Why the Best AI Engineers Leave Companies That Are Winning

The naive way to read the news is that Google is losing its AI edge.

Four of the most influential AI researchers in the industry — Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le — left on August 5 to start a company called Discovery Loop. Demis Hassabis stepped down as DeepMind's CEO to become Alphabet's newly created chief scientist. Two weeks earlier, Alphabet had reported its best quarter in history: revenue grew 24 percent to $119.8 billion, earnings per share hit $9.11, and Google Cloud grew 82 percent.

The stock dropped 4 percent on the announcement. Investors treat leadership departures as a signal of internal trouble. But the financial data says the opposite. So which frame is right?

I think the answer isn't about whether Google is winning or losing. It's about what happens when a company wins at the wrong thing.

Dean and his team spent the last two decades building the infrastructure that made Google possible. Dean co-founded Google Brain in 2011. Ghemawat co-authored the foundational papers on MapReduce and the GFS distributed filesystem that underlie Google's entire computing stack. They weren't hired; they were the people who figured out how to make the machine work at planetary scale.

Then Google figured out how to sell ads on top of that machine, and the machine became enormously profitable, and the incentives shifted. The company's priorities became what kept a $3 trillion revenue engine humming. Frontier research — the kind of exploratory, high-risk work that produces new architectures — started competing with product optimization for compute, attention, and organizational bandwidth.

Oriol Vinyals said it directly. Large organizations have inertia that makes radical changes difficult. The team wanted to build something different.

This isn't the first time. In 2025, Google lost at least 11 AI and cloud executives, mostly to Microsoft. In June 2026, Nobel Prize winner John Jumper left for Anthropic and Noam Shazeer — who had returned in 2024 after Google paid roughly $2.7 billion to license his Character.AI — left for OpenAI. A venture capital firm's analysis found that Google DeepMind engineers were 11 times more likely to move to Anthropic than Anthropic employees were to join Google.

The pattern is not that Google is failing. It's that Google is too big to be interesting to the people who build new things. The kind of researchers who want to discover new transformer architectures or automate the scientific method don't need a platform with 950 million monthly active users. They need compute, freedom, and a problem that hasn't been solved yet. A $120 billion quarterly revenue machine, however well-funded, is not that environment.

The interesting detail is what Google did about it. Instead of blocking the departures or claiming the losses don't matter, Alphabet invested in Discovery Loop and committed to being its cloud partner for the first year. Sundar Pichai attempted multiple meetings to retain them and, when that failed, structured a relationship that keeps the team connected to Google's infrastructure without constraining their research direction.

That's the right response to this kind of loss. It treats frontier researchers not as employees to be retained but as independent operators to be courted. The people who matter in frontier AI don't respond to equity grants. They respond to the shape of the work.

The real question for investors isn't whether these departures signal decline. It's whether Google's commercial momentum can sustain itself even as its most creative researchers keep leaving. The Q2 numbers suggest it can. Cloud revenue reached $24.8 billion with $8.8 billion in operating income, up from $2.8 billion a year earlier. Nearly 90 percent of the Fortune 100 now use Gemini Enterprise. Model APIs are processing 22 billion tokens per minute.

Google has built a distribution and infrastructure moat that doesn't depend on any single researcher. The problem is that the people who could build the next architectural breakthrough — the kind that makes current moats irrelevant — have increasingly little reason to stay inside the machine rather than outside it.

The test is simple. Watch where the people who discover new architectures choose to work. If they keep leaving companies that are commercially successful but organizationally large, then the next generation of breakthroughs will come from small teams, not from the companies that currently own the market. And if that's true, then Google's current commercial dominance is a leading indicator of where the next disruption will come from — not a shield against it.

I haven't seen enough evidence to call Google a sell. But I suspect the investors who treat this as a headline problem are asking the wrong question. The question isn't whether Google will lose more people. It's whether anyone has figured out how to make frontier research possible inside a company that's also responsible for keeping the world's search engine, ad platform, and cloud infrastructure profitable every quarter.