DeepMind leadership overhaul ends frontier-lab status

DeepMind underwent a full leadership change on 5 August. Demis Hassabis, co-founder and former CEO, is no longer involved in day-to-day operations. Jeff Dean, former Google chief scientist and Gemini lead partner, is leaving to create a new lab called Discovery Loop. Dean founded Google Brain and the TPU program. Alongside him depart Sanjay Ghemawat, Quoc Le and Oriol Vinyals. Ghemawat and Le were Google Fellows, a title reserved for a handful of senior technical contributors, and Vinyals was a lead partner on Gemini. Former DeepMind CTO and the last remaining Gemini lead, Koray Kavukcuoglu, replaces Hassabis in leading DeepMind and Gemini.

From frontier lab to lagging

According to an analysis by Semi Analysis, these are not the actions of people who are enthusiastic about Gemini 4 Pro. The authors state that DeepMind is no longer a frontier lab. They attribute this assessment to Tokenomics customers a few months ago, citing talent drain from reinforcement-learning teams and poor compute allocation. The conclusion is that Google’s chance of returning to state-of-the-art (SOTA) leadership is effectively zero. Google will continue to release models, but the path to the summit is closed.

Who really benefits from the shake-up

The biggest beneficiary is Google Cloud (GCP), not Anthropic or OpenAI. In the past Gemini and GCP competed for compute allocation. Now it is clear that Thomas Kurian, CEO of GCP, won. The outlook is that GCP’s revenue growth will accelerate sharply, with the division’s annual revenue growth already above 100 percent. At the same time, senior AI researchers are choosing to leave Google, raise billions from external investors and Google Ventures, and spend the money on Nvidia graphics processors inside GCP, in a model already tested by David Silver of Ineffable Intelligence last November.

From Gemini 3 Pro to quiet cancellation

At the end of 2025, Anthropic, OpenAI and Google were considered the big AI trio. In November 2025 Gemini 3 Pro was, according to many, the world’s best model, prompting Sam Altman to declare “red code” at OpenAI. Since then the gap between DeepMind and OpenAI and Anthropic has only widened. Gemini 3.5 Flash failed, and industry rumors suggest the near-release Gemini 3.5 Pro would have been roughly on par with Opus 4.5. That is below GLM 5.2, and certainly below Fable 5 and GPT 5.6. Following an institutional report published on 9 July, Google quietly cancelled Gemini 3.5 Pro and replaced it with Gemini 3.6 Flash as an interim model, but it is inferior to Muse Spark 1.2, Grok 4.5 and first-tier Chinese open-source models. Semi Analysis’s comparison of Spark 1.2 and Grok 4.5 to 3.5 Flash is confusing to them because it pits Tera models against Luna.

Talent flight, model delays

The talent exodus from DeepMind does not stop with Dean and his team. Noam Shazeer and John Jumper, together with most senior reinforcement-learning experts on Gemini, have also left the company. Although Google backed most of the foundational breakthroughs of the current large-language-model paradigm, it can no longer retain the leading AI talent. Meanwhile OpenAI has solved its pre-training bottlenecks, and a larger model code-named “Doug” is under active development. The assessment is that both MSL and SpaceXAI will surpass Google in model quality by year-end.