Career 2 min read

Jeff Dean Leaves Google for $650M Science AI Startup Axiom

Jeff Dean has resigned from Google DeepMind alongside key researchers to launch Axiom Research, a startup backed by $650 million to build scientific AI models.

Jeff Dean has officially resigned from Google DeepMind to launch an independent startup called Axiom Research. The August 5 departure marks the most significant leadership transition at Google since the 2023 consolidation of the Brain and DeepMind units. The new venture shifts focus away from standard language modeling toward applied artificial intelligence for the hard sciences.

Foundation Models for Physical Laws

Axiom Research is developing Autonomous Discovery Engines. Rather than optimizing for text output, these foundation models are grounded directly in physics, chemistry, and biology data. The technical objective is to accelerate the Design-Build-Test-Learn cycle in laboratory settings.

By pairing these models with closed-loop robotic automation, the startup aims to deploy software that can hypothesize, run experiments, and refine theories without human intervention. This mirrors recent industry pushes where an AI agent automates scientific discovery code, moving commercial value toward physical outputs like new drugs and advanced battery materials.

The Leadership Exodus

The founding team includes several primary architects of Google’s modern AI infrastructure. Demis Hassabis will remain the sole lead of Google DeepMind, while Zoubin Ghahramani will absorb several of Dean’s long-term strategy responsibilities.

ExecutiveFormer Google RoleKey Contributions
Jeff DeanChief ScientistGoogle Brain co-founder, infrastructure
Oriol VinyalsVP of ResearchGemini and AlphaStar
Quoc LeResearcherAutoML and Seq2Seq
Douglas EckPrincipal ScientistMagenta project

Funding and Infrastructure Scale

To support the computational demands of scientific modeling, Axiom Research has secured an estimated $650 million seed round led by Thrive Capital and Andreessen Horowitz (a16z). This capital immediately flows into compute resources.

The startup has established a multi-year partnership with Oracle Cloud Infrastructure (OCI) to access NVIDIA Blackwell-2 (B200) clusters. Training physics-based models requires massive data ingestion from laboratory sensors and simulations, necessitating tight integration between the training clusters and experimental data pipelines.

If you are building multi-agent systems or evaluating AI infrastructure, this launch underscores a rapid capitalization shift toward domain-specific vertical models. Generative discovery requires fundamentally different evaluation frameworks and compute architectures than conventional conversational applications.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading