In his CES 2025 keynote, Jensen Huang, CEO of Nvidia, delivered this widely quoted observation: “The ChatGPT moment for general robotics is right around the corner.” To support his “right around the corner” prediction, Huang listed three specific robotic embodiments that can function immediately without special environmental accommodations:
- AI agents perform tasks just like any other information worker
- Self-driving vehicles operate on roads that are already in place
- Articulated humanoid robots fit directly into our physical surroundings, sensing and manipulating objects as people do
For enterprises with significant physical assets, I want to point out a fourth embodiment:
- Industrial automation systems become increasingly robotic as physical AI expands autonomy beyond machines to include larger-scale processes. With this shift, we’re seeing the rise of robotic enterprises.
Combining robotics (automation), operational technologies (often called OT) and IT systems into a companywide, multimodal, real-time data estate connects the AI world with the physical world. AI-OT-IT fusion generates the real-time ground truth that upgrades decision making, enhances process efficiencies, provides a holistic context for advanced process automation (industrial robotics) and transforms ERP, SCM and BI analytics from reactive to proactive. Hence, physical AI is the new North Star driving industrial automation and enterprise digital transformation.
Nvidia’s “right around the corner” time estimate for physical AI is vague because two technical barriers prevent Huang (and me) from providing a specific timetable. First, physical AI requires new, real-world, physics-aware models and unique development platforms. Second, despite compelling business cases and a decade of IoT development, most data produced by OT is still inaccessible to IT systems and AI-powered business applications.
Driven by industrial customers eager to accelerate business transformation, OT suppliers are seeking expedient ways to bridge the OT-IT gap. Likewise, AI suppliers (including Nvidia) are doubling down on physical AI. Let’s take a look at some recent developments.
Bridging The Physical AI Gap: Nvidia Cosmos And Omniverse
LLMs such as ChatGPT and Llama do not model the physical world. Developing accurate physical AI models for equipment such as robots, autonomous vehicles and industrial systems requires collecting, filtering, tagging and curating enormous amounts of real-world training data. To speed up this labor-intensive process, Nvidia has developed Cosmos, which it announced at CES 2025. Cosmos is a development platform for physical AI with that has a set of world foundation models trained on 20 million hours of video. The focus is on physical dynamics — teaching AI about the physical world so that virtual objects behave like real ones and obey the laws of physics. Nvidia says Cosmos will do for robotics and industrial AI what Llama 3 has done for enterprise applications.
Here’s how it works. Cosmos works with Omniverse, Nvidia’s graphics collaboration platform, to create realistic simulations for training physical AI systems. Development begins by using Omniverse to build realistic 3-D models of real-world facilities, machinery, robots and other equipment. Cosmos then uses generative AI to populate Omniverse scenes, leveraging its WFMs to generate photorealistic, geospatially accurate scenarios. Cosmos then synthesizes additional scenarios to create a multiverse of training data with many combinations of diverse and unexpected situations. Omniverse simulates these scenes, capturing visual data from various points of view, which enables developers to train, validate, test and optimize the target model.
Nvidia’s physical AI development and deployment platform comprises three distinct workloads running on three different types of computers:
- AI model training and fine-tuning — Nvidia DGX supercomputer platform
- Physical AI development, simulation, visualization, testing and optimization — Nvidia OVX servers
- Deployment platforms — Nvidia AGX robotics computers
These workloads are not practical on traditional CPUs, because all three require AI acceleration. Nvidia has optimized its AI development toolchain for DGX and OVX platforms, much like the company optimized its CUDA software for its GPUs.
Similarly, AGX is the native, optimized robotics platform target for Nvidia’s physical AI models. However, industrial customers need the flexibility to run AI applications across diverse physical AI embodiments. Platform targets vary from microcontroller-based sensors with modest ML inference acceleration to robotics computers capable of running large generative AI models, so AGX platforms aren’t always appropriate deployment targets. To put it another way, platform choice is use-case dependent, not a product selection decision like which GPU to buy.
Although Nvidia has offered cross-platform AI model deployment options for years, physical AI tools are new, and we don’t yet have practical experience using these workflows for non-Nvidia deployment platforms. I encourage Nvidia to tackle this issue head-on and build heterogeneous deployment targets into the physical AI toolchain from the start. Robust cross-platform support removes toolchain adoption barriers by enabling customers to use Nvidia’s tools across a broad spectrum of deployment hardware.
Bridging the OT-IT Data Gap: The “Data First” Mindset
Although AI creates extremely compelling integration business cases, most operational data remains isolated from large-scale AI because of the complexity, cost, and security risks of connecting OT systems with mainstream IT. This is the OT-IT gap — the chasm between the heterogeneous, chaotic world of industrial IoT and the uniform, managed world of IT.
Today, AI is driving greater demand for operations data. Motivated by this, enterprise software suppliers are scrambling to find efficient and expedient ways to bridge the OT-IT gap. The solution is surprisingly simple. Instead of trying to push IT technologies into the OT world, suppliers are now moving to a straightforward “data first” mindset, which is a new way of thinking about OT integration. For years, developers have struggled to turn customized, complicated, expensive, hard-coded, application-specific device management and connectivity mashups into “end-to-end solutions.” Unfortunately, the past decade of IIoT projects taught us that this approach does not scale.
Developers are now turning to a better alternative: bridging OT and IT with simple interfaces for device identity, security, data, events and status. This approach simplifies OT data access and enables embedded OT software to evolve independently from cloud-native IT systems. Multimodal AI applications further reduce OT-IT integration costs by ingesting diverse types of machine data as-is, reducing the need for costly data transformation.
Recent product announcements and integration demos from AWS, Google, Honeywell, Microsoft, Qualcomm and other major cloud frameworks and ERP suppliers confirm this trend. (I’ll cover this topic in a follow-up article.) The goal is clear: feed the rapidly growing market for AI-enhanced business transformation with massive amounts of OT data via standard protocols and simple APIs. In other words, grab OT data without redesigning or intrusively modifying IIoT devices.
Looking Ahead To A World Of Robotic Enterprise
“I have a perfect model of the world. It’s actual size.” Comedian Steven Wright’s one-liner metaphorically describes building virtual replicas (i.e., digital twins) of physical objects, systems, and processes that look real and accurately simulate real-world behavior. It’s not so far-fetched — physical AI models really do appear to be “actual size” when viewed in 3-D. Adding OT data brings these models to life, accurately simulating complex scenarios in near real time. Physical models trained with dynamic simulations should enable a new generation of AI-powered applications to deliver step-function improvements in process efficiency, worker safety, equipment uptime, product quality, decision making and other high-value use cases. These use cases deliver impressive ROI — at least on paper and in the lab. However, physical AI technology the likes of Nvidia Cosmos is new, and interfacing with OT equipment is often problematic for all the reasons touched on above, so the timeline is still uncertain.
Here’s my take on a reasonable schedule for adopting these technologies. Nvidia’s AI ecosystem is solid, and the company has a lot riding on physical AI. Plus the company has become a prime mover in AI, and it is blessed with a mountain of money and a very, very long list of enterprise customers thirsty for its wares. All of that makes Cosmos and Omniverse good bets, and indeed customers are already developing solutions on the platform. For instance, my firm’s CEO and chief analyst Patrick Moorhead recently wrote about how Nvidia, Accenture and KION are collaborating to digitize warehouse operations. He shares my optimism about Nvidia’s physical AI platform. Enterprises with physical infrastructure can begin using these tools immediately while planning on large-scale deployments within the next year or two.
However, I am less optimistic about the economics and timelines for connecting OT data sources to physical models. The IIoT connectivity barrier is coming down, but not fast enough to keep up with physical AI growth. The problem arises from the diversity of OT and IIoT. IT systems have uniform architectures, but OT systems do not. In line with what I explained earlier, I recommend bridging the OT-IT gap with simple interfaces that gather OT data as-is rather than intrusively customizing IIoT devices. The good news here is that the focus on physical AI by Nvidia and others provides more impetus to achieve these connections. I predict significant improvement in AI-OT-IT connectivity this year as physical AI makes enterprises increasingly robotic and increases the urgency of bridging the OT-IT gap.
