Caltech Professor Walked Away From Bezos-Backed AI Venture to Build Physics Model
A research team led by Caltech professor Anima Anandkumar has unveiled a new kind of artificial intelligence designed not to predict words, but to predict physical phenomena — after the researchers declined an offer to help lead Jeff Bezos-backed Project Prometheus.
Anandkumar and her husband and business partner, Benedikt Jenik, are the co-founders of Accelerated Understanding Inc., which says its AI system handled 5 trillion pieces of data in a single prompt during testing. Reuters reported the system can process a volume of information vastly beyond the typical context capacity of leading language models.
The company is betting that AI built around physics rather than language could unlock applications ranging from semiconductor design and robotics to weather forecasting and energy exploration.
A Different Approach to AI
Unlike ChatGPT-style systems, which are primarily trained to understand and generate language, Accelerated Understanding’s technology is designed to model physical processes across space and time.
The company uses neural operators, a technology Anandkumar helped advance during her work in scientific machine learning. The approach also moves away from the Transformer architecture that underpins most modern large language models.
Anandkumar described the philosophy behind the technology as a shift from a human-centered view of intelligence toward one focused on nature and the physical world.
That puts Accelerated Understanding in a growing field of companies developing so-called world models — AI systems intended to understand environments, physical relationships and real-world processes rather than simply manipulate text.
Why 5 Trillion Data Points Matter
The company’s reported 5-trillion-data-point test is notable because conventional AI systems are generally constrained by how much information they can process within a single context.
Accelerated Understanding says its system was able to work with an amount of data roughly 5 million times larger than the context capacity typically associated with leading models from companies such as Anthropic and Google, according to Reuters.
That does not mean the system is a larger or better chatbot. Its purpose is fundamentally different.
Instead of reading enormous quantities of text and predicting what comes next, the model is intended to identify patterns in physical systems and predict how those systems behave.
Chip Design Could Be a Major Test
One of the company’s most important commercial targets is semiconductor design.
AI is already being used in the chip industry to write software, automate engineering tasks and assist with technical reasoning. Accelerated Understanding believes a deeper understanding of physics could help engineers optimize materials, temperatures and other variables that determine how chips perform.
The potential benefit is reducing the amount of trial and error required in laboratories before a design can be validated.
If the technology works at commercial scale, physics-focused AI could become another layer in the semiconductor design process — complementing existing AI tools rather than simply replacing them.
Robotics, Weather and Energy Are Also Targets
Accelerated Understanding says the same underlying technology could potentially be applied to robotics, extreme-weather prediction and geological analysis for energy companies.
The broader ambition is to avoid creating a separate mathematical model for every individual physics problem.
Instead, the company wants one AI system capable of responding to a wide range of physical-world questions.
That approach puts the startup in competition with a broader movement toward AI systems that can understand and reason about the physical world. Other companies and research groups are also pursuing world-model technology, although they are taking different technical approaches.
Anandkumar Was Once Offered a Role at Bezos’ Prometheus
The company’s independent path is particularly notable because Anandkumar and Jenik were previously approached about joining Project Prometheus, the ambitious AI venture backed by Amazon founder Jeff Bezos.
According to Reuters, biotech entrepreneur and investor Vik Bajaj discussed a collaboration with the pair during a late-2024 meeting in the Los Angeles area. Bajaj would later co-found Prometheus with Bezos.
Anandkumar had previously worked as a scientist at Amazon and spent five years as a director at Nvidia. Jenik is an AI infrastructure engineer, and the pair had already begun building their own company.
A proposal reviewed by Reuters described Anandkumar as a potential public face of Prometheus, board member and leader of its scientific vision. The proposal also offered Anandkumar and Jenik a combined 35% stake and salaries totaling $1 million annually, with the compensation potentially increasing to $2 million after three months of work.
The proposal also outlined more than $2 billion in planned financing through Series B rounds, including funding from investors such as Bezos, according to Reuters.
Prometheus declined to comment to Reuters.
Bezos Took Prometheus in a Different Direction
Anandkumar and Jenik ultimately continued building Accelerated Understanding independently.
Bezos and Bajaj went ahead with Prometheus, which has since become one of the most heavily funded AI startups focused on physical-world applications.
In June 2026, Prometheus announced a $12 billion Series B, bringing its reported valuation to about $41 billion. The company is pursuing AI designed to accelerate engineering and manufacturing of complex physical systems.
Prometheus has described its ambition as building an “artificial general engineer” — AI capable of helping compress the lengthy process of designing and manufacturing sophisticated physical products.
That makes the decision by Anandkumar and Jenik especially significant: two teams with overlapping interests in physical AI are now pursuing substantially different strategies.
The Idea Started at Nvidia
Anandkumar’s interest in physics-based AI has roots in her work at Nvidia.
She joined the chipmaker in 2018 and led scientists working on how Nvidia’s GPUs could be used for advanced AI research. One early project explored using AI to accelerate weather prediction while maintaining accuracy comparable with traditional computational methods.
Nvidia CEO Jensen Huang reportedly became enthusiastic about the work and showcased Anandkumar’s research on neural operators during Nvidia’s 2021 GTC conference. Reuters reported that Huang encouraged her to pursue the broader idea.
Anandkumar has since taken that research into her own company.
She has not disclosed Accelerated Understanding’s funding, but said the startup has partnerships with computing providers that have supplied hardware clusters for developing and operating its models. She declined to identify those partners, and Nvidia did not respond to Reuters’ question about whether it is backing the company.
What Happens Next
Accelerated Understanding is initially targeting enterprise customers rather than consumers.
Its biggest test will be whether the physics-focused approach can move beyond impressive technical demonstrations and deliver measurable advantages in real-world industries.
Chipmakers, energy companies, robotics developers and other engineering-heavy businesses could have strong incentives to adopt AI that reduces expensive simulations and physical experimentation.
But the technology remains an ambitious bet. Handling enormous datasets is only one part of building useful scientific AI; the models also need to make reliable predictions that translate into real engineering decisions.
For now, Accelerated Understanding is betting that the next major leap in AI will not come from teaching machines to read more text — but from teaching them to understand how the physical world works.