Google Just Lost Four of Its Best People in One Day. That's Not Bad Luck.

Here’s what actually happened this week, in order. Sundar Pichai posted a memo announcing that Demis Hassabis is leaving day-to-day control of DeepMind to become chairman of the unit and chief scientist of all of Alphabet, with DeepMind’s CTO, Koray Kavukcuoglu, taking over as SVP and reporting straight to Pichai. Buried in the same announcement: Jeff Dean, Google’s chief scientist and a 27-year veteran who was one of the company’s first thirty employees, is leaving to co-found a new venture called Discovery Loop. He’s not going alone. Sanjay Ghemawat, the engineer who co-built Bigtable and MapReduce, the systems that made Google’s scale possible in the first place, is going with him. So are Oriol Vinyals, a DeepMind VP, and Quoc Le, a co-founder of Google Brain. Four of the people most directly responsible for Google’s entire AI research identity left in a single announcement, and the stock dropped 5 percent, briefly erasing close to $190 billion, before the day even settled.

That alone would be a bad week for any company. But it isn’t a standalone event, and pretending it is would be letting Google off easy.

Start with Noam Shazeer, because his timeline is the clearest evidence of the pattern. Shazeer co-invented the Transformer architecture, the paper that every model from GPT to Gemini to Claude is built on, while at Google. In 2021 he left, reportedly because Google wouldn’t ship the conversational AI he wanted to build, and co-founded Character.AI. In August 2024, Google paid $2.7 billion, structured as a licensing deal so it wouldn’t look like an acquisition, largely to get him back. He came back as co-lead of Gemini, the company’s flagship answer to ChatGPT. Two years later, in June 2026, he left again, this time for OpenAI. Google didn’t just lose Noam Shazeer once. It identified the exact problem, paid nearly three billion dollars to fix it, and watched the fix walk out the door on the same timeline as the first time.

Now look at where all of this talent is actually landing. Shazeer is at OpenAI. Google DeepMind’s AlphaFold lead, John Jumper, has been reported to be headed to Anthropic. Jeff Dean and three of his closest collaborators just built their own lab from scratch, and Alphabet is reportedly providing it with cloud and compute support, which tells you Google would rather fund a friendly spinout than find a way to keep this team inside its own walls. Every frontier AI company that matters right now has a meaningful chunk of Google or DeepMind DNA in its senior ranks. That’s not a coincidence and it’s not simply the free market doing its thing. Something inside Google is pushing these people out the door.

We don’t think competition is the problem here. A world where OpenAI, Anthropic, and a handful of well-funded independent labs are all racing hard against Google is a better world than one where Google coasts on its distribution advantage. What’s frustrating is watching the company that trained an enormous share of the people building that competition fail to hold onto them. Google invented the transformer. Google built the infrastructure the entire industry now runs on. It has more data, more compute, more distribution, and more cash than nearly anyone else on the planet, and it keeps handing its most important people a reason to leave and build somewhere else instead.

The Discovery Loop framing deserves real credit. Structuring this as a spinout Alphabet still has access to, rather than a clean break, is a genuinely smart piece of corporate strategy. It keeps some upside, reduces direct headcount and management complexity, and probably softens some of the antitrust scrutiny that comes with concentrating too much AI talent under one roof. Whoever designed that deal deserves credit for softening the blow. But strategic packaging doesn’t answer the actual question, which is why the company’s most senior AI research talent keeps deciding the best version of their work happens somewhere else, repeatedly, under the same leadership, across multiple product cycles.

That’s the question a board exists to ask, and it’s a fair one to put on Sundar Pichai directly. He’s been CEO since 2015 and has now watched Shazeer leave twice, Jumper reportedly head for a competitor, and his own chief scientist walk out alongside three of DeepMind’s most senior researchers, all inside a leadership structure he built and runs. Maybe Kavukcuoglu stabilizes DeepMind and this turns out to be the last of it. Maybe the incentive structure gets fixed and the next generation of Google’s AI researchers stays put. But four departures in one day, on top of a $2.7 billion round-trip that didn’t hold, is the kind of pattern that should have a board asking whether the person running the company is the right person to fix it, not whether the next press release can spin it well enough.

We want Google in this race, building great products, pushing the frontier labs to move faster because Google is moving fast too. Right now the biggest thing standing between Google and that outcome is Google’s own inability to keep the people who could actually get it there, not OpenAI, not Anthropic.