Deciphering the "West Point" of AI: 10 People Prop Up a Valuation of $70 Billion
When the innovation dividend of a top - tier laboratory approaches its upper limit, the spill - over of talent and ideas becomes almost a natural law.
Just like the "PayPal Mafia" in Silicon Valley back then, since the explosion of ChatGPT, OpenAI is becoming the "Whampoa Military Academy" in the AI world.
According to the incomplete statistics of Crow Intelligence, between 2022 and 2025, a total of 25 people left OpenAI. Among them, 9 chose to start their own businesses and founded 8 AI companies. Even if we don't count the three companies whose valuations have not been disclosed, the cumulative valuation of the remaining 6 companies is close to $70 billion. In addition, another 16 people joined other AI companies such as Meta.
These people cover almost all key positions at OpenAI: model R & D, training systems, alignment and security, product engineering, and toolchains. They are not only the builders of the GPT series but also witnessed the entire process from research prototypes to products with hundreds of millions of users.
This is a round of talent spill - over with strong organizational penetration.
In the business world, they didn't choose to "copy OpenAI" but tried to reconstruct the system logic that once only existed within OpenAI: some emphasized security first, some rebuilt the toolchain, and some directly implemented agent applications; some companies reached a valuation of $5 billion within three months of establishment, and some completed financing of over $100 million before even having a product.
To some extent, the departure of these people didn't cut off OpenAI's influence. Instead, it spread its technological path and organizational experience to a broader industrial level through new companies.
OpenAI - affiliated entrepreneurs support a $70 - billion market
When the innovation dividend of a top - tier laboratory approaches its upper limit, the spill - over of talent and ideas becomes almost a natural law.
Just like the "PayPal Mafia" in Silicon Valley back then, OpenAI from 2022 to 2025 is also becoming the "Whampoa Military Academy" in the AI world.
According to the incomplete statistics of Crow Intelligence, during these three years, 9 core members have left OpenAI and founded 8 AI companies. Even if we don't count the two companies whose valuations have not been disclosed, the cumulative valuation is already around $70 billion.
They are not ordinary engineers. Most of them held positions such as research leaders, chief scientists, or core team members before leaving. The directions they led cover model structures, training systems, security mechanisms, and product deployments, almost covering OpenAI's technological core.
Judging from their entrepreneurial directions, they mainly focus on AI security, agents, and AI applications.
First, there is a wave of entrepreneurship around "AI security".
In May 2024, Ilya Sutskever, the co - founder of OpenAI and the long - serving chief scientist, chose to leave and founded Safe Superintelligence (SSI). This is a pure research - oriented company that advocates "regulation as a service" as a prerequisite for superintelligence and provides global AI developers with capability assessment, risk modeling, and an interpretability framework.
▲ Ilya Sutskever, Paul Christiano, Daniel Kokotajlo (from left to right)
The founding team of SSI includes Paul Christiano, the former head of Alignment, and Daniel Kokotajlo, a strategy researcher. It received joint investment from Sequoia Capital and Founders Fund within a few months of its establishment, with the first - round financing exceeding $500 million, making it one of the world's highest - valued AI security companies.
Meanwhile, Mira Murati, the former CTO, and John Schulman, a co - founder of OpenAI, jointly founded Thinking Machines Lab, trying to rebuild the infrastructure of "research as a platform" for universities and enterprises.
▲ Mira Murati
This company reused the concept of OpenAI's toolchain, emphasizing data governance, model reproducibility, and AI responsibility tracking. It completed a $2 - billion seed - round financing in July this year, and its valuation reached $20 billion in October.
The second category is entrepreneurship around "agents" and human - computer interaction.
Adept AI was founded by David Luan, the former vice - president of engineering, and focuses on "AI assistants that can operate computers." He once led the training systems of GPT - 2 and GPT - 3. After leaving OpenAI, he quickly assembled a team and received financing of over $400 million.
Inflection AI was founded by Suleyman, the co - founder of DeepMind, and Simonyan, the former strategic advisor at OpenAI. The 35 - member core team includes many engineers from the GPT project. The company emphasizes "conversation as an agent," and its product Pi is considered the "AI assistant with the warmest personality." Currently, its valuation is nearly $4 billion.
Aravind Srinivas, the founder of Perplexity AI, was in charge of the inference system and multimodal search at OpenAI. Most of the team he leads is from OpenAI's toolchain group. It has completed $1.5 - billion financing, and its valuation exceeds $20 billion. Its model of "conversational search + citation tracing" is regarded as a key turning point in AI search.
The third direction is transferring the capabilities of general models to vertical scenarios.
Eureka Labs was founded by Karpathy and focuses on AI education and adaptive learning systems, creating a teaching platform that can automatically generate courses, feedback, and evaluations. Most of the team members are from OpenAI's toolchain. The first - round financing reached $400 million, and the valuation exceeds $5 billion.
Covariant was founded by Pieter Abbeel and specializes in general robot operating systems; Periodic Labs focuses on materials science and laboratory AI automation and completed Series A financing in 2025, with a valuation of $800 million.
Compared with other startups, entrepreneurs from OpenAI are more likely to achieve high valuations in a short time.
Ilya Sutskever's SSI, without a product or users, completed $1 - billion financing and reached a valuation of $5 billion in just three months;
Thinking Machines Lab, founded by Mira Murati, the former CTO, received $2 - billion seed - round financing five months after its establishment;
Periodic Labs, founded by Liam Fedus, the former vice - president of research at OpenAI, received $200 - million financing led by a16z just three months after its establishment.
The common feature of these companies is that they don't have a clear product path yet, but their founders are from OpenAI's core management team. They haven't even started building a revenue model, but their valuations have been pushed to billions of dollars.
This is a rare market signal. In the eyes of capital, as long as the starting point is close enough to OpenAI, it's worth betting on.
From Meta to xAI, why has OpenAI become the global AI talent pool?
Besides entrepreneurship, OpenAI is quietly becoming the most important talent "reservoir" in the entire AI industry. According to the incomplete statistics of Crow Intelligence, since 2022, at least 16 core members have left OpenAI and joined other AI companies.
Many enterprises have regarded OpenAI as a "supply source" of top - notch technological capabilities. In the past six months, the most aggressive one has been Meta.
From June to July, a well - organized team moved from OpenAI's Zurich and San Francisco research teams to Meta en masse - this was not an individual act but a collective transfer of an organized team.
Statistics show that as many as 11 people from OpenAI joined Meta's newly established "Superintelligence Labs," including Shengjia Zhao, Jason Wei, Lu Liu, Shuchao Bi, Allan Jabri, Alexander Kolesnikov, Xiaohua Zhai, Jiahui Yu, Lucas Beyer, Hongyu Ren, etc.
They cover almost all key capabilities at OpenAI in multimodality, model alignment, training optimization, and underlying systems:
Shengjia Zhao became Meta's chief scientist and rebuilt the core research route of the team - from model alignment, inference frameworks to retraining of visual Transformers;
Jason Wei took over the model science work, focusing on multi - task generalization and inference consistency;
Allan Jabri and Jiahui Yu continued the research on DALL·E image generation and vision - language fusion, grafting OpenAI's multimodal experience onto the Llama system.
Xiaohua Zhai and Lucas Beyer, from the Zurich team, deeply optimized distributed capabilities such as FSDP/DTensor in PyTorch, enabling Meta to catch up with OpenAI's internal architecture in distributed training and data sharding.
This is a "pure - blooded OpenAI team," and Meta is using it to replicate and upgrade its own AGI research system.
Meta is not the only "beneficiary."
Kyle Kosic, one of the first founding members of xAI, jumped from OpenAI to xAI in 2023 and served as the head of infrastructure, leading the development of relevant models. He helped Elon Musk's team build an inference framework similar to OpenAI's in a short time. However, in May 2024, he chose to return to OpenAI.
At DeepMind, Logan Kilpatrick, the former head of the developer ecosystem at OpenAI, took over as the head of developers and the community for Gemini. He once led the ecological construction of the GPT API and now continues a similar path, strengthening the developer interface and commercial feedback mechanism of the Gemini product.
Why are people from OpenAI the most sought - after in the market?
The answer is not complicated. They are among the few who witnessed the entire process of models such as GPT - 4, GPT - 4.5, GPT - 5, and Sora from training, evaluation, security alignment to global launch. They know how to transform cutting - edge algorithms into commercial systems for hundreds of millions of users, and this ability is scarce and cannot be replicated quickly.
More importantly, OpenAI's extremely flat organizational structure provides them with a highly comprehensive practice field.
Inside OpenAI, there are two main branches: the research team and the engineering team. The research team is responsible for model prototypes, security strategies, and alignment mechanisms, while the engineering team builds stable online systems.
There is no obvious gap between the two. Researchers can directly influence product decisions, and developers also participate in model verification. The teams operate in a "group system," and each group has almost full - process permissions from research to deployment, similar to a mini - startup unit.
This R & D system with high freedom and high coupling has spawned a group of "versatile and in - depth" talents: they are familiar with underlying algorithms and also have engineering implementation and productization thinking.
To find such people, OpenAI's employment criteria are significantly different from those of mainstream research institutions. It has two clear "don't cares":
First, it doesn't care about academic degrees. A doctoral degree is not a prerequisite for entry, and many core researchers only have a bachelor's degree. For example, Aditya Ramesh, the author of DALL·E, only has a bachelor's degree from New York University.
Second, it doesn't care about seniority. OpenAI is used to letting newcomers take on important roles. Bill Peebles, the person in charge of the Sora project, is a newly - graduated doctor in 2023 and started leading the team less than a year after joining.
This mechanism has forged a group of people with interdisciplinary knowledge structures, a strong implementation orientation, and a willingness to be responsible for the final product. They are familiar with cutting - edge technologies and know how to turn technologies into large - scale products.
For Meta, xAI, and many emerging companies, what they are competing for is not just the technical resumes themselves but the key talents nurtured by OpenAI's organizational mechanism and product philosophy.
These people can integrate the mission - driven research spirit with deliverable product standards. And this is exactly the most needed ability for building the next - generation AI companies.
This article is from the WeChat official account "Crow Intelligence Talk", author: Intelligent Crow. Republished by 36Kr with permission.