AI chip frenzy worth $148 billion, but this business has stalled: NVIDIA's bet on manufacturing yields little over four years

Despite NVIDIA generating nearly $148 billion in revenue from its AI chip business in the nine months ending in October, far exceeding the $27.5 billion in the same period of 2023, the company's key attempt to transition to an integrated hardware and software platform has faced significant setbacks.

As CEO Jensen Huang's core bet to enter the $50 trillion manufacturing and logistics market, its Omniverse software business has yielded minimal returns after four years of high investment, with the commercialization process severely lagging.

According to insiders who spoke to the media, due to "almost nonexistent" demand since its launch in 2022, NVIDIA has decided to shut down the Omniverse Cloud service by August 2025. Previously, the company had spent hundreds of millions of dollars renting thousands of GPUs from Oracle, Google, and Microsoft to support the project, but for most of the time, due to a lack of external customers, the company had to scramble to find internal uses to avoid chip idleness.

Although Jensen Huang continues to promote the business at this week's CES consumer electronics show, portraying it as a multi-trillion-dollar "physical AI" opportunity, he is frustrated with the slow progress of the department internally. Media reports citing insiders state that Jensen Huang has repeatedly criticized the team for wasting engineering resources on "demos" rather than "products," and that the business has not been widely adopted by large enterprises as expected.

Despite the explosive growth of the AI chip business, the market has not yet reacted sharply to the revenue gap of Omniverse, but this exposes the difficult challenge the hardware giant faces in trying to carve out a second growth curve. If the issues of software usability and industry adaptation cannot be resolved, NVIDIA's ambitions in robotics and industrial digitalization may face long-term shelving.

Demand Plummets and Cloud Service Shutdown

NVIDIA officially launched Omniverse in 2021, positioning it as a platform for designers to collaborate on 3D design in a virtual space, and attempted to expand it as a core tool for building digital twins of cars, factories, and robots. Jensen Huang's vision was to drive companies to simulate before physical manufacturing, thereby leading to billions of dollars in chip procurement and software licensing revenue of up to $4,500 per chip annually.

However, there is a significant gap between reality and vision. According to former NVIDIA employees who spoke to the media, despite the company listing clients including BMW, Siemens, Foxconn, and Boston Dynamics, very few clients have actually signed on to use the Omniverse Cloud servers for large-scale simulations. Software developers generally report that the platform is "difficult to use," has incomplete features, and is prone to crashing. Software developer Valentin Forager stated that when trying to simulate human activities in a virtual environment, the system crashes directly, "as soon as you try to do something slightly beyond its preset limits, it breaks."

Additionally, the scene creation tools on the platform are complex to operate and the documentation is outdated, making it difficult to resolve many technical issues. At an event last November, even a NVIDIA representative admitted that the software was not ready to meet specific needs and suggested clients turn to competitor Unity's software. This lack of product maturity directly led to the termination of the cloud service project

Internal Anxiety and Pressure from Management

Although the revenue from Omniverse accounts for a very small portion of NVIDIA's massive financial statements, Jensen Huang has shown significant anxiety about its performance. According to sources familiar with his thoughts, he has long been concerned that competitors will seize the initiative, and thus he is vigorously pushing the company to find new sources of income.

This frustration has erupted multiple times in internal meetings. According to attendees, during an all-employee meeting in 2023, Jensen Huang publicly questioned Rev Lebaredian, the vice president responsible for Omniverse, about whether the team had achieved profitability. In another conference call, faced with the team's request to increase personnel to develop new products, Huang angrily reprimanded them for nearly an hour, accusing the team of wasting engineers' time on demonstrations and emphasizing that the existing manpower was already sufficient.

The actual implementation of collaborative projects has also sparked dissatisfaction among management. Media reports citing informed sources stated that Huang had a fit of rage upon learning that the scale of the partnership announced with BMW was far smaller than he had initially expected. Although NVIDIA executives claimed that over 20,000 planners at BMW globally use the software, they did not specify how much this translated into actual sales.

Long-term Bets Face Industry Barriers

In response to the doubts, NVIDIA executives likened Omniverse to CUDA— the latter took over a decade of investment to fundamentally change the field of deep learning.

Rev Lebaredian stated that Omniverse is the foundational software to unlock the vast market of "physical AI," and the current long-term vision is being rewarded through adoption by robotics and automotive companies. To promote ecosystem development, NVIDIA has not only involved Huang's children in the department but has also invested in companies like Synopsys (which recently acquired Ansys) and MetAI, attempting to integrate semiconductor design and industrial simulation tools.

However, the field faces fierce competition and structural barriers. In the robotics simulation space, Unity Technologies and the open-source simulator Gazebo are strong competitors. More critically, many large companies, such as Tesla, prefer to develop internal simulation software rather than rely on NVIDIA's general platform.

Industry-specific technical barriers are also a significant challenge. The co-founders of two robotics startups pointed out that NVIDIA's Isaac Sim tool is impractical for handling complex objects with changing shapes, such as clothing. Additionally, cost-effectiveness is a factor hindering widespread adoption; Dan Cole, the COO of industrial automation equipment design company Loupe, candidly stated that renting cloud servers is not cost-effective.

As Daniel Yu, CEO of MetAI, noted, Omniverse is currently just a horizontally open platform for developers to build on, rather than a complete application, which means NVIDIA's attempt to create a market from scratch still requires a long nurturing period