in MetAI, a Taiwan-based startup that builds digital twins for semiconductor foundries, including construction planning and factory layout simulation for facilities at Taiwan Semiconductor Manufacturing Co.
“I think people underestimate how many resources are required to build these simulations with real domain knowledge,” said Daniel Yu, CEO of MetAI. “It takes a lot of time. Omniverse is not a complete application —it’s a horizontal, open platform for developers to build on.”