Pasteur Labs & ISI

Here is a fact about science that sounds like a management consultant invented it but is actually true: there is a quadrant named after Louis Pasteur. The political scientist Donald Stokes drew it. One axis is "does this advance fundamental understanding," the other is "is this useful in the real world," and most research picks a corner. Pure theory here, applied grunt-work there. Pasteur's Quadrant is the box where both are true at once - where you cure rabies and also invent microbiology on the way.

Pasteur Labs & ISI named itself after that box. Which is either the most pretentious thing a startup has ever done or a fairly precise statement of intent, and having read a lot of their material I lean toward the second one. The whole company is organized around refusing to choose between "interesting" and "useful."

What they actually do is a thing they call Simulation Intelligence, which they define as "the systematic merger of AI, simulation, and scientific computing." The word "merger" is doing real work there. Lots of companies bolt a neural network onto a physics simulator and call it AI-for-science. Pasteur's claim is that the bolt-on is the problem - that you have to fuse the three into one system, integrated at scale, or you don't get the benefits.

The benefit, if it works, is enormous and slightly boring to describe, which is often how you can tell something is real. Traditional engineering simulation answers one question at a time: will this part hold, will this reactor stay stable, will this wing flex. You run it, you read the result, you tweak an input, you run it again. It is guess-and-check with a supercomputer.

Pasteur's pitch is that its simulators are differentiable end-to-end - meaning you can take the derivative of the whole thing. That is a nerdy sentence with a large consequence. If the simulator is differentiable, you don't have to guess-and-check. You can follow the gradient straight downhill to the answer. Optimization instead of search. Design instead of trial and error. The company's stated goal is to drive "the digital-physical delta to zero" - to move so much real-world testing into software that the gap between the model and reality basically closes.

Their simulators, they say, are data-driven, GPU-native, and automatically differentiable, wrapped in "patent-pending end-to-end differentiable physics programs" and something called CAxML data interfaces. Translated: they want computational engineers, physicists, and machine-learning people to build in the same room, on the same system, instead of throwing files over a wall to each other.

The most interesting structural fact about Pasteur Labs is that it is a public-benefit corporation - they say the only one of its kind pursuing science at the intersection of physics and AI - and is chartered to align 50% of its resources with its declared public mission. That is an unusual thing to promise investors. It is also, apparently, an excellent recruiting tool.

The roughly 43-person team is stitched together from DeepMind, NASA, CERN, Tesla, Nvidia, Meta, AFRL, and Ansys, distributed across four continents, with a New York headquarters and satellite offices in Copenhagen and Seattle. The company openly models itself on Bell Labs - the mid-century research shop that produced the transistor, the laser, and information theory by putting brilliant people from different fields in one building - but with, as they put it, startup agility.

The FOSAI turn

In August 2025 the company bought FOSAI, a digital-platform developer for aerospace and defense autonomous systems whose clients include the US Space Force and DARPA. FOSAI's CEO Gregory Falco framed it in exactly the terms Pasteur likes: "The future of autonomous systems depends on our ability to model multi-scale phenomena with unprecedented accuracy and speed."

What that acquisition signals is that Simulation Intelligence is not staying in the lab. It is heading toward autonomous systems that have to work the first time, in places you can't test twice - which is, when you think about it, the purest possible case for driving the digital-physical delta to zero.

And then there's Tesseract, which the company open-sourced in 2025 and even threw a hackathon for. Giving away your core plumbing is a curious move for a venture-backed company, until you remember that standards win. If everyone's simulators speak the same differentiable interface, the whole field speeds up - and the company that authored the standard is well positioned in the middle of it.