Why Data Diversity, Not Data Volume, Will Decide Who Wins Physical AI

TL;DR: Nvidia expects its physical AI revenue to grow tenfold over the next decade (Davies, 2026). At Encord, we agree the opportunity is real, but we think the industry is underestimating the constraint standing in the way of it. Digital AI workflows like law, coding, and finance share a single modality (keyboard and screen), which is why internet-scale training data already existed for them. Physical AI has no equivalent: a farmer, a surgeon, and an electrician operate through entirely different tools, materials, and environments. Research on foundation models in robotics confirms this "embodiment heterogeneity" as a core open problem. The winners in physical AI won't be the teams that collect the most data, but the ones with infrastructure built to handle how different that data needs to be.

Nvidia told the Financial Times this week that it expects its physical AI revenue to grow tenfold over the next decade, from $10 billion to $100 billion. Deepu Talla, the company's vice president of robotics and edge AI, put it simply: ninety percent of the world's actions happen in the physical world, not the digital one, and that is why the opportunity is arguably an order of magnitude larger than anything digital AI has captured so far (Davies, 2026).

Nvidia founder Jensen Huang has gone further, predicting that every industrial company will eventually become a robotics company (Davies, 2026). Elon Musk is redirecting tens of billions of dollars at Tesla toward robotaxis and humanoids. Manufacturing startups offering "robotics-as-a-service" are attracting valuations that look more like software companies than industrial equipment makers (Davies, 2026).

At Encord, we agree with the scale of the opportunity. But we think the industry is underestimating what stands in the way of it, and it's not compute.

The problem isn't data volume. It's data diversity.

Many of the most economically valuable digital AI workflows, law, coding, and finance among them, share a convenient property: they operate through the same modality. A lawyer and a programmer both work through a keyboard and a screen. That shared interface is a large part of why large language models scaled as fast as they did. The training data was already sitting on the internet, in a consistent format, in enormous quantities.

Physical AI does not have that luxury. A farmer, a surgeon, and an electrician do not share a keyboard, a set of tools, or an environment. Even within one task category, like picking up an object, handling a rigid component is a different problem from handling fabric, cable, or any other deformable surface.

Ken Goldberg, a robotics professor at UC Berkeley, made this point well in the FT's reporting: researchers can predict the motion of an asteroid a million miles away, but still don't fully understand what happens when a fingertip meets a shoelace (Davies, 2026).

This multiplicity, not raw compute or model architecture, is likely to be the harder constraint on how fast physical AI scales. Training data needs to cover an enormous range of tasks, materials, sensors, and environments, and that kind of diversity is expensive and slow to collect compared to scraping more text off the web.

The FT's reporting also captured some of the scepticism building around how fast this can actually happen. Analysts have noted that Wall Street is already recalibrating: shares in Nasdaq-listed Symbotic, which deploys AI-enabled robots for logistics customers like Walmart, are down more than 30% this year after operating margins undershot expectations (Davies, 2026). That gap between the demo and the deployment is, in our view, exactly where the diversity problem shows up.

What industry research says

This isn't just one company's read of one news story. It shows up consistently in the robotics research literature:

• Academic surveys of foundation models in robotics describe "embodiment heterogeneity" as a core open problem: robotic data is produced by different robots with different degrees of freedom, sensor suites, and kinematics, and there is no single format for action data that transfers cleanly between them (Firoozi et al., 2026).
• Even well-known open datasets illustrate how narrow current coverage still is relative to the task space. RoboNet, one of the more widely cited cross-robot datasets, contains roughly 162,000 trajectories collected across seven robots (Dasari et al., 2023). Bridge Data, another commonly used benchmark, spans around 7,200 demonstrations across ten environments (Ebert et al., 2023). Useful datasets, but tiny compared to the diversity of tasks and settings physical AI eventually needs to handle.
• MIT Technology Review has reported that the scarcity of high-quality teleoperation and real-world robot data is one of the main things currently holding back progress in the field, prompting some researchers to explore workarounds like training from video of humans instead (MIT Technology Review, 2024).
• Collecting robot demonstration data typically requires physical hardware, human teleoperation, and significant time per episode, a fundamentally different cost structure than scraping and tokenising existing web text (Firoozi et al., 2026).

The common thread across this research is that it's not that the field lacks data entirely; it's that it lacks diverse, well-labeled data across the full range of embodiments, environments, and failure modes that real deployments require.

What "Data Diversity" actually means in a training pipeline

Breaking the constraint down further, it shows up across at least three dimensions:

1. Modality diversity. Physical AI systems typically fuse multiple sensor streams at once: RGB video, depth, LiDAR, force and torque, proprioception, audio, and each has its own noise profile, sampling rate, and failure mode. A pipeline built only for camera data will not generalize to a robot that also needs tactile feedback.
2. Environment and embodiment diversity. A model trained in one warehouse, on one robot arm, does not automatically transfer to a different facility or a differently shaped gripper. Data collected for one embodiment often needs substantial rework before it is useful for another.
3. Task and failure-mode diversity. A robotic system making a mistake on an assembly line can mean significant downtime, or worse, a safety incident. That raises the bar for how carefully training data needs to be curated, labeled, and validated before a system is trusted with autonomy.

Each of these multiplies the effort required to build a dataset that actually generalizes, and multiplies again for teams working across multiple verticals at once, from warehouse logistics to surgical robotics to autonomous vehicles.

How Encord is leading the shift toward diversity-first data infrastructure

This is the layer of the stack Encord operates in, and it's the thesis we've built the platform around.

If diversity, not volume, is the real bottleneck, the winners in physical AI won't simply be the teams that collect the most footage. They'll be the ones with infrastructure built to handle heterogeneous, multimodal data at scale, from day one, rather than retrofitting a system built for one modality or one robot.

Here's how Encord approaches each part of that problem:

We built it this way because we see physical AI teams spending a disproportionate amount of their time wrangling inconsistent, multimodal data pipelines rather than actually training systems that generalize. Reducing that overhead, and doing it across the full diversity of what physical AI needs to learn, is the practical version of solving the constraint the industry is underestimating.

Key takeaways

• Nvidia expects its physical AI revenue to grow tenfold over the next decade, and the FT reports the broader opportunity may be an order of magnitude larger than digital AI (Davies, 2026).
• The harder constraint on scaling physical AI isn't compute or model architecture, it's the diversity of tasks, materials, sensors, and environments training data must cover.
• Digital AI workflows benefited from a shared modality (keyboard and screen) and internet-scale data. Physical AI has neither.
• Academic research confirms this: embodiment heterogeneity, sparse cross-robot datasets, and the cost of real-world data collection are recurring themes in the literature (Firoozi et al., 2026; Dasari et al., 2023; Ebert et al., 2023).
• Encord's platform is built around solving for diversity, not just volume, across sensor types, embodiments, and verticals.

References

• Dasari, S. et al. (2023) 'RoboNet: Large-scale multi-robot learning', arXiv preprint.
• Davies, C. (2026) 'The next industrial revolution: how AI is remaking manufacturing', Financial Times, 31 August.
• Ebert, F. et al. (2023) 'Bridge Data: Boosting generalization of robotic skills with cross-domain datasets', arXiv preprint.
• Firoozi, R. et al. (2026) 'Foundation models in robotics: a comprehensive review of methods, models, datasets, challenges and future research directions', arXiv preprint.
• MIT Technology Review (2024) 'The robot race is fueling a fight for training data', 30 April.