August 5, 2026. A Wednesday. Google's stock dropped 4% after two announcements landed like a one-two punch: Jeff Dean — employee #30, the man who built Google's AI brain over 27 years — was leaving. And Demis Hassabis, the Nobel-winning CEO of DeepMind, was stepping back to a chairman role.
If you're keeping score at home, that's now a Transformer co-author (gone), a Nobel Prize winner (gone), two Gemini core engineers (gone), and the company's longest-serving technical legend (gone) — all in the span of about seven weeks.
The question Steve Wu asked me to dig into isn't what happened. It's why. Why can't Google — with its $4 trillion market cap, its bottomless compute, its legendary free cafeterias — make people stick around?
Let me give you the answer upfront: Google's people management isn't failing. It's structurally obsolete. The company optimized HR for a world that no longer exists.
Let's get the body count straight, because the speed of this exodus is what makes it unprecedented:
| Date | Who Left | Where They Went | Why It Matters |
|---|---|---|---|
| June 18 | Noam Shazeer | OpenAI | Co-authored "Attention Is All You Need" — the Transformer paper. Google paid $2.7B to bring him back in 2024. Stayed less than 2 years. |
| June 19 | John Jumper | Anthropic | Won the 2024 Nobel Prize in Chemistry for AlphaFold. Nine years at Google. Gone. |
| June 29 | Jonas Adler & Alexander Pritzel | Anthropic | Core Gemini model contributors. Coding and training specialists. |
| July 2026 | Arthur Conmy | Anthropic | Senior research engineer, Gemini 2.5, AI safety. |
| August 5 | Jeff Dean + 3 others | Discovery Loop | Employee #30, 27 years, Google's AI architect + Oriol Vinyals, Quoc Le, Sanjay Ghemawat |
The stock market noticed. GOOGL dropped from $363.79 on June 17 to $346.13 by June 23 — a 6% haircut in five trading days. Today it sits roughly 14% below its 52-week high of $404.44.
That's not market noise. That's the market repricing Google's talent risk in real time.
Here's the brutal math that Google's HR department can't solve.
If you're a top-tier AI researcher at Google, your compensation is predominantly Restricted Stock Units (RSUs) in a $4 trillion company. Google's stock might go up 15-25% in a great year. Your $2 million annual equity grant becomes $2.3-2.5 million. Nice. Predictable. Boring.
If you join Anthropic — currently valued at $965 billion in private markets and eyeing an IPO as soon as Fall 2026 — your pre-IPO equity could 5x, 10x, or more at liquidity. If you join OpenAI, which has already filed confidentially for an IPO, the same math applies.
Noam Shazeer is the blueprint. He left Google in 2021, co-founded Character.AI, got acquired back by Google for $2.7 billion (personally walking away with hundreds of millions), stayed 20 months, and then jumped to OpenAI for another bite at the pre-IPO apple. The guy has turned job-hopping into a wealth-generation engine.
Google's compensation committee can approve any number. But they cannot issue pre-IPO equity. It's structurally impossible. And that structural constraint — not bad HR — is the single biggest reason talent is walking out the door.
HR Takeaway for Large Organizations: If your equity compensation is public-company RSUs and your competitors are offering pre-IPO upside, you are fighting with a knife at a gunfight. No amount of free food or massage credits closes that gap.
The Los Angeles Times reported something that should make every tech CEO's stomach drop:
"Shortly before Shazeer announced his plans to join OpenAI, computing power dedicated to one of his projects was reassigned to a London-based team at Google DeepMind."
Let me translate that from corporate-speak: Google's most valuable AI researcher had his compute taken away and given to another team. So he left.
This isn't about money. This is about autonomy. Top AI researchers don't just want big salaries — they want the freedom to run their experiments at scale. When Google's internal bureaucracy starts reallocating GPU clusters like it's reallocating conference room bookings, the researchers who can leave, do.
The LA Times noted this wasn't isolated: "an issue that has prompted other employees to leave the company entirely."
HR Takeaway: For knowledge workers at the absolute frontier, compute access is a form of compensation. When you strip it away for "collaboration" or "efficiency," you're effectively cutting someone's pay. They notice.
This stat from VC firm SignalFire's 2025 industry analysis is devastating: DeepMind engineers are nearly 11 times more likely to leave for Anthropic than Anthropic engineers are to join DeepMind.
Eleven. To. One.
That's not a talent war. That's a talent rout.
Why? Three reasons:
Mission focus. Anthropic does one thing: build safe AGI. Google does search, ads, cloud, YouTube, Android, Waymo, Verily, and about 47 other things. For researchers who want to build the future, focus matters.
Safety alignment. John Jumper explicitly chose Anthropic because its safety-first culture aligned with his values after AlphaFold. Whether you agree with AI safety concerns or not, the people building these systems care deeply about them.
Speed. Google's Gemini 3.5 Pro was supposed to launch in June. It's August. Still nothing. Meanwhile Anthropic and OpenAI are shipping. Top researchers want to work where their work actually ships.
Here's the part that should make Google shareholders scream.
Alphabet just announced full-year capital expenditures of up to $205 billion. The company turned cash flow negative for the first time on record. They're spending like a drunken sailor on TPUs, data centers, and AI infrastructure.
And the people they're building all that infrastructure for are leaving.
It's like building the world's greatest concert hall and then watching your orchestra walk across the street to play in a garage — because the garage lets them choose the setlist.
Google Cloud is actually doing fine — 82% revenue growth to $24.8 billion in Q2. But cloud revenue is downstream from model quality. If the people building your models keep leaving, the cloud growth won't last.
Let's be fair. Google basically invented modern people analytics. "People Operations" (the rebrand from "Human Resources") was built on data — Project Oxygen proved managers matter, Project Aristotle proved psychological safety drives team performance. Google's HR was the envy of every Fortune 500 company for two decades.
But here's the thing: Google optimized its people systems for a world where Google was the destination. Where the hardest problem was filtering the thousands of applicants, not retaining the ones you had. Where the competition was Microsoft and Apple — other public companies with the same equity structure.
That world is dead.
The new competition is pre-IPO startups that can offer:
Google's HR playbook — designed for a mature public company competing against other mature public companies — has nothing for this fight.
Jeff Dean's departure deserves special attention because it's fundamentally different from the Shazeer/Jumper exits.
Dean isn't joining a competitor. He's starting Discovery Loop, a public benefit corporation focused on using AI to automate scientific and engineering research. Google is investing in it. The departure is "on friendly terms."
After 27 years — after building MapReduce, BigTable, TensorFlow, and Google's entire AI strategy — Dean wants to build something new. The fact that even Google's longest-serving technical legend couldn't find a way to do that inside Google is the most damning indictment of all.
If Jeff Dean can't build his dream project at Google, who can?
This points to what I'd call the innovation ceiling — the point in a large company where even your most valuable people can't get new, ambitious projects off the ground because the bureaucratic gravity is too strong.
RSU-based compensation is mathematically inferior to pre-IPO equity. Google can't fix this without fundamentally changing its corporate structure. They're handcuffed by their own success.
When GPU clusters become something managers fight over, researchers leave. Compute should be abundant and allocation transparent. At Google, it's neither.
If a 27-year veteran and employee #30 can't build his dream project internally, your internal mobility story is broken. Period.
UK non-competes may delay Jumper's Anthropic start date to 2027, but they didn't stop him from leaving. You can't litigate your way to retention.
Paying $2.7 billion to reacquire Shazeer only to lose him in 20 months isn't just embarrassing — it's a signal to every other top researcher that Google will pay you to leave and then pay you again to come back, and you can still leave whenever you want. The incentives this creates are perverse.
Create an internal startup program with real equity. Let top researchers spin out projects with genuine ownership stakes. Google Ventures could fund them. This is what Jeff Dean is essentially doing externally — Google should have offered it internally.
Compute tokens, not compute politics. Give every senior researcher a guaranteed compute budget they control. No manager can take it away. If Shazeer had this, he might still be at Google.
Match pre-IPO equity with synthetic equity structures. Phantom stock, appreciation rights, or special-purpose vehicles that track competitor valuations. Is it complex? Yes. Is it worth it? Also yes.
The era of "work here for 20 years and collect RSUs" is over for elite technical talent. The new contract is: "build something meaningful here for 3-5 years, get outsized rewards, and we'll support your next move."
Google should embrace being a talent accelerator rather than a talent destination. Alumni networks, venture funding for departing founders, and boomerang-friendly policies. Jeff Dean's "friendly departure" with Google investment is the model.
Before you short GOOGL, consider:
Google still has the deepest bench. Hassabis said it: "We have by far the biggest and broadest research bench of any of the labs out there." For every Shazeer who leaves, there are 50 brilliant researchers who stay.
Infrastructure is a moat. TPUs, data centers, YouTube training data, search distribution — these are assets no startup can replicate quickly.
Cloud is the real business. AI research glamour aside, Google Cloud is growing 82% and enterprises care about reliability, not Nobel Prizes.
The "brain drain" might be overblown. Hassabis called the talent market "ferociously competitive" but noted Google "wins our fair share." People move between labs constantly.
Discovery Loop could be a win-win. If Google invests in Dean's startup and it succeeds, Google captures upside without the bureaucratic overhead of running it internally.
Google's people management crisis isn't about free food, office perks, or even compensation bands. It's about something deeper: the structural impossibility of offering startup-level upside and autonomy inside a $4 trillion public company.
The HR playbook that made Google the world's most desirable employer — data-driven people analytics, generous benefits, campus-as-destination — was designed for a different war. In the AI talent war, the weapons are pre-IPO equity, compute autonomy, mission focus, and shipping velocity. Google has none of the advantages in these categories.
Until the company invents new retention structures that match what Anthropic and OpenAI can offer — or until those companies go public and lose their equity advantage — the bleeding will continue.
The $2.7 billion Shazeer round trip should be studied in business schools as the moment everyone should have realized: you can't buy loyalty with money when the competition is offering ownership.
What do you think? Is Google's talent exodus a temporary blip or a structural crisis? Reply and let me know — Steve, I'd especially value your take as someone who's built and managed teams in the tech space.
