While the companies that defined AI's past are losing the people who defined its future.On June 18, Noam Shazeer, core author of the Transformer paper and co-head of Google Gemini, announced on X that he was leaving Google and joining OpenAI, which had secretly submitted an IPO application to the SEC.. He is one of the eight equally contributing authors of "Attention is All You Need" in 2017, which laid the technical foundation for modern large language models. Sam Altman immediately retweeted and commented, "Noam is one of the people I have most wanted to work with since the first day OpenAI was founded. It only took ten years."
48 hours later, on June 19, John Jumper, winner of the 2024 Nobel Prize in Chemistry and core leader of AlphaFold, announced that he would leave Google DeepMind, where he had worked for nearly nine years, to join Anthropic.
The two resignations of top talents that occurred almost simultaneously were enough to shock the AI circle. And if you extend the timeline, you will find a clearer direction. On May 19, former OpenAI founding member Andrej Karpathy announced that he would join the Anthropic pre-training team. Although he has never worked at Google, his choice also illustrates one thing. Top talents are concentrating on OpenAI and Anthropic, and Google is becoming the main exporter in this talent reorganization.
Three resignations, not an isolated case, but a trend
Jumper is no ordinary researcher. In 2024, he won the Nobel Prize in Chemistry together with Demis Hassabis and David Baker for leading the AlphaFold project, using AI to predict the three-dimensional structure of proteins in a very short time, overcoming a problem that has plagued the biological community for fifty years.
John Jumper (right) takes a photo with Demis Hassabis
Shazeer is a key figure in the history of modern AI development. He joined Google in 2000 and co-authored "Attention is All You Need" in 2017. The Transformer architecture proposed in this paper is the technical cornerstone of all current large language models. In 2021, because Google refused to release the AI chat product it co-developed with Daniel De Freitas, he chose to leave and founded Character.AI in 2022. Three years later, Google brought him back for about US$2.7 billion and appointed him as co-head of Gemini. However, less than two years after returning, he chose to leave again, this time to OpenAI.
Noam Shazeer and another AI executive
And Karpathy’s choice further confirms the larger trend. In May 2026, after concluding his educational startup project Eureka Labs, the founding member of OpenAI announced that he would join the Anthropic pre-training team and be responsible for "empowering Claude with core knowledge and capabilities through large-scale training operations." He never worked at Google, but his whereabouts illustrate where top talent is concentrating.
Andrej Karpathy
Extending the horizons, this talent flow trend has already emerged. After the merger of Google Brain and DeepMind in April 2023, a large number of hardcore researchers moved to OpenAI, Anthropic and xAI. Tracking the authorship of cutting-edge AI papers on ArXiv, we can find that the names of institutions on the profile pages of more and more top researchers have changed from "Google" to "OpenAI" or "Anthropic".
OpenAI and Anthropic are bringing together the most influential talent in the AI field. And Google is becoming the main exporter of this talent flow.
Misplaced mission
This is the most essential difference, and its importance goes beyond salary and computing power.
Nearly 80% of Google parent company Alphabet's revenue comes from advertising. This means that all investments in the field of AI must ultimately answer a product-oriented question, how will this serve the advertising business.
Shazeer soon discovered after returning in 2024 that Google's core logic had not changed. The fundamental constraint he faces at Gemini is that catching up with ChatGPT is always a constrained task under the advertising business-first structure. The goal is not to redefine the boundaries of AI capabilities, but to maintain advertising market share.
In contrast, OpenAI's charter clearly takes the benefit of AGI (artificial intelligence) as its core mission. Anthropic has been built around AI security since its establishment. It is registered as a public benefit company (PBC) and is legally obliged to balance the interests of shareholders and social interests. At both companies, top researchers don't have to answer the question, "How can we help our advertising departments increase revenue?" They only need to focus on one goal, how to continue to push the boundaries of model capabilities.
Many researchers who transferred from Google to these two institutions repeatedly mentioned the same word, "focus" in subsequent interviews. At Google, key performance indicators are search click-through rate, ad conversion rate, and YouTube watch time. At Anthropic, the key performance indicator is Claude's performance in pre-training and post-training. For scientists like Jumper who have devoted nine years of their academic and professional lives to protein folding problems, this high level of focus is irreplaceable. At Anthropic, AI for Science is not a fringe project, but one of the core research directions.
Mission is push, and capital is pull. In terms of salary incentives, Google is at a structural disadvantage.
OpenAI has secretly submitted an IPO application to the SEC in 2026, and Anthropic is also in the IPO preparation queue. Employees of both companies hold large stakes that are expected to be cashed out in the public markets. Jumper and Shazeer chose to join before this window, and the timing is no coincidence. In contrast, Google's market value has exceeded US$2 trillion, and its stock price has limited room to double in the short term. The explosive power of equity incentives is at least an order of magnitude different.
What deserves more attention is the completely different pricing logic of the two types of companies in the capital market. The leaked OpenAI audited financial report shows that its GAAP net loss in 2025 will be approximately US$38.5 billion to US$39 billion (including approximately US$30 billion in non-cash conversion expenses), and its operating loss will expand from US$8.78 billion in 2024 to approximately US$20.9 billion. However, the capital market reaction remains positive. During the same period, OpenAI's revenue soared from US$3.7 billion to US$13.07 billion, an increase of 253%. In the first quarter of 2026, the company's revenue was US$5.7 billion and operating expenses were US$3.7 billion. Investors are willing to pay for a "loss-for-growth" strategy.
At Google, the same scale of investment in AI raises a question in the capital market: "What impact will this have on profit margins?" The same large-scale investment in the AI field is called a strategic investment in OpenAI, and it is regarded as a cost center expansion at Google.
From the perspective of a top researcher, the logic behind this choice is not complicated. On one side is a company that is about to IPO and whose equity may achieve a nine-digit value within two years. All employees focus on optimizing model capabilities. On the other side is a mature giant with a market capitalization of two trillion. The work of researchers needs to be continuously coordinated with the quarterly goals of the advertising and search teams.
DeepMind merger creates new centrifugal force
In April 2023, Google Brain and DeepMind merged into Google DeepMind, under the unified leadership of Demis Hassabis. The official narrative at the time was to "concentrate our forces." But looking back three years later, the actual effect of the merger is clearly controversial.
The merger failed to fundamentally solve the problem of restructuring the voice of transforming research results into products.
DeepMind's basic research results need to be implemented through the product team, and the product team has its own independent timeline and priority considerations. Gemini is a typical case. Shazeer was appointed as co-head, but the product release rhythm and commercialization path are still highly constrained by the search and cloud business units. This is in sharp contrast to the model in which all members of OpenAI operate around the same core product goals.
The merger also caused tensions over cultural identities. Google Brain is more focused on engineering and commercial implementation, while DeepMind is more focused on basic science and long-term exploration. After the merger, the long-term research-oriented culture was seen to have eroded under the pressure of a "service product roadmap".
A former Google researcher wrote on X, "When we were asked to align our research direction with the product roadmap, I knew it was time to go."
Jumper's departure can be seen as a statement on the cultural direction after the merger. He has worked at DeepMind for nearly nine years, going through the independent research period, the integration period after the merger, and the current stage of increasing pressure for productization. When the research environment increasingly demands alignment with search engine key performance indicators, leaving becomes a calculated but not difficult decision to make.
The deeper problem is that less than two years after Shazeer's return, the pace of AI product releases has not accelerated significantly. Gemini narrowed the capability gap with ChatGPT, but never became a leader in the niche. He did not publicly express his dissatisfaction, and his statement on X was standard professional phrasing, but the actions spoke for themselves.
The talent landscape is undergoing irreversible restructuring
This brain drain is no longer just a matter of a few people changing jobs.
Google can bring back top researchers, but it can’t change the most fundamental thing: its core business model is advertising, and AI is an enabling tool, not its ultimate mission. Money can buy someone back, but money can't make Google stop being Google. This means the exodus is not stopping and is a structural trend rather than a few isolated departures.
On the other side, OpenAI and Anthropic are taking their own paths. OpenAI has the strongest competitiveness in large language model research, while Anthropic combines AI security and scientific applications. The two companies have clear boundaries and each has its own moat. Google is stuck in the middle. It has neither the explosive product power of OpenAI nor the brand differentiation of Anthropic in the security field.
What really tilts the talent balance irreversibly is the IPO window period. When top researchers can earn nine or even ten figures of wealth through equity redemption within one or two years, the compensation system of any mature giant cannot compete on the same dimension. The year 2026 is likely to be remembered not because of the rapid advancement of a certain AI capability, but because the talent landscape completed a structural reorganization during this year. In this round of competition, talent density determines model capabilities, model capabilities determine market share, and market share determines the list of winners.
It's not impossible for Google to make a comeback. It has one of the world's largest computing infrastructures, the largest user data reserves, and continues to lead in the number of AI academic paper publications. But all these advantages are based on the premise that you have to have good enough people to use them. And what Google is losing is precisely these people.
This may be the quietest crisis since Google was founded. There are no major product mistakes, no heavy regulatory fines, and no financial explosions. It was just a group of the smartest people who, one by one, chose to leave. In the field of AI, the real moat has never been data, not computing power, or even the model architecture itself. It’s the people who stay and push the boundaries of technology day in and day out. And Google is discovering that retaining these people is much more difficult than training a model with trillions of parameters.