What are the top operators and VCs saying about the future of Robotics & AI?

A consensus is forming that the primary bottleneck in robotics has shifted from hardware to software, with the core challenge now centered on developing "physical intelligence" using AI foundation models [8, 15, 23, 25]. Hardware is increasingly viewed as commoditized and sufficiently capable for many tasks [15, 22]. Consequently, progress is gated by the ability to acquire and leverage massive datasets to train these models . The industry is moving away from slow, unscalable data collection methods like teleoperation, which some experts predict will become a negligible data source in the **next year or two** [5, 19]. The emerging paradigm is "sensorized human data," primarily using egocentric video from humans performing tasks to create datasets at a scale that could enable a neural scaling law for physical dexterity, similar to what has been achieved with LLMs [4, 5].

To process this new scale of data, model architectures are evolving significantly. A proposed paradigm shift is underway from Vision Language Action (VLA) models, which are criticized for prioritizing language over an understanding of the physical world, to World Action Models (WAMs) [4, 9]. WAMs are designed to learn physics emergently by simulating and predicting future world states from video, enabling a more robust and generalizable foundation for physical action . This learning process is heavily reliant on simulation, with a long-term vision of moving beyond classical physics engines to fully neural simulators that learn physics directly from video data, converting compute into a nearly infinite source of training environments [3, 10]. However, a key limitation is that current generative simulation technology is not yet useful for training complex manipulation tasks because it fails to accurately model contact physics, making real-world interaction data a crucial competitive moat [17, 13].

While the focus is on AI, strategic disagreements persist regarding hardware form factors and market structure. One school of thought argues the future lies not in a single humanoid model but in **thousands of different, specialized** hardware form factors, each optimized for energy efficiency and specific tasks in domains from manufacturing to medicine [1, 3, 6, 21]. An alternative view favors generalist, platform-based approaches, envisioning a future where intelligence can be infused into hardware almost instantly, analogous to software development [7, 20]. This debate extends to market dynamics, where some analysts believe incumbents with deep capital and supply chain expertise are structurally favored to win in sectors like autonomous vehicles, presenting a significant hurdle for startups . This contrasts with the view that the current moment is ripe for new ventures, as the convergence of commoditized hardware, advanced edge compute, and generalizable foundation models creates new opportunities .

The investment landscape reflects both immense optimism and significant uncertainty. VCs and operators foresee robotics becoming the largest category within AI and potentially expanding the total addressable market for technology by a **factor of 100** [26, 30]. The industry is often described as being in a pre-ChatGPT moment, where a scalable recipe exists but has not yet been fully productized, with some predicting a "GPT-3 moment" of large-scale deployment is imminent [24, 27]. However, this enthusiasm is tempered by the lack of predictable scaling laws that can connect investment in data and compute to specific capability improvements, a formula that has been crucial for attracting massive capital to LLMs [8, 14, 16]. This makes robotics a higher-risk research endeavor, with some predicting that many recent investments will fail because the slow, capital-intensive learning cycles of the physical world do not align with software-based investment playbooks [16, 28, 29]. Investment strategies therefore diverge, with some advocating for "picks and shovels" plays like simulation platforms and others for software-centric, platform-based approaches [12, 18, 20].