AI founders walked away from Bezos-backed Prometheus. Now they’re building physics AI to model the universe

Accelerated Understanding says its neural-operator AI processed 5 trillion data points in a single prompt, roughly 5 million times the typical context size of flagship models from Google and Anthropic.

Two AI researchers once courted to help lead Jeff Bezos-backed Project Prometheus have taken a very different path. Instead of joining the heavily funded startup, Anima Anandkumar and Benedikt Jenik kept building their own company, Accelerated Understanding Inc., and are now unveiling an AI system that models physics rather than language, Reuters reported.

The scale of its first results is hard to ignore.

Accelerated Understanding says its model handled 5 trillion data points in a single prompt during testing. The company puts that at roughly 5 million times the typical context capacity of flagship models from Anthropic and Google.

The comparison comes with an important distinction. This isn’t another large language model trying to read more books, documents or code. Anandkumar and Jenik are building AI that predicts physical phenomena across space and time.

“The language-centric view of intelligence is humans at the center. Putting physics at the center is a nature-centric view,” Anandkumar, a Caltech professor of computing and mathematical sciences, told Reuters.

Neural operators are at the center of the technology, an approach Anandkumar helped pioneer. Accelerated Understanding has moved away from the Transformer architecture behind modern language models such as ChatGPT, betting that AI built around physical systems can tackle problems that text-trained models aren’t naturally equipped to solve.

That could include semiconductor design, robotics, extreme-weather prediction and geological analysis. In chip development, for example, the system could model how materials and temperature affect performance, potentially reducing expensive trial-and-error experiments.

The larger bet is that companies won’t need a separate mathematical model for every physical problem. Accelerated Understanding wants one AI architecture that can work across many types of physics data. The startup plans to focus first on enterprise customers rather than consumers.

They walked away from Bezos-backed Prometheus

The company’s origin story makes that ambition more interesting.

According to Reuters, investor and biotech entrepreneur Vik Bajaj met Anandkumar and Jenik over dinner in the Los Angeles area in late 2024 to discuss working together. Bajaj later co-founded Project Prometheus with Amazon founder Jeff Bezos.

A proposed offer reviewed by Reuters would have made Anandkumar the public face of Prometheus, a board member and leader of its scientific vision. Jenik would have become a board observer. Together, they were offered a 35% stake and a combined annual salary of $1 million, rising to $2 million after three months.

The proposal outlined more than $2 billion in planned Series B financing, including capital from Bezos.

They passed.

Anandkumar and Jenik continued building Accelerated Understanding. Bezos and Bajaj went forward with Prometheus, which raised a massive $12 billion Series B in June 2026 to pursue AI capable of automating the manufacturing of complex physical systems.

Another major name behind Accelerated Understanding’s scientific roots is Nvidia CEO Jensen Huang.

Anandkumar joined Nvidia in 2018 and spent five years as a director, leading researchers exploring how GPUs could power frontier AI. Her team demonstrated that AI could dramatically speed up weather prediction while maintaining accuracy comparable with computational forecasting methods.

Huang showcased her neural-operator work at Nvidia’s GTC conference in 2021. Anandkumar recalled telling him that the technology could eventually eat physics theorists’ lunch.

“I want it to eat all their lunches,” Huang replied, Anandkumar said.

She declined to disclose Accelerated Understanding’s funding or identify the computing providers supplying hardware clusters for its models.

The startup now enters a growing race to move AI beyond language. Researchers including Yann LeCun and Fei-Fei Li are pursuing world models that give machines a richer representation of physical and spatial reality.

Accelerated Understanding is making a different wager: intelligence capable of predicting the physical universe may require an architecture built around physics from the start.

Five trillion data points in one prompt is an eye-catching demonstration. The bigger question is whether that scale can translate into better chips, robots, weather forecasts and industrial systems. That’s the next test.