Alphabet Funded the Startup That Took Jeff Dean

On August 5, 2026, Google announced a reorganization of its AI leadership. On the same day, four of the people most responsible for building that AI announced they were leaving to start a company. The two announcements are the same announcement.

Jeff Dean — Google's 30th employee, hired in 1999, most recently chief scientist of Google DeepMind — is leaving after 27 years to be CEO of Discovery Loop, Inc., a Delaware public benefit corporation. He is taking Sanjay Ghemawat, the senior fellow who co-designed most of the distributed systems Google runs on; Quoc Le, a founding member of Google Brain; and Oriol Vinyals, a senior research scientist at DeepMind whose name is on a decade of the lab's most cited work.

The funding round is co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, and Doerr Capital. The amount has not been disclosed.

Alphabet is also in the round.

That last detail is the story. Google did not lose these people to a competitor it can fight. It capitalized them.

What Discovery Loop says it is building

The stated goal is to automate the experimental loop itself — not to write papers faster, but to run science as a throughput problem. The company describes initiating and iterating thousands of experiments simultaneously under algorithmic control, compressing the cycle of hypothesis, execution, measurement, and revision until the bottleneck stops being human attention.

It is also, per the founders, exploring recursive self-improvement: using the system to build better versions of the system.

Whether that works is an open question. What is not open is whether this team can build the substrate for it. Dean and Ghemawat wrote MapReduce, Bigtable, and Spanner — the papers that taught the industry how to run computation across tens of thousands of machines. If the constraint on automated science is orchestration at scale rather than model quality, this is a founding team assembled precisely against that constraint.

The public benefit corporation structure matters less than the market read behind it. A PBC signals that the founders expect to be judged on scientific output over a long horizon rather than on ARR in eighteen months. That is a bet only a team that can raise on reputation alone can make.

The reshuffle underneath

The same week, Demis Hassabis stepped out of the Google DeepMind CEO role. He becomes chairman of the unit and Alphabet's first chief scientist, saying he wants to focus on strategic and global AGI questions, advise his successor, and give more time to Isomorphic Labs, the AI drug-discovery company he co-founded.

His successor does not get the title. Koray Kavukcuoglu — DeepMind's CTO and Alphabet's chief AI architect — takes over day-to-day operations as senior vice president of Google DeepMind, reporting directly to Sundar Pichai. There is no standalone CEO of Google DeepMind anymore.

Read structurally, that is a demotion of the lab's independence and a promotion of its integration. DeepMind was acquired in 2014 and spent a decade defending a distinct research identity. An SVP reporting into the CEO is an org chart that says the research lab is now a product organization. Kavukcuoglu's prior title — chief AI architect for all of Alphabet — tells you which direction the reporting line is meant to pull.

Both moves are defensible in isolation. Hassabis has been public about wanting to work on AGI strategy and biology rather than run a 2,000-person org. Kavukcuoglu is an internal promotion, not a hire, and Gemini's product surface is now Google's search surface. Consolidation has a real argument behind it.

But the timing is not a coincidence, and neither is the direction of the departures.

The second exit wave in seven weeks

In June, DeepMind lost Noam Shazeer to OpenAI and John Jumper to Anthropic inside a single week — the Transformer's co-author and the lab's Nobel laureate, gone to the two competitors that most want to beat it. We covered that at the time as a retention problem.

The August wave is a different problem. Shazeer and Jumper went to rivals. Dean, Ghemawat, Le, and Vinyals did not go to a rival — they went to a company that does not exist yet, in a category Google was already funding internally, and Google wrote a check.

That is not a talent loss. That is a spin-out Alphabet decided it preferred to own a slice of rather than lose entirely or fund at full cost inside.

There is a coherent corporate logic here. Automated science is a bet with a long payback and an unclear product; it competes badly for resources inside an organization whose quarterly narrative is Gemini's ad and cloud monetization. Alphabet's capex is running north of $200 billion, and every dollar of it has to defend a revenue line. Putting the moonshot outside the building, on someone else's balance sheet, with an equity stake attached, is the same move Google has made with quantum spin-offs and Isomorphic — externalize the option, keep the upside.

The cost is that the people who know how to build the compute substrate now build it somewhere else, for everyone.

What this actually signals

Three things.

First, the research-lab-as-independent-entity model is closing at Google. No CEO title, an SVP reporting to Pichai, a chairman focused on strategy. DeepMind is being absorbed into the product org, and the people who preferred the research posture are leaving to reconstruct it elsewhere.

Second, "AI for science" has graduated from a lab initiative to a fundable category. Radical and Khosla co-leading, with Kleiner, Lightspeed, and Doerr following, is not seed-stage curiosity — it is a consensus that automated experimentation is the next surface after coding agents. Isomorphic proved the drug-discovery slice. Discovery Loop is arguing the general case.

Third, the frontier labs have started shedding founders the way Google shed its own in 2015. The compensation and compute needed to keep four principal engineers building one thing inside a public company now exceeds what a board will approve — but not what a venture round will.

Google spent $2.7 billion in 2024 to bring Noam Shazeer back. Two years later it is cheaper to fund the exit than to prevent it. That is a repricing of what a research organization is worth relative to what it costs, and it happened in one week.