Google’s AI lab is losing the researchers who built its foundations, and the departures are landing at the same time internal morale questions are surfacing.
Gemini co-lead Noam Shazeer, a co-author of the 2017 Transformer paper that underpins nearly every large language model in use today, announced he is leaving Google for OpenAI (Axios). Days earlier, Nobel laureate John Jumper, the AlphaFold researcher who spent almost nine years at Google DeepMind, announced he was joining Anthropic (Axios). Both moves add to a run of senior DeepMind departures that US outlets are now describing as a talent exodus.
If you run a business that depends on any of these three labs, the story isn’t the gossip. It’s what happens to release cadence, pricing, and model behavior when the people who ship the roadmap change employers mid-cycle.
What happened
According to Axios reporting based on conversations with current and former Google DeepMind employees:
- “Several top researchers have departed for competing AI labs,” including Shazeer to OpenAI and Jumper to Anthropic (Axios).
- Poor internal morale is contributing to delayed model releases at DeepMind, with sources linking some of the frustration to Google’s Pentagon technology deal and to a sense of falling behind competitors on agentic coding (Axios).
- One DeepMind employee told Axios, “We’re behind,” referring to Google’s model capabilities relative to rivals (Axios).
- Google disputes the morale narrative, saying its first-half AI attrition is lower than a year ago and that more than 90% of people offered an AI role accept it (Axios).
Shazeer’s exit is notable because Google spent roughly $2.7 billion less than two years ago to bring him back to the company through a licensing deal tied to Character.AI, the startup he co-founded after first leaving Google in 2021.
Why it matters
Model labs are not interchangeable vendors. The people who lead pretraining, alignment, and coding-agent work shape what a model is good at, how fast it improves, and how conservative its safety tuning is. When several of those people move to a competitor within weeks of each other, that’s a leading indicator for:
- Release timing. Roadmaps built around specific research leads can slip when those leads leave, which is part of why Google’s next flagship Gemini model has faced delays.
- Competitive positioning between labs. Anthropic gaining a scientist known for applying AI to biology research signals a push beyond chat and coding into scientific applications. OpenAI gaining an architecture-level researcher reinforces its core model research bench.
- Retention economics across the industry. If a $2.7 billion re-signing didn’t keep a top researcher in place for two years, it tells you compensation alone isn’t what’s driving these moves — mission focus and equity upside at pre-IPO labs are also pulling talent.
None of this means Gemini becomes unusable or that Google stops shipping. It means the pace and character of releases from all three labs are more exposed to key-person risk than most operators assume when they pick a primary model vendor.
What operators should do
You don’t need to pick a side in a lab rivalry. You need a vendor strategy that survives one.
- Don’t single-thread your stack on one lab’s roadmap. If your product depends on a specific model’s behavior (tone, tool-calling style, context window), test your workflows against at least one alternative every quarter so a competitor’s model can absorb the load if your primary vendor’s release slips.
- Track model changelogs, not just headlines. Talent departures are a signal, not a schedule. Watch actual release notes and benchmark results from the labs you depend on rather than reacting to every departure story.
- Re-evaluate pricing and rate limits after a wave of departures. Labs sometimes adjust pricing or free-tier limits when competitive pressure increases — that can work in your favor if you’re paying for API access.
- Keep a short list of two models you trust for your core use case. If you’re choosing between ChatGPT and Claude for business workflows, run the same prompts through both this month and note where each one is stronger — you want that answer before a migration is forced on you, not during one.
If you want a structured way to compare the two leading assistants for day-to-day business use, this comparison walks through cost and capability trade-offs in detail.
Related tool: Compare the two most widely deployed assistants side by side before you standardize your stack — see how Claude and ChatGPT differ on pricing, context window, and agentic task support.
Related links
- Learn: /learn/chatgpt-vs-claude-for-business-2026
- Learn: /learn/how-to-build-your-ai-stack-2026
- Tool: /tools/claude
Primary source: Axios (Axios).
