When we sat down to whiteboard @Propheusai, we knew the answer was to build in the Physical AI space. The problem: The world is messy and we saw how enterprises across the world were struggling to capture pieces of the physical world in chunks - it's like gathering all these ingredients, but they all expire before the pie is ever made. Demographics, weather, people sentiment, people movement, events, competitor activity - all critical to spot opportunities that have topline and bottomline impact for enterprises. The numbers back what we felt in that room. The global Physical AI market is projected to cross $1 trillion by 2035, growing north of 33% a year. All these robots running around us - but how do they make sense of the messy physical world around them, in real time? Capturing this meant building our Digital Atlas - Agents that not only collect but can make sense of the physical world in ways that uniquely suit enterprise needs. And this intelligence isn't just for robots - Agents and humans both act in the physical world too, and need it now.  This is Physical AI, at its quietest, but most potent. Link in the comments to learn more!

Physical AI: An Ecosystem that means so much more than Robotics If you've been following the Physical AI space, you must have noticed that VC funding in the sector has seen a meaningful uptick lately. But behind the scenes of the current glamor and spectacle around Physical AI, those who have keenly followed the industry know that the category has always been a difficult one. There was a time when Physical AI simply meant robotics. And when that was true, it meant building narrow, pre-programmed machines - brittle, single-task, controlled-environment experiments. But today, the promise of Physical AI is that robots, embodied machines, and even software agents can act in the real world, conduct meaningful tasks, and do so at scale. Physical AI is playing out in the physical world - where, as @NextBigTeng puts it, "AI perceives, decides, and acts." And this time, the potential seems more promising than ever, because critical parts of this system are being built in parallel, compounding toward several real-world use cases, all at once. Here's what's really compounding: 1. Foundation Models for the Physical World We're seeing the emergence of a new class of AI models purpose-built for physical action - Vision-Language-Action (VLA) models that let a robot see an environment, understand a command in natural language, and translate that into a physical motor action, in real time. Key players: Physical Intelligence (π), Figure AI (Helix), Google DeepMind (Gemini Robotics), NVIDIA (GR00T), Amazon 2. Breaking the Training Data Bottleneck This was the hardest wall in robotics for years. LLMs got trained on the entire internet. Robots don't have that luxury. Teaching a robot to pick up a cup requires actually doing it thousands of times - capturing every sensor signal, every motor torque, every failure. That's slow, expensive, and doesn't generalize. The wall is finally breaking, from multiple directions at once. Simulation lets companies train robots in photorealistic virtual environments at massive scale. Teleoperation platforms let humans demonstrate tasks that get recorded as training data. Synthetic data generation creates realistic scenarios programmatically. Investment in world models, which enable robots to predict and plan autonomously - surged from $1.4B in 2024 to a record $6.9B in 2025. Deployed robots continually feed back data to retrain and improve the model, forming a continuous learning loop. Key players: NVIDIA (Cosmos, Omniverse, Isaac Sim), Cortex AI, Scale AI, Skild AI, Axis Robotics, World Labs 3. The Inference Layer Making Intelligence Fast Enough to Act Even the best robot brain is useless if it can't act in real time. A robot on a factory floor or a surgical table cannot wait 2 seconds for a cloud server to respond. It needs to think and move in milliseconds, locally, on the device itself. This is the edge inference problem, and it's being solved now. New chips and on-device compute architectures are making it possible to run complex foundation models directly on the robot, with no cloud dependency. Key players: NVIDIA (Jetson AGX Thor), Figure AI, Tesla (custom AI chips for Optimus), Qualcomm 4. The Hardware Layer: Bodies Getting Cheaper and Better None of the above matters if the physical robot itself costs $500K, breaks constantly, or can only do one thing. Hardware has historically been the expensive, slow-moving constraint. That's changing. Goldman Sachs reports humanoid manufacturing costs dropped 40% between 2023 and 2024. Bank of America projects material costs falling from ~$35,000 in 2025 to $13,000–$17,000 per unit within the decade. Multiple form factors are maturing simultaneously - humanoids, quadrupeds, arms, mobile manipulators - each suited to different environments. Key players: Tesla (Optimus), Unitree, Figure AI, Boston Dynamics (electric Atlas), 1X Technologies, Agility Robotics 5. The Cognition Layer: Understanding the World They Operate In This is the most underappreciated layer, and in my view, very important for real-world deployment at scale. A robot can have a great foundation, fast inference, cheap hardware, and rich training data, and still fail the moment it steps into an environment it wasn't trained on. Why? Because the physical world is dynamic. It changes constantly. Weather shifts operations. People move differently at different times of day. Events disrupt normal patterns. Sentiment affects how humans interact with machines. This is exactly the gap @Propheusai is building for, a live knowledge representation of the real world (weather, demographics, people movement, sentiment, events) that gives both embodied AI systems and software agents the contextual grounding to truly act, not just execute. Physical AI systems can only fully compound when this layer is complete. And our team including @Sumanth_n83 , @sachinjose, @michellemzhou and @sreejan_c are excited to be building this layer - the Digital Atlas - the most comprehensive knowledge representation of every place on earth. Stay tuned. I'm also digging into what I'm calling the blue collar vs. white collar applications of Physical AI, and why I believe the ROI of Physical AI isn't decades away. It's already here - and we’re seeing it everyday. Enabling this is the cognition layer and its foundation is our Digital Atlas. More on this coming soon!

We had done a benchmarking study a few months ago testing 20+ models for Enterprise AI performance at @Propheusai and even then, while Qwen was not mainstream (and Kimi/GLM didn’t mean anything), it was still neck and neck with Claude and GPT. It’s no surprise now that it has become a model of choice given its open source roots and highly versatile family of models. If someone had said in 2023 that China would lead in open source models while US would be primarily closed source, we’d have laughed. Yet, here we are.. @michellemzhou @Sumanth_n83

Great essay by @drfeifei on world models. At @Propheusai, @michellemzhou and I have a strong belief that this is the next frontier of AI innovation in the coming decade. A major application of using real world representations is to guide and simulate the next state (actions and impact) on Enterprise businesses that exist in the physical world - Retail, Consumer Goods, Real Estate. LLMs lack the context and the grounding needed to solve these large scale problems. Excited to see what comes next from @theworldlabs !