Veeda AI Raises $90M to Build Simulated Reality for Physical AI

Former NVIDIA Toronto AI lab leader Sanja Fidler launched Veeda AI with $90 million to build world models and simulated environments for training and evaluating Physical AI before deployment.

What changed

Sanja Fidler, the former head of NVIDIA’s Toronto AI lab, launched Veeda AI with longtime NVIDIA colleagues Zan Gojcic and Huan Ling.

The company confirmed that it has raised $90 million from investors including Radical Ventures and Khosla Ventures. Veeda’s stated mission is to build “simulated reality for Physical AI”: world models and environments where robots and embodied systems can learn and be evaluated before interacting with the real world.

Public reporting describes the financing as seed funding, although Veeda’s first-party site currently emphasizes the mission and team rather than a detailed round announcement.

Why it matters

Physical AI has a data problem that cannot be solved by copying the language-model playbook.

Text and code are abundant, searchable, and cheap to duplicate. High-quality embodied data is slower and more expensive to collect. A robot must encounter geometry, motion, contact, occlusion, uncertainty, failure, and safety constraints across many environments.

World models attempt to reduce that bottleneck by learning representations of how physical scenes evolve and by generating environments where agents can practice before real deployment.

The potential value is broader than synthetic training data. A useful simulation and world-model stack could support:

• policy training
• rare-event generation
• safety evaluation
• hardware and sensor variation
• domain randomization
• counterfactual testing
• pre-deployment acceptance tests
• regression testing after model updates

This places Veeda in an infrastructure layer between foundation models, simulation engines, data systems, and robot deployment.

The hard part is not visual realism

A simulated environment can look convincing and still teach the wrong behavior.

For robotics, the relevant standard is whether simulated experience transfers to real machines. That requires credible dynamics, contact, object properties, sensor behavior, timing, uncertainty, and failure modes—not only photorealistic video.

The company therefore faces several simultaneous proof burdens:

1. predictive quality: does the world model represent the physical transitions that matter for control?
2. controllability: can users construct and vary scenarios relevant to their robots?
3. evaluation integrity: do simulated scores predict real-world performance?
4. transfer: do policies trained or tested in Veeda’s environments improve deployment outcomes?
5. economics: is the simulated loop cheaper and faster than collecting equivalent real-world experience?

The founding team’s NVIDIA and academic background is relevant, but personnel pedigree is not deployment evidence.

Value-chain read-through

Veeda reinforces the view that simulation, synthetic data, and world models may become one combined market rather than three isolated categories.

A robotics company does not only need a renderer. It needs a repeatable pipeline connecting task definitions, environment generation, policy training, evaluation, and real-world feedback. The company that closes this loop can benefit from every deployment because field failures reveal which simulated scenarios were missing.

That feedback loop may be more defensible than a generic world-model API.

What to watch next

• initial product access and target customer segment
• supported robot embodiments, sensors, and control interfaces
• benchmark tasks with disclosed real-world transfer results
• customer evidence showing lower field-test cost or faster deployment
• coverage of contact-rich manipulation and long-horizon tasks
• evaluation tools that detect regressions before rollout
• integration with existing simulation, VLA, and robot-learning stacks
• whether Veeda captures proprietary field feedback or remains a standalone synthetic-data supplier

Interpretation

The $90 million seed is a large bet on the layer robots encounter before the factory floor, warehouse, road, or home.

If Physical AI scales, the industry will need more than larger models and more robot bodies. It will need environments where systems can fail cheaply, measure progress consistently, and rehearse edge cases that are dangerous or rare in reality.

Veeda’s opportunity is to make simulated experience predictive enough to change real deployment decisions. The decisive metric will not be the visual quality or total hours of generated data. It will be whether customers see measurable improvements in task success, safety, intervention rate, and time to deployment on real machines.

Not investment advice. Research notes only.