Google Loses Jeff Dean and Three AI Founders as Alphabet Sheds $185B and Funds the Exodus
Two men who almost never gave interviews just cost Alphabet more than the company plans to spend building data centers all year.
That is the number worth sitting with. When Jeff Dean and three colleagues resigned on Wednesday to start Discovery Loop, Alphabet reportedly shed roughly $185 billion of market value in a single session — close to the entire midpoint of its 2026 capital budget. The market treated four departures as a $185 billion event.
The market got the size right and the reason wrong.
The market got the size right and the reason wrong. Shareholders sold the departures. They should have looked at why the departures happened.
Start with the trade Google ran in the other direction. In August 2024 it reportedly paid $2.7 billion for a non-exclusive license to Character.AI, a structure built to bring Noam Shazeer — co-author of the 2017 transformer paper — back into DeepMind. In July 2025 it reportedly paid about $2.4 billion for Windsurf's license and founders. Call it $5.1 billion across 24 months to import researchers by checkbook.
Shazeer, the name the $2.7 billion deal was built around, reportedly resigned 22 months later for OpenAI.
Here is the accounting nobody priced. That $2.7 billion bought an intangible asset Google could book. But you cannot capitalize a human being, and therefore you cannot impair one. When Shazeer walked, there was no write-down, no charge, no line item. The asset was never on the balance sheet to begin with.
So the acquihire is a purchase you can't unwind. You pay up front and pray the person stays. When they don't, the loss is real but invisible.
Now look at what Google signed this week. Alphabet is a founding investor in Discovery Loop, one participant in a seed round co-led by Radical Ventures and Khosla Ventures that hasn't even closed. It is also the startup's cloud partner, reportedly with a commitment to supply compute for at least the next year, plus a collaboration on machine-learning systems research.
That is not a purchase. That is an option.
Alphabet writes a check into a round it does not control, and instead of booking a cost, converts the relationship into Google Cloud revenue. If Discovery Loop's bet on automating the experimental loop pays off, Alphabet owns a piece of the upside and has been selling it compute the whole way. If it fails, the exposure is a seed stake, not a $2.7 billion intangible.
Financially, Alphabet upgraded the trade. It turned an uncontrollable retention problem — the thing that produced no write-down but plenty of pain — into a cheap call option that runs through the fastest-growing part of its business.
Which is why the panic was misdirected. Shareholders sold the departures. They should have looked at why the departures happened.
Because the real crack is not the talent market. It is inside the building.
CNBC reported that some Google researchers have grown frustrated over access to the computing capacity they need for ambitious projects — while watching Google Cloud sell TPUs to outside customers, including Anthropic, whose models compete directly with Gemini. Every chip is a choice among competing priorities, and researchers are watching those chips go to a rival.
That is a capital-allocation problem wearing a talent-exodus costume.
Google is spending more than almost any company on earth on data centers and chips, with a 2026 capex plan of $195 billion to $205 billion. And yet its own scientists say the constraint is access. The scarcity isn't dollars. It's who inside the machine gets the compute.
Layer on the bureaucracy. Moving research into product at Google requires layers of approval — exactly the friction that makes OpenAI, Anthropic, or a four-person startup more appealing to people who prefer lab work to balance sheets.
Oriol Vinyals put the pull side plainly: extreme focus works best when the thing you care about is the only thing the company does.
Gil Luria of D.A. Davidson named the deeper motive. "They're not interested in commercializing AI," he reportedly said of the departing cohort. "They're interested in being part of history."
That line should worry Alphabet more than the stock chart. It says the people who built the company's technical advantage increasingly believe the next breakthrough happens outside it — and no 82% cloud growth number changes that calculus, because commercialization is precisely what they are leaving to escape.
And this cohort has been leaving all year. David Silver went in February. Shazeer and Nobel laureate John Jumper reportedly left within days of each other in June, a departure that was reportedly around the time Alphabet fell 5.1% in a single session. Peter Norvig, reportedly 25 years the head of Google research, is gone. All eight authors of the transformer paper have reportedly now left.
Wednesday was the second people-driven drawdown in six weeks.
So set aside the number that dominated the week — the 82% cloud growth, the 24% revenue, the $8.8 billion cloud operating income. Those tell you the commercial machine is compounding. They tell you nothing about whether the frontier lab still works.
The metric that matters now is Gemini 4.
Look at what the lab just lost in a single day. Dean was the model program's overall co-technical lead. Vinyals co-led the flagship models. Senior leadership of the Gemini effort turned over in twenty-four hours, in the middle of a race the company itself calls ferocious.
The product backdrop is already shaky. Gemini 3.5 Pro was reportedly targeted for June and, per Bloomberg, fell short of internal marks, with low morale cited as a factor by Axios. Gemini 4 is a much larger model still in training, with no announced release date.
And it now ships under Koray Kavukcuoglu, DeepMind's technology chief of 13 years, elevated to senior vice president reporting straight to Pichai. Demis Hassabis moved up to chairman and Alphabet chief scientist, pulling toward AGI strategy and Isomorphic Labs. Google DeepMind no longer has a CEO.
Kavukcuoglu is not an outsider — he'd reportedly already absorbed much of Hassabis's model-development remit over the past year. But he now runs the flagship program having lost its co-technical lead and its longtime CEO on the same morning.
That is the falsifiable test. If Gemini 4 ships on time and competitive, the thesis that these departures were a manageable option trade holds, and the $185 billion tantrum looks like exactly that. If it slips, the compute-rationing and bureaucracy complaints stop being grievances and start being the explanation for why Google can't hold the frontier.
Notice one small tell. Dean's farewell post on X reportedly initially didn't thank Google or mention his 27 years — an omission conspicuous enough that observers flagged it before he added a message later.
People don't leave a place that works for them to write a seed check they helped fund. They leave because the alternative — thousands of parallel experiments, no approval layers, all the compute pointed at one thing — is the version of the job they can no longer get inside the company they built.
Alphabet made the smart accounting move. It converted a retention problem it could not win into an option it can afford to lose.
But the option only pays if the lab it left behind can still ship. Watch Gemini 4. Everything else this week was noise.