Google DeepMind Restructures Leadership as Demis Hassabis Steps Back and Jeff Dean Departs to Launch Startup
In a historic and sweeping reorganization on Wednesday, August 5, 2026, Alphabet Inc. announced a complete leadership overhaul at Google DeepMind. Driven by persistent delays of its flagship Gemini 3.5 Pro model and intensifying pressure from OpenAI and Anthropic, the restructuring marks the end of Google's era as a unified, research-led frontier AI developer.
The Leadership Overhaul and the Discovery Loop Exodus
The restructuring is defined by the departure of Google's most legendary technical minds:
- Demis Hassabis, DeepMind co-founder and former CEO, has stepped back from day-to-day operations.
- Jeff Dean, Google's Chief Scientist, co-founder of Google Brain, and Gemini co-lead, is leaving Google after 27 years to launch Discovery Loop, an independent AI-for-science startup.
- Joining Jeff Dean at Discovery Loop are three other Google AI icons: Sanjay Ghemawat (Google Senior Fellow and co-creator of MapReduce, Bigtable, and Spanner), Quoc Le (Google Fellow and co-inventor of sequence-to-sequence learning), and Oriol Vinyals (Gemini co-lead and pioneer of AlphaStar).
- Koray Kavukcuoglu, former DeepMind CTO and the sole remaining Gemini co-lead, has been appointed to replace Hassabis as the sole leader of DeepMind/Gemini.
On the same day, Discovery Loop officially launched as a public benefit corporation (PBC) based in Palo Alto, focused on automating the experimental loop in scientific research, biology, chip design, and recursive machine learning. The startup announced an initial seed funding round co-led by Radical Ventures and Khosla Ventures, with participation from Lightspeed, Kleiner Perkins, Doerr Capital, and Alphabet itself.
The Death of Gemini 3.5 Pro and Pivot to Gemini 4
Following months of development setbacks and underwhelming performance relative to competitors like Anthropic's Claude Fable 5 and OpenAI's GPT-5.6, Google has silently canceled Gemini 3.5 Pro. The company is now attempting to cope by shifting its entire research focus and pre-training compute to Gemini 4.
As a bridge model, Google released Gemini 3.6 Flash, but industry analysts note that its capabilities fall behind rival open-weight models, leaving Gemini in 8th or 9th place globally.
The Financialization of Google Cloud (GCP)
Rather than prioritizing frontier research, Alphabet's management has pivoted toward the financialization of Google Cloud Platform (GCP). Google has committed over 20% of its total TPU shipments from 3Q26 to 4Q27 directly to Anthropic under long-term contracts.
While this has starved DeepMind of the compute necessary to remain competitive, it has supercharged GCP's financial performance. GCP grew 82% in the latest quarter, driven by TPU system sales accounted for on a gross basis at ~$35 billion per gigawatt. While DeepMind's status as a frontier lab has effectively collapsed, GCP is on track to reach over $73 billion in third-party AI ARR by the end of 2027, adding ~$3 to Google's EPS.