NVIDIA introduced AI into the physical world at CES 2025

Gensen fan, a CEO of NVIDIA, stated this view, which is widely quoted in a keynote speech at CES 2025. “The moment of Chatgpt in general robot engineering is coming to the moment.” To support his predictions, Huang is concrete without any special environmental considerations. Three specific examples of robots are listed.

AI agents execute tasks like other information workers. Automatic cars run on the roads that are already installed. I will detect and operate the object like that

For companies with important physical assets, I would like to point out the fourth embodiment.

As the physical AI expands its autonomy to expand its autonomy beyond the machine, the industrial automation system is increasingly robotized. With this change, the rise of robot companies is seen.

By combining robotics (automation), operation technology (often called OTs), and IT systems to company -wide multimodal real -time data assets, the AI ​​world and the physical world are connected. The fusion of AI-OT-IT has upgraded decision-making, improves process efficiency, provides comprehensive context for advanced process automation (industrial robot), and responds to ERP, SCM, and BI analysis. Generates a real -time ground truth that converts from a type to a proactive type. Thus, physical AI is a new pole star to promote industrial automation and corporate digital transformation.

The time estimation of NVIDIA’s physical AI “right away” is ambiguous. Huang (and me) cannot provide specific schedules due to two technical barriers. First, physical AI requires a new model that recognizes the real world’s physics and a unique development platform. Second, most of the data generated by OTs, despite the convincing business case and the 10 -year IoT development, is still unable to access business applications using IT systems and AI.

OT suppliers are seeking appropriate ways to fill the gap between OT and IT, with the support of customers in the industry who wants to accelerate business transformation. Similarly, AI suppliers (including NVIDIA) are focusing on the development of physical AI. Let’s look at some recent trends.

Fill the gap between physical AI: NVIDIA COSMOS and Omniverse

LLMs such as Chatgpt and LLAMA do not model the physical world. To develop an accurate physical AI model for devices such as robots, autonomous driving cars, industrial systems, etc., you need to collect, filter, tag, and curse a huge amount of real world training data. To speed up this labor -intensive process, NVIDIA has developed Cosmos and announced at CES 2025. COSMOS is a physical AI development platform with a series of world infrastructure models trained in 20 million hours of video. The focus is in physics, teaching AI about the physical world, the virtual objects operate like a reality, and follow the laws of physics. NVIDIA states that Cosmos will do the same for robot engineering and industrial AI, as LLAMA 3 has done an enterprise application.

The mechanism is as follows. COSMOS creates a realistic simulation to train physical AI systems in cooperation with OmniverSE, a NVIDIA graphic collaboration platform. Development begins by using Omniverse to build realistic 3D models for real world facilities, machinery, robots, and other devices. Next, Cosmos uses the generated AI to set the Omniverse scene, and uses WFM to generate a real and geographical accurate scenario as shown in the photo. Next, COSMOS synthesizes additional scenarios to create a multidimeter world of training data that combines a variety of and unexpected situations. Omniverse simulates these scenes, captures visual data from various perspectives, so that developers can train, verify, test, and optimize target models.

NVIDIA’s physical AI development and deployment platform consists of three different workloads run on three different types of computers.

Training and fine -tuning AI model – NVIDIA DGX Super Computer Platform Physical AI Development, Simulation, Visualization, Test, Optimization – NVIDIA OVX Server Introduction Platform – NVIDIA AGX Robotics Computer

All three of these three workloads require AI acceleration, so it is not realistic in conventional CPUs. NVIDIA has optimized AI development tool chains for DGX and OVX platforms, just as the CUDA software optimized for GPUs.

Similarly, AGX is a native and optimized robot platform target of NVIDIA’s physical AI model. However, industrial customers need the flexibility to execute AI applications in various forms of physical AI. The target of the platform is that the AGX platform is always the right deployment target because it varies from microcontroller -based sensor with moderate ML inference high speed to robot computers that can run large -scale generated AI models. Not limited. In other words, the selection of the platform depends on the use case, and it is not a decision to select a product such as which GPU to buy.

NVIDIA has been offering a cross -platform introduction option for many years, but the physical AI tools are new and have no actual experience using these workflows on non -NVIDIA platforms. I encourage NVIDIA to work on this problem from the front and build a heterogeneous target target in the physical AI tool chain. The robust cross -platform support allows customers to use NVIDIA tools for a wide range of introduction hardware, removing barriers to the tool chain introduction.

Fill the OT-IT data gap: The concept of “data first”

AI creates a very convincing integrated business case, but most of the operation data has been separated from large AI due to complexity, cost, and security risk when connecting OT systems to the mainstream IT. It is left. This is the Gap between the OT-IT gap, that is, the world of an industrial IoT that is chaotic with a mixed mixture, and the uniformly managed IT world.

At present, the demand for operation data is increasing by AI. The motivation for this is that the Enterprise Software Supplier is excited to find an efficient and appropriate way to fill the OT and IT gap. The solution is surprisingly easy. Rather than trying to push IT technology into the OT world, suppliers are now shifting to a simple concept of OT integration, a simple “data first” concept. Developers have been struggling to change the mashups of the application -specific device management and connection that have been customized, complex, expensive, hard -coded into an end -to -end solution. Unfortunately, the IIOT project in the last 10 years has shown that this approach cannot be expanded.

Developers are now looking at better alternatives to bridge OT and IT with a simple interface for device IDs, security, data, events, and status. This approach simplifies the OT data access, allowing the built -in OT software to evolve independently from the cloud native IT system. Multi-modal AI applications will further reduce OT-IT integrated costs and reduce the need for cost-consuming data conversion by incorporating various types of machine data.

This tendency is backed by AWS, Google, Hanewell, Microsoft, Quarcom, other major cloud frameworks and recent ERP suppliers. (This topic will be featured in a follow -up report.) The goal is clear. Supply a large amount of OT data via standard protocols and simple APIs to the rapid growth of AI business transformation. In other words, you can get OT data without redesiting or complicated IIOT devices.

Looking at the world of robot companies

“I have the perfect model of the world. It’s a real size.” In the comedian’s Steven light one liner, it looks real and simulates the movement of the real world accurately (system, system, and virtual replica of processes ( It explains a formal explanation of building digital twins. It is not so far. The physical AI model looks like a “actual size” when displayed in 3D. When you add OT data, these models are lively and simply simulate complex scenarios in real time. With a physical model trained in dynamic simulation, applications that utilize new generation AI are process efficiency, workers’ safety, equipment operating hours, product quality, decision -making, and other high -value use cases You can achieve improvement. These use cases bring impressive ROI at least on paper and in the laboratory. However, physical AI technology such as NVIDIA COSMOS is new, and the schedule is still uncertain because the interface with OT devices often occurs for all reasons mentioned above.

My views on rational schedules for introducing these technology are as follows: NVIDIA’s AI ecosystem is strong, and the company uses a lot of physical AI. In addition, the company is an AI promoter, blessed with a list of mountains like mountains and a very long corporate customer who is hungry for its products. Considering all of these, Cosmos and Omniverse are an advantageous option, and customers are already developing solutions on this platform. For example, my company’s CEO and Chief Analyst Patrick Moore Head recently wrote about how NVIDIA, ACCENTURE, and KION cooperate to digitize warehouse operations. He shares my optimistic perspective on NVIDIA’s physical AI platform. Companies with physical infrastructure can immediately start using these tools, planning a large -scale introduction within one or two years.

However, I am not very optimistic about economics and schedules for connecting OT data sources to physical models. The barrier of IIOT connection is falling, but it is not enough to keep up with physical AI growth. This problem comes from the diversity of OT and IIOT. IT systems have unified architectures, but not in the OT system. In accordance with what I mentioned earlier, it is recommended that you fill the OT-IT gap with a simple interface that collects OT data as it is, rather than customizing the IIOT device. The good news here is that NVIDIA and others focus on physical AI, which has further promoted to realize these connections. I predict that physical AI has become robotized for companies and the urgency to fill OT-IT gaps will greatly improve this year.