Noam Shazeer is returning to Google in 2024 thanks to the $2.7 billion acquisition of Character.AI, after founding his own startup following his departure from the empire in 2021. John Jumper, recently awarded the 2024 Nobel Prize in Chemistry for AlphaFold, has signed on with Anthropic just two years later. Between them, a dozen top scientists have chosen to compete or create their own companies. The pattern is clear: something is broken at DeepMind, and the models are not the issue.
The question arises: why does Google have so much talent but struggle to retain it when it needs it most? The answer seems to go beyond salaries or equity; it is tied to something key: how AI is developed and deployed under the pressure of a corporate giant.
The Shazeer Case: When Buying Your Own Startup Is Cheaper Than Changing the Culture
Shazeer is no ordinary researcher. He is a co-author of the influential 2017 paper "Attention Is All You Need," a technical cornerstone of the transformers that power ChatGPT, Claude, Gemini, and virtually all relevant language models. When he left Google in 2021 to co-found Character.AI with fellow ex-Google employee Daniel De Freitas, his message was clear: he preferred starting from scratch to dealing with internal bureaucracy.
Interestingly, three years later, Google paid $2.7 billion for Character.AI, largely to bring Shazeer and his team back. They didn't acquire the technology—Gemini already existed—but rather the talent lost due to their own policies. The deal included a $2.5 billion licensing fee, which essentially acted as a golden parachute for Shazeer to return.
The irony is striking: Google spent nearly three billion dollars to regain what it could have kept by offering more product autonomy. Character.AI had previously raised $193 million, with a valuation of $1 billion. However, Google multiplied that figure by 2.7 just to bring its own people back.
What this reveals: The issue wasn't financial compensation. Shazeer likely had considerable equity at Google. The real challenge lay in the speed of execution, the autonomy to experiment, and the ability to translate research into practice without passing through five layers of corporate approval.
Jumper and AlphaFold: When a Nobel Isn't Enough Reason to Stay
John Jumper offers an even more revealing example. In 2024, he won the Nobel Prize in Chemistry alongside Demis Hassabis for AlphaFold, an AI system that solved the problem of protein folding. Without exaggeration, it's one of the most significant scientific achievements of the decade. However, just six months after the Nobel, Jumper announced his move to Anthropic.
AlphaFold was not just an academic paper; it transformed real biomedical research. Over 2 million researchers have utilized the predicted structures database. This system has accelerated drug discovery, deepened understanding of diseases, and facilitated the design of industrial enzymes. Google had Jumper as its star in biocomputation.
And it lost him.
Anthropic is the startup founded by brothers Dario and Daniela Amodei, also ex-OpenAI, responsible for building Claude. Their focus is on "Constitutional AI" and model safety. For Jumper, the decision means trading Google's corporate scale for a team of 500 people dedicated to fundamental research with immediate application.
The economic signal: Anthropic raised $7.3 billion in its latest round in 2024, achieving a valuation of $40 billion. This enables them to compete in terms of compensation, but what they truly promise is something Google cannot offer: quick decision-making, less internal politics, and real ownership over the direction of their research.
The Broader Pattern: Why DeepMind's Meritocracy Collides with Google's Reality
DeepMind was acquired by Google in 2014 for $500 million, with the promise of operational autonomy. For years, it maintained considerable organizational distance, with offices in London and an academic culture that promoted open publications. The agreement included specific clauses regarding research independence.
However, in 2023, Google merged DeepMind with Google Brain to create Google DeepMind, consolidating all AI efforts under one umbrella. The corporate logic was clear: eliminate redundancies, share resources, and accelerate the development of commercial products like Gemini.
But the reality was different. Researchers who had signed on to work in a British-style academic organization found themselves reporting within Google's product structure. Feedback cycles lengthened. Projects began requiring alignment with Alphabet's goals, and open publication of research became more selective.
Documented cases:
- Shane Legg, co-founder of DeepMind, retains his title but has decreased his profile since the merger.
- Laurent Sifre, principal scientist, moved to Mistral AI in Paris in 2025.
- Victoria Krakovna, AI safety lead, left the company for Redwood Research.
- David Silver (AlphaGo) founded his own autonomous agents startup.
None of these cases involved scandals; simply, world-class talent choosing different environments.
The Empire Paradox: Infinite Resources, Limited Speed
Google DeepMind has advantages that any startup would kill for:
- Access to cutting-edge TPUs with no budget constraints.
- Unique datasets from products with billions of users.
- Exceptional technical talent (not everyone has left).
- A brand that attracts top candidates.
So why is it losing the talent war to Anthropic (500 people), OpenAI (1,700 people), and even startups like Mistral (350 people)?
The answer lies in the cost of coordination. Google has 180,000 employees. Alphabet answers to public shareholders. Any major product decision must align with multiple stakeholders: Cloud, Android, Search, YouTube teams, privacy policies, legal and ethical AI considerations.
A researcher at Anthropic can propose a risky experiment on a Monday and get the green light by Tuesday. At Google, that experiment must validate that it doesn't cannibalize existing products, meets internal ethical guidelines, doesn't create reputational vulnerabilities, and has a clear path toward commercialization or impactful publication.
Concrete example: When Anthropic decided to launch Claude with a 100k token context window in March 2023, it took weeks from decision to deployment. Google had the technical capacity to achieve the same; in fact, its researchers had published papers on long contexts earlier. However, Gemini launched that functionality months later because it required coordination between teams, product testing, and validation of inference costs at a scale of millions of users.
What DeepMind is Doing to Fight Back (And Why It Might Not Be Enough)
Google is not ignoring the problem. It has implemented several strategies:
Google Labs: A semi-autonomous unit where small teams can iterate on experimental AI products without going through complete approval processes. Bard (now Gemini) emerged from here, as did NotebookLM and several multimodal AI experiments.
Aggressive compensation packages: According to reports from levels.fyi, research scientists at DeepMind are receiving packages ranging from $800k to $1.2M annually (base + bonus + RSUs), competitive with offers from Anthropic and OpenAI.
Continuous open publication: DeepMind continues to publish high-impact papers. In 2025-2026, significant works have emerged on:
- Gemini 2.0 and native multimodal architectures.
- AlphaFold 3 for complex protein-ligand interactions.
- Mathematical reasoning systems (AlphaProof, AlphaGeometry).
- Autonomous agents for robotic control.
However, there are structural limits:
-
Equity incentives are different: In a pre-IPO startup, a senior researcher can hold between 0.1% and 0.5% of the company. If Anthropic reaches a valuation of $100B at its IPO, that could mean between $100-500M. Google's RSUs are tied to Alphabet's performance, where individual contributions have a very low marginal impact.
-
Autonomy can't be simulated: Google can create "labs" and semi-autonomous units, but ultimately, everything resides within the corporate hierarchy. A researcher knows that their project can be canceled if it stops aligning with top management's priorities.
-
Impact timeline: At DeepMind, you may work five years on a fundamental problem and publish an extraordinary paper. At Anthropic, you can work six months and see your research implemented in a product used by millions. For certain profiles, that immediate tangibility is worth more than articles in Nature.
The Uncomfortable Questions No One Is Asking Out Loud
The talent exodus from DeepMind raises deeper questions about the future of AI research:
Is the corporate research model sustainable in the long term? Bell Labs, Xerox PARC, Microsoft Research: all have had golden ages where fundamental research coexisted with commercial objectives. But all faced increasing pressure to demonstrate immediate returns on investment. DeepMind is at that critical juncture.
What happens when the best scientists optimize for equity instead of publications? The traditional academic race focuses on citations, grants, and tenure. However, the race in corporate AI is increasingly oriented toward the value captured in equity of startups with potential billion-dollar exits. This shifts the priority of the problems being addressed.
Is the concentration of talent in startups better for AI advancement? Some argue that having the best scientists distributed among OpenAI, Anthropic, Mistral, and others generates more experimentation and diversity of approaches than if they were all at Google/DeepMind. However, this also fragments resources and complicates projects that require massive scale.
Google's Bet: Gemini and Ecosystem vs. Individual Talent
Google's strategic response seems to be: build a robust ecosystem that transcends individuals. Gemini does not rely on a single star scientist. It is integrated into Search, Gmail, Docs, YouTube, and Android, products that boast billions of users.
This strategy has a precedent: Google has lost many of the original creators of its most important products (Gmail, Maps, Chrome), and those products have not only survived but thrived. The bet is that AI research will follow the same pattern.
The risk is clear: AI is not a traditional software product. It is fundamental research in real-time. Breakthroughs arise from the individual insights of exceptional people. AlphaFold exists because John Jumper had the idea to apply attention mechanisms to 3D protein structures. There is no product roadmap that generates those kinds of leaps.
If Google continues to lose the people generating those insights, can it maintain its technological leadership, no matter how many TPUs it has?
The Final Irony: Everyone Is Building the Future, Just in Different Places
What is fascinating about the DeepMind exodus is not that people are leaving—this is common in any organization—but where they are headed and what they are doing there. Shazeer returned, but under different conditions. Jumper joined Anthropic to work on... language models applied to proteins, essentially a continuation of AlphaFold with more autonomy.
They are not abandoning AI research. Rather, they are seeking places where they can have a greater impact, faster, and with less organizational friction. By 2026, those places are increasingly less likely to be established tech giants.
Google DeepMind will remain a pillar of research. It will continue to publish important papers and attract exceptional talent. However, the era in which it could retain all of its exceptional talent has come to an end. That, undoubtedly, changes the game.
The question for founders and scientists is: What kind of organization do you want to do your best work in? Because by 2026, you have more options than ever. And it seems that more and more people are choosing places that are not called Google.