2026-06-30 · Article

Google Lost Six Minds in a Fortnight

The people who built modern AI just left Google for its rivals.

By Tech Sight Intelligence

When the architects of the modern AI era leave for your competitors, the industry shifts under your feet

Noam Shazeer co-wrote the paper that made modern AI possible. Google paid $2.7 billion to get him back. He just left for OpenAI. He was the first of six.

In the last fortnight of June 2026, six of the most consequential researchers in AI history departed Google DeepMind for its direct competitors. The list reads like the founding document of the modern AI era.

Who Left, and Where They Went

Shazeer's 2017 paper, "Attention Is All You Need," introduced the Transformer architecture underpinning every large language model in existence. Google acqui-hired him from Character.ai in 2024 for $2.7 billion. He left for OpenAI.

John Jumper, who shared the 2024 Nobel Prize in Chemistry for AlphaFold 2, the protein-folding model that changed structural biology, joined Anthropic. Jonas Adler and Alexander Pritzel, both key contributors to the development of Gemini, also moved to Anthropic. Denny Zhou, who founded Google Brain's reasoning research team and built the technical foundations for multi-step thinking in language models, updated his LinkedIn to Meta Superintelligence Lab. David Silver, one of DeepMind's founding researchers and the world's foremost reinforcement learning researcher, launched his own startup: Ineffable Intelligence.

Google's share price fell more than 5% on the news. It was not an overreaction.

"Shazeer and Jumper are not merely talented researchers. They are architects of the modern AI era. Now they are building for someone else."

Why They Left

The departure of six leading researchers in a fortnight reflects a convergence of factors that Google cannot address with a single intervention. Anthropic and OpenAI are both reported to be approaching initial public offerings, offering pre-IPO equity that publicly traded Alphabet cannot match structurally. The financial upside at a company whose valuation is still being set is a different proposition from the upside at a company already in the NASDAQ 100.

Inside DeepMind, a strategic pivot toward commercial coding applications has shifted resource allocation. Researchers whose work is aligned with longer-horizon AI research, including world models, fundamental reasoning, and reinforcement learning, have found that alignment weakened. The people who left are, by and large, the ones whose work sat furthest from Google's near-term coding productivity agenda.

For Shazeer, the draw to OpenAI was scale of ambition. For Jumper, Anthropic's safety-first culture aligned with his values following AlphaFold's applications in drug discovery. These are not the same motivation. Google is losing talent for multiple reasons simultaneously, and there is no single fix.

What Departs With Them

The practical consequences operate at two levels. The first is immediate: six people with deep expertise in the architectures and training dynamics of frontier models are now building for Anthropic, OpenAI, and Meta. Those organisations have acquired institutional knowledge of Google's approaches that no conventional competitive intelligence operation could have obtained.

The second is harder to quantify. Senior researchers carry not just technical skills but institutional memory: the accumulated understanding of what was tried, what failed, and why certain architectural choices were made. That knowledge does not transfer in a handover document.

Enterprise buyers should note the implication. The Gemini models that Adler and Pritzel helped build continue to be developed. But the people who best understood their foundations have moved on. That is not a reason to abandon Google AI products. It is a reason to track capability trajectories across providers more carefully than a static vendor assessment suggests.

What It Means for AI Reliability

A small number of people understand how frontier AI models work at the level needed to anticipate failure modes, design evaluation methodologies that catch genuine risks, and make judgment calls about deployment readiness under competitive pressure. When that expertise migrates, it does not disappear. It moves.

Anthropic and OpenAI have just become meaningfully stronger in their ability to develop and evaluate frontier models. Google's frontier work faces a period of rebuilding capability it has lost.

For enterprise buyers assessing the long-term reliability of AI providers, the question is not who publishes the best benchmark results this month. It is who retains the foundational research capability to understand why their models do what they do, and to correct course when they do not.

"Benchmarks tell you where a model is today. Research depth tells you where the organisation will be in two years."

Sources: LinkedIn profile updates for Shazeer, Jumper, Adler, Pritzel, Zhou, and Silver (all confirmed via public posts); Fortune and TechCrunch reporting on departures; Alphabet share price data, June 2026; Nobel Committee records (Jumper, 2024 Chemistry Prize).