Where the Architects Went

Two of Google's most consequential AI researchers announced this past week that they're leaving. Noam Shazeer, the Transformer co-author and Gemini co-lead, posted to X on June 18 that he's joining OpenAI. Hours later, John Jumper announced his own move to Anthropic after nearly nine years at Google DeepMind, where he co-created AlphaFold and shared the 2024 Nobel Prize in Chemistry with Demis Hassabis. Neither lab has said what its new hire will work on.

Google paid $2.7 billion in 2024 to bring Shazeer back from Character.AI, and he was gone again inside two years. Jumper had a Nobel, a nine-year tenure, and the AlphaFold team he'd led since six months out of his PhD, and still chose to leave. Whatever Google offered to keep them, it wasn't what they wanted to do next.

Shazeer co-designed the Transformer architecture every frontier model on earth runs on today, and was running pretraining for Gemini. Jumper's AlphaFold has predicted more than 200 million protein structures and reset how computational biology is done at the lab bench. Either replacement is a decade-long research question, well above what a recruiting cycle solves.

Google still has Demis Hassabis, the broadest compute footprint of any frontier lab, and the deepest internal model lineup. Two researchers leaving don't break a roadmap, but they signal which roadmaps the field's most senior names think will get them where they want to go fastest, and twice this week the signal pointed away from Google.

What each rival is buying is specific. Anthropic gets a researcher whose career-defining work is computational biology, at a lab that has been quietly building life-sciences capability. Whatever Anthropic ships in protein, drug discovery, or molecular work through 2026 and 2027 will carry Jumper's fingerprints. OpenAI gets a pretraining co-lead from the Gemini team on the eve of its next flagship, at a lab whose pretraining stack is the one Shazeer was building against. The hire is a peek into how OpenAI's next round of models gets designed.

These moves don't change what ships next quarter. Senior researchers take time to ramp into new orgs, and Jumper said publicly he plans to take time off before joining Anthropic. Where these hires show up in lab capability is a 2027 question, not a 2026 one. Nothing about the next model from Anthropic or the next flagship from OpenAI is going to carry these names on it.

Reading lab capability in 2026 means looking past the product announcements to the X feeds that previewed them two years earlier. By the time the next flagship from OpenAI or the next biology release from Anthropic ships, this week's two departures will already be priced in.

What to Do With This

If you use a frontier model for biology, chemistry, or any work where the answer depends on a specific lab's lead in a specific scientific domain, watch what Anthropic ships in that domain through Q4. The hire moves the medium-term capability curve there, not anywhere else.

If you lead a team that picks vendors by current capability, add a one-line "who works on this now" annotation to your vendor map. Public leaderboards lag the research by a year, and senior hires are the part that doesn't.

If you own a product roadmap that depends on a single lab continuing to lead in a specific area, treat senior researcher moves as the leading indicator they already are. The Q3 and Q4 shifts in capability ranking will reflect this week's announcements before they reflect anyone's earnings call.

Want More Than This Newsletter?

Alcreon publishes a daily AI briefing, long-form dossiers, and an analysis feed for the teams actually shipping AI in production. This newsletter is one read out of the full library.

Read the daily feed or browse the editorials.

Also on the Radar

The EU Picks a Winner for Its Open-Source Frontier Model Bet

The European Commission selected the EUROPA consortium on June 19 as the winner of its Frontier AI Grand Challenge. The Italian-led group, headed by Domyn, will build an open-source frontier model spanning the 24 official EU languages, designed to run on European infrastructure. Europe's answer to "the frontier sits inside two US labs" is a public-funded open model with sovereignty baked into the procurement.