Google DeepMind Lost 8 Researchers in One Summer. The Budget Was Never the Problem.

8 named researchers left Google DeepMind and its parent company this summer, and the model they were building missed 3 deadlines in a row. Demis Hassabis stepped down as CEO of Google DeepMind and moved to Chairman, telling staff he needed "time and space to focus on the big picture." Koray Kavukcuoglu, previously the CTO, became Senior Vice President reporting directly to Sundar Pichai, a departure from how the lab had been structured. Gemini 3.5 Pro missed an original June target, then a revised mid-July target, and entered August 2026 without shipping.

Behind that record sits more computing capacity and more capital than almost any research organization on earth. Alphabet purchases chips at a scale its competitors cannot approach, and it can compensate nearly anyone almost anything. Neither advantage delivered a model in June, and neither retained the people who understood which model deserved to ship.

The reshuffle is the visible half of the story

Noam Shazeer, a co-lead on Gemini, went to OpenAI, while John Jumper, who co-invented AlphaFold and holds a Nobel Prize for that work, went to Anthropic, as did Jonas Adler and Alexander Pritzel. On August 5, Jeff Dean left Google after 27 years as its chief scientist, having helped build both the search infrastructure the company still operates on and Gemini's multimodal models. He departed alongside Sanjay Ghemawat, a senior fellow and one of Google's top engineers, Quoc Le, a founding member of Google Brain, and Oriol Vinyals, a senior research scientist at DeepMind.

Those 4 founded Discovery Loop, a public benefit corporation designed to run thousands of scientific experiments simultaneously, with the stated goal of recursive self-improvement: using AI to discover better AI. Radical Ventures and Khosla Ventures co-led the funding round, with Kleiner Perkins, Lightspeed, and Doerr Capital also participating. Alphabet supported it too, so Google's own parent company is now helping finance a venture founded by the researchers who just left it.

Everything technical stayed behind. The weights, the checkpoints, the data centers, the patents, and the published papers all remain Google property, and nothing on the balance sheet walked out the door.

What left Google DeepMind was judgment

What left was accumulated taste for which experiments deserve resources, which results deserve confidence, and who answers for the consequences when a model turns out to be wrong. Developing that inside one person takes decades, and developing it across a team that argues productively takes longer. It appears on no asset register, and no acquisition restores it on a predictable schedule.

There is a name for that layer of the work, and it is the governing idea behind everything I write about innovation culture and intrapreneurship: Above the Algorithm. Judgment, taste, trust, and accountability are the responsibilities that still have to belong to a person. A company can rent intelligence by the hour now, and it cannot rent the colleague who recognizes when that intelligence is confidently wrong.

Organizational strain of this kind surfaces in the same place every time, in the calendars of the people carrying it. One DeepMind engineer, speaking anonymously to Fortune, described the condition plainly:

I've been pushing 60-hour weeks for quite a long time because there is so much to do.

Hear that as a capacity statement rather than a complaint, because somebody working 60-hour weeks has time to execute and no remaining time for the harder question of whether the work is even correct. Judgment is the first thing a schedule like that spends, and the last thing a budget can repurchase.

Roles transform, and the human part is always the judgment

James Bessen's research on ATMs remains the cleanest version of this pattern: tellers per branch declined after the machines arrived, branches multiplied, and the occupation shifted toward relationship work. VisiCalc did something comparable to bookkeeping in 1979, eliminating the arithmetic and expanding the analyst profession around whatever remained. The machine took the procedure. The person kept the judgment.

The half-life of skills is short in every industry now, and shortest inside the industry manufacturing the thing that shortens it. Model capability converges, because serious laboratories arrive at comparable hardware and comparable architectures within a release cycle or two. Judgment never converges. It concentrates in particular people, and it relocates when they do.

Cutting your people and losing them arrive at the same destination

Oracle reduced its workforce to finance AI capital spending, a decision I worked through in a separate piece. Google DeepMind made no such decision and arrived somewhere adjacent anyway: the people whose judgment was the actual advantage now work at OpenAI, at Anthropic, and at a startup Alphabet is helping to finance. One organization spent the asset deliberately. The other strained it until it walked.

The enterprise version of this is the pattern Federal Reserve data keeps pointing toward, which I examined here: the returns from AI follow the people who redesign the work around it. Compute is a purchase order. Redesign is human work, performed by people who earned the standing to be trusted with it.

I built Cisco's Innovate Everywhere Challenge with a small team, a company-wide intrapreneurship program that put real budget and executive attention behind employee ideas. The lesson that stayed with me had little to do with the winning ideas. Experienced people remain where their judgment gets used, they leave when it gets routed around, and almost none of them name that as the reason on the way out.

Your company has a smaller version of this situation and less money to absorb the mistake. Somebody there understands which AI output to distrust, which customer promise the system cannot honor, and which number in the quarterly deck is optimistic. That person rarely appears on anyone's retention list, and is being asked to carry more this quarter than last.

Google DeepMind will be fine in the way large institutions are fine. It has the capital to rebuild, a London operation Google says it remains committed to, and a bench deeper than almost anyone's. The invoice arrives later, at the next transition, when the people who could have said "that result is wrong" are saying it somewhere else. So here is the question I am carrying into my own week, and it travels comfortably into yours: who here holds the judgment our AI plan depends on, and when did anyone last ask them what they think? If answering that takes longer than a minute, better to discover it now than a summer from now. And if you read this differently, I am easy to find.

Sources: Fortune, August 10, 2026; TechCrunch, August 5, 2026.

Here is what makes Alex a credible voice on this topic: he built Cisco's Innovate Everywhere Challenge, the company-wide intrapreneurship program, and shaped a $1.1B innovation portfolio across 20 years of keeping experienced people and their judgment inside a company instead of at a competitor.

If your AI plan depends on 5 people nobody has asked lately, that is worth an hour. book a conversation →