Jeff Bezos offered Anima Anandkumar and her husband $1 million to $2 million a year, plus a 35% equity stake, to help lead Prometheus, his $41 billion bet on AI that understands the physical world. They turned him down.
It is a sign of how much confidence is riding on what’s known as a world model — an AI system built to understand the physical world directly, rather than through text — that two of the field’s top researchers would bet on their own version of one over Bezos’s checkbook.
Prometheus — backed by Bezos along with JPMorgan, Goldman Sachs, BlackRock and DST Global, and staffed by employees drawn from OpenAI, DeepMind, Meta and Nvidia — is trying to build AI that can design and manufacture physical systems, from jet engines to drug compounds, automating everything from prototyping to pre-production machinery. To lead that effort, it courted Caltech professor Anandkumar, a former AI research leader at Nvidia and Amazon, and her husband Benedikt Jenik, an AI infrastructure engineer.
Instead, the pair launched an independent company in August called Accelerated Understanding, with a different bet: rather than a Transformer-based model — the “T” in ChatGPT, invented at Google — their AI processes physics data using neural operators, a technique Anandkumar helped pioneer years ago. Where systems like ChatGPT learn from text, predicting the next word in a sentence, Accelerated Understanding’s model has learned to predict physical phenomena in space and time. “Our models understand the world directly in 4D (3D + time) and across physical phenomena,” Anandkumar wrote in a LinkedIn posting. “Going full 4D requires massive context length, we have pushed it to a trillion in training and exceeding 5 trillion at inference.”
It’s the latest example of the race between some of the brightest and most decorated names in AI to build world models, AI systems that learn the physical and spatial dynamics of reality such as gravity, object persistence, light, and causality. Along with Anandkumar, Bezos and Prometheus co-founder Vik Bajaj, a Stanford scientist and former co-founder of Verily, Google’s life sciences unit, there is Yann LeCun, a Turing Award winner and Meta’s former chief AI scientist and founder of France’s AMI Labs (Advanced Machine Intelligence); and Fei-Fei Li, the Stanford professor and “Godmother of AI” now building her own AI systems that learn how the physical world behaves at World Labs.
It is no surprise then that the World Economic Forum, in a June report produced with the open-science publisher Frontiers, named world models as one of the ten emerging technologies of 2026.
World models are expected to have big impact in the next five years on both industry and science. They are part of a larger category called Physical AI, which includes other types of AI that interact with the physical world such as autonomous cars, robots and Self-Driving Labs, which combine AI, robotics, and human expertise to accelerate scientific workflows in pharmaceuticals, chemicals, energy, and materials. (see The Innovator’s Startup of the Week story on Self-Driving Labs company Atinary).
Today’s large language models have done pretty well by using text, video and sensors (so-called multi-modal) data. The biggest proof point is that autonomous cars are already on the road. But there is a need to go further to solve complex problems in science and industry, Hiroaki Kitano, a professor at Okinawa Institute of Science and Technology (OIST) as well as president and CEO of Sony Computer Science Laboratories, said in an interview.
Most world-model and many robotics companies are pre-revenue or early-revenue but preliminary results appear promising and could have significant practical impact.
“By using AI that understands physics, we have already built models that predict weather, plasma [simulation] in nuclear fusion reactors, how drug molecules and materials behave, how to design a medical catheter that reduces bacterial infection etc.,” Anandkumar said in a LinkedIn posting.
Making The Impossible Possible
Kitano, a co-author of the Forum report, sees opportunities for world models in industrial design, manufacturing, and other physical industries as well as scientific discovery.
World models “will really change the way we do science,” says Kitano. “Right now, the problem with scientific discovery is that human scientists generate a very small number of hypotheses. We are not good at creating a very large number of hypotheses because it is very complex, meaning the hypothesis space is non-linear and has high dimensionality. If an AI can generate hundreds of hypotheses for these questions, and then we can create an autonomous self-driving lab to test and verify 100 hypotheses almost simultaneously, then this will absolutely speed up discovery in drug research and materials and environmental science.”
Self-Driving Labs, an earlier form of physical AI, do not need to wait for world models. Instead, they design, run and learn from real-world chemistry and materials experiments in a closed loop. “Our algorithms learn the ground truth from the lab experiments,” says Hermann Tribukait, co-founder and CEO of Atinary. Atinary, a World Economic Forum Pioneer, says its Self-Driving Labs can generate as much data in five days as a PhD does in five years.
“I would argue that the biggest impact of physical AI in the lab is that it will allow scientists to run experiments that would otherwise be too expensive and too hard,” says Tribukait. “It will make the impossible possible.”
As world models mature, Self-Driving Labs may start allowing world models to reason about chemistry it hasn’t physically tested yet — the same kind of leap Anandkumar and LeCun are betting on for engineering and robotics.
Follow The Money
Money is pouring into physical AI: Prometheus has raised $12 billion with a valuation of $41 billion, World Labs has raised $1.23 billion, with participation from tech giants Nvidia, AMD ventures and Autodesk, giving the company a valuation of $5 billion while France’s AMI Labs has raised $1.03 billion at a pre-money valuation of $3.5 billion.
Meanwhile Physical AI funding (which Crunchbas News defines as robotics, autonomous vehicles, drones, industrial automation and sensors), has quadrupled in a year: global venture funding reached $47.4 billion across 521 deals in the first half of 2026, almost four times the $12 billion raised in the second half of 2025 and up nearly 80% from the $26.4 billion in the first half of 2025. It is important to note that autonomous car company Waymo’s $16 billion Series D funding round alone was close to one-third of the entire first half of 2026 total.
Robotics and robot foundation models dominate: robotic foundation models and general purpose robots together account for roughly 77.6% of pure-play physical AI capital in one 12-month dataset, with defense/autonomy, materials, drug discovery and industrial simulation as smaller but fast growing adjacent categories, according to Crunchbase.
Europe represents one of the largest concentrations of Physical AI outside of the U.S. and China. Along with AMI Labs, Germany’s NEURA Robotics has raised $1.4 billion, the UK’s Humanoid has raised $152 million at a $1.35 billion valuation; PhysicsX, another UK company has raised $300 million at a $2.4 billion valuation and Germany’s Helsing has raised $1.8 billion at a valuation of $18 billion.
The Forum report notes that the first platforms built on an understanding of the physical world are now in developers’ hands. “NVIDIA’s Cosmos platform, launched in 2025 and trained on 20 million hours of physical-world data spanning robotics, industrial environments and driving, is the most significant early deployment of these ideas,” says the report. “Robots trained on Cosmos generalize to physical situations they have never encountered because they reason from an internal model of how things behave, not a library of situations they have seen before.”
Analysts Are Bullish Long-term, Cautious Near-Term
McKinsey projects at least $1 trillion in economic value from Physical AI and robotics by 2040, while Forrester, Gartner, BCG and Bain all flag limited near-term enterprise value and warn against mistaking hype for deployable value. BCG goes so far as to warn of a potential “misallocation of industrial capital” if humanoid forecasts don’t materialize.
The disagreement is largely one of timeline, not direction: McKinsey and BCG are weighing the payoff out to 2040, while Forrester is judging what Physical AI can deliver in the next two to five years — a shorter window in which the integration and safety hurdles loom largest.
How Physical AI Could Cause Material Harm
As world models move AI from observing or simulating operations to actively informing decisions in real-world physical settings organizations will need clearer ways to test, govern and hold these systems accountable, the Dubai Future Foundation says in its contribution to the Forum report.
The first changes would be felt in sectors where operations already generate continuous data, including manufacturing and logistics. In these settings, the model in production can learn from the operation as it runs. This could extend automation into forms of physical work that depend on reading a situation in the moment, rather than following fixed instructions. As the technology matures, industries built on skilled physical judgement, including logistics, construction, manufacturing and elder care, could see their work forces change. The institutions that retrain and redeploy people as roles evolve will help determine how productivity gains are shared.
World models also introduce a different kind of risk: they may not only reproduce data bias, but also build flawed assumptions about how the world behaves, says the report. A model can be internally consistent and still wrong, creating blind spots when systems move from controlled environments into real-world settings, says the report. Catching this kind of error would require clear assumptions up front, rigorous stress testing and audits that track data drift, possible tampering and changes in the model’s underlying assumptions.
The challenge is that many institutions deploying world models are built around stability and standardization. They are not designed for systems that continuously revise their assumptions.
“Physical AI and robotics will eliminate labor bottlenecks in every industry and make way for systems to adapt on the fly, but the technology will deliver limited near-term value until organizations overcome integration, scaling, safety, data, and workforce challenges,” says an April Forrester report which places physical AI and humanoid robotics in the two-to-five-year payoff horizon.
Research institutions could help by operating validated model libraries as shared infrastructure rather than proprietary assets, says the Dubai Future Foundation. No single organization can generate enough edge cases on its own to make these models reliable; reliability will need to accumulate across the field.
Policy-makers would also need governance frameworks that support accountability, transparency and safe deployment in critical sectors. Those frameworks would need to develop quickly enough to keep pace with adoption.
If they don’t, the consequences could be catastrophic, says Kitano, the report’s co-author. Physical AI could cause material harm through uncontrolled robots, hazardous chemical synthesis, or engineered biological agents, he says.
In August, the New York Times reported that scientists have already used AI to create new kinds of viruses, raising hopes for medical advances while also raising the disturbing possibility that the technology could someday be used to invent dangerous pathogens.
Kitano proposes safeguards that may include capability restrictions, auditing, AI regulation, bio safety rules, synthesis-level controls, and access verification for sensitive biological tools.
How to regulate dual-use biological capabilities without blocking legitimate research, remains an open question, says Kitano. The line between what could be good or bad for humanity is not always clear-cut.
The challenge is that audit frameworks, liability standards and validation benchmarks are still being built, even as deployment accelerates
“The impact of world models will depend on whether organizations can identify flawed assumptions before they influence decisions with real-world consequences,” says the Forum report. “If they can, these systems could make critical operations safer, faster and more resilient. If they cannot, errors may be harder to detect and spread further before they are caught.”
With such high stakes getting it right is crucial. Anandkumar and Bezos are each chasing that goal from different ends of the same race. A lot of money and public safety are riding on whether AI can learn to understand the physical world as well as it has learned to talk about it.
This article was reported and written by the author. AI was used to assist in research and to copy-edit.
