Veeda AI Raises Over US$90m for Robotics World Models

An orange industrial robot arm holding a metal fixture in an engineering laboratory, illustrating a report on funding for robot training simulation. Photo by Diego Martinez on Pexels

BUSINESS · 21 Aug 2026 —

TORONTO, 21 AUG 2026 — Veeda AI, founded by Nvidia's former vice-president of AI research, has raised more than US$90 million in seed funding to build world models for robotics. The company was incorporated in early June, which makes the round roughly ten weeks old.

Ninety million dollars before a product is a bet on a specific technical claim, and it is worth understanding what the claim is.

The round

>US$90mSeed, from Khosla Ventures and Radical Ventures

Early JuneIncorporation date, under three months before the raise

Sanja FidlerFormer Nvidia VP of AI research, with Zan Gojcic and Huan Ling

TorontoBase, and among Canada's largest seed rounds

The company, legally Veeda Innovation, is developing world models and simulation systems intended to let robots learn before they are deployed physically. The co-founders worked together at Nvidia on simulation, computer vision and generative AI.

The round is described as among the largest seed financings raised by a Canadian company.

What a world model is, and why robotics wants one

A world model is a system that predicts what happens next in a physical environment. Push this, and it falls that way. Grip here, and it slips.

Robotics needs them because robots and language models learn in fundamentally different ways. A language model trains on text that already exists in enormous quantity and costs nothing to read twice. A robot learns by acting, each attempt takes real time, hardware wears out, and failures break things. The data that would make robots capable is the most expensive data in machine learning.

Simulation is the standard answer and it has a standard problem, which practitioners call the reality gap. A policy trained in a simulator learns the simulator, including its approximations, and behaviour that works perfectly in software frequently fails on contact with an actual surface, an actual friction coefficient and an actual imprecise motor.

The bet behind a company like this is that generative models can close that gap — producing simulated environments varied and physically faithful enough that what a robot learns inside them transfers. If that works, robot training stops being bounded by hardware time. If it does not, it is an expensive rendering system.

There is a second use for a world model that gets less attention than training and may arrive sooner. A machine that can predict the next few seconds of physical consequence can also check its own intended action before executing it — refusing a grip that its model says will drop the object, or a path that ends in a collision. That is a safety and reliability function rather than a learning one, and it is the kind of capability an industrial buyer will pay for long before it cares how the robot was trained.

The founder premium is the largest line item

A company ten weeks old with no product does not have a valuation derived from anything measurable. What it has is a team whose previous work is known and a thesis several very large investors want exposure to.

That is not irrational. Simulation and neural rendering research is a small field, the people who led it at the largest hardware company in it are few, and if the thesis is right the option value is enormous. Paying up for the small number of teams who could plausibly execute is how early-stage investment in a hot field works.

The number, then, describes the market for founders, not the market for a product. We made the same observation about River AI's US$1.1 billion raise at two months old, and this is the same structure at a tenth of the size. The pattern is now common enough that a seed round in this field should be read as a talent price rather than a company valuation.

The physical AI thesis is having a fortnight

Three stories in three days point at the same conviction from different angles.

We reported that Unitree listed in Shanghai and rose sharply, having shipped more than 5,500 humanoid robots last year — more than anyone, and a small number in absolute terms. We wrote about TDK putting inference onto an industrial sensor. And this round funds the training layer underneath both.

Read together, the sector is being built in the wrong order from an investor's point of view and the right order from an engineer's. The machines exist and ship in small numbers, and the sensing is maturing. The bottleneck — what this money is aimed at — is teaching a general-purpose machine a new job without months of hand-tuning.

Why the region should care about the training layer

Southeast Asian manufacturing is the most likely large customer for general-purpose robots outside China, and the constraint here is not the hardware price.

A fixed automation line is already affordable and already deployed where volumes justify it. What the region's contract manufacturers cannot easily buy is a machine that can be retasked when a customer changes a product, because the retasking currently requires specialist integration work that costs more than the robot and is not available locally.

A world model that lets a machine learn new tasks in simulation would make general-purpose robots economic in the high-mix, medium-volume plants common to the region. That is the shape of most electronics and consumer-goods manufacturing across Malaysia, Vietnam and Thailand.

Because it would arrive as software, independent of supply chains, it is also the part of the robotics stack a regional buyer could adopt fastest.

What we could not establish

Anything about the technology beyond the description. No published results, no benchmark, no demonstration and no statement of what the models can currently do, which is normal at this stage and means the round is not evidence about the approach.

Also unestablished: the valuation and what share the round represents; whether Khosla and Radical led jointly or one led; whether Nvidia has any relationship with the company; whether the products will be sold to robot makers, to manufacturers or as a research platform; the intended timeline to a commercial product; and how the approach differs from the simulation work already published by the founders' former employer.

What to watch

Watch for a transfer result. The only measurement that matters in this field is a policy trained in simulation performing on real hardware it has never touched, published with enough detail to reproduce. Everything before that is a rendering demonstration.

Then watch who partners with them. A world model company that signs a robot manufacturer has a distribution path and a source of real-world data to close its own loop; one that stays independent has to build both.

Finally, watch whether the former employer competes directly. The founders built simulation tooling at a company that already sells it. This round is an implicit bet on what a small team can do that a very large one cannot.

Eva Chin’s Verdict Business

Ninety million dollars into a company incorporated ten weeks earlier, with no product and no published results. Before dismissing that, understand what is being bought. A world model predicts what happens next in a physical environment — push this and it falls that way, grip here and it slips — and robotics needs one because robots cannot learn the way language models did. A language model trains on text that already exists in vast quantity and costs nothing to read twice. A robot learns by acting: each attempt takes real time, hardware wears, and failures break things. That makes robot data the most expensive data in machine learning. Simulation is the standard answer with a standard problem, the reality gap: a policy trained in a simulator learns the simulator, approximations included, and fails on contact with a real friction coefficient and an imprecise motor. The bet is that generative models can produce environments faithful enough that learning transfers. If it works, robot training stops being bounded by hardware time. If not, it is expensive rendering. Read the number honestly, though. A ten-week-old company has no valuation derived from anything measurable, so this prices a team and a thesis — the same structure as River AI's raise at two months old, at a tenth of the size. What makes it regionally interesting is that this is the missing layer, not the hardware. High-mix, medium-volume plants across Malaysia and Vietnam cannot buy retasking; a machine that learns a new job in simulation is exactly what would make general-purpose robots economic there.

Sources & cross-checks

Primary: SiliconANGLE, 19 August 2026 — for Veeda AI raising more than US$90 million in seed funding to build world models for physical AI, founded by Sanja Fidler, formerly Nvidia's vice-president of AI research.
Also reported: Unite.AI — for the round being backed by Khosla Ventures and Radical Ventures. BetaKit — for the co-founders Zan Gojcic and Huan Ling, longtime collaborators with experience in simulation, computer vision and generative AI; the legal name Veeda Innovation; incorporation in early June 2026; the Toronto base; the development of world models and simulation systems to help robots learn before physical deployment; and the round ranking among the largest seed financings raised by a Canadian company.
Our earlier reporting: The same founder-premium structure is River AI's US$1.1 billion raise at two months old; the surrounding physical-AI stories are Unitree's Shanghai debut and TDK's on-device industrial sensor.
What could not be established: ⚠ Anything about the technology beyond its description — no published results, benchmark or demonstration, which is normal at this stage and means the round is not evidence about the approach. Also unestablished: the valuation and the share the round represents; whether Khosla and Radical led jointly or one led; whether Nvidia has any relationship with the company; whether products will be sold to robot makers, to manufacturers or as a research platform; the timeline to a commercial product; and how the approach differs from simulation work already published by the founders' former employer.
Method and limits: Compiled 21 August 2026. The explanation of why robot training data is uniquely expensive and what the reality gap is, the framing of the investment as a bet on simulation-to-reality transfer, the argument that a seed round of this shape prices a team and a thesis rather than a company, the observation that the sector is being funded hardware-first while the training layer is the bottleneck, and the assessment that retasking rather than hardware price is the regional constraint are RECATOOLS analysis rather than claims by Veeda AI or its investors. This article reports a private financing and is not investment advice.