Noam Shazeer went to OpenAI. John Jumper joined Anthropic. Both left Google within a seven-day span. Yet neither needed the money; both were among the highest-paid researchers in Mountain View and decided it was time to go. This isn't just a story about compensation or equity. Rather, it is about product decisions, strategic vision, and how Google is losing the most important battle in its history—not the one about building better AI, but the one about deciding what to do with it.
Photo: Adi Goldstein on Unsplash
The timing of these departures is no coincidence. Shazeer co-led Gemini, the project that was supposed to eclipse GPT-4. Jumper, on the other hand, won the Nobel Prize in Chemistry for AlphaFold, the system that solved protein folding. Both joined Google in the last decade believing that this was where the significant work would happen. Ironically, they left because Google systematically chose not to launch what they built, or did so late or without conviction. This is the story of how the brightest minds in AI lose faith in your company, not due to technical shortcomings, but because of strategic inaction.
The Paradox of Shazeer: Building the Best Model and Not Being Able to Launch It
Noam Shazeer is no ordinary researcher. He co-invented the Transformer mechanism in 2017, the architecture that enables ChatGPT, Claude, Gemini, and practically every modern LLM. When Google acquired Character.AI (the startup Shazeer founded after leaving Google the first time) for $2.7 billion in August 2023, many thought it was a triumphant return. Shazeer was coming back home. He was set to co-lead Gemini, the project that would put Google back on the map.
However, the reality turned out to be different. Gemini launched in December 2023, six months after GPT-4. When Gemini 1.5 Pro finally arrived in February 2024, it was technically superior in many benchmarks, even with a context of 1 million tokens, better multimodal reasoning, and lower latencies. But the timing had already slipped. OpenAI had captured the market's imagination, and developers were already building on GPT-4. Not to mention, enterprise deals were already signed.
What truly frustrated Shazeer was not losing the benchmark race. It was Google's inability to decide what product to build around the technology. Gemini underwent three rebrands in 18 months. First, it was Bard with Gemini, then Gemini standalone, and finally Gemini Advanced. Each pivot confused the market further. Meanwhile, ChatGPT maintained a simple message: "the most capable conversational AI in the world." One product. One narrative. A clear decision.
Shazeer spent two years watching product decisions undermine world-class technology. In March 2026, when OpenAI launched GPT-5 and announced that its revenue run rate had crossed $10 billion annually, Shazeer made his decision. He left for OpenAI not because they had better researchers, but because they knew what to build and how to launch it.
Jumper and AlphaFold: The Nobel That Google Let Slip Away
Photo: Steve A Johnson on Unsplash
John Jumper is different. He’s not in the world of LLMs. He focuses on proteins and computational biology, an area where AI can save lives rather than just generate memes. AlphaFold solved one of the great challenges of biology: predicting the 3D structure of proteins from their amino acid sequences. This feat earned him the Nobel Prize in Chemistry in 2024. However, it gave DeepMind (and, by extension, Google) an unprecedented scientific victory, but the problem was that Google didn’t know what to do with that victory.
AlphaFold launched as an open-source project, which is incredible for science but terrible for business. While Jumper and his team published the code and model weights freely, startups like Insilico Medicine, Recursion Pharmaceuticals, and Exscientia raised billions using exactly that technology for drug discovery. Google gained prestige while startups captured value.
Jumper repeatedly proposed creating a commercial division for AlphaFold within DeepMind, aiming to offer drug discovery-as-a-service and form partnerships with pharmaceutical companies. In his proposal, he suggested a freemium model where academic institutions would have free access, but companies would pay for computing capacity, fine-tuned models, and enterprise support. Google's response was consistent: "it's not aligned with our core strategy."
What is that core strategy? Ads, Cloud, and Search. Even in 2026, with over $80 billion in annual revenue from Google Cloud, the mentality remains the same: AI as a feature, not as a product. AlphaFold could have been the foundation for a complete division of "Google Life Sciences AI," but it ended up as a spectacular paper and a GitHub repository that others are monetizing.
When Anthropic launched "Claude BioResearch" in April 2026, an LLM specifically trained for biological analysis and drug discovery, with partnerships already signed with Roche and Novartis, Jumper saw the future that Google rejected. Two weeks later, he was in San Francisco closing his contract with Anthropic.
The Systematic Pattern: Top Technology, Second-Rate Execution
What ties Shazeer and Jumper together isn't bad luck or unfortunate timing. It's a systematic pattern that has repeated at Google since 2018:
1. They build superior technology. Google Brain invents Transformers. DeepMind solves AlphaGo, AlphaFold, AlphaCode. Google develops TPUs, the most advanced ML infrastructure on the planet.
2. They hesitate to launch. Gemini gets delayed six months. Bard arrives late and confused. AlphaFold is given away. Each launch comes with disclaimers, endless beta versions, and corporate messaging that no one understands.
3. Competitors execute. OpenAI launches quickly, iterates in public, and captures mindshare. Anthropic builds enterprise relationships from day one. Both have product clarity that Google struggles to achieve.
4. Researchers become frustrated. Watching your technically superior work lose in the market because your company doesn't know how to sell it is demoralizing. The best leave.
This pattern isn't new. It happened with Google+, Google Glass, Inbox, and Stadia. But in AI, the consequences are different. We're not talking about failed products; we're talking about the next key technological platform. And Google is letting it slip through its fingers.
DeepMind's headcount fell by 12% between January 2025 and March 2026. Not because people were laid off, but because the best left. To OpenAI, to Anthropic, and to startups that no one knows yet, but will raise Series B rounds of $200 million in 18 months.
Why OpenAI and Anthropic Attract Google's Talent
The easy narrative is to think it’s about money. "OpenAI pays more." "Anthropic offers better equity." But honestly, it’s more complex than that. Shazeer and Jumper didn’t need more money. Both had made a fortune at Google and could retire tomorrow if they wanted.
What OpenAI and Anthropic offer is mission clarity and decision-making speed.
At OpenAI, if you build a better model, you launch it. There’s a safety review process, sure, and red-teaming. But if you pass those filters, the model goes out. There aren’t three layers of approval from product managers who want to "align it with Search strategy." There’s no legal blocking launches because it "could cannibalize existing revenue."
At Anthropic, if you identify a market opportunity (like Claude BioResearch), you have the autonomy to pursue it. You don’t need to convince a cloud division that your project "fits into its roadmap." You just build, launch, and then iterate.
This difference in decision-making culture is lethal. The best researchers in AI don’t just want to do science; they want to see their science impact the world. They want the models they build to reach millions of users. They want to close the loop between research and product.
Google, with all its infrastructure and resources, has lost that capability. It has the best researchers, the best data centers, and the best datasets. But it lacks the launch culture that turns research into impact.
The Real Cost of the Exodus: It’s Not Talent, It’s Momentum
Losing Shazeer and Jumper hurts. But the real cost isn’t individual talent. Google could hire 50 excellent PhDs tomorrow, but the real cost is momentum.
When the co-leaders of your most important projects leave, you send a message to the rest of the organization: "this isn’t going anywhere." Those who remain start updating their LinkedIn profiles. Those who were considering joining look away. And newcomers question their decisions.
Momentum in AI is measured in innovation cycles. OpenAI is in its GPT-5 cycle, planning GPT-6. Anthropic launched Claude 4 and is already testing Claude 5 internally. Google launched Gemini 2.0 in February 2026 and... what’s next? No one knows. Not even internally.
This exodus also affects future recruiting. The best ML PhDs from Stanford, MIT, and Berkeley will watch who is innovating faster. They’ll look at who is launching products that matter. And increasingly, that answer is not Google.
The irony is brutal: Google invented Transformers. Google put AI on the map with AlphaGo. Google has more resources than OpenAI and Anthropic combined. And, unfortunately, it is losing the race that it started.
The Question Google Must Answer (Before Losing More Leaders)
It's not just about paying more. It's not about better cafes or workspaces. It’s about answering a key question: Does Google want to win in AI or just participate?
If the answer is "win," they need profound structural changes:
-
Product autonomy for DeepMind and Google Brain. Let them decide what to build and when to launch, without approvals from Search or Ads.
-
Clarity in monetization. AlphaFold cannot be a prestige project. Gemini cannot be "Bard with another name." Each project needs a clear business model from day one.
-
Speed of launch. If a model is ready in January, it launches in January. Not in July "after aligning it with the Cloud roadmap."
-
Real equity in specific projects. Not Google stock options (which are already sky-high), but equity in the AI divisions that researchers build.
If the answer is "participate," then Google should be honest about it. It could become the best infrastructure provider for others building AI. Sell TPUs, GCP, and ML tools. Let OpenAI and Anthropic win the application layer. This isn’t the worst strategy, but it requires honesty about what battle they are fighting.
The problem is that Google seems to be caught in the middle. Investing billions in research, but lacking the culture to turn that research into winning products. Hiring the best, but losing them because it can't execute.
Shazeer and Jumper didn’t leave because Google was a bad place to research. They left because Google is a bad place to launch. And in 2026, in AI, research without launch is irrelevant.
How many more leaders does Google need to lose before it changes how it decides what to build?