In March 2025, Causal Labs announced a $6M Seed round. I'm grateful Kelsie Zhao and Dar Mehta let me participate.

The round was led by Kindred Ventures, with participation from Refactor Capital, BoxGroup, Factorial, Otherwise, Karman Ventures, and a strong group of angel investors. But the round isn't really the story. The story is what Kelsie and Dar are actually building.

Causal Labs is building what its founders call a Large Physics Model — a foundation model for cause-and-effect reasoning in the physical world. Their first deployment is in weather prediction and weather control. The longer arc is something more fundamental: the missing piece of intelligence that today's AI systems do not yet have.

What Causal Labs Does

The current generation of AI is extraordinarily good at pattern recognition over text and images. It is much less good at reasoning about physical cause and effect in the real world.

If you've ever wondered why self-driving cars are still hard despite a decade of investment, or why robotics has lagged behind language models, the answer lives in this gap. A model trained to predict the next token can recognize patterns in language. It cannot, on its own, predict what will happen if you push a glass off a table, divert a river, or seed a cloud. Those predictions require an internal model of physics, of causality, of how an action propagates through a system over time.

Causal Labs is building that layer.

The bet — and it's a bet I find compelling — is that the right way to teach an AI system causality is not to scale up text training. It's to train on a domain that is physics-rich, data-rich, and inherently chaotic. Weather satisfies all three. The atmosphere is governed entirely by physical law. There are petabytes of multi-sensor data — satellites, weather stations, balloons, radar — collected continuously. And weather is chaotic enough that prediction errors compound rapidly, which forces any successful model to learn real causal structure rather than surface correlations.

Solve weather, and you have a foundation model for physics. Solve physics, and you have one of the missing pieces of general intelligence.

Who Built It

Kelsie Zhao and Dar Mehta are not weather researchers. They are former safety-critical AI engineers from Cruise, Waymo, and Google Brain — people who spent the last decade building deep learning systems for autonomous vehicles, where being wrong about cause and effect is not a benchmark question, it's a fatality.

That background is the entire premise of the company. The hardest unsolved problem in self-driving is out-of-distribution generalization — what happens when the car encounters something it has never seen before. Today's AI, including state-of-the-art language models, does not handle this well. It interpolates within its training distribution. It does not reason from first principles. The Causal Labs founders watched this problem block deployment timelines at Cruise and Waymo, and they came away with a clear thesis: pattern recognition alone will not get us to systems that act safely in the real world. You need physics. You need causality.

So they left to build it.

Why Weather, Why Now

A few reasons this team's bet on weather is especially well-timed:

The data is there. Weather is one of the few physical domains with the volume, density, and modality diversity required to train a frontier physics model. Satellite imagery, ground stations, radar, balloons, ocean buoys — petabytes of multi-modal sensor data, continuously generated, publicly available. Robotics, by comparison, has nothing like this scale of structured physical data.

The customers are there. Aviation, agriculture, energy, logistics, insurance, and federal/state/local governments all make multi-billion-dollar decisions every day based on weather forecasts. The current state of the art — physics-based numerical weather prediction — is hitting its limits. There is real, immediate enterprise demand for higher-resolution, real-time, hyperlocal forecasts and the decision-support that comes with them.

The climate stakes are rising. Extreme weather events are becoming more frequent and more economically consequential. The infrastructure of human resilience — from disaster response to grid management to agricultural planning — depends on better models. This isn't a speculative future market. It's a market that needs better tools today.

The longer arc is weather control. This is the piece that gets less attention but is, structurally, the most interesting. Causal Labs' eventual ambition is not just to predict weather but to enable safe, steerable interventions: cloud seeding for drought relief, hurricane intensity reduction, wildfire suppression. That's a category of capability that doesn't really exist yet, and it lives downstream of getting the physics modeling right.

A Few Things I Admire About Dar and Kelsie

Sitting on the cap table of a company like this is a privilege, not a victory lap. A few things that stand out about the way they're building:

They earned the right to take this swing. They are not academics parachuting into a flashy problem. They are operators who built safety-critical AI for a decade, watched the field hit a wall on out-of-distribution generalization, and identified the underlying gap themselves. The company is downstream of their diagnosis, not the other way around. That's rare.

The thesis is structurally contrarian, in the right way. Most of the AI ecosystem is building on top of LLMs. Kelsie and Dar are building underneath them — the physics layer the language layer is missing. They saw something most of the field hasn't yet, and they're acting on it.

They've structured the company sensibly. Weather forecasting and decision support is a near-term enterprise business with paying customers. The physics foundation model is the long-term platform play. They're aware that one funds the other, and they're not pretending otherwise.

Safety is a first principle, not an afterthought. They built the company around safety from day one, drawing on their AV experience. In a category where the long arc involves intervening in real-world atmospheric systems, that posture isn't optional — it's what makes the whole thing buildable. I trust them to be careful.

A Closing Thought

Most of the AI conversation right now is about scale: more compute, more parameters, more tokens. Causal Labs is one of a small number of companies betting that the next breakthrough won't come from more of the same — it'll come from teaching machines something they don't currently know how to learn. Cause and effect. Physics. The structure of the world.

If they're right, the next decade of AI looks very different from the last one. I think they're right.

‍

‍