Talent is the real moat in AI — and Google is losing it

Google invented some of the math behind modern AI. The 2017 paper "Attention Is All You Need," written by Google researchers, is the foundation almost every major AI system today is built on. Yet Google's own flagship model, Gemini 3.5 Pro, has now missed three launch deadlines in a row — first June 2026, then July 17, and now with no new date announced. The company that invented the technology is struggling to ship the product.

The technical problems have piled up. Google scrapped the first version of Gemini 3.5 Pro and rebuilt it from scratch after it kept failing tests involving code generation and complex multi-step instructions. The rebuilt version cleared those problems but hit new ones: wrong answers too often, and outputs that changed unpredictably. Meanwhile, OpenAI's GPT-5.6 and Anthropic's Fable 5 are already in customers' hands. Every week Gemini misses its target, rivals pull further ahead.

While the model slipped, the team started leaving. In one week in late June, four senior Gemini researchers announced their departures. Noam Shazeer — a Gemini co-lead and co-author of the original Transformer paper — joined OpenAI. John Jumper — whose AlphaFold research earned a Nobel Prize in Chemistry — joined Anthropic, along with two other senior team members. The day Shazeer's departure became public, Alphabet's share price fell roughly 5%, wiping $225 billion from the company's value in one day.

The lesson is not just about Google. It is about what actually determines who wins in AI. More computing power, more data, a bigger brand name — none of that stops talented people from choosing a different employer. The researchers who understand why a model works, and how to fix it when it does not, are very hard to replace. If you are building AI products or choosing which AI company to partner with, the team behind the model matters more than the name on the door.