Generalist robotics models trained on diverse, cross-platform data are fundamentally superior to specialist models trained on single platforms.
A cloud-hosted AI model is the most effective and scalable architecture for controlling physical robots today, and latency can be managed through techniques like action pre-fetching.
The primary bottleneck for AI advancement is the lack of physical world interaction data, a gap that robotics is uniquely positioned to fill.
The economics of the robotics industry are rapidly changing, lowering barriers to entry and creating a massive market opportunity that could significantly impact GDP.
True autonomy will be achieved through a mixed-autonomy approach, where systems are deployed in the real world to learn incrementally from edge cases and human intervention.
▶The Supremacy of Generalist, Cross-Embodiment ModelsApr 2026
Wong champions the idea that the path to capable robotics AI lies in training large, generalist models on data from many different robot types. He cites the OpenX project, which showed a 50% performance improvement over specialist models, and the emergent zero-shot capabilities of Physical Intelligence's models as evidence for this cross-embodiment approach.
This theme suggests that companies with access to the most diverse robotics datasets, rather than those with the most units of a single hardware platform, may hold a significant competitive advantage in developing foundational AI for robotics.
▶Cloud-First Robotics ArchitectureApr 2026
Wong reveals that Physical Intelligence's strategy is to run its AI models in the cloud, rather than on the robot's local hardware. He addresses the potential latency issues by explaining a pre-fetching technique and highlights the benefit of being able to integrate with partner hardware without deep, system-specific knowledge.
This architectural choice decouples the AI brain from the physical body, potentially accelerating AI development and allowing for a more hardware-agnostic business model focused on providing intelligence as a service.
▶The Business of General-Purpose RoboticsApr 2026
Wong frames the mission of Physical Intelligence as building a single AI to control any robot for any task, a goal he believes could contribute 10% to US GDP. He acknowledges the primary risk is the problem taking decades to solve, but points to accelerated progress, such as considering real-world deployment years ahead of schedule, as a reason for optimism.
Investors should note the high-risk, high-reward nature of this venture; success hinges on a fundamental breakthrough in general AI, but the potential economic impact is framed as transformative.
▶An Impending 'Cambrian Explosion' in RoboticsApr 2026
Wong predicts a massive proliferation of specialized robotics companies due to the falling costs and reduced expertise needed to enter the market. He sees this as a major business opportunity, not just for vertical applications but also for services like tele-operation and data collection that support this growing ecosystem.
This prediction implies that the most valuable long-term plays in robotics may not be in building specific hardware, but in creating the platforms, tools, and foundational models that empower this new wave of robotics entrepreneurs.