Noam Shazeer co-authored “Attention Is All You Need” - the paper that made modern AI possible. Google paid $2.7 billion to bring him back in 2024 through the Character.AI acquisition. He left after 22 months.
The Deadline That Said Too Much
Gemini 3.5 Pro was committed for a June launch at Google I/O on May 19. By late June, the model was still in limited Vertex AI enterprise preview. The stated reasons: token efficiency shortfalls, coding performance below flagship tier requirements, and multi-step reasoning gaps (The Agent Report, 2026).
Technical delays happen. But Claude Sonnet 5 and GPT-5.6 Sol both shipped on time in the same window. When competitors deliver on schedule and you don’t, the question shifts from “what’s wrong with this model?” to “what’s wrong with the development pipeline?”
Gemini 3.5 Pro is now targeting July 17, 2026. That’s its second missed deadline.
Four Departures in Ten Days
Between June 18 and 27, four senior Google DeepMind researchers announced their exits:
- Noam Shazeer - Gemini co-lead, Transformer architecture co-author - to OpenAI
- John Jumper - Nobel Prize 2024, AlphaFold co-creator - to Anthropic
- Jonas Adler - Gemini coding lead, AlphaFold contributor - to Anthropic
- Alexander Pritzel - Pretraining specialist - to Anthropic
Three of four chose Anthropic. These aren’t just domain experts - they carry direct knowledge of Gemini’s internal architecture, and they’re now working at the company building Claude to compete against it. That’s not talent flight - it’s a capability transfer to the competition.
Markets responded: Alphabet shares fell from a peak of $402.38 to $345.29 over six weeks. $225 billion in market cap erased (The Agent Report, 2026).
The $2.7 Billion Question
Shazeer’s departure cuts deeper than the headline. Google didn’t lose a researcher - they lost someone they paid $2.7 billion to retain, less than two years ago. He lasted 22 months.
When the person you paid $3 billion to keep decides to leave in under two years, the question stops being about compensation. It becomes a culture question. And culture doesn’t show up in quarterly earnings until it’s already costing you years of compounding momentum.
John Jumper’s move to Anthropic is equally significant. AlphaFold - arguably DeepMind’s most celebrated achievement of the decade - was his work. The person who built it is no longer at Google.
What This Means for Enterprise AI Buyers
What doesn’t change: Google Cloud infrastructure remains best-in-class. TPU, global network reach, and Vertex AI’s multimodel support - which already lets you run Claude and GPT-4 alongside Gemini - are unaffected by researcher departures.
What changes: Model roadmap reliability. When a flagship model misses its launch window twice, procurement teams planning 12-month AI roadmaps can’t treat Google’s shipping dates as fixed. That uncertainty carries real cost - engineering timelines, integration budgets, go-live commitments.
DeepSeek V4 Flash is outperforming Gemini on coding benchmarks at a fraction of the price. Anthropic just added three researchers who know how Gemini works from the inside. OpenAI now has the man who co-wrote the architecture both companies are based on.
The Asia Angle
For enterprises across Vietnam and Southeast Asia building on Google Cloud, this situation calls for attention - not panic. Vietnam’s AI market reached $1.2 billion by late 2026, growing 76% year-over-year (InCorp Vietnam, 2026). Major players including FPT, Viettel, and VNG run significant workloads on Google Cloud infrastructure.
That doesn’t need to change. Infrastructure stickiness is real, and Vertex AI’s multimodel support means you’re not locked into Gemini specifically - you can access Claude, GPT-4, and Gemini within the same platform.
But if your Q3-Q4 2026 roadmap has hard dependencies on Gemini 3.5 Pro features, build contingency now. Not because Google AI is collapsing - but because a model that has missed two consecutive commitments doesn’t earn unconditional planning trust.
Diversify your model dependencies. Vertex AI already makes that easy.
NateCue's Take
Three of the four departing researchers chose Anthropic - and they're not just bringing expertise, they're bringing internal knowledge of Gemini's architecture directly to the team building Claude against it. That's a capability transfer, not just a talent departure. Google's infrastructure advantages remain class-apart: TPU, global network, Vertex AI's multimodel support aren't touched by this. For enterprise teams in Asia: don't abandon Google Cloud. But if your Q3-Q4 roadmap has hard dependencies on Gemini 3.5 Pro, run Claude Sonnet 5 and GPT-5.6 in parallel now. A model that has missed two consecutive deadlines doesn't earn unconditional trust.