Caltech Duo Launches Accelerated Understanding, a Physics AI That Ditches Transformers

While the industry’s attention remains locked on ever-larger language models, a startup launched today is betting that the next frontier of AI has nothing to do with language at all. Accelerated Understanding Inc, founded by Caltech professor Anima Anandkumar and AI infrastructure engineer Benedikt Jenik, officially debuted on Tuesday with an AI system built for physics — one that dispenses with the Transformer architecture that powers virtually every famous model from ChatGPT to Claude.

A Different Kind of Intelligence

The contrast with mainstream AI is stark. Today’s flagship models were trained on the world’s text, learning to predict the next word in a sentence. Accelerated Understanding’s model instead learns to predict physical phenomena in space and time. In tests, the company says, it ingested 5 trillion data points in a single prompt — roughly 5 million times what Anthropic’s and Google’s flagship models can typically consume. Reuters offered a vivid analogy: it is like reading Tolstoy’s War and Peace not once, but 5 million times in one sitting.

The technical foundation is neural operators, a class of AI methods Anandkumar helped pioneer. Rather than mapping sequences of tokens to other sequences of tokens, neural operators learn mappings between entire function spaces — mathematical objects that can represent temperature fields, fluid flows, or stress distributions across space and time. Because they operate on continuous fields rather than discrete words, they can swallow physics data at a scale no text-oriented context window can approach.

“The language-centric view of intelligence is humans at the center. Putting physics at the center is a nature-centric view,” said Anandkumar, a professor of computing and mathematical sciences at Caltech who previously spent five years as a director at Nvidia and worked as a scientist at Amazon.

What It’s For

The company is targeting enterprise deals rather than a consumer product, and its pitch is generality: instead of brittle, bespoke mathematical models for each engineering problem, one AI system can handle any physics query a business throws at it. The founders highlighted four initial domains:

• Chip design. Where other companies apply text-trained reasoning models to semiconductor problems, Accelerated Understanding argues that an intrinsic grasp of physics is what actually optimizes materials and thermal behavior for chip performance — cutting expensive trial-and-error in the lab.
• Extreme weather prediction. Anandkumar’s early neural-operator work at Nvidia showed AI could match the accuracy of complex numerical weather computations at a fraction of the cost.
• Robotics. A model that genuinely predicts physical phenomena could give robots a native understanding of contact, friction, and dynamics.
• Geological analysis. Energy companies could use it to sift through subsurface data that would overwhelm conventional pipelines.

The Prometheus Backstory

The launch closes one of the more intriguing personnel sagas in recent AI history. In late 2024, over dinner at an upscale restaurant in greater Los Angeles, investor and biotech entrepreneur Vik Bajaj — who would go on to co-found Project Prometheus with Jeff Bezos — discussed a collaboration with Anandkumar and Jenik, according to meeting records seen by Reuters.

The subsequent offer letter, entitled “Project Prometheus,” proposed that Anandkumar become the public face of the company, a board member, and owner of its scientific vision. She and Jenik — her husband, who would have joined as a board observer — were offered a 35% stake in the company plus a combined $1 million annual salary that would double to $2 million after three months. The letter outlined more than $2 billion in committed financing through Series B, from investors including Bezos himself.

They turned it down. Bezos and Bajaj went on to raise a $12 billion Series B for Prometheus in June 2026 as it pursues AI that automates the manufacturing of complex physical systems. Anandkumar and Jenik kept building solo — and now hold 100% of their own vision instead of 35% of someone else’s.

The Jensen Huang Connection

Nvidia looms over the story in more ways than one. Anandkumar was hired by the chipmaker in 2018 and led a team pushing its GPUs toward frontier AI research. An early project demonstrating AI-accelerated weather prediction amazed CEO Jensen Huang, who personally presented her neural-operator work at Nvidia’s GTC conference in 2021. “He just got so excited,” she recalled. When she remarked that AI could eat physics theorists’ lunch, Huang replied: “I want it to eat all their lunches.”

It was Huang, she says, who originally encouraged her to pursue the idea. Anandkumar declined to discuss the startup’s funding, saying only that she has partnerships with computing providers who furnished hardware clusters for development. Nvidia did not reply when Reuters asked whether it is backing the endeavor — though observers on Bluesky note it “looks like Jensen backed them, but nothing is official yet.”

The World-Model Race

Accelerated Understanding enters a heating race. Startups overseen by AI luminaries Yann LeCun and Fei-Fei Li are pursuing so-called world models — systems that understand spatial reality better than text-trained AI. The shared thesis is that language is a thin slice of intelligence, and that models grounded in physical reality will unlock applications language models cannot touch.

Anandkumar’s wager is more specific: that a generalized form of neural operator — predicting phenomena even the eye cannot see — is the best architecture for that future, and that it can pay for itself in business terms rather than research prestige alone. Whether enterprises will pay for physics-native AI the way they now pay for text-native copilots is the question the next year will answer. But a 5-million-fold jump in per-prompt data appetite is the kind of capability gap that makes the industry sit up and take notice.