As Jeff Bezos pours billions into AI designed to build the physical world, two researchers once offered a major role at his project have walked away to build an AI that tries to understand physics itself.
Two AI researchers were once being lined up to help Jeff Bezos build one of the most ambitious AI companies on Earth. They turned the offer down. Now they have unveiled what they were building instead: an AI designed not to understand language, but to understand physics and predict what happens in the physical world.
The researchers are Anima Anandkumar and Benedikt Jenik, founders of a new company called Accelerated Understanding Inc. Reuters reports that the pair had been approached about joining Bezos-backed Project Prometheus, which is aiming to use AI to automate the manufacturing of complex physical systems. Instead, they continued working independently and have now revealed an AI model that they say processed 5 trillion pieces of data in a single prompt during testing. That’s roughly five million times the amount of information that leading language models from companies such as Anthropic and Google can typically process at once.
This AI Isn’t Trying to Be ChatGPT
The easiest way to understand what Anandkumar and Jenik are building is to forget about chatbots for a moment. ChatGPT and similar systems were built around language. Feed them huge amounts of text and they learn to predict what comes next, allowing them to write, answer questions, generate code and reason through problems.
Accelerated Understanding is taking a very different route. Its system is designed to predict physical phenomena across space and time. Instead of treating the world primarily as text, it is being trained around physics data and mathematical relationships. Anandkumar describes this as moving from a human-centred view of intelligence to a “nature-centric” one, where physics rather than language sits at the centre.
The technology underneath it is also different. Most of today’s major AI systems rely on the Transformer architecture, the technology behind the “T” in GPT. Anandkumar’s system instead uses neural operators, a technology she helped pioneer. Neural operators are designed to learn relationships between complex physical systems, potentially allowing AI to predict how those systems behave without requiring a separate mathematical model to be built for every individual problem.
That could be a big deal outside the world of chatbots. The company sees possible applications in semiconductor design, robotics, extreme-weather prediction and geological analysis. For chip manufacturers, for example, the AI could model how materials, temperatures and other physical variables affect performance before engineers have to run expensive physical experiments. In weather forecasting, the same basic approach could be used to predict phenomena across enormous datasets.
Bezos Wanted Them. They Said No.
The strangest part of the story may be how close Anandkumar and Jenik came to becoming part of Bezos’ enormous AI gamble.
According to Reuters, the pair had already started their own company when investor and biotech entrepreneur Vik Bajaj met them over dinner in the Los Angeles area in late 2024. Bajaj would later become one of the co-founders of Project Prometheus alongside Bezos.
The two researchers were subsequently presented with a proposal to join Prometheus. The offer was not exactly small. Anandkumar could become the public face of the company, join its board and take responsibility for its scientific vision. Jenik would become a board observer. Together, they were offered a 35% stake in the company and a combined salary of $1 million a year, rising to $2 million after three months. The proposal also outlined more than $2 billion in committed financing through Series B.
They walked away.
Instead of taking Bezos’ money and building Prometheus, Anandkumar and Jenik kept developing Accelerated Understanding. Bezos and Bajaj went ahead with Prometheus, which raised a staggering $12 billion Series B in June 2026. Prometheus is pursuing AI that can automate the manufacturing of complicated physical systems, making it one of the biggest bets yet on AI moving beyond screens and into the physical world.
There is an interesting Nvidia connection here too. Anandkumar previously worked at Amazon and spent five years as a director at Nvidia, where she led research into using GPUs for advanced AI. One of her early projects showed that AI could accelerate weather prediction while matching the accuracy of much more complicated conventional calculations.
Nvidia CEO Jensen Huang was apparently impressed enough to encourage her to pursue the idea. Reuters reports that when Anandkumar told Huang that AI could eventually outperform traditional physics experts, Huang replied: “I want it to eat all their lunches.”
The AI Race Is Moving Into the Real World
The 5-trillion-data-point claim is the headline-grabbing part of Accelerated Understanding’s launch, but the bigger story is what it represents. The AI industry has spent years racing to build systems that understand human language, images, video and code. Now another race is emerging around systems that understand the physical world itself.
That shift could have enormous consequences. An AI that can accurately model how materials behave could help design better chips. One that understands complex physical systems could improve robots. Another could predict extreme weather or help energy companies analyse geological formations. Instead of creating a specialised mathematical model for every individual problem, Accelerated Understanding believes one general physics-focused AI could eventually tackle many of them.
The company plans to focus initially on enterprise customers rather than launching a consumer chatbot. And there is still a long way between processing an enormous amount of physics data and actually producing useful commercial predictions. The 5-trillion figure comes from the company’s own testing, not from an independent benchmark proving that it is five million times more capable than ChatGPT.
But the direction is fascinating. Bezos is betting billions on AI that can help build the physical world. Anandkumar and Jenik were offered a place at the table and chose to build something different. They bet that the next big AI breakthrough may not come from teaching machines to speak more like humans.
It may come instead from teaching them to understand nature and how the physical world works.
In case you missed:
- Scientists Just Uncovered Quantum Effects That Could Send Messages Back in Time
- AI vs Your Poker Face: The Battle Unfolding at ESPN’s World Series of Poker 2026
- China Wants a Future Without Windows. It Starts This September.
- Moltbook: AI agents now have their own Reddit, and humans aren’t allowed to post!
- AI Is Bringing Dead Nuclear Plants Back to Life
- India Brings AI to Its Aatmanirbhar Weather Forecasting
- These Indian Workers Are Teaching Robots How to Be Human
- Amazon Caught Destroying Rare Books to Feed AI: Historians and Book Lovers Furious
- OpenAI Just Revealed Its AI Went Rogue. The Rest of the Story Is Even Stranger.
- Japan just made Remote Quantum Computing a reality!
