Google DeepMind bled dry: Nobel Prize winner John Jumper joins Anthropic, Transformer architect Noam Shazeer flees to OpenAI — the AI talent war enters a brutal phase

Google DeepMind bled dry: Nobel laureate John Jumper joins Anthropic, Transformer architect Noam Shazeer bolts for OpenAI — the AI talent war enters a brutal phase

🔎 Two departures, one earthquake

On June 19, 2026, Google DeepMind loses John Jumper, 2024 Nobel Prize in Chemistry laureate and co-creator of AlphaFold. Two days earlier, it was Noam Shazeer — co-author of the foundational paper "Attention Is All You Need" and co-lead of Gemini — who announced his departure for OpenAI.

These two departures are not accidental. They are part of a talent hemorrhage that has been going on for six months: Andrej Karpathy joined Anthropic on May 19, 2026, and David Silver, the architect of AlphaGo and AlphaZero, left DeepMind in January 2026 to found Ineffable Intelligence with a $1B seed round.

The paradox is striking. Demis Hassabis publicly declares that we are "at the foothills of the singularity," while Gemini 3.5 Pro is slated for release before June 30. But his most brilliant lieutenants are voting with their feet.

The key points

• John Jumper (2024 Nobel Prize in Chemistry, AlphaFold) leaves Google DeepMind for Anthropic on June 19, 2026, according to Bloomberg and The Next Web.
• Noam Shazeer (co-inventor of the Transformer, co-lead Gemini) joins OpenAI in mid-June 2026, reports 9to5Google and The Next Web.
• Google acquired Character.AI for $2.7B in late 2024 primarily to bring Shazeer back. Less than two years later, he leaves again.
• These departures follow those of Andrej Karpathy to Anthropic (May 2026, CNBC) and David Silver to Ineffable Intelligence (January 2026).
• Anthropic is valued at $965B. OpenAI is preparing for an IPO. DeepMind remains an internal division of Alphabet.

Recommended tools

John Jumper at Anthropic: when a Nobel Laureate leaves applied science for safety

Jumper is not joining Anthropic to do computational biology. He is joining to apply his expertise in molecular modeling to the safety of advanced AI systems.

This is a strong signal. AlphaFold revolutionized structural biology — the revue Nature called this breakthrough a "transformer of biology" in 2021. But Jumper is aware that the models he helped build are becoming more powerful than the tools they were meant to serve.

Anthropic, with its $965B valuation, can offer equity packages that Google simply cannot match. Not because Alphabet lacks cash, but because an Anthropic share has a much steeper growth trajectory than a GOOG share. A senior researcher at DeepMind receives Alphabet RSUs. At Anthropic, he receives equity in a company that is already worth nearly a trillion and is aiming to double before the IPO.

Geopolitical context also plays a role. Anthropic recently opened an office in Seoul and signed an MOU with South Korea on AI safety. AI safety is no longer an abstract research topic — it is a diplomatic issue. Jumper is joining this momentum.

Noam Shazeer at OpenAI: Google's $2.7B reduced to nothing

The irony is cruel for Google. In late 2024, Alphabet acquired Character.AI for around $2.7B. The deal was widely interpreted as a disguised "acqui-hire": the real price tag was the return of Noam Shazeer.

Shazeer is no ordinary researcher. He is a co-author of the "Attention Is All You Need" (2017) paper, which introduced the Transformer architecture. Without this paper, there is no GPT, no Claude, no Gemini. He is literally the man who invented the engine behind all of generative AI.

Less than two years after his forced return (or rather: bought-out return), he is leaving again. And not for just anywhere: to OpenAI, Google's direct rival in the model race. OpenAI is also entering what Sam Altman describes as a "Phase 3" — a personal AGI for every human, with an automated AI researcher.

Why OpenAI? The IPO approach is a massive magnet. Engineers and researchers joining OpenAI today are betting on a post-IPO valuation that could exceed $300B. It's a rational calculation, not an ideological choice.

Shazeer takes with him an intimate knowledge of the Gemini architecture, the technical trade-offs made during development, and the weaknesses he knows better than anyone. It is a considerable transfer of strategic know-how.

Karpathy and Silver: the exodus begins well before June 2026

The departure of Jumper and Shazeer makes the headlines, but the movement started earlier.

Andrej Karpathy joined Anthropic on May 19, 2026, according to CNBC. Karpathy is a unique figure: co-founder of OpenAI, former AI director at Tesla, trainer of a generation of developers through his Stanford courses. His arrival at Anthropic is not a surprise — he has always been close to Dario Amodei — but the timing is a message.

David Silver left in January 2026 to found Ineffable Intelligence, with a $1B seed round. Silver is the man behind AlphaGo, AlphaZero, and much of DeepMind's reinforcement learning heritage. His departure to create a new entity rather than join an existing competitor is significant: he likely judges that DeepMind's research model no longer allows him to innovate as he wishes.

Besides, Google DeepMind's work on reinforcement learning in strategic games and Atari continues to be published, but without Silver's signature. The expertise remains, the architect has left.

Why DeepMind Can't Keep Its Stars

The problem is structural. DeepMind is a division of Alphabet. Despite relative autonomy, researchers don't have the same upside as elsewhere.

A senior engineer at Anthropic or OpenAI holds equity in a company that has not yet realized its full market value. At Google, RSUs are capped by Alphabet's market capitalization — already colossal, meaning there is less room for multiplication.

Culture also plays a role. Hassabis is a scientist. Altman and Amodei are entrepreneurs. For a researcher who wants to see their work transform the world at maximum speed, the framework of a $965B startup is more stimulating than that of a $2,000B tech conglomerate.

There is also a size factor. DeepMind now counts thousands of researchers. The "small team突破" effect that produced AlphaGo and AlphaFold is diluted. Recent departures suggest that the best want to regain that agility.

Anthropic vs OpenAI: two competing attraction models

The two major departures of June 2026 perfectly illustrate the bifurcation of the market.

Anthropic attracts with safety and fairness. Jumper and Karpathy choose Anthropic because the "safety-first" mission resonates with their backgrounds. Anthropic sells the idea that building AI responsibly is just as profitable as building it fast. Its $965B valuation proves them right in the market. The recent Anthropic acquisition of Stainless for over $300M also shows that the company is investing massively in the infrastructure around Claude — a signal of confidence for talent who want to build, not just research.

OpenAI attracts with scale and an IPO. Shazeer chooses OpenAI because it's where models are deployed at the largest scale, and where the promised IPO creates irresistible financial momentum. OpenAI is also in Phase 3 of its roadmap — personal AGI. For an architecture engineer like Shazeer, it's the most ambitious project in the world.

The agentic ranking of June 2025 already set the tone: OpenAI's GPT-5.5 dominated at 98.2, followed by Gemini 3 Pro Deep Think at 95.4 and Claude Opus 4.7 Adaptive at 94.3. The war of scores is also a war for talent.

The Hassabis paradox: "foothills of the singularity" but the team is emptying out

Demis Hassabis uttered a phrase that will circulate throughout Silicon Valley: we are at the "foothills of the singularity." This is an extraordinarily strong statement from a scientist known for his caution.

Yet, at the exact moment he says this, his lieutenants are leaving. This is the central paradox: the greater the promise, the stronger the pressure on individuals, the more irresistible the temptation to leave.

Gemini 3.5 Pro is scheduled for release before June 30, 2026. In the June 2025 overall ranking, Gemini 3.1 Pro was in the lead with 92 points, ahead of GPT-5.5 at 91 and GPT-5.4 Pro at 91. But without Shazeer to steer the architecture, and with a probably tense internal climate, the timing is delicate.

Hassabis must also contend with the discussions at the G7 summit in Evian, where Altman, Amodei, and himself were brought together for the first time around the table of global AI governance. Three CEOs, but two of them just stole his best people.

Consequences for science: AlphaFold without Jumper

The most underestimated impact of this leak concerns science, not just tech.

AlphaFold has predicted the structures of over 200 million proteins. Jumper's work accelerated biological research by several decades. But AlphaFold is not a finished product — it is an ongoing research program.

Without Jumper, the question is: who carries the vision? DeepMind has teams, certainly. But the difference between an excellent team and an excellent team led by a Nobel Prize winner is measurable. Related modeling work, such as large-kernel convolutional neural networks for continuous gravitational wave searches, shows that expertise in cutting-edge computational modeling is rare and precious.

Google released VaultGemma, a differentially private LLM, which shows that DeepMind continues to innovate. But external researchers' confidence in the AlphaFold program could erode if departures multiply.

Implications for the model race: Gemini 3.5 Pro under pressure

The June 2025 LLM model rankings showed a tight race at the top:

Gemini 3.5 Pro is expected to maintain or widen this lead. But losing the chief architect two weeks before release creates an objective risk. The remaining engineers must execute a vision whose main designer is now at the competitor.

Shazeer knows Gemini's architectural trade-offs better than anyone. If he shares even intuitively this knowledge with OpenAI teams, Gemini's competitive advantage shrinks.

The geopolitical stake: AI as a power weapon

This war for talent is not just a Silicon Valley affair. It has direct geopolitical implications.

The United States recently ordered the blocking of certain Anthropic models, showing that AI models are now treated as sensitive technologies on par with advanced semiconductors. In this context, every talent movement between American labs alters the balance.

DeepMind is based in London, with a strong presence in the United Kingdom. Anthropic and OpenAI are based in San Francisco. When Jumper and Shazeer leave London for California, it is not just an inter-company transfer — it is a transfer of technological capability from a European ally to the heart of the American technological complex.

The G7 summit in Évian, where Altman, Amodei and Hassabis met, was supposed to address these imbalances. But how can you regulate an ecosystem where talent circulates faster than laws?

The broader context: an unprecedented war for talent

The departures from DeepMind are part of an industry-wide dynamic. Senior AI researcher salaries have skyrocketed: a principal researcher can now negotiate a total package of $10M to $50M over four years, between salary, bonus, and equity.

This is not new in itself — Silicon Valley has always seen talent wars. But the scale is different. You're not poaching a good engineer. You're poaching a Nobel Prize winner, the inventor of the Transformer, the father of AlphaGo.

Senior-level AI researchers likely number fewer than 500 in the world. Each departure is a systemic loss for the institution that suffers it. Google can hire 1,000 junior engineers — that will not make up for the departure of Shazeer.

The phenomenon recalls the research on modélisation des dunes mouvantes: the most imposing structures are also the most vulnerable to shifts in their foundation. DeepMind has built a massive dune. The sand is moving underneath.

What Google can still do

Google is not powerless, but its options are limited.

The first would be to spin off DeepMind into an independent entity with its own valuation. This would make it possible to offer more attractive equity. But Alphabet has refused this option on several occasions — probably because DeepMind's value is largely captured by the Google ecosystem (Search, Cloud, Android).

The second would be to promote a new generation of internal leaders. Researchers like those who worked on the linearization of finite abelian subgroups of the Cremona group or on the reduction of two-dimensional atomic models show that mathematical depth still exists at DeepMind. But public visibility is lacking.

The third, more cynical option: accelerate model releases to prove that DeepMind is doing fine without its stars. Gemini 3.5 Pro before June 30 is precisely this strategy. The risk is releasing a suboptimal product and retrospectively validating the departure of the architects.

❌ Common mistakes

Mistake 1: Thinking Google is "fine" because it's rich

Alphabet has $80B in cash. But money doesn't create Nobel Prize winners. Wealth allows you to hire, not to retain. Confusing financial resources with talent retention is the classic mistake of analysts who don't understand the dynamics of startups vs. conglomerates.

Mistake 2: Reducing these departures to salary issues

Equity is a factor, but not the only one. Karpathy could have negotiated anywhere. Jumper had immense scientific prestige at Google. The choice is also ideational: researchers want to believe they are building the future, not optimizing a search engine.

Mistake 3: Overestimating the short-term impact on products

Gemini 3.5 Pro will likely release with competitive scores. The remaining teams are excellent. The real impact will be measured in 12-18 months, on next-generation architectures, where the absence of visionaries will be cruelly felt.

❓ Frequently Asked Questions

Who is John Jumper?

Researcher at Google DeepMind, co-creator of AlphaFold, 2024 Nobel Prize in Chemistry laureate for revolutionizing protein structure prediction. He joins Anthropic on June 19, 2026.

Why is Shazeer so important?

Co-author of "Attention Is All You Need" (2017), the paper that introduced the Transformer architecture — the foundation of all modern LLMs. He was also co-lead of the Gemini models at Google.

How much did Google pay to get Shazeer back?

Around $2.7 billion via the acquisition of Character.AI in late 2024. He leaves for OpenAI less than two years later, making this investment largely symbolic.

Is Anthropic really worth $965 billion?

That is its latest reported valuation (June 2026). This figure reflects market expectations regarding Claude's trajectory and a potential IPO, not current revenues.

What is the impact on AlphaFold?

In the short term, limited — existing models continue to function. In the medium term, uncertain: the research program loses its scientific leader and its long-term vision.

✅ Conclusion

Google DeepMind is losing its most important architects at the worst possible time — just days before the release of Gemini 3.5 Pro and while Hassabis himself acknowledges the imminence of a technological tipping point. The AI talent war is no longer a metaphor: it is a reality redrawing the Silicon Valley hierarchy in real time. To follow the evolution of the models emerging from this reshuffling, test the free AI APIs and judge for yourself.
```