Cut Dilution in a Physical AI Seed Raise

Veeda AI raised a $90 million USD seed round 3 months after founding. The round was co-led by Khosla Ventures and Radical Ventures, with participation from 2077 Ventures. The company plans to use the funding to hire researchers and build simulation infrastructure.

Veeda builds multimodal world models that simulate physical reality for embodied AI agents to learn in virtual environments. Its three co-founders come from Nvidia, and CEO Sanja Fidler established Nvidia’s Toronto research lab in 2018. The company is building out its team across four locations, including Toronto and Zurich. World models are described as a heavily funded AI category in 2026, with AMI Labs raising $1.03 billion in Europe’s largest-ever seed round at a $3.5 billion valuation.

For seed-stage founders building embodied AI, robotics, simulation, or world-model companies, the takeaway is not “raise a big seed.” It is: prove the layers that reduce uncertainty, and use that proof to cut dilution.

The three-layer physical AI evidence checklist

Physical AI is a stack of evidence. If you are raising pre-revenue, your deck should separate what you have proven in simulation, what your data pipeline can defend, and what an enterprise pilot can validate. Each layer should reduce uncertainty enough to lower the cost of the next raise.

Simulation: fidelity, agent learning speed, evals

Simulation is the first layer because it is where embodied AI teams can compress trial and error. The question is not whether you have a simulator. It is whether the simulator is useful for learning. Investors should see fidelity, agent learning speed, and evals.

Fidelity: Does the environment preserve the physical constraints that matter? A warehouse robot needs collision, friction, occlusion, and object deformation in the right places. If the simulation is too clean, the model will look impressive in demo and fail in deployment.
Agent learning speed: Can your agents improve faster in simulation than in the real world? If the simulator is only a rendering tool, it is not an asset that reduces uncertainty. If it changes sample efficiency, it is.
Evals: Do you have repeatable benchmarks that separate progress from noise? A good eval suite should show which capability improved, which failure mode remains, and whether the next training run is worth the compute.

The trade-off is obvious: high-fidelity simulation is expensive to build and maintain. A lower-fidelity environment can be faster, but it may teach the wrong behaviors. The dilution question is whether your simulation layer is reducing time to a deployable policy enough to justify the capital. The number I would put in the deck is episodes-to-threshold: the number of simulated episodes needed to reach a fixed success rate, with wall-clock time as the secondary number. If that number does not fall as the model improves, the simulator is not changing sample efficiency.

Data: pipeline, curation, defensibility

Data is the second layer because physical AI models are only as good as the distribution of experiences they can ingest. A data pipeline is not a folder of videos. It is a system that captures, labels, filters, stores, and reuses experience to improve the model.

Pipeline: Can you move from raw sensor data or simulation traces to training-ready data without a manual bottleneck? If every dataset requires a team of annotators to clean it, the pipeline is fragile.
Curation: Do you know which data is useful, redundant, or dangerous? Curation is where a data moat starts. A company that can identify high-value edge cases is more valuable than one that simply has more bytes.
Defensibility: Is the data tied to a proprietary environment, a unique deployment, a customer workflow, or a simulation stack competitors cannot easily replicate? If the data is generic, it is a cost. If it is tied to a hard-to-copy source, it is an asset.

For pre-revenue teams, the practical test is whether your data pipeline can improve the model without waiting for a full commercial rollout. If it cannot, the data story is a promise. The number I would put in the deck is curation failure rate: the share of labeled episodes rejected for mislabel, unsafe behavior, or unusable sensor quality. If the rate is high, the pipeline is not producing training-ready data.

Pilots: enterprise use case, cost-to-value, deployment constraints

The third layer is the enterprise pilot. A pilot should not be a free demo. It should be a controlled experiment that tests whether the system can deliver value under real constraints.

Enterprise use case: Is the task painful enough that a customer will keep using the system? The use case should have a clear owner, a measurable outcome, and a reason to replace the current process.
Cost-to-value: Can you show the path from deployment cost to operational value? You do not need revenue to prove this, but you need a credible model of labor, downtime, safety, or throughput that makes the value legible.
Deployment constraints: What breaks in the real environment? Network latency, lighting, safety rules, maintenance access, operator training, and facility restrictions all matter. A pilot that ignores them will produce a beautiful demo and a bad business case.

The dilution guardrail is simple: each pilot should reduce time-to-value or customer acquisition cost before the next raise. If a pilot only proves that the technology works in a lab, it has not done its job. The number I would put in the deck is a pilot cost-to-value threshold: the pilot must reduce manual handling time, error rate, or cost per task below a stated incumbent baseline. If it cannot, the pilot is a demo.

How to use this to cut dilution

The Veeda round is a data point, not a template: $90M, 3 months after founding, co-led by Khosla and Radical, with 2077 participating, and a stated plan to hire researchers and build simulation infrastructure. The dilution guardrail is to tie the next ask to a named simulation eval or pilot cost reduction, not to the size of the last round.

Make the capital ask match the layer. A simulation ask should be justified by a named sample-efficiency eval, such as episodes-to-threshold at a fixed success rate. A pilot ask should be justified by a stated cost reduction, such as lower manual handling time or cost per task.
If the ask cannot be tied to a named eval or cost reduction, it is a bet on narrative, not operating progress.

This article is general information, not investment or legal advice.