Accelerated Understanding Inc, a startup spun out of Caltech, has unveiled an enterprise physics AI built on neural operators rather than the Transformer architecture that powers most modern AI systems. The system ingested 5 trillion data points in a single prompt during tests — roughly 5 million times what current frontier AI models handle.
The company was co-founded by Caltech professor Anima Anandkumar, a former Nvidia scientist named one of TIME's 100 most influential people in AI, and Benedikt Jenik. The pair reportedly turned down an offer from Prometheus — a Bezos-backed venture — that included a $1-2 million annual salary, a 35% equity stake, and $2 billion in committed Series A and B financing. Prometheus subsequently closed its own $12 billion Series B in June.
Neural operators are mathematical models that learn to map between function spaces, making them particularly suited for physical simulations where the underlying equations are known but computationally expensive to solve. Unlike Transformers, which process sequential token data, neural operators can directly learn continuous physical fields like fluid dynamics, electromagnetic propagation, and structural mechanics.
Target applications include chip design optimization, robotics control, weather prediction, and geological analysis. The company says its approach can reduce simulation times from hours to seconds for complex physical systems.
The launch comes amid growing interest in AI for scientific computing, with Nvidia, Google DeepMind, and Microsoft all investing heavily in physics-informed AI models. The key differentiator for Accelerated Understanding is its ability to handle massively larger datasets than existing approaches, potentially enabling simulations at scales previously considered computationally intractable.
