Google DeepMind Faces Talent Exodus as Key Researchers Depart for Anthropic and OpenAI
Several prominent AI researchers from Google DeepMind have moved to rivals Anthropic and OpenAI, while OpenAI unveiled its first custom AI chip and Anthropic launched a new Slack integration.
The story
The AI software landscape saw a significant shift this week as Google DeepMind experienced a notable talent exodus. Key researchers, including Nobel laureate John Jumper, who led the AlphaFold team, and Gemini contributors Jonas Adler and Alexander Pritzel, have departed for Anthropic.
Arthur Conmy also joined Anthropic, and Noam Shazeer, a co-author of the Transformer paper, moved to OpenAI. These departures, occurring within a few days, have raised concerns about Google's ability to retain top AI talent and its impact on the development of models like Gemini 3.5 Pro, which has seen a delayed release. The moves underscore the intense competition for foundational AI expertise, particularly as Anthropic and OpenAI gear up for potential IPOs, offering substantial equity incentives to attract talent.
Who moved
Google DeepMind
What Changed: Several key AI researchers, including Nobel laureate John Jumper (AlphaFold lead) and Gemini contributors Jonas Adler, Alexander Pritzel, and Arthur Conmy, departed for Anthropic, with Noam Shazeer moving to OpenAI.
Consequence: This talent drain has raised investor concerns about Google's AI talent retention and execution, contributing to a reported intraday stock drop for Alphabet, and potentially impacting the delivery of models like Gemini 3.5 Pro.
OpenAI
What Changed: OpenAI, in partnership with Broadcom, unveiled 'Jalapeño,' its first custom AI chip designed for large language model inference.
Consequence: The chip aims to reduce compute costs by approximately 50% compared to typical AI GPUs while matching the performance of Nvidia's Blackwell and Google's TPUs, signifying OpenAI's move to own more of its infrastructure stack.
Anthropic
What Changed: Anthropic launched 'Claude Tag,' an AI agent designed to embed directly into Slack workflows to accumulate organizational knowledge.
Consequence: This positions Claude to absorb company context over time, creating a data flywheel that could make the model progressively harder to displace within enterprise environments.
Qualcomm
What Changed: Qualcomm announced its acquisition of Modular for approximately $3.92 billion in an all-stock deal.
Consequence: This acquisition strengthens Qualcomm's position in AI software, as Modular's technology allows developers to run AI applications across various chips without extensive code rewriting.
New models
GLM-5.2
Lab: Zhipu AI
What: This open-weight model has achieved a new benchmark record for open-weight AI models.
Use: It is freely available for download, fine-tuning, and commercial deployment, drawing comparisons to closed frontier models in terms of capabilities.
Market signals
Mirendil, a new AI lab founded by former researchers from Anthropic, OpenAI, Google DeepMind, and xAI, raised $200 million in seed funding at a $1 billion valuation.
Implication: This significant early-stage funding for a new entrant, backed by prominent investors like Andreessen Horowitz and Kleiner Perkins, highlights continued investor confidence in specialized AI talent and the agentic AI market.
A new bipartisan U.S. nonprofit, RAISE US, launched with over $500 million in commitments from Amazon, Anthropic, Microsoft, and the OpenAI Foundation.
Implication: The initiative aims to help workers adapt to job changes driven by AI, signaling a growing focus among major AI players and policymakers on workforce transition and the broader societal impact of AI adoption.
AI Weekly highlighted a trend of 'closed loop' financing, where major AI labs are funding each other and acting as customers for one another's services.
Implication: This dynamic raises questions about the true underlying demand in the AI economy versus growth driven by recursive funding, a factor that public investors will scrutinize as some labs prepare for IPOs.
What we'll be watching
- The AI Engineer World's Fair is scheduled from June 29 to July 2, 2026, featuring over 4,000 builders and 400+ sessions on LLMs, agents, and applied AI.
- GITEX AI Europe will take place from June 30 to July 1, 2026.
- OpenAI's DevDay 2026 is scheduled for September 29 at the Fort Mason Center in San Francisco.
- Google's Gemini 3.5 Pro, initially targeted for June general availability, is still expected before June 30, with its release being closely watched after recent talent departures.
Reporting + analyst voices: grounded via Google Search at publish time.