Key Takeaways

  • Roughly 95% of enterprise AI pilots fail because corporate buyers deploy generic foundation models against standard workflows without proprietary data behind the firewall.
  • Inspiren deploys low-cost optical and sensory hardware across senior living communities, cutting overnight resident falls by 60% to 70%.
  • Inspiren uses continuous 24/7 video monitoring over thousands of resident days to build training data that identifies early markers of cognitive decline.
  • Evercoast repurposed 4D volumetric capture technology originally built for Olympic broadcasts into a primary motion training pipeline for humanoid robotics.
  • Investor capital in physical AI is shifting toward companies that control proprietary real-world data collection rather than pure-play robotics assemblers.

Inspiren Proves Physical AI Starts With Sensor Infrastructure

Most enterprise software deployments of machine learning hit a ceiling quickly. Studio VC Managing Partner Joseph Coyne notes that roughly 95% of enterprise AI pilots fail because they attempt to automate generic white-collar tasks using broad foundation models. The software wrappers lack proprietary context, and the underlying algorithms commoditize rapidly.

Physical AI operates on different economics. Instead of scraping public text from the open internet, it roots machine learning models in continuous physical observations. Nvidia CEO Jensen Huang projects physical AI will become a $50 trillion market, yet capturing that value requires direct access to physical environments.

Coyne points to portfolio company Inspiren as a working blueprint for how sensory hardware creates defensive software moats. Inspiren installs low-cost devices equipped with computer vision and environmental sensors inside senior living communities and hospital rooms to monitor patient movement.

“Inspiren, one of the first use cases was both in hospital and senior living was they were able to show they could eliminate 60 to 70% of falls that would happen often times in the night,” (play from 10:37) Coyne said.

The immediate operational return is simple: fewer catastrophic injuries, lower liability claims, and reduced strain on night-shift nursing staff. But the real enterprise asset is the video archive accumulated over years of live deployment.

“Well, wouldn't it be great if you had a 24/7 data set of video capture for thousands of patients over thousands of days?” (play from 14:12) Coyne noted. “That sounds like the perfect training data set for that type of use case.” (play from 14:20)

By tracking continuous human motion over extended periods, Inspiren converts operational fall detection into a predictive diagnostic platform. The longitudinal data enables the software to detect subtle gait changes and physical degradation that indicate cognitive decline long before clinicians notice symptoms.

Evercoast Turns Media Production Into Robotics Training Data

A similar data dynamic governs Evercoast, another business backed by Coyne. Evercoast specializes in 4D volumetric motion capture, recording human bodies moving through three-dimensional space over time. The company did not launch as an industrial robotics supplier. Its early revenue came from media production and high-profile sports projects, including Olympic broadcasts.

That commercial work yielded an unexpected asset: an enormous library of photorealistic human movement. Humanoid robotics companies cannot train autonomous systems on static images or synthetic simulations alone. Physical robots need high-density, real-world data on balance, spatial reaction, and human kinetics to function around people.

Evercoast redirected its volumetric capture engine to serve as a training feed for robotics developers. Companies building physical autonomous systems can buy structured, multi-angle motion capture data generated in physical production environments rather than spending tens of millions of dollars building proprietary capture studios.

Why It Matters

Software companies that rely strictly on foundation model APIs face compressed margins and weak retention as basic reasoning tools become commodities. Physical AI reverses that dynamic by anchoring software value in proprietary sensor infrastructure. Enterprise value will accrue to operators who control real-world physical environments and the longitudinal sensory data required to train autonomous systems.