Veeda Raises $90M for World Models, Competition Heats Up

Investments in physical AI are accelerating: a $90 million seed round has been raised around Sanja Fidler, while players are considering acquisitions. In Canada, Veeda is emerging with the support of Khosla Ventures and Radical Ventures, aiming for multimodal world models. These movements are part of a growing reliance on Chinese open-weight models, even as the exact contribution of these approaches remains uncertain in light of the limitations of LLMs.

Massive Seed Rounds and Acquisition Appetite

At the end of summer and the beginning of fall 2026, embodied AI is experiencing notable enthusiasm. On August 19, Sanja Fidler teamed up with former colleagues from Nvidia, and a $90 million seed round was established around this project. A recognized Slovenian-Canadian researcher, Fidler is part of a landscape that already includes several startups in the world model space benefiting from what is considered very solid funding. At the same time, Anthropic is preparing for what is being touted as a record IPO, and indicators suggest a marked interest in world models, with the possibility of an acquisition of Physical Intelligence.

Veeda Bets on Multimodal World Models

Khosla Ventures and Radical Ventures are leading the seed round for Veeda, a round presented as one of the largest ever raised by a Canadian startup. Veeda, founded by three former Nvidia employees linked to the University of Toronto, plans to focus its efforts on multimodal world models. The team asserts its belief that robotics will transform the ways people and goods are moved, manufactured, and constructed, as well as operations in the physical world.

Chinese Open Weights and Technical Uncertainties

China has established itself as a major player in open-weight models, to the point of influencing how companies design AI. Western companies are relying on these advancements, as seen with Harvey, which is building a model based on Moonshot AI, and the race for world models is intensifying. Technically, a world model aims to simulate physics, space, geometry, and real-world dynamics, at a time when the computational infrastructure for a future in AI is being built. However, the limitations of LLMs remain, and the scope and utility horizon of physics-based world models for accelerating the training of humanoid-like robotic projects still need to be clarified.