The Evolved Strategy: Kohli on AlphaFold's Dispersal and DeepMind's Gemini Turn
A public-record reconstruction of Google DeepMind's AlphaFold team breakup, Pushmeet Kohli's pivot language, and the Nobel-winning talent now heading to Anthropic.
Google DeepMind did not hold a press conference to explain what happened to the team behind AlphaFold. Instead, the story assembled itself from a Financial Times report, a handful of confirmed reassignments, and the public LinkedIn moves of researchers who helped win the 2024 Nobel Prize in Chemistry.
This piece reconstructs that public record — what DeepMind confirmed, what Pushmeet Kohli said about strategy, and where the people went.
What DeepMind confirmed
According to the Financial Times on July 29, 2026, Google DeepMind has broken up the core team behind AlphaFold. Most original AlphaFold paper authors were reassigned over the past year rather than kept on a single dedicated protein-folding unit.
DeepMind confirmed the moves to the FT. Researchers were redirected toward:
- Gemini-powered scientific assistant systems — tools meant to help scientists and, over time, automate parts of research workflows
- Enzyme design, nuclear fusion, and genomics programs elsewhere in the lab
- Isomorphic Labs, Alphabet's drug-discovery subsidiary spun out of DeepMind in 2021 (Novartis and Eli Lilly partnerships on record)
Nearly one quarter of the full-time Google DeepMind authors on the original AlphaFold papers have left the company entirely, the FT reported.
Kohli's pivot language
Pushmeet Kohli — vice president of research at Google DeepMind, founder of the AI for Science team — gave the FT a framing quote that reads less like a personnel update and more like a doctrine change:
"Our strategy over the last nine years has been to focus on grand challenges . . . a concrete goal every project is focused on. The strategy has evolved."
Read plainly: the era of organizing elite researchers around one scientific moonshot — protein structure prediction being the emblematic case — is giving way to frontier model competition. DeepMind is now explicitly building Gemini systems that assist scientists while racing OpenAI and Anthropic on agentic AI.
Kohli did not, in the public record cited here, describe AlphaFold as finished. The FT's reporting is about team structure and talent routing, not a product shutdown.
The departures that matter
The inner circle moves are what turned a restructuring story into a talent-market signal.
- John Jumper (AlphaFold co-lead, DeepMind VP and engineering fellow) and Jonas Adler moved earlier in 2026 to Google's internal "Code Strike" team — assembled to improve AI coding capabilities as Google chases Anthropic and OpenAI.
- In July 2026, Jumper announced he is joining Anthropic, with Adler and AlphaFold colleague Alexander Pritzel following.
An unnamed DeepMind employee told the FT all three were "instrumental, important, core members" and that their exits sparked internal surprise.
Anthropic, for its part, launched Claude Science for biology and drug discovery earlier in 2026 — a direct commercial lane for the expertise those researchers carry.
What DeepMind says it is proud of
DeepMind's statement to the FT emphasized legacy rather than retreat: it is "incredibly proud" of AlphaFold's scientific heritage and global impact, noting the project "inspired the launch of Isomorphic Labs with the mission to reimagine drug discovery."
That framing matters. AlphaFold remains a reference implementation of AI-for-science credibility. The question is whether organizational continuity — the dedicated team that iterated from CASP breakthrough to Nobel — survives the pivot to general models and agents.
The strategic read (grounded in public facts only)
Three verified threads intersect:
- Grand-challenge teams that produced AlphaFold are dispersing into Gemini science tooling, other scientific domains, or Isomorphic.
- Frontier lab competition is pulling Nobel-adjacent talent toward Anthropic (and previously toward internal Google coding efforts).
- Kohli's language explicitly marks a strategy evolution — from concrete scientific goals per project toward model-and-agent platforms that span many goals.
What we cannot claim from public sources: whether AlphaFold model quality or release cadence will degrade, or whether Isomorphic absorbs enough of the bench to compensate. Those are open questions the public record has not answered.
What to watch next
- Whether DeepMind publishes new AlphaFold-family science without the original author bench
- Isomorphic Labs pipeline disclosures in the Novartis / Eli Lilly partnerships
- Anthropic's Claude Science outputs now that Jumper, Adler, and Pritzel are on record as incoming
- Any DeepMind Gemini science-agent product launch that replaces the "single grand challenge team" workflow Kohli described