Nvidia Invests in Verkada to Scale Physical AI on 2.4M Devices

Nvidia's pitch that it will be the default stack for 'physical AI' has mostly been slide-deck material so far, so a strategic investment in a company that already runs 2.4 million cameras and sensors in the field is a more concrete data point than the usual keynote demo. SiliconAngle reports that the chipmaker took an undisclosed stake in Verkada, the physical security company, and signed a technical partnership to run Nvidia's Cosmos world foundation models and its Physical AI Data Factory toolkit across Verkada's fleet.

The scale is what makes the tie-up interesting. Verkada's kit is deployed at 30,000 organizations across 170 countries, with more than 100 of them in the Fortune 500, and the company crossed $1 billion in annualized bookings around its last funding round. That was a CapitalG-led round seven months ago at a $5.8 billion valuation, so Nvidia is buying into an already-hot cap table rather than seeding an early bet.

The technical claim to watch sits on video search. Verkada is putting Cosmos and the Physical AI Data Factory to work behind the job of finding a specific person, object, or moment across thousands of hours of recordings, with the Data Factory spinning up synthetic footage to fill gaps in training data. Verkada says accuracy on mean average precision for spatial-temporal queries has climbed 68% since the work began. Co-founder and chief executive Filip Kaliszan framed the work as 'AI that keeps students safe in schools, protects workers on factory floors, helps retailers prevent theft.'

The honest caveat is that the 68% figure is Verkada's number, and the reporting does not give you a dollar size for Nvidia's stake or the data-sharing terms between the two companies. That matters for a sensor footprint deployed in hospitals and schools, especially given that in 2021 hackers breached live feeds from about 150,000 of Verkada's cameras, including systems in hospitals, schools and workplaces. Pushing more AI across a larger fleet raises the stakes on whatever the current security posture actually is.

For Nvidia, though, Verkada is exactly the kind of design partner Cosmos needs. World foundation models get better on real deployments with real edge cases, not in the lab. Retail loss prevention, factory floor safety, and school security are not the sexiest use cases for physical AI, but they are the ones where synthetic-plus-real training loops start to make sense at the scale a fleet like Verkada's can provide.