The essence of the talent retention crisis is not salary, but the simultaneous imbalance of organizational structure, computing power allocation rights, and research autonomy.

Author and source: 0x9999in1, ME News



TL;DR

• From June 19 to 24, 2026, Google DeepMind lost five core researchers in six days: Noam Shazeer joined OpenAI, while John Jumper, Jonas Adler, Alexander Pritzel, and Arthur Conmy collectively moved to Anthropic, and Rishub Jain founded a nonprofit focused on AI safety.
• Shazeer's departure was triggered by the reallocation of computing resources; Google redirected his team's computational power to DeepMind's London pretraining team, causing his Transformer variant project to be shelved.
• This is not an isolated incident. According to SignalFire’s 2025 Talent Report, DeepMind engineers are 11 times more likely to move to Anthropic than vice versa, and OpenAI employees are 8 times more likely to move to Anthropic.
• Anthropic completed its Series H round on May 28, raising $65 billion, resulting in a post-money valuation of $965 billion—exceeding OpenAI’s $852 billion—and secretly filed for an IPO on June 1, opening the door for a fall listing.
• Non-compete agreements in the UK could delay the actual start dates for researchers in London, such as those from Jumper, until the end of 2026 or even 2027, but what Anthropic is competing for isn't just time.
• The essence of the talent retention crisis is not salary, but the simultaneous imbalance of organizational structure, computing power allocation rights, and research autonomy.

One, six days, five people, a massive crack

First, clearly understand what actually happened.

On June 19, 2024, John Jumper, co-creator of AlphaFold and a 2024 Nobel Prize laureate in Chemistry, announced his departure from DeepMind, where he had worked for nearly nine years, to join Anthropic. In the same week, Noam Shazeer, co-lead of Gemini—the researcher who co-authored the seminal 2017 paper introducing the Transformer—moved to OpenAI to serve as Head of Architecture Research.

In the following days, Bloomberg struck again on June 24: Jonas Adler and Alexander Pritzel, two senior research scientists regarded by Google as key contributors to Gemini, are also set to join Anthropic. Adler leads AI coding initiatives, while Pritzel has long specialized in foundational system architecture.

It’s not over. Senior research engineer Arthur Conmy, who worked on Gemini 2.5 and AI coding, has announced his move from London to San Francisco to join Anthropic, focusing on alignment research during model training. Canadian researcher Rishub Jain, who spent seven years at DeepMind, is bypassing the big tech route—he’s founding a nonprofit AI safety organization focused on scalable supervision and human alignment.

Six days, five people, different directions, but the same vector—leaving DeepMind.

Is this a normal personnel fluctuation? Clearly not. This is a chain reaction of organizational signals.

II. How Shazeer Left: When Computing Power Became Politics

Many people are watching the salary slips, thinking the battle for AI talent is being fought there—but they’re looking at the wrong battlefield.

What ultimately convinced Shazeer to leave was compute power. According to multiple media reports, Google reallocated his team’s computing resources to DeepMind’s pretraining team in London to “optimize cross-team collaboration.” The research on his variant Transformer architecture was abruptly disrupted when the compute resources were withdrawn.

What is the nuance of this matter?

It exposes a truth: in the era of large models, control over computing power is control over research. Whether a researcher can run experiments, the scale at which they can run them, and how many iterations they can perform determine whether they are a scientist or merely a PowerPoint presenter. When computing power is reallocated by organizations, research autonomy is like having a load-bearing beam removed.

Shazeer left Google in 2021 to found Character.AI, only to be pulled back in 2024 by a $2.5 billion reverse acquisition. This time, his departure is less dramatic—just exhaustion.

Who still wants to be held back by hash rate quotas? Who is willing to make way for someone else’s pre-training plan?

The answer is written on the resume update button.

III. Anthropic’s Siphoning: Valuation, Equity, and Research Space

Why Anthropic?

Hold the data for now. According to SignalFire’s 2025 State of Talent Report: Engineers from DeepMind are about 11 times more likely to move to Anthropic than vice versa, and the flow from OpenAI to Anthropic is about 8 times greater. Anthropic’s two-year retention rate is 80%, DeepMind’s is 78%, and OpenAI’s is slightly lower.

What does 11x mean?

This isn't liquidity; it's one-way drainage. It's like pumping water on one side while leaking on the other.

Money is certainly one factor. On May 28, 2026, Anthropic closed a $65 billion Series H funding round led by Altimeter, Dragoneer, Greenoaks, and Sequoia, achieving a post-money valuation of $965 billion. This figure surpassed OpenAI’s $852 billion valuation from March, making Anthropic the highest-valued private AI company for the first time.

What does 965 billion mean? It means the book value of early employees' equity has doubled or tripled over the past twelve months, indicating that new hires receiving restricted stock now have an implied liquidity expectation. Coupled with Anthropic’s secret filing for an IPO on June 1st and the opening of an autumn listing window, the leverage effect of pre-IPO options is at its peak.

But if you think Anthropic is only poaching talent with money, you’re mistaken.

Conmy focused on alignment, Jumper pursued science and life sciences, and Rishub Jain left entirely to work on security for a nonprofit—these individuals chose Anthropic or the Anthropic path because they valued clarity in research priorities. Anthropic’s clear focus on safety, alignment, and interpretability, combined with the commercial success of the Claude series generating substantial cash flow—reported by multiple media outlets to have reached an annual operating revenue of approximately $47 billion—has provided a far more definitive answer to the question: “Does what I’m doing actually matter?”

Money solves "whether to stay or go"; the issue solves "where to go." Anthropic holds both levers simultaneously.

IV. Google’s Retention Dilemma: It’s Not About Lack of People, But Structural Issues

Demis Hassabis’s remarks at the Cannes Lions Festival were intriguing. He openly acknowledged that the competition for AI talent is currently the most intense in the history of the tech industry, while insisting that Google is “still winning.”

This is classic crisis communication—maintaining stability externally while reassuring internally.

But can it hold steady?

Google doesn’t lack money. Alphabet’s cash and short-term investments as of Q1 2026 remain in the hundreds of billions of dollars. Google also doesn’t lack brand strength—DeepMind’s reputation, AlphaFold’s Nobel Prize, and Gemini’s product portfolio are all still intact.

What's missing?

What’s lacking is organizational flexibility. This wave of departures has sharply highlighted three structural issues:

First, the internal power struggle over compute allocation. The Shazeer incident is not an isolated case but a continuation of longstanding tensions following the merger of DeepMind and Google Brain. In a large research organization, once resource allocation becomes "politicized," frontline researchers begin to lose patience.

Second, the decision-making chain is too long. How many layers of review, how many cross-regional coordinations, and how many OKR alignments are required for a底层architecture overhaul that Pritzel wants to推进? In comparison, Anthropic is one order of magnitude smaller, with a much shorter decision radius.

Third, non-compete agreements under UK law have backfired as negative PR. These London researchers from Jumper won’t be able to officially start working at Anthropic until the end of 2026 or even 2027. What was intended as an employer protection mechanism has instead fueled a narrative that “DeepMind can’t retain talent and can only delay departures through contracts,” further reinforcing negative perceptions of Google in the talent market.

The essence of the talent retention crisis has never been salary, but whether the organization’s collective momentum still accelerates the realization of individual researchers’ value. When that momentum becomes resistance, even Nobel laureates will vote with their feet.

Five: The Dissolution of the AlphaFold Team: The Path to the Breakup of a Dream Team

Jumper left—what now for the AlphaFold team?

Scattered observations over the past few months have begun to paint a picture: some core members have moved to Isomorphic Labs—a drug discovery company spun out of DeepMind, with Hassabis serving as CEO—indicating that at least these talents remain within the Google ecosystem. Others have followed Jumper to Anthropic, betting on Anthropic’s initiatives in life sciences and healthcare.

This is worth a second look.

AlphaFold is DeepMind’s most scientifically impactful achievement over the past decade, with its protein structure database cited millions of times by biologists worldwide. The 2024 Nobel Prize represents both scientific recognition and valuable brand equity. The dissolution of this team not only signifies the diffusion of technical expertise but also indicates a crack in DeepMind’s moat in the field of scientific AI.

Anthropic acquired Jumper not to enhance Claude’s next-generation conversational abilities, but to open a door—into protein design, drug screening, and computational life sciences—fields backed by hundreds of billions of dollars in pharmaceutical markets and immense scientific influence.

As for Isomorphic Labs, in a sense, it serves as Google’s “lightning rod”: diverting scientists who don’t want to stay at DeepMind into an independently structured subsidiary, ensuring that talent doesn’t directly flow to competitors. This is a defensive move, not an offensive one.

Six: Impact on AI Companies: Three Things That Will Change

Talent migration has never been a neutral event. After this round of upheaval, at least three things will change.

First, flat organizational structures will once again become an internal topic for major tech companies. Google will not immediately dismantle DeepMind, but Hassabis must answer one question: How can the allocation of computing resources and research priorities among Gemini, Search, and Cloud no longer become a source of internal friction?

Second, valuation and IPO timelines will accelerate in reverse. Once Anthropic’s IPO materializes in fall 2026, the entire AI industry’s private valuation benchmarks will be recalibrated. OpenAI, xAI, and Mistral will be forced to respond—either accelerating their own IPO plans or continuing to raise private round valuations to maintain employee equity appeal. The 11x multiple cited in SignalFire’s report may not shrink over the coming quarters; instead, it could expand further.

Third, research topics will continue to diversify. In the areas of safety and alignment, Anthropic has already established a significant talent advantage. In infrastructure, OpenAI has filled a key gap with the addition of Shazeer. Google’s remaining strengths may lie more in productization and system integration—but these advantages appeal far less to top-tier research talent than the opportunity to conduct cutting-edge research.

As for second-tier players like Meta, xAI, and Mistral, it’s hard to say whether they’ll benefit from this wave of volatility. They’re more likely to attract mid-tier talent, while the top-tier candidates continue to primarily flow toward Anthropic, with some going to OpenAI.

Seven: The True Turning Point in the Talent War

As I write this, I look back on these six days.

Shazeer chose the path of rivalry, Jumper chose the path of issue selection, Adler and Pritzel chose the path of organizational fatigue, Conmy chose the path of passion for the domain, and Jain chose the path of returning to their original purpose. Five people, five reasons—but the unspoken message behind them all is just one: At DeepMind, I can no longer do what I want to do.

This is the real question Hassabis needs to answer—not more zeros on a paycheck, not an extra word of encouragement in a meeting, but answering the question every researcher asks themselves before opening their IDE each morning: Will the code I write today, the experiments I run, and the PRs I submit become meaningless labor if compute power is pulled away, topics are rewritten, or decisions are delayed?

Anthropic is currently in a delicate position. Its $965 billion valuation is inflated, its fall IPO was driven by hype, and although its annualized operating revenue has reached $47 billion, the multiple between revenue and valuation remains astonishing. Will the market eventually correct this figure with future earnings reports? Very likely.

But that’s another story.

This week in June 2026, the talent landscape of the AI industry quietly shifted its anchor. The anchor is no longer on salary, not on brand, and not even entirely on valuation. It’s on this one phrase: “Can I continue doing the research I want to do here?”

The first person to answer this question correctly wins the next round.

The rest are footnotes.

Reference materials

1. Bloomberg, "Google Poised to Lose Two More Senior AI Staffers to Anthropic," June 24, 2026.
2. Reuters / CNBC, "John Jumper to Leave Google DeepMind for Anthropic," June 19, 2026.
3. TechCrunch, "AI researchers continue to leave Google for its rivals," June 24, 2026.
4. Semafor, "DeepMind Chief Demis Hassabis Says Google's Still Winning the AI Talent War," June 23, 2026.
5. SignalFire, "2025 State of Talent Report" (on the 11x flow ratio from DeepMind to Anthropic and two-year retention data).
6. Anthropic official announcement, "Anthropic raises $65B in Series H funding at a $965B post-money valuation," May 28, 2026.
7. The New York Times, "Anthropic Files to Go Public, Setting Stage for Huge I.P.O.", June 1, 2026.
8. CNBC, "Anthropic Surpasses OpenAI as Most Valuable AI Startup, Nears $1 Trillion Valuation," May 28, 2026.