The next NVIDIA in robotics may not exist yet—but the companies racing to build foundational AI models and hardware for physical automation are drawing investor attention at unprecedented scales. Figure AI, valued at $39 billion after NVIDIA’s participation in a Series C funding round exceeding $1 billion in September 2025, represents the closest parallel to NVIDIA’s dominance in computing infrastructure, positioning itself as the platform layer for humanoid robotics. Where NVIDIA created the architectural advantage for AI training through GPUs, Figure AI and select competitors are attempting to own the base models and simulation layers that will govern how robots learn, move, and solve problems in the real world.
The distinction matters because NVIDIA has already staked its claim in the robotics layer itself. NVIDIA released its Isaac GR00T foundation model—an open model that brings humanlike reasoning to robots, enabling them to break down complex instructions and execute tasks using prior knowledge and common sense. The company also rolled out Newton, a GPU-accelerated physics engine released in 2026 that enables simulation of complex robot actions like walking through snow or handling delicate objects. CEO Jensen Huang articulated the vision clearly: “Physical AI has arrived—every industrial company will become a robotics company.” This shift fundamentally changes the competitive landscape, where the next dominant player won’t necessarily look like the last one.
Table of Contents
- Who Is Solving the Hardest Problems in Robotics Right Now?
- The Physics Simulation Bottleneck and Why It Matters
- Foundation Models for Reasoning vs. Foundation Models for Control
- Investment Patterns and Who’s Betting on What
- Why Simulation and Real-World Performance Still Don’t Align Perfectly
- The Role of Open Models in Competitive Advantage
- The Next Five Years and Market Consolidation Ahead
- Conclusion
- Frequently Asked Questions
Who Is Solving the Hardest Problems in Robotics Right Now?
Figure AI represents the most visible bet on humanoid robotics, but the problems it’s solving aren’t unique to that company. Boston Dynamics, acquired by Hyundai, continues advancing bipedal and quadrupedal locomotion. Caterpillar is integrating nvidia technology into autonomous heavy machinery. Franka Robots and NEURA Robotics, both NVIDIA partners, are pushing dexterous manipulation—the ability to handle fragile or complex objects—further into practical deployment.
The “hard problems” these companies face overlap significantly: real-world sensor noise, dynamic environments where simulation doesn’t match reality, and the economics of scaling production. What separates the leadership contenders is their approach to the foundation model layer. NVIDIA’s release of Isaac GR00T demonstrates that the company views robotics not as a hardware problem but as a reasoning problem. A robot that can interpret “pick up the egg without breaking it” requires far more than mechanical precision; it needs to understand material properties, force application, and task decomposition. Figure AI’s $39 billion valuation reflects investor belief that humanoid form factors will dominate industrial and service robotics, though this premise remains contested by specialists who argue that task-specific morphologies (wheeled bases for logistics, articulated arms for manufacturing) will remain superior for decades.
The Physics Simulation Bottleneck and Why It Matters
One of the most underestimated barriers to deploying robots at scale is the gap between simulation and reality—often called “sim-to-real transfer.” A robot trained entirely in a physics simulator will fail catastrophically on real surfaces, real objects with imperfect geometry, and real lighting conditions. NVIDIA’s Newton physics engine attempts to solve this by providing GPU-accelerated simulation accurate enough that policies trained in simulation transfer more effectively to hardware. However, this remains incomplete. Robots trained on Newton simulations still require real-world adaptation, fine-tuning, or reinforcement learning on actual hardware to achieve production reliability.
The limitation here is critical: even with state-of-the-art physics simulation, deploying a new robot to a new task still requires empirical validation. A manufacturing facility can’t rely entirely on simulated data to validate that a robot arm won’t drop parts on a production line. This means companies solving these problems must either accept longer deployment timelines, invest heavily in high-fidelity sensor data collection, or develop hybrid approaches where simulation accelerates learning but hardware testing remains mandatory. The economic pressure to reduce deployment time is intense, which is why partners like Caterpillar and Franka Robots are moving quickly—but not recklessly.
Foundation Models for Reasoning vs. Foundation Models for Control
Isaac GR00T occupies an interesting position in the robotics stack. It’s a reasoning model, not a control model. GR00T breaks down high-level instructions (“organize this workbench”) into subtasks and intermediate goals, but it doesn’t directly command motors or joints. That layer remains specialized, often relying on traditional robotics software or smaller, task-specific neural networks.
This separation reflects a genuine architectural insight: the models that understand language and reasoning are different from the models that need real-time, low-latency motor control. Companies like Figure AI are building across both layers—reasoning for task planning and control policies for physical execution. This dual-stack approach mirrors NVIDIA’s strategy of providing tools at multiple levels (foundation models, physics engines, simulation platforms) rather than betting on a single architectural paradigm. The risk of this approach is fragmentation; if the reasoning layer and control layer don’t integrate smoothly, a robot might generate a valid plan that its motor controllers can’t execute. Real deployments have encountered this problem repeatedly, where AI planning assumes capabilities the underlying hardware doesn’t actually have.
Investment Patterns and Who’s Betting on What
Figure AI’s $39 billion valuation places it in rare air. For context, Boston Dynamics—arguably the most advanced robotics company from a technical standpoint before Hyundai’s acquisition—never reached that valuation independently. The difference is strategic focus: Figure AI is explicitly building humanoid robots as a platform, betting that task generality matters more than task specificity.
NVIDIA, by contrast, is distributing its robotics bet across multiple partners and companies, reducing single-company risk while maintaining leverage across the ecosystem. This divergence in investment strategy reflects genuine uncertainty about market structure. Will robotics follow a platform winner-take-most model like smartphones (one or two dominant form factors dominating 80% of market value), or will it remain vertically fragmented (different industries adopting different robot morphologies and suppliers)? NVIDIA’s multi-partner approach hedges this uncertainty, but Figure AI’s bet on humanoids as the general-purpose default is bolder and potentially more valuable if correct. The tradeoff is that Figure AI’s success depends on humanoid robots becoming economically viable at scale, a threshold the industry hasn’t yet crossed.
Why Simulation and Real-World Performance Still Don’t Align Perfectly
The robotics industry has spent decades building simulation environments—Gazebo, V-REP (now CoppeliaSim), MuJoCo—but the gap between simulated and real performance persists. NVIDIA’s Newton is more physically accurate than most open-source alternatives, but “more accurate” doesn’t mean “accurate enough.” A robot trained in Newton to stack blocks might perform flawlessly in simulation yet struggle with real blocks that have slightly different surface friction, slight warping, or dust. This limitation has profound implications for deployment timelines and costs.
A company deploying Figure AI or another advanced robot to a manufacturing floor must budget for an adaptation period, where the robot learns the specific quirks of that environment. This isn’t a flaw unique to these companies—it’s inherent to robotics—but it means that the “next NVIDIA” in robotics will likely win by dramatically reducing this adaptation cost. Companies that can deliver robots requiring minimal or no real-world retraining will command pricing power and market share. Current contenders haven’t yet solved this; they’ve only made incremental progress.
The Role of Open Models in Competitive Advantage
NVIDIA’s decision to release Isaac GR00T as an open model reflects a strategic choice: capture the ecosystem advantage rather than monopolize the capability. This mirrors NVIDIA’s approach to CUDA; by making powerful tools accessible, NVIDIA increased the size of the market for the GPUs those tools require. Figure AI, by contrast, is building proprietary models and hardware, betting on vertical integration and proprietary advantage.
The open-model approach has a historical advantage in software (Linux defeated proprietary Unix variants), but robotics adds a hardware dimension that changes the calculus. A robot is both software and physical machinery. Figure AI’s humanoid robots are hardware products; they can’t be commoditized by openness the way CUDA-competing frameworks haven’t displaced NVIDIA from AI infrastructure. This suggests both strategies will coexist: NVIDIA providing the foundational tools and simulation layers (where openness drives adoption), while companies like Figure AI capture value through specialized hardware and integrated solutions.
The Next Five Years and Market Consolidation Ahead
The robotics industry is entering a phase where consolidation seems inevitable. NVIDIA is investing across multiple companies; Hyundai owns Boston Dynamics; Microsoft, Amazon, and Google are building robotics divisions. The question isn’t whether a “next NVIDIA” will emerge, but rather whether the market will support multiple dominant players or converge on a handful of winners. Current valuations suggest investor confidence that humanoid robotics will represent a multi-trillion-dollar market, large enough for several major players.
The companies best positioned to dominate in the next five years are those solving the sim-to-real gap most effectively and building integration advantages across reasoning, control, and deployment. Neither pure software play nor pure hardware play appears sufficient. NVIDIA’s advantage lies in serving all players; Figure AI’s advantage lies in vertical integration. The eventual winner may simply be the company that deploys robots at scale first, proving that the economics work, and then defends that position through service, software updates, and ecosystem lock-in.
Conclusion
The next NVIDIA in robotics won’t be determined by technical capability alone—NVIDIA itself has ensured that multiple companies now have access to excellent foundation models and simulation tools. It will be determined by execution: which companies can take the hard-solved problems (reasoning, physics simulation, motor control) and integrate them into deployable systems that solve real customer problems at viable cost. Figure AI represents the most visible bet in this direction, but the outcome remains genuinely uncertain.
The robotics industry is at an inflection point where the platform layer is shifting from pure hardware (chips) to hybrid software-hardware stacks (models, simulators, specialized robots). Companies that win this transition will need excellence in AI, robotics engineering, manufacturing, and customer integration simultaneously. That’s a genuinely hard problem, which is precisely why the next dominant player in this space will be as noteworthy as NVIDIA has been in AI infrastructure.
Frequently Asked Questions
Is Figure AI the next NVIDIA in robotics?
Figure AI represents the most visible bet on becoming a dominant platform in robotics, with a $39 billion valuation and NVIDIA backing. However, it competes in a different layer than NVIDIA (hardware and integrated systems vs. foundational AI tools), so the comparison is imperfect.
Why did NVIDIA release Isaac GR00T as an open model?
Open models increase ecosystem adoption and expand the market for complementary NVIDIA products like the Newton physics engine and GPU compute resources. This mirrors NVIDIA’s strategy in AI infrastructure.
What’s the sim-to-real problem in robotics?
Robots trained in physics simulators often fail in real-world conditions due to differences in sensor noise, surface properties, object geometry, and environmental variables. Even advanced simulators like Newton don’t fully eliminate this gap.
Which companies are leading in robotics besides Figure AI?
Boston Dynamics (Hyundai-owned), Franka Robots, NEURA Robotics, and Caterpillar are all advancing robotics with NVIDIA partnerships. Each focuses on different applications and form factors.
Will humanoid robots dominate industrial robotics?
This remains contested among specialists. While humanoid form factors offer task generality, task-specific robots (articulated arms, wheeled bases) may remain superior for many industrial applications for decades.
How long until deployed robots require no real-world adaptation?
Significant progress is being made, but current systems still require adaptation to new environments. The company that dramatically reduces this requirement will likely achieve major competitive advantage.
