Yes—but it might not be a new company at all. The next defining platform in robotics isn’t likely to be about building better robot hardware. It’s about building the software layer that controls any robot, regardless of its mechanical design. NVIDIA is already ahead in this race with its Physical AI stack. But companies like Skild AI and Physical Intelligence are raising enormous sums to challenge that dominance, suggesting the control layer market is big enough for multiple winners—and that whoever owns this layer will have outsized influence over the entire robotics industry. Consider this: Skild AI just raised $1.835 billion to deploy generalized robot intelligence across ABB Robotics and Universal Robots systems.
That funding level mirrors what infrastructure platform companies raise, not robot manufacturers. The companies betting the biggest money are betting that control software will be the next indispensable layer. The distinction matters because hardware manufacturers have been fragmented for decades. KUKA competes with FANUC. Universal Robots competes with ABB. But if one company can write software that makes a KUKA arm work the same way as a FANUC arm, or that lets the same AI model control both—suddenly the hardware differences become commoditized. That’s what happened in computing when operating systems became valuable than processors, and it’s starting to happen in robotics right now.
Table of Contents
- Why Control Software Is Becoming More Important Than Hardware
- The Race to Build Universal Robot Control Stacks
- How NVIDIA Built Its Physical AI Advantage
- The Emerging Challenger Companies
- The Limits of Control Layer Dominance
- Industrial Integration: Where the Money Is Made
- The Future of Robot Control Software
- Conclusion
Why Control Software Is Becoming More Important Than Hardware
For most of industrial robotics’ history, competitive advantage came from mechanical precision, speed, and reach. A robot arm that could maintain tolerances of 0.03 millimeters won customers. Build a faster six-axis arm, and manufacturers bought it. But the robotics industry has matured to the point where dozens of companies can build mechanically competent arms. The differentiator has shifted to what those arms can be instructed to do, and how quickly they can be reprogrammed for new tasks. nvidia‘s CEO Jensen Huang made this explicit at GTC 2026: “Every industrial company will become a robotics company.” What he meant is that companies like General Motors or Foxconn don’t care whether they own a KUKA, a FANUC, or a proprietary arm in five years. They care about the software that tells the robot what to do.
Right now, ABB Robotics, FANUC, KUKA, and Yaskawa operate an installed base of roughly 2 million industrial robots worldwide. All four of these traditional manufacturers are now integrating NVIDIA Omniverse libraries into their production systems, signaling that even legacy robotics companies accept the control software layer is becoming the competitive battleground. The financial numbers support this shift. Skild AI and Physical Intelligence combined have raised more than $2.3 billion in recent funding rounds to build control software—more than some entire robot hardware companies are worth. Neither Skild AI nor Physical Intelligence manufactures robots. They’re building the brain that lives between the sensor inputs and the motor commands. That’s where the power is consolidating.
The Race to Build Universal Robot Control Stacks
NVIDIA’s approach is to release increasingly capable foundation models designed specifically for robot control. Their Isaac GR00T N1.6 and N1.7 are open reasoning vision-language-action models that enable full-body control across different humanoid robot platforms. Unlike traditional robot programming, which requires writing unique code for each hardware configuration, these models aim to work across different mechanical designs. NVIDIA also released Cosmos 3 and Alpamayo 1.5 as frontier prediction models—these are trained to anticipate how robots will affect their environment, which is essential for planning complex multi-step tasks. But here’s a significant limitation: universal control is harder in practice than the funding rounds suggest. A model that works on a Boston Dynamics humanoid might not immediately transfer to a KUKA industrial arm because the sensing hardware is different, the kinematics are different, and the deployment environments are different. NVIDIA’s Isaac Lab-Arena is an open-source simulation framework designed to address this by letting companies safely test robot capabilities in virtual environments before deploying to physical hardware.
The problem is that sim-to-real transfer still isn’t solved. A robot trained in simulation often fails on real hardware because the virtual world is never quite accurate enough. Companies are building data factories to address this—NVIDIA announced its Physical AI Data Factory Blueprint, which is an open reference architecture for automating the generation of training data and evaluation metrics. The architectural question is whether one company’s control layer can ever truly dominate this space the way operating systems dominated computing. In computing, Windows and Linux were so useful that they made hardware nearly invisible. In robotics, the diversity of hardware, use cases, and safety requirements may be too large for a single winner. KUKA’s new iiQKA.OS2 operating system, unveiled in 2026 with a virtual robot controller and AI-readiness, suggests even traditional manufacturers aren’t willing to cede complete control to an external software layer.
How NVIDIA Built Its Physical AI Advantage
NVIDIA’s dominance in robotics software stems from a deliberate vertical integration strategy. They don’t just provide models—they provide the entire pipeline: simulation (Isaac Sim), foundation models for control (GR00T), prediction models (Cosmos), and hardware specifications (Jetson Thor as the edge AI compute backbone). This mirrors how NVIDIA dominated AI graphics processing. Just as every large language model needed GPUs, every advanced robot system is increasingly built around NVIDIA’s Jetson processors running their software stack. A concrete example illustrates this advantage: when Texas Instruments announced its collaboration with NVIDIA on March 5, 2026, to integrate mmWave radar with NVIDIA Jetson Thor and Holoscan, they weren’t announcing a partnership of equals. They were announcing that to build low-latency 3D perception for humanoid robots, you need TI’s sensor expertise combined with NVIDIA’s compute and robotics software stack.
The integration works because NVIDIA has already written the Holoscan SDK—the software framework that connects sensors to compute to control models. Companies building robots can now buy TI radar, Jetson Thor compute, and NVIDIA’s software as an integrated solution. This lock-in effect is powerful. The challenge for NVIDIA is that this vertical integration can become a vulnerability. Companies like Skild AI and Physical Intelligence are explicitly building heterogeneous systems—control software that claims to work with any robot hardware from any vendor. If they succeed, they could commoditize NVIDIA’s proprietary advantage. But NVIDIA’s ecosystem advantage is self-reinforcing: the more robots run on Jetson hardware, the more companies build tools for that hardware, and the more expensive it becomes to build alternatives.
The Emerging Challenger Companies
Skild AI’s $1.835 billion raise and partnerships with ABB Robotics and Universal Robots represent a direct bet against NVIDIA’s vertical integration model. Skild is positioning itself as hardware-agnostic—their generalized robot intelligence should theoretically work across industrial arms from different vendors. Universal Robots and ABB Robotics, which compete with each other, both backed Skild. That’s unusual. Neither vendor would normally fund a third party that could disrupt their proprietary advantages unless they believed that control software standardization was inevitable and that being left out of it would cost them more. Physical Intelligence took a different path with its $470 million raise. Rather than targeting industrial robots, they’re focused on building universal AI control for embodied systems broadly—humanoids, mobile manipulators, even autonomous vehicles. Their explicit goal is to create a control stack that’s general enough to handle the diversity of hardware in robotics without requiring hardware-specific tuning. The tension between Skild’s industrial focus and Physical Intelligence’s broader embodiment thesis mirrors the historical tension between specialized and general-purpose computing.
Skild is betting that industrial robotics is a large enough market to deserve dedicated optimization. Physical Intelligence is betting that generalization across more hardware categories will eventually prove more valuable. The tradeoff these challengers face is real. Specializing in one domain (industrial arms) lets you optimize for that domain’s constraints and safety requirements, but it limits your total addressable market. Being general-purpose lets you address a larger market, but you risk being less competitive in any specific domain. FANUC, the world’s largest industrial robot vendor, is taking a middle path with their partnership with Inbolt, a French AI startup. Together, they deployed precision robots on continuously moving assembly lines—a notoriously difficult control problem—and General Motors was the first adopter. FANUC isn’t betting its business on a single software vendor. They’re integrating AI partners to enhance their own offerings while maintaining proprietary control.
The Limits of Control Layer Dominance
Before declaring a winner in the robotics control layer race, understand that this isn’t like the operating systems war. An operating system is abstract—it works the same whether you’re computing on a laptop, server, or phone. Robots are physical. A control layer that works for humanoid robots might be fundamentally incompatible with a SCARA arm or a mobile manipulator because the kinematics, dynamics, and safety constraints are completely different. Building truly universal control software requires solving problems that computing platforms never had to solve. There’s also a practical limitation: hardware manufacturers have decades of IP locked into their motion control architecture. KUKA’s new iiQKA.OS2 includes a virtual robot controller—this isn’t an accident. It’s an architectural choice to create a boundary between the underlying proprietary control layer and third-party software that sits on top.
This means even if you have a perfect control model, you still have to integrate it with each manufacturer’s virtualization layer. That integration work becomes a barrier to entry that favors established vendors with good relationships to robot manufacturers. Safety certification is another bottleneck often overlooked. Industrial robots must comply with ISO 10218:2026 and similar standards. Any control layer that touches safety-critical functions has to go through certification for each robot type and use case. This certification burden is expensive and time-consuming. It’s one reason why FANUC, KUKA, and ABB haven’t been displaced despite their mediocre software—the certification moat is real. Startup companies can release software that claims to be universal, but until they have certified implementations working on production hardware in real factories, they remain theoretical advantages.
Industrial Integration: Where the Money Is Made
The proof that control layer businesses can scale comes from looking at where integrations are actually happening in production. FANUC and Inbolt’s work with General Motors on continuously moving assembly lines is significant because it solves a real, hard problem. Previous generations of robots couldn’t reliably track and manipulate parts on moving conveyor belts without constant human reprogramming. The AI control system that FANUC and Inbolt deployed can learn from examples and adapt to variations in part position and orientation.
General Motors using this in production means it’s not a research project—it’s generating real revenue and solving real manufacturing problems at scale. This is where the value actually gets captured. Companies like Skild AI and Physical Intelligence can raise massive funding rounds because they’re addressing a pain point that manufacturers understand viscerally: reprogramming robots for new tasks is expensive and slow. A universal control layer that dramatically reduces programming time has clear ROI. But the path from “capability” to “production revenue” is long and expensive, which is why the companies succeeding in this space are either well-capitalized startups or partnerships between startups and established hardware vendors.
The Future of Robot Control Software
Over the next three to five years, expect the robotics control layer to look more like the graphics hardware market than the operating systems market. NVIDIA dominated GPU computing not by having a perfect product but by having an ecosystem that other companies couldn’t easily replicate. Multiple control software companies will likely succeed by specializing in different domains: one for industrial arms, one for humanoids, one for mobile manipulation. The real competition won’t be winner-take-all. It will be about which companies can best integrate with hardware vendors and build the trust required for safety-critical deployments.
NVIDIA’s position is strong but not invulnerable. Their weakness is that they’re building top-down—starting with a general foundation model and trying to make it work on specific hardware. Companies like Skild AI and Inbolt’s partnership with FANUC are building bottom-up—starting with specific manufacturing problems and generalizing from there. Bottom-up approaches often win in industrial markets because they solve real problems that customers understand. The control layer winner in robotics might not be a single company. It might be a consortium or an ecosystem where multiple companies specialize in different domains and hardware categories work well with each.
Conclusion
The next NVIDIA in robotics isn’t necessarily a new company, and it may not have a single identity. NVIDIA itself is a legitimate candidate because they’re investing in the entire stack and integrating with hardware manufacturers globally. But the size of the opportunity and the diversity of robotics applications create space for challengers like Skild AI and Physical Intelligence to win in specific domains.
The outcome will depend not on which company has the best research, but on which companies can move fastest from capability to production-grade, certified deployments that actual manufacturers trust. For people working in robotics, the key insight is this: if you build hardware, control software is becoming the factor that determines whether your product is competitive. If you build software, the path to market success isn’t through generic generalization—it’s through solving specific, hard problems that manufacturers understand and will pay to solve. The control layer businesses that win will be the ones that stay closest to actual production manufacturing problems, not the ones that claim the most general capabilities.
