AI startup accelerating physical understanding launches with massive scale

Anima Anandkumar, Caltech professor and former Nvidia AI research leader, has launched Accelerated Understanding, a new artificial intelligence company developing models designed to understand and simulate physical phenomena rather than generate language.

The startup, co-founded by Anandkumar and Benedikt Jenik, a technology engineer, emerged publicly last week with claims that its AI system can process up to five trillion pieces of data in a single prompt, a scale the company said could open new possibilities for scientific discovery, chip design, robotics and weather prediction.

The launch comes as the artificial intelligence industry looks beyond large language models and chatbots towards systems capable of understanding how the physical world works.

Unlike ChatGPT-style systems, which are trained largely on text and designed to predict language, Accelerated Understanding is building AI around physics.

Its models are designed to predict how physical phenomena evolve across space and time, potentially allowing researchers and companies to simulate complex systems before testing them in the real world.

A different approach to artificial intelligence

At the heart of Accelerated Understanding’s technology is an approach known as neural operators, which Anandkumar helped pioneer.

The technology differs from the transformer architecture that underpins many of today’s leading AI models, including systems developed by companies such as Google and Anthropic.

Instead of processing language and predicting the next word in a sentence, Accelerated Understanding’s models are designed to learn how physical systems behave.

This means the technology could potentially be used to predict changes in weather systems, understand the behaviour of materials, optimise semiconductor designs or model complex robotic environments.

The company said its technology can understand physical phenomena across three-dimensional space and time, effectively creating what Anandkumar describes as a four-dimensional representation of the physical world.

Five trillion pieces of data

One of the startup’s most striking claims is the scale of information its model can process. In tests, Accelerated Understanding said its AI handled five trillion pieces of data in a single prompt.

Reuters reported that this is roughly five million times the amount of information that leading AI models from Google and Anthropic typically process.

Read also: Nigeria’s culture producers need enterprise development and industrial capacity

The comparison illustrates the different ambitions behind the technology. Large language models are capable of processing books, documents and long conversations. But modelling physical systems often requires AI to analyse enormous volumes of data across space and time.

Weather patterns, for example, involve the movement and interaction of enormous numbers of variables. Designing a semiconductor also requires understanding complex interactions involving materials, temperature and energy.

Accelerated Understanding believes traditional AI architectures may not be sufficient for these types of problems.

The founders who walked away from Bezos-backed Prometheus

The story behind Accelerated Understanding has also attracted attention because Anandkumar and Jenik were previously approached to lead a major artificial intelligence venture backed by Amazon founder Jeff Bezos.

According to Reuters, the two researchers received an offer connected to Project Prometheus, an ambitious AI initiative focused on automating the manufacturing of complex physical systems.

The proposed deal included significant equity, salaries, and access to billions of dollars in committed financing, but Anandkumar and Jenik chose to continue independently.

While they built Accelerated Understanding, Project Prometheus went on to raise a $12 billion Series B funding round in June 2026.

The decision has made Accelerated Understanding one of the more unusual new entrants in the expensive race to build frontier AI. Rather than joining a heavily funded AI project backed by one of the world’s richest technology entrepreneurs, Anandkumar and Jenik decided to pursue their own vision.

Anandkumar’s Nvidia connection

Anandkumar brings significant experience in both academic research and the commercial AI industry. She previously worked at Nvidia, where she led research exploring how the company’s graphics processing units could be used for advanced AI applications.

One of the projects she worked on demonstrated how AI could accelerate weather forecasting while maintaining the accuracy of more computationally intensive scientific methods.

The work attracted the attention of Nvidia chief executive Jensen Huang, who later presented Anandkumar’s research on neural operators at Nvidia’s GTC conference.

The bigger bet: AI that understands reality

Accelerated Understanding has entered a growing market for what is often described as physical AI or world models. The idea is to move beyond AI systems that primarily understand digital information and build systems capable of understanding objects, environments, and physical interactions.

This could become important for industries where errors in the physical world are expensive, and if Anandkumar and Jenik succeed, the company could help push AI beyond language and into the laboratories, factories, and physical systems where the next generation of scientific discoveries may be made.

            More from our Technology Column
        

        
    Join BusinessDay whatsapp Channel, to stay up to date

Open In Whatsapp