In one week, Google lost the inventor of the Transformer to OpenAI and the creator of AlphaFold to Anthropic. But the exodus didn’t start this week. It started in 2017 — the same year the Transformer paper was published.
The Scoreboard — June 2026
7 of 8
Transformer paper authors who left Google
$2.7B
Google paid to get Shazeer back. He left again in <2 years.
1
Nobel laureate lost to Anthropic this week
6
Companies founded by ex-Google Transformer authors
The Week That Made It Undeniable
On June 18, Noam Shazeer — co-inventor of the Transformer architecture and co-lead of Google Gemini — announced he was leaving Google for OpenAI. One day later, John Jumper — Nobel laureate and creator of AlphaFold — announced he was leaving Google DeepMind for Anthropic.
Two departures in 48 hours. One to the company Google is trying to beat. One to the company Google is trying to contain. Both scientists who defined what Google was supposed to be.
The Shazeer departure carries a special sting. Google paid approximately $2.7 billion in 2024 to bring him back from Character.AI — the startup he founded after leaving Google in 2021. He was appointed co-lead of Gemini. Less than two years later, he walked out the door again — this time to the rival Google most fears.
$2.7 billion. That is the price Google paid for less than two years of Noam Shazeer’s time. It may be the most expensive failed retention in corporate history.
The Transformer Eight — Where They All Went
In 2017, eight Google researchers published “Attention Is All You Need” — the paper that introduced the Transformer architecture and made modern AI possible. Every large language model — GPT, Claude, Gemini, Llama — is built on their work. Nine years later, almost none of them are still at Google.
The Transformer Eight — Status June 2026
Noam Shazeer
Google → Character.AI (founded) → Google ($2.7B return) → OpenAI (June 2026)
Ashish Vaswani
Google → Adept AI (co-founded) → Essential AI (founded, $1B valuation)
Llion Jones
Google → Sakana AI (founded, Tokyo)
Aidan Gomez
Google → Cohere (co-founded, enterprise AI)
Jakob Uszkoreit
Google → Inceptive (founded, RNA design)
Lukasz Kaiser
Google → OpenAI (researcher)
Illia Polosukhin
Google → NEAR Protocol (co-founded, blockchain)
Niki Parmar
Google → Adept AI (co-founded with Vaswani) → Status unclear
Seven of the eight authors have left. Six founded or joined competing companies. Two — Shazeer and Kaiser — went directly to OpenAI. The paper that enabled $10 trillion in market value was written entirely at Google. Almost none of the people who wrote it stayed.
Beyond the Paper — The Broader Exodus
The Transformer Eight are the most visible departures, but they are not the only ones. Google has been hemorrhaging senior AI talent for years:
John Jumper (AlphaFold, Nobel Prize) → Anthropic — June 2026
Dario Amodei (Google Brain → OpenAI VP Research) → Founded Anthropic
11 executives left Google for Microsoft in 2025 alone
Apple’s Siri lead left for Google DeepMind in Jan 2026 — Google can still attract, but can’t retain its own
The pattern is consistent: Google hires brilliant researchers, gives them resources to do foundational work, publishes the results — and then watches as those researchers leave to commercialize the insights elsewhere. Google is not losing a talent war. Google is funding one.
Why They Leave
The departures cluster around three structural forces:
1. THE PRODUCT GAP
Google publishes research. OpenAI and Anthropic ship products. Shazeer wanted to build consumer-facing AI — Google’s ad-revenue model made that structurally difficult. He left twice over the same friction.
2. THE MISSION GAP
Jumper moved to a safety-focused lab. Dario Amodei founded one. Multiple DeepMind researchers have cited the post-merger shift away from fundamental research toward applied AI as their reason for leaving. When the mission changes, the missionaries leave.
3. THE FOUNDER GAP
Six of the Transformer Eight founded companies. They didn’t leave for better salaries — they left for ownership. The researchers who invented the most valuable architecture in computing history were employees at Google. They watched others build empires on their work. Then they decided to build their own.
The Structural Read
Google’s problem is not compensation. Google can match any offer. The problem is structural: Google is organized to protect an advertising business, not to ship AI products. Every breakthrough Google produces must pass through a filter: does this help ads, or does it threaten them?
That filter is invisible to users but perfectly visible to researchers. It is why Shazeer left the first time — Google wouldn’t let him ship a chatbot that might cannibalize Search. It is why he left the second time — even after $2.7 billion and the co-lead of Gemini, the structural constraint didn’t change.
The Core Problem
Google doesn’t have a talent problem.
Google has a product-culture problem
that manifests as a talent problem.
Anthropic ships Claude. OpenAI ships ChatGPT. Cohere ships enterprise AI. Essential AI ships agents. All founded or staffed by ex-Google researchers. Google published the Transformer paper — and then spent nine years watching everyone else build on it faster.
Harness Theory
Google Had the Harness. It Couldn’t Keep It Tightened.
A harness is not just capability — it is capability wrapped in a system that directs it toward outcomes. Google had the talent, the compute, the data, and the research. It lacked the system to direct them toward products. The researchers who built the harness at Google left to build their own — at companies where the harness connects directly to users, not to an ad server. The lesson: capability without product direction is a research grant, not a moat.
What It Means
Google is not dying. It still has Gemini, the largest compute infrastructure in the world, and billions in AI revenue. But the talent exodus reveals something the stock price doesn’t: Google has become the training ground for the AI industry, not its destination.
The researchers come for the resources, do their best work, publish it — and then leave to build companies that compete with Google using the ideas Google paid them to develop. The Transformer was invented at Google. AlphaFold was invented at Google. The founders of Anthropic, Cohere, Essential AI, Sakana AI, and Inceptive all trained at Google. And now OpenAI’s newest hire is the man Google spent $2.7 billion to bring back.
Google built the foundation of modern AI. Then it lost everyone who made it. That is not a talent-market story. That is a strategy story — and it is the most expensive one in tech history.
Business Engineer Deep Dive
The AI Supercycle — Where Talent Flows, Value Follows
The Transformer Eight created six companies across five layers of the AI stack. The Map of AI shows why talent migration is the leading indicator of where value concentrates — and which layers are about to get disrupted.
Sources: Attention Is All You Need (2017), Calcalist, CB Insights, Axios — June 19, 2026