Google DeepMind leadership change: Gemini now reports to Pichai
Demis Hassabis stepped back from CEO. Koray Kavukcuoglu runs Google DeepMind reporting to Sundar Pichai — and owns the Gemini API you build on.
The lab that ships the Gemini models just moved one rung closer to Alphabet's CEO. In a Google DeepMind leadership change announced August 5, Demis Hassabis gave up day-to-day control, and Koray Kavukcuoglu now runs the unit reporting straight to Sundar Pichai. If your product calls the Gemini API, the interesting part isn't who got the corner office. It's which org chart the roadmap now sits on.
What actually happened
Per Pichai's own post, Hassabis becomes Chair of Google DeepMind and Chief Scientist of Alphabet. He keeps working on AGI strategy, stays in London, and gives more time to Isomorphic Labs. He is not leaving.
Kavukcuoglu, previously Google DeepMind's CTO, becomes SVP of Google DeepMind and reports directly to Pichai. His remit is the part you care about: Gemini model development, frontier research, the Gemini app, and Google's AI developer platforms. Note the title — SVP, not CEO. The unit that was run at arm's length now sits inside the reporting line of the company that has to monetize it.
Jeff Dean is also out after 27 years, launching an independent public benefit corporation with Sanjay Ghemawat, with Google investing. We wrote about that departure on Tuesday. Fortune reports Alphabet shares fell about 5% on the news.
We are deliberately not printing a "what this means for AGI" take. Nobody outside the building knows. What is observable is the structure.
Why a vendor's leadership change matters for your business
Research orgs optimize for papers. Product orgs optimize for shipping and margin. When the group that owns a model API starts reporting to the CEO of the company selling cloud, you should expect the things product orgs do: faster release cadence, faster deprecation of old model IDs, more aggressive tiering, and price changes that arrive on a business calendar rather than a research one. That is not a criticism. It is a planning assumption.
Concretely, two things belong on your list this month. First, pin your model IDs. If your code says gemini-latest anywhere, a routine model refresh becomes an unannounced behavior change in your production app. We have cleaned up that exact incident for a client, and the fix takes an hour before it happens and a week after.
Second, keep a tested second route. Not a spreadsheet listing alternatives — an environment variable that swaps the provider and a set of your own evals you can run against both. The point isn't that Google will do something hostile. It's that a vendor reorganizing its AI leadership twice in one week is telling you the roadmap is in motion, and your leverage is being able to leave.
Key takeaways
- Announced August 5: Hassabis becomes Chair of Google DeepMind and Alphabet's Chief Scientist
- Koray Kavukcuoglu is SVP of Google DeepMind, reporting directly to Sundar Pichai
- He owns Gemini models, frontier research, the Gemini app, and the AI developer platforms
- Jeff Dean is leaving after 27 years; Alphabet shares fell roughly 5% on the news
- Pin your model IDs, keep your own evals, and make provider swap a config change
Your model layer should survive someone else's reorg. We build AI systems where the provider is one line of config, the model ID is pinned, and your own evals gate every upgrade. See how we build vendor-agnostic AI or tell us what you're building on.
- #google-deepmind
- #gemini
- #vendor-risk
- #ai-models
- #model-portability
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Suno will watermark AI songs and cap your downloads
Suno is adding watermarking, fingerprinting, and download limits to AI-generated music. Your rights to distribute AI output are a vendor setting, not a deed.
Read itSapiom's $35M: your agent bill is a routing decision
Sapiom raised $35M Series A to route AI agent calls to the cheapest capable model. The lesson isn't the vendor — it's that model choice belongs in config, not code.
Read it