Google DeepMind Lost Four Elite Researchers in Six Days. Gemini 3.5 Pro Missed Its Public Deadline Too
TL;DR
Noam Shazeer, John Jumper, Jonas Adler, and Alexander Pritzel all left Google DeepMind within six days. Alphabet shed over $100B in market cap in a single session. Gemini 3.5 Pro missed its June deadline. Is this a talent blip or a structural signal?
June 18: Noam Shazeer announces he’s joining OpenAI. June 20: John Jumper confirms he’s going to Anthropic. June 24: Jonas Adler and Alexander Pritzel follow Jumper to Anthropic. Four researchers, six days, all out of Google DeepMind. And today, June 30, Google confirms Gemini 3.5 Pro will not ship this month as promised.
Shazeer co-authored “Attention Is All You Need” (2017), the paper that introduced the Transformer architecture powering nearly every modern LLM. Jumper shared the 2024 Nobel Prize in Chemistry for AlphaFold. Adler and Pritzel were core contributors to Gemini model training. The week couldn’t have looked worse for Google’s AI leadership story.
Here’s where I’ll stake a position: this is a compensation structure problem, not a capability problem. Pre-IPO equity at Anthropic and OpenAI is outcompeting Google’s salary and RSU packages on expected value. If that hypothesis is wrong, I’d expect Google to retain talent even after Anthropic and OpenAI go public. Feel free to push back if you have a different model.
What Happened in Six Days
Shazeer’s departure carries extra weight because it’s his second exit. He left Google in 2021 to co-found Character.AI, describing the company as “too bureaucratic, slow-moving, and risk-averse.” Google later spent $2.5 billion to license Character.AI’s technology and brought Shazeer back. He’s now heading to OpenAI.
Jumper’s path is different. His work at DeepMind was scientific AI: AlphaFold solved the protein structure prediction problem that had stumped biology for 50 years. Choosing Anthropic over OpenAI is notable. Anthropic has been quietly expanding its scientific AI and biosecurity research portfolio.
Adler and Pritzel hit closer to home for Gemini specifically. Both were central to Gemini’s training infrastructure and system design. Their joint announcement on June 24 landed in the same news cycle as the first wave of Shazeer and Jumper coverage.
TechCrunch reported that departures from DeepMind have been accumulating across 2026, including David Silver (reinforcement learning pioneer) launching Ineffable Intelligence. This week just compressed the pattern into a single news event.
The Numbers Behind the Story
A 5.7% single-day drop on Alphabet’s roughly $2 trillion market cap works out to approximately $114 billion in erased market value. Four people. One trading session.
That reaction looks extreme in isolation. The market isn’t pricing the four individuals. It’s pricing the probability that Google’s organizational culture has become structurally unable to compete for frontier research talent. That’s a different and more durable problem.
On model performance: the LMSYS Chatbot Arena currently ranks both Gemini 3.5 Flash and Gemini 3.1 Pro outside the top five by community vote. Arena uses preference-based voting from real users in open-ended conversational tasks, so it doesn’t capture specialized benchmarks where Gemini may perform better. Still, it tracks user perception in the exact deployment context where most commercial value gets realized.
The Gemini 3.5 Pro delay has a concrete technical explanation: token consumption in extended agentic tasks exceeded acceptable thresholds. This is a solvable engineering problem, context windowing adjustments or cost guardrails could close it in weeks. The issue is that the deadline miss landed on the same week as the research exits.
Worth flagging on scale: Alphabet closed an $84.75 billion infrastructure equity raise in June, the largest AI infrastructure financing in corporate history, with Berkshire Hathaway committing $10 billion via private placement. Google does not have a compute problem. The question is whether compute alone determines outcomes, or whether research direction matters enough that losing people who set that direction is a real capability risk.
Three Indicators Worth Watching
Gemini 3.5 Pro’s Arena ranking at launch. This is the cleanest short-term signal. If Pro ships and lands in the Arena top three, the talent exits look like short-term noise. If it launches outside the top five, the organizational thesis gets reinforced in a very public way.
Anthropic’s next scientific AI move. Jumper brings AlphaFold’s full methodology into Anthropic: problem framing, training design, and evaluation at the scientific frontier. Three to six months from now, watch for whether Anthropic publishes anything in biology, chemistry, or materials science. That would be the second-order effect of this hire, and it matters more than the benchmark comparison.
Google’s organizational response timeline. Shazeer’s 2021 memo cited bureaucracy and risk aversion by name. If Google’s visible response to this week consists of silence or slow internal shuffles, the conditions that caused these departures will persist. The observable indicator is whether any public restructuring or compensation package announcement comes from DeepMind leadership before Q3 earnings.
If this was useful, subscribe to the newsletter for weekly AI PM insights and GenAI case studies.
Related Articles
White House Demands Zero Jailbreaks for Fable 5: Security Experts Say It's Impossible
Day 7 of the Fable 5 ban: the White House demands the model be completely jailbreak-proof before it relaunches. Security experts are unanimous: that's technically impossible for any frontier LLM, and Dario Amodei has already refused both of the government's proposed fixes.
US Orders Anthropic to Pull Fable 5 and Mythos 5: A Narrow Jailbreak That Took Down Its Most Powerful Models
The US Commerce Department ordered Anthropic to suspend its two most capable models, Fable 5 and Mythos 5, citing a narrow jailbreak tied to cybersecurity capabilities. Anthropic complied. Then it pushed back.