Veeda Seed

Veeda AI Deep Dive: A Multimodal World-Model Company Founded by Three Ex-NVIDIA Research Leaders to Train Robots, with a USD 90 Million Seed in August 2026

Veeda AI was founded by three ex-NVIDIA research leaders including former NVIDIA VP of AI Research Sanja Fidler to build a multimodal foundation world model for robot training, and closed a USD 90 million seed in August 2026 led by Khosla Ventures and Radical Ventures, still pre-customer.

Executive Summary

Veeda AI (Veeda Innovation) is a Toronto-based company founded in June 2026 by three former NVIDIA research leaders: Sanja Fidler, the chief executive and a former NVIDIA vice president of AI research who also holds a professorship at the University of Toronto; Zan Gojcic, the chief technology officer and former director of NVIDIA's Zurich research; and Huan Ling, the chief scientist and a former NVIDIA generative-AI researcher. The company builds a multimodal foundation world model for robot training, combining action-conditioned video, latent dynamics, a rectified-flow diffusion transformer, inverse-dynamics and latent-action models, and diffusion forcing for long-horizon consistency, with compatibility claimed for Isaac Lab and Carla and the ability to post-train pi0 and GR00T-N. It announced a USD 90 million seed on 19 August 2026 led by Khosla Ventures and Radical Ventures, described as one of the largest seed rounds in Canadian history, and is still pre-customer.

Company Overview

Veeda AI was incorporated in June 2026 and is headquartered in Toronto, with offices listed in Mountain View, Singapore and Zurich. The founding team is unusually concentrated in senior NVIDIA research talent: Fidler ran NVIDIA's Toronto AI lab and is a well-known figure in scene understanding and generative models; Gojcic led 3D and autonomous-driving research in Zurich; Ling worked on generative AI. The company is privately held at the seed stage and, by its own description, has no customers yet.

Technology

The technical thesis is that robot policies should be trained inside a learned, multimodal world model rather than purely from teleoperation or from task-specific imitation. Veeda's architecture couples action-conditioned video generation with latent dynamics, uses a rectified-flow diffusion transformer as the generative backbone, and adds inverse-dynamics and latent-action models so that actions can be inferred from observations and vice versa. Diffusion forcing is used to keep long-horizon rollouts consistent. The company claims compatibility with Isaac Lab and Carla and the ability to post-train existing robot models such as pi0 and GR00T-N, positioning its world model as a training substrate rather than a standalone policy.

A notable adjacent publication is OmniDreams, an autonomous-vehicle world model released by the founders' research circle in June 2026, which demonstrates the same family of world-model techniques applied to driving. This suggests the company's method is intended to generalise across embodiments and sensor suites rather than to be tied to one robot form factor.

Product and Commercial Model

Veeda's product today is a research and platform stack rather than a shipped commercial offering. The intended customers are robotics developers and manufacturers who want to train policies in simulation-like world-model rollouts before deploying to hardware. No pricing, no paying customer and no revenue have been disclosed. The commercial logic is to monetise the world model as training infrastructure, in the same spirit as Physical Intelligence and Skild AI, but the company is earlier and pre-customer.

Competitive Landscape

Veeda competes directly with Physical Intelligence, Skild AI, AMI Labs and the robot teams at NVIDIA itself. Its differentiation is the depth of NVIDIA research pedigree across all three founders and a clearly specified architecture. The difficulty is that the field is exceptionally well funded and fast moving, and that NVIDIA, the founders' former employer, is both a potential partner and the most formidable competitor in robot foundation models through Cosmos, Isaac and GR00T.

Financing and Capital Structure

The USD 90 million seed announced on 19 August 2026 was led by Khosla Ventures and Radical Ventures and is described as one of the largest seed rounds in Canadian history. For a company incorporated only two months earlier and with no customers, the round signals strong conviction from two of the most prominent deep-tech and AI investors, but also prices in substantial execution and commercialisation risk.

DISCREPANCIES

1. Incorporation versus seed timing. The company was incorporated in June 2026 and announced a USD 90 million seed in August 2026, a two-month gap that some coverage treats as effectively simultaneous. Source grade: OFFICIAL.

2. Funding amount and basis. The seed is reported as USD 90 million led by Khosla Ventures and Radical Ventures; the pre- or post-money basis and exact per-investor allocation are not disclosed. Source grade: MEDIA.

3. Founder titles. Sanja Fidler is described as CEO and former NVIDIA VP of AI research and a University of Toronto professor; the precise relationship between her academic and company roles is variously framed. Source grade: OFFICIAL versus MEDIA.

4. OmniDreams relation. OmniDreams, an AV world model published in June 2026, is attributed to the founders' research circle; its institutional relationship to the company is described narratively, not documented. Source grade: OFFICIAL versus MEDIA.

5. Customer status. The company states it is pre-customer; some coverage implies strong inbound interest without quantification. Source grade: OFFICIAL, non-disclosure.

6. Compatibility claims. Compatibility with Isaac Lab and Carla and the ability to post-train pi0 and GR00T-N are company claims without a public demonstration benchmark. Source grade: OFFICIAL, unverified.

Risks and Diligence Gaps

The central gap is the absence of customers and deployed systems. Veeda is a pre-customer, seed-stage research company whose credibility rests on founder pedigree and a clearly specified architecture rather than on shipped products. A second gap is technical verifiability: the world-model claims, including post-training of pi0 and GR00T-N and Isaac Lab and Carla compatibility, need independent reproduction. A third gap is the NVIDIA overlap: the founders' former employer is simultaneously a potential partner and the most formidable competitor. A fourth gap is commercialisation, since monetising a world model as training infrastructure requires proving it reduces real-robot data cost and improves policy transfer.

Overall Assessment

Technology scores 8.0. The concentration of senior NVIDIA research talent across all three founders and the clearly specified rectified-flow diffusion-transformer architecture with inverse-dynamics and latent-action components are strong signals. Deducted points: no independent benchmark, pre-product maturity, claims of pi0 and GR00T-N post-training unverified.

Commercialisation scores 5.0. The training-infrastructure logic is sound but there is no pricing, no customer and no revenue. Deducted points: pre-customer, undefined go-to-market, no deployment evidence.

Financing scores 8.5. A USD 90 million seed led by Khosla Ventures and Radical Ventures, among Canada's largest, is an exceptional seed outcome and a strong conviction signal. Deducted points: two-month gap between incorporation and the round, pre- or post-money basis undisclosed.

Product maturity scores 6.0. The architecture is described in concrete terms and the OmniDreams precedent shows the method works in a related domain, yet no commercial product is shipped. Deducted points: research-stage, no customer reference, packaging evolving.

Market position scores 8.0. Robot foundation models are a large and hot theme, and the founder profile is among the strongest in the field, but NVIDIA's own Cosmos, Isaac and GR00T are direct competitors. Deducted points: intense competition including the former employer, Canada-based but globally contested market.

The composite view is 7.1. Veeda AI is a founder-pedigree bet of unusual quality in the robot world-model race, with a clear architecture and one of the largest seed rounds in Canadian history, but with the standard pre-customer deficits of no product, no revenue and claims awaiting independent reproduction. The most informative next disclosures will be a public benchmark against pi0 or GR00T-N, a first commercial deployment, and a clearer statement of which customers the world model serves first.

References

  1. Veeda AI announcement and coverage, 19 August 2026: USD 90 million seed led by Khosla Ventures and Radical Ventures, described as one of Canada's largest seeds; incorporation in June 2026; offices in Toronto, Mountain View, Singapore and Zurich.
  2. Founder profiles, 2026: Sanja Fidler, CEO and former NVIDIA VP of AI research and University of Toronto professor; Zan Gojcic, CTO and former director of NVIDIA Zurich research; Huan Ling, chief scientist and former NVIDIA generative-AI researcher.
  3. Technical descriptions, 2026: multimodal foundation world model for robot training combining action-conditioned video, latent dynamics, rectified-flow diffusion transformer, inverse-dynamics and latent-action models, diffusion forcing for long-horizon consistency; claimed compatibility with Isaac Lab and Carla and ability to post-train pi0 and GR00T-N.
  4. OmniDreams publication, June 2026: an autonomous-vehicle world model from the founders' research circle demonstrating the same world-model family applied to driving.
  5. Coverage of Khosla Ventures and Radical Ventures deep-tech and AI investments, 2026.
  6. Analysis of the robot foundation-model field, 2026: comparison with Physical Intelligence, Skild AI, AMI Labs and NVIDIA's Cosmos, Isaac and GR00T.
  7. Coverage of diffusion-forcing and rectified-flow diffusion transformers in long-horizon video and world models, 2026.
  8. Startup databases and company filings, 2026: incorporation date, office locations and seed-stage status.
This report is compiled from public sources and does not constitute investment advice. The company is private with no audited financials; the funding amount, the founder titles, the OmniDreams relationship and the Isaac Lab, Carla, pi0 and GR00T-N compatibility claims are drawn from media coverage and company statements with no independent reproduction, and the company is pre-customer with no disclosed revenue; uncertainties are flagged throughout. Data is current as of 5 November 2026.

Related tools

This report is compiled from public sources. It is for reference only and does not constitute investment advice. Data is current as of the publication date; please refer to the company’s official filings.