The most legendary couple turned down an offer worth 100 billion yuan.
At the end of 2024, at an upscale restaurant in Los Angeles, Vik Bajaj, a former Google X Lab executive, investor and biotech entrepreneur, treated a couple to a lavish meal and extended an offer that is now valued at hundreds of billions of RMB.
This couple is far from ordinary: the wife is Anima Anandkumar, the Bren Professor of Computing and Mathematical Sciences at the California Institute of Technology, and the husband is Benedikt Jenik, an AI infrastructure engineer.
This offer came from "Project Prometheus", with very sincere terms: Anandkumar would serve as a director of Prometheus to lead the scientific vision, Jenik would take a seat as a board observer, and the couple would hold a combined 35% stake in the company. Their total annual salary would reach 1 million USD, and rise to 2 million USD after three full months of employment.
At that time, Prometheus was nothing more than a code name, and no one knew how much it would be worth in the future. What mattered, however, was that investors including Jeff Bezos had already committed in writing to more than 2 billion USD in funding covering up to the Series B round.
What made this couple stand out was that they did not agree to the offer on the spot, but did not reject it either. To be precise, the two sides engaged in prolonged back-and-forth negotiations. Meanwhile, the couple had already quietly prepared to launch their own company, Accelerated Understanding.
Around the middle of 2025, the negotiations between the two sides broke down. In November 2025, Bajaj and Bezos officially established Prometheus together, with Bezos stepping in to serve as co-CEO. In June 2026, the company completed a 12 billion USD Series B financing round, with a post-money valuation of approximately 41 billion USD. Its total cumulative financing has exceeded 18 billion USD.
That 35% stake is now worth more than 100 billion RMB. The point of this story is not how much money the couple "lost" — after all, no one knows how much that 35% stake has been diluted through these rounds of financing, and Bezos' company does not seem to follow the traditional valuation logic for tech firms anyway.
The core point lies in the ongoing shift of power between capital and talent amid the current AI wave.
The core disagreement between the couple and Bezos was over development paths. As previously covered by ChinaVenture in the report *700 Billion, Bezos Raised a "Vulture Fund"*, Project Prometheus was positioned to pursue AI + manufacturing mergers and acquisitions: using large amounts of cash from financing to acquire traditional manufacturing enterprises, and directly deploy AI in physical factories.
The couple, by contrast, advocated going all-in on foundational physical model training. They judged that foundational models were the real bottleneck, and premature acquisition of physical manufacturing assets would turn the company into a capital holding group that weakens R&D capabilities.
It is a typical power struggle between technology and capital. And the period when this offer was turned down happened to be exactly when the discourse power of AI scientists began to rise sharply.
From the second half of 2024 to the whole of 2025, a number of new labs sprung up in Silicon Valley. Almost all the founders were technical leaders who left OpenAI, DeepMind and Anthropic, including Ilya Sutskever, Mira Murati, Liam Fedus and Ekin Dogus Cubuk. They were backed by wealthy individuals, big tech firms and top-tier VCs who only invested in people, without setting any commercialization targets.
According to statistics from Deedy, a partner at Menlo Ventures, most of these new labs reached a 1 billion USD valuation at the seed round stage.
Top-tier capital bets on talent, but not just talent — it bets on the ticket to AGI represented by these top geniuses. When capital is deployed at this scale, the whole point is to let these AI scientists call the shots themselves.
The game between the couple and Bezos took place in this specific context. In the end, neither side conceded. The couple eventually founded their own company, and Bezos pushed Prometheus to a valuation of 42 billion USD. They both found their respective niches, and everything seems like the best possible outcome.
What Makes This Standout Couple So Valuable
First, let's look at what makes this couple so exceptional that Jeff Bezos, the former world's richest man, was willing to offer them a 35% stake.
Anandkumar's academic resume is top-tier. In 2004, she studied electrical engineering at IIT Madras, obtained her PhD from Cornell University in 2009, and then completed postdoctoral research at MIT. After that, she worked at the University of California, Irvine, rising from assistant professor to associate professor.
2017 was a key node in her career: she was appointed the Bren Professor at Caltech. The "Bren" title comes from the donation fund of Donald Bren, a school trustee and real estate tycoon. It is the highest-level named tenured professorship at Caltech, with scarcity and prestige far exceeding that of ordinary full professors. She became one of the youngest Bren Professors in the history of Caltech.
While building her academic career at Caltech, she also made remarkable achievements in her industry positions.
From 2016 to 2018, Anandkumar served as Chief Scientist at Amazon AWS, and was one of the core architects of AWS's early cloud AI system. She participated in the incubation of the industry's earliest cloud-based AI training and inference products.
Since 2018, she officially joined NVIDIA as Director of AI Research, and served in the role for nearly five years. She led the core team to tackle the adaptation of GPUs to cutting-edge AI algorithms and the deployment of large-scale deep learning, laying the underlying technical paradigm for NVIDIA's computing power to adapt to modern large models and scientific computing models.
It can be said that Anandkumar is one of the very few all-around top experts who "understand foundational models, industrial deployment, and computing infrastructure" — with both top-tier academic background in physical AI and hands-on experience in deploying cloud AI and GPU underlying AI systems.
What makes her even more exceptional is that Anandkumar is one of the important pioneers of the Neural Operators research direction.
Neural Operators are not just a simple model name, but a set of methods that attempt to directly learn the laws of physical systems. Many phenomena in nature have complex multi-scale features: how fluids flow, how materials respond to force, how waves propagate, and how weather evolves cannot be fully described by simple rules.
Traditional methods often rely on complex physical equations and large-scale numerical calculations, while the idea of Neural Operators is to let AI directly learn the relationships in these physical systems, and then predict their evolution in space and time.
This is a completely different technical path from Transformers. The core of Transformers is to predict the next token, which essentially looks for statistical patterns in symbolic sequences. Neural Operators, by contrast, face the continuous physical world: what they learn is not language probability, but the structure, changes and evolution of real physical systems.
In other words, Neural Operators are aimed at "understanding and predicting nature". The open-source academic ecosystem for Neural Operators is already quite mature, but most Silicon Valley physical AI startups are following paths of PINN (Physics-Informed Neural Networks), JEPA world models or vertical industrial simulation acceleration, rather than taking Neural Operators as the core foundational model of their entire company. Among the current star startups in Silicon Valley, only Accelerated Understanding takes the native Neural Operators physical AI path.
One of Anandkumar's most iconic achievements is FourCastNet, a high-resolution weather model trained with Neural Operators. It can perform weather forecasting at a speed tens of thousands of times faster than traditional numerical forecasting methods, while maintaining accuracy close to that of traditional physical forecasting methods.
The value of Neural Operators goes far beyond weather forecasting. The methodological framework behind it is also applicable to complex physical problems that have long plagued the scientific and industrial communities, such as plasma evolution simulation in nuclear fusion, safer and more reliable autonomous flight of drones, design of new materials and medical equipment, and even prediction of molecular behavior in drug R&D.
Compared with ordinary large models that only process text, images and code, Anandkumar's technical path cuts into a more underlying and real physical world.
What best illustrates the weight of her technical path may be a conversation she once had with Jensen Huang.
At the 2021 NVIDIA GTC conference, Jensen Huang specifically introduced the work of Anandkumar's team in the direction of Neural Operators. She told Huang at the time that this technology might "eat physicists' lunches" in the future.
Huang's reply was: "I want it to eat everyone's lunch."
This bold statement means that if AI can learn to predict the evolution of physical systems, its impact will not be limited to a single discipline, but will reshape the entire boundary of modern technology spanning meteorology, energy, manufacturing, materials and drug R&D.
It is easy to see why such an all-around talent who "understands foundational models, industrial deployment, and computing infrastructure" — whose expertise perfectly aligns with Bezos' Project Prometheus, not to mention her mastery of the Neural Operators "killer app" — would be offered such a generous package.
Unlike Anandkumar, who focuses on academic theory, her husband Benedikt Jenik is an infrastructure engineer deeply experienced in large-scale machine learning. He once participated in autonomous driving simulation research at MIT, then joined QuantCo to lead the development of high-concurrency machine learning infrastructure covering large-scale engineering systems such as algorithmic pricing and healthcare billing compilation.
He rarely publishes academic papers, and his core competence is to turn physical model algorithms into training and operation systems that can support super-scale data. In Prometheus's offer, Anandkumar would serve as a director to lead the scientific direction, while Jenik would be arranged as a board observer.
While the arrangement may seem like the woman is more prominent than the man, their skill sets complement each other perfectly. The combination of "academic brain plus engineering hands" — one responsible for defining cutting-edge technical directions, the other capable of making those directions operational — is undoubtedly one of the scarcest configurations, which explains why Bezos offered such a generous offer.
Unfortunately, the couple ultimately declined the offer. Different paths make no good partners: their disagreements with Bezos were not trivial frictions, but disputes over development paths and core strategies.
Prometheus
In November 2025, Prometheus was officially established, with Bezos serving as co-CEO. This was his first return to a front-line position after stepping down as Amazon's CEO in 2021. The other co-CEO is Vik Bajaj, a physicist and chemist who worked closely with Larry Page at Google X in the early years, and later co-founded Foresite Labs and Verily.
At the time of its establishment, the company completed a 6.2 billion USD first round of financing, and its team was almost entirely poached in batches from Meta, OpenAI and Google DeepMind.
In April 2026, Prometheus was close to completing a financing round of approximately 100 billion USD with a valuation of around 38 billion USD. On June 11, the 12 billion USD Series B round was finalized, pushing its valuation to approximately 41 billion USD. The list of investors in this round is quite notable: Bezos himself, JPMorgan Chase, Goldman Sachs, BlackRock, DST Global and ARCH Venture Partners.
It is worth noting that there are almost no traditional VCs among the investors. A single round of 12 billion USD can only be absorbed by balance sheet entities that treat the capital allocation as part of their asset management business, with banks and asset management firms stepping in and tech companies taking a back seat.
How did Bezos explain the positioning of this company? He said on CNBC on June 11: What drives the wealth of nations? What drives the wealth of civilization? The answer is invention.
It is thus easy to understand why Bezos named the company Prometheus, the fire stealer. He clearly hopes that these "inventions" combining AI and traditional industries can lead humanity to a new level of civilization.
What Bezos wants to build is called an "Artificial General Engineer". To put it plainly, it is an AI system that can compress the time from inventing a technology to manufacturing a physical product. Chips, jet engines, batteries, solar panels and drugs all fall within its scope.
Bezos specifically clarified that this has nothing to do with robots. It is more like an AI-powered CAD software: humans only need to put forward a goal, such as increasing the thrust of a jet engine by 10%, and this AI CAD will accomplish the task on its own.
As the world's richest man, Bezos certainly does not only focus on a single product or technology.
As covered in the report *700 Billion, Bezos Raised a "Vulture Fund"*, Bezos plans to use this fund to acquire traditional industrial enterprises in aerospace, chip manufacturing, national defense and other fields, and then use Prometheus' AI to transform them. The acquired factories feed their operational data to the model, and the more powerful model in turn optimizes the operation of these factories.
The logic looks familiar. It is the Amazon Virtuous Cycle flywheel, which Bezos wants to replicate in the physical world.
Put Anandkumar's scientific research path side by side with Bezos' grand vision, and you can roughly understand why the couple turned down this sky-high offer.
If Prometheus wants AI to automate the "manufacturing of complex physical systems", its focus will inevitably fall on manufacturing, engineering, and faster production of physical products. It needs factories, supply chains, and data from real production lines.
What Anandkumar wants to do is to let AI understand the laws of physics first, so that the model can predict how physical phenomena evolve in time and space, instead of rushing to manufacture products.
Both paths are centered on physical AI. But one uses AI to transform the manufacturing industry, while the other first lets AI understand nature. The former is an engineering and capital-intensive business, while the latter is a scientific problem. But what if the scientific problem also becomes a matter of money?
The Power Shift Between Capital and Technology
At the end of August, Accelerated Understanding, the company that the couple had prepared for nearly two years, was officially unveiled to the public.
Anandkumar and Jenik believe that intelligence is no longer the real bottleneck for scientific and engineering progress. The amount of creative ideas that current AI can generate has long exceeded the upper limit that any lab can actually verify. No matter it is a new chip layout scheme or a new material concept, the real bottleneck that slows down progress is the ability to verify these ideas in the real world.
This judgment is very important, and it will and has been influencing the venture capital trends for many years to come, which we will discuss later.
In short, if the laws of physics can be simulated with sufficient accuracy, part of real-world experiments can be replaced. Moreover, simulation can provide something that laboratories cannot: tell researchers which direction to improve their designs.
Prometheus's bet is that AI needs to learn in real factories and production lines, and data must come from the manufacturing process. Accelerated Understanding's bet is that AI must first learn the laws of physics before talking about applications.
No one knows which one is right for now. But the couple's persistence points to a dilemma that has repeatedly played out in the technology industry: for technical talents of this caliber, they cannot go far without discourse power and decision-making authority.
This script has been staged many times. Ever since Dario Amodei led 14 core OpenAI researchers to collectively leave and found Anthropic, a large number of top scientists employed by big tech firms have left their positions one after another — Ilya Sutskever from OpenAI, Noam Shazeer, core contributor to Transformer, Yann LeCun from Meta, Jeff Dean, chief scientist leading Google Research — every departure becomes a trending headline in financial and tech media.
The same thing is happening in China. Lin Junyang, the core leader of Alibaba's Qwen, left his position in March 2026 and later founded Pragma Tech. As I covered in the reports *The Genius Becomes Alibaba's Strategic Abandoned Child* and *An Experiment Worth 14 Billion*, the core conflict was not about money, but about strategic direction and organizational power.
These people have their own "North Star". It may be academic influence, verifying a technical path that has not yet become mainstream, or solving a problem they believe will affect the future of humanity. But enterprises have organizational constraints: products need to be launched on schedule, huge investments need to be justified, and research directions need to align with the company's overall strategy.
If the goals of the enterprise and scientists cannot align, conflicts will spread from performance evaluation to technical route, computing power allocation, and organizational power.
At this point, money becomes useless. For example, Noam Shazeer was paid 2.7 billion USD by Google in 2024 to return to the company, but he still left less than two years later. Even compensation at that level cannot retain top talent.
Therefore, Anandkumar and Jenik's choice, viewed in this context, is very sober.
Zoom out the perspective, and the current venture capital ecosystem in the United States is completely different from what it was before the launch of GPT.
Around 2022, model capabilities had been proven, but there was no commercial validation yet. Anjney Midha, early financing consultant of Anthropic,