Bonsai Robotics is running 430 autonomous agricultural machines and building its own world model on Nvidia Cosmos

Bonsai Robotics

Bonsai Robotics has 430 autonomous machines deployed across agricultural operations — not prototypes, but commercial units running real applications on farms today. CEO Tyler Niday made that point from a Sonoma County vineyard, where five different robots were running simultaneously during the conversation.

The company describes itself as a full-stack physical AI company for rugged outdoor environments. It builds both retrofit autonomy kits, including a John Deere retrofit, and autonomous-first form factors like a 60-foot boom hybrid electric machine. Niday claims the proprietary form factors cut CapEx by half compared to conventional alternatives while also reducing OpEx.

“We have 430 machines deployed. Our first commercial deployment was last year. These aren't just prototypes — these are real applications deployed on the farm impacting farmers in agriculture right now. We trained a world model on Cosmos 3. We have a million acres of data collected.”

World model

The technical core is a proprietary world model built on Nvidia Cosmos, trained on what Niday calls one million acres of collected data. Nvidia's Cosmos 3 was originally optimized for on-road environments; Bonsai used its agricultural dataset to adapt it to outdoor and rugged terrain. The result is a single model stack that runs across all the company's platforms and environments.

Niday's case for owning the full stack is grounded in data ownership. Because Bonsai collected the data early and at scale, it now has what he believes is the most diverse dataset in the rugged autonomy space. Controlling the world model, the user interface, the hardware, and the customer relationship is, in his view, the only defensible position in agricultural autonomy — vertical-specific use cases require too much customization for a general-purpose model provider to serve adequately from the outside.

Operations today

One person can manage an entire fleet, handling navigation, operational monitoring, spray volumes, and model parameters from a single interface. That operator-to-machine ratio is what Niday says makes the unit economics work. The company's first commercial deployments were last year, and it is now being pulled into verticals beyond agriculture.