Jensen Huang's vision for artificial intelligence extends far beyond predicting the next word in a sentence. The Nvidia CEO encouraged two AI researchers to pursue a radically different approach to machine learning, one that abandons the language-focused models dominating the industry today in favor of systems that understand and predict physical phenomena. That bet has now materialized into a startup called Accelerated Understanding Inc., which unveiled an AI model capable of processing 5 trillion pieces of data in a single prompt, roughly 5 million times the capacity of flagship models from Anthropic and Google.

The two co-founders, Anima Anandkumar and Benedikt Jenik, turned down a lucrative offer from Jeff Bezos-backed Project Prometheus to pursue this vision independently. Bezos and biotech entrepreneur Vik Bajaj had proposed a deal that would have given Anandkumar and Jenik a 35% stake in Prometheus plus $1 million in annual salary, doubling to $2 million after three months, with over $2 billion in committed financing through Series B. Instead, the pair chose to build their own company, while Prometheus went on to raise $12 billion in Series B funding in June 2026.

What Makes This AI Model Different From ChatGPT?

The fundamental difference lies in what the AI learns to predict. While systems like ChatGPT are trained on the world's text data to predict the next word in a sentence, Accelerated Understanding's model learns to predict physical phenomena across space and time. The company dispensed with the Transformer architecture, the foundational technology behind most modern large language models (LLMs), and instead uses neural operators, a technology that Anandkumar helped pioneer years earlier.

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

Anima Anandkumar, Professor of Computing and Mathematical Sciences at Caltech

This shift in perspective opens new possibilities for enterprise applications. Rather than relying on brittle, custom-built mathematical models for each specific problem, Accelerated Understanding believes a single physics-aware AI can handle diverse business challenges more efficiently and with less trial-and-error experimentation.

How Could Physics-Based AI Transform Industry?

• Chip Design: While other companies use AI trained on text to reason through semiconductor applications, Accelerated Understanding argues that intrinsic understanding of physics is key to optimizing materials and temperatures for chip performance with fewer lab iterations.
• Robotics and Automation: Physics-aware AI can better predict how robots will interact with their environment, improving control and precision in manufacturing and other physical tasks.
• Weather Prediction and Climate Modeling: The model can predict extreme weather patterns and atmospheric phenomena, potentially improving forecast accuracy and lead times for severe events.
• Energy Exploration: Geological data analysis for oil, gas, and renewable energy companies becomes more efficient when the AI understands the underlying physics of subsurface structures and fluid dynamics.

Anandkumar's journey to this moment reveals how Huang's influence shaped her thinking. When hired by Nvidia in 2018, Anandkumar led a team of scientists exploring how the company's graphics processing units (GPUs) could accelerate frontier AI research. An early project demonstrated how AI could speed up weather prediction with accuracy matching complex traditional forecasting methods. Huang was so impressed that he presented Anandkumar's work on neural operators at Nvidia's annual GTC Conference in 2021.

"He just got so excited," Anandkumar recalled. When she mentioned how AI could outperform physics theorists, Huang replied: "I want it to eat all their lunches."

Anima Anandkumar, Professor of Computing and Mathematical Sciences at Caltech

That enthusiasm translated into encouragement for Anandkumar to pursue the idea independently. Although Anandkumar declined to discuss specific funding details, she noted that Accelerated Understanding has partnerships with computing providers who furnished hardware clusters to develop and run the model. Nvidia did not respond to questions about whether it was backing the endeavor, though Anandkumar credited Huang with inspiring the entire direction.

Why Did They Walk Away From a $12 Billion Opportunity?

The decision to decline Prometheus's offer reflects a fundamental difference in vision. Prometheus is targeting AI that can automate the manufacturing of complex physical systems, a broader mandate. Accelerated Understanding, by contrast, is focused on enterprise deals rather than consumer offerings, betting that specialized physics-aware AI will find a more immediate market among companies that need to optimize physical processes.

This move also comes as Nvidia itself is expanding its AI infrastructure bets. The company is investing approximately $30 billion in AI startup Perplexity, valuing it over $30 billion, and has agreed to a multibillion-dollar deal with Poolside AI involving a $6 billion license payment plus a $1 billion investment. Nvidia's Vera Rubin rack-scale system is now in full production and delivers roughly four times faster performance than the nearest competitor on standard benchmarks.

Meanwhile, Nvidia CEO Jensen Huang has been actively signaling confidence in the AI infrastructure market. In June 2026, during a visit to Seoul, Huang told investors to "buy at a discount" during a sharp global technology stock sell-off, reiterating that the AI infrastructure build-out was still in its early stages. Nvidia's stock has risen about 2.2% since those comments, though the company's valuation multiples have remained relatively stable as Wall Street simultaneously raised earnings expectations.

The emergence of Accelerated Understanding represents a broader trend in AI development, where companies like those overseen by AI leaders Yann LeCun and Fei-Fei Li are pursuing "world models" that understand spatial reality better than systems trained purely on text. Anandkumar's bet is that physics-aware neural operators represent the most promising path forward, capable of delivering both better performance and more practical business value than language-centric approaches alone.