Key Takeaways:
A Caltech professor's physics-first AI startup, which turned down a $2 billion Bezos-backed offer, has unveiled a model that processes 5 trillion data points in a single prompt.
Accelerated Understanding Inc, the physics-first AI startup founded by Caltech professor Anima Anandkumar, unveiled a model that processed 5 trillion data points in a single prompt — roughly 5 million times the context window of Anthropic's Claude or Google's Gemini. The system, built to predict physical phenomena across space and time rather than the next word in a sentence, dispenses with the transformer architecture that powers OpenAI's ChatGPT.
"The language-centric view of intelligence is humans at the center. Putting physics at the center is a nature-centric view," Anandkumar, co-founder and professor of computing and mathematical sciences at Caltech, said.
The model processes physical data using neural operators, a technique Anandkumar helped develop, and targets chip design, robotics, extreme weather forecasting, and geological analysis for energy companies. Instead of building bespoke mathematical models for each use case, a single AI can handle any physics query in the business world, Anandkumar said. The founders plan to pursue enterprise deals before any consumer offering.
The launch pits the startup against well-funded rivals chasing "world models" that understand spatial reality better than text-trained AI, including ventures overseen by AI leaders Yann LeCun and Fei-Fei Li, and Project Prometheus, the Jeff Bezos-backed venture that raised $12 billion in June after the couple rejected a 35% equity offer.
A $2 Billion Offer, Declined
In late 2024, Vik Bajaj, who went on to co-found Project Prometheus with Bezos, dined with Anandkumar and Jenik at an upscale restaurant near Los Angeles to discuss a collaboration. The offer letter, seen by Reuters, proposed Anandkumar as the company's public face and board member, with the couple receiving a combined 35% stake, a $1 million annual salary doubling to $2 million after three months, and more than $2 billion in committed financing through Series B from investors including Bezos.
The couple chose independence. Bezos and Bajaj went on to raise a $12 billion Series B for Prometheus in June 2026, targeting AI that automates the manufacturing of complex physical systems. Prometheus declined to comment.
Nvidia's Long Shadow
Anandkumar's path to physics-first AI runs through Nvidia, where she was hired in 2018 and led a team pushing how the chipmaker's graphics processing units could power frontier AI. An early project showed AI could speed weather prediction with accuracy matching forecasters' complex computations. Nvidia CEO Jensen Huang presented her neural operator work at the company's GTC conference in 2021.
"When she mentioned how AI could eat physics theorists' lunch, Huang replied: 'I want it to eat all their lunches,'" Anandkumar recalled.
Anandkumar declined to discuss funding but said Accelerated Understanding has partnered with a computing provider that supplied the hardware cluster for developing and running its AI. She declined to name the partner, and Nvidia did not respond to questions about whether it backs the venture.
The model's claim of 5 trillion data points in a single prompt remains unverified against independent benchmarks, and the company did not disclose test conditions. If confirmed, the neural-operator approach could shift how enterprises model physical systems, challenging transformer-based incumbents that dominate the AI market. For investors, the open question is whether a physics-first architecture can win enterprise contracts against better-funded rivals — and whether the capital-intensive computing requirements erode the startup's runway before it lands its first major deal.
This article is for informational purposes only and does not constitute investment advice.
