Two DeepMind departures in 48 hours drop Google shares 5% and revive AI race anxiety

Welcome to Eye on AI. Google DeepMind’s “stay forever” era just got punctured: Noam Shazeer announced he is leaving to join OpenAI, and John Jumper said he is leaving to join Anthropic, with both moves landing within 48 hours. The market noticed. Google’s shares tumbled more than 5% on Monday after the news.

This is not just HR drama. It’s a credibility test for Alphabet’s AI leadership at a time when observers say DeepMind’s top models are slipping on benchmark leaderboards, and its model release pace is lagging. Shazeer’s exit matters for a very specific reason: he helped build Google’s early LLM-based chatbot system, LaMDA, in 2021, and he had previously returned to Google in 2024 after cofounding Character.ai. Jumper’s Nobel-level credibility also raises the stakes: he shared the 2024 Nobel Prize in chemistry with DeepMind CEO Demis Hassabis for AlphaFold, then stayed focused on protein predictions and the idea of using large language models as tools for science.

So what changed? Start with Shazeer. He announced his departure on Thursday to go to OpenAI. The source traces a backstory: before he left Google the first time, Shazeer is thought to have authored an anonymous memo that later leaked, criticizing Google for being too bureaucratic, slow-moving, and risk-adverse. The narrative then connects to the widely understood turning point: OpenAI launched ChatGPT in November 2022, jump-starting the category in public imagination and putting rivals on the scoreboard. Shazeer and Daniel de Freitas cofounded Character.ai, and then were lured back to Google in 2024 in a deal where Google licensed Character’s technology for a reported $2.7 billion payment.

Now Shazeer is leaving again, and the “why” is the most dangerous part for Alphabet executives. Neither Shazeer nor Jumper have said publicly why they’re leaving, and the source walks through the most obvious explanation: money. Shazeer is thought to have made hundreds of millions of dollars from the Character.ai licensing deal that brought him back to Google, and the implication is that any earn-out period has ended. The piece also suggests both could be beneficiaries of accelerated-vesting Google stock options that the company has used to keep top talent from being lured away by big pay packages at places like Meta’s Superintelligence Lab.

But the story challenges that simplistic read by arguing money may not be the main driver. It notes that Shazeer was likely already extremely well-compensated, and it characterizes Jumper as not appearing primarily motivated by money. That shifts the core anxiety from compensation to mission. If researchers at the center of DeepMind are exiting for what they see as a better scientific opportunity, then Alphabet’s problem is not just retaining individuals. It is whether DeepMind’s culture, incentives, and release rhythm are still aligned with the pace of frontier AI.

Then there is Jumper. He announced his departure after the Shazeer news, moving to Anthropic. Jumper shared the 2024 Nobel Prize in chemistry with Demis Hassabis for work creating AlphaFold, the system that can predict protein shape from DNA sequences and solved a 50-year grand challenge in biochemistry. Post-Nobel, he continued working on models that predict other protein properties, like how they bind to one another and how small molecules used in pharmaceuticals might bind.

The source adds a strategic wrinkle: it suggests science may be getting less priority at DeepMind than in the years leading up to ChatGPT. Google DeepMind still maintains a large team applying AI to fundamental science challenges, and it has created a Gemini-powered system that can act as an “AI scientist” assisting researchers across scientific domains. But it frames a sense that “science” is taking a back seat relative to earlier years. Isomorphic Labs, spun out of DeepMind in 2021 and also led by Hassabis, is described as heavily focused on commercial science rather than “blue-sky” research.

Now connect those internal priorities to external competition. The source says watchers question whether DeepMind is dropping out of the lead pack in the AI race. It points to Gemini 3.5 Flash and Gemini 3.1 Pro often ranked outside the top five on AI benchmark leaderboards, behind models from Anthropic and OpenAI and also Chinese labs such as Zhipu AI and MiniMax. It also flags release cadence: at Google I/O in May, Google said it was preparing Gemini 3.5 Pro for wide release targeting June general availability. If that lands in June, it would be about four months after the last frontier wide release, Gemini 3.1 Pro in February. Meanwhile, Anthropic is described as having released two significant Claude Opus updates in that same time period, plus a new class of AI models called Mythos noted for world-leading long-range autonomous task completion, especially in coding and cyber domains.

The story’s most executive-relevant section is the tension around strategy and risk. Alphabet’s defenders say Google can’t afford the same kinds of chances OpenAI and Anthropic can take, because Alphabet has billions of users and profits to defend. There’s also a fiduciary duty angle: Alphabet cannot make risky bets that could jeopardize public market returns. The piece notes it’s plausible some inside Google believe boldness is unnecessary because distribution can carry the company as long as Google can “more or less match” technological advances.

That’s the tradeoff. If Google downshifts into “not losing,” it can lean on search, ads, and user distribution to outlast pure-play labs. But leaders at other firms, and investors on every cap table with AI exposure, will notice what this looks like on the ground: when elite researchers leave at the exact moment external benchmarks and release timing come under scrutiny, the signal is harder to contain than an earnings miss. For Alphabet’s board and for peers watching the frontier, the message is simple and uncomfortable. In AI, pace and talent are strategy, not inputs. And a 48-hour double departure is a loud datapoint that forces every executive asking the same question: are you playing to win, or just trying not to fall behind?