Caltech Researchers Unveil Physics-First AI Processing 5 Trillion Data Points, Rejecting Bezos Offer

Two prominent AI researchers have taken an unconventional path by rejecting a lucrative offer from Jeff Bezos's Project Prometheus to launch their own venture. Anima Anandkumar, the Caltech Bren Professor of Computing and Mathematical Sciences, and Benedikt Jenik unveiled Accelerated Understanding Inc. on August 25, 2026, introducing an AI model fundamentally different from mainstream generative AI systems. The company's core innovation lies in its architectural approach. Rather than using the Transformer architecture that powers systems like ChatGPT, Accelerated Understanding employs neural operators—technology Anandkumar helped pioneer for modeling physical phenomena. This distinction is fundamental: while language models predict sequences of text, this system learns relationships within physical systems and predicts how they evolve across space and time. In testing, the model processed 5 trillion data points in a single prompt, roughly 5 million times the typical context capacity of flagship models from Anthropic and Google. The decision to pursue independence was not made lightly. In late 2024, Anandkumar and Jenik received an offer from Project Prometheus that included a combined 35% equity stake, initial compensation of $1 million annually, and over $2 billion in committed financing through Series B rounds. Yet they declined, choosing instead to continue developing their physics-centric vision independently. Prometheus subsequently raised $12 billion in Series B funding in June 2026 to pursue AI for automating complex physical system manufacturing. Accelerated Understanding targets enterprise applications including semiconductor chip design optimization, robotics, weather prediction, and geological analysis. The founders indicate they have partnerships with computing providers furnishing hardware clusters for development and deployment, though they declined to name these partners. Notably, Nvidia CEO Jensen Huang had previously encouraged Anandkumar to pursue this direction after she led his company's research into how GPUs could power frontier AI between 2018 and 2023, demonstrating how neural operators could accelerate weather prediction while maintaining computational accuracy.