On August 5, 2026, Google made a change that affects three strategic fronts at the same time: who leads the development of Gemini, who thinks about Alphabet's long-term research and where some of the company's most influential researchers will work from now on.
Demis Hassabis stepped down from daily operational responsibilities at Google DeepMind to become Chair of the unit and Chief Scientist at Alphabet. Koray Kavukcuoglu, until then CTO of DeepMind and already Chief AI Architect at Google, assumed day-to-day leadership as Senior Vice President. In the same movement, Jeff Dean and Sanjay Ghemawat left Google to participate in the creation of Discovery Loop. The independent company also counts Quoc Le and Oriol Vinyals among its four co-founders.
The first reading may be that of a stampede of talent. It is incomplete. Google lost four enormously important researchers and engineers, but decided to invest in the new company, provide infrastructure through Google Cloud and maintain collaboration on machine learning systems. At the same time, it concentrated under Kavukcuoglu the operation of Gemini, frontier AI research, the Gemini app, and developer-facing teams.
The result is less a simple change of executives and more a redistribution of Alphabet's bets: execution of models and products on the one hand, long-term strategic research on the other and, outside the company, a new company that wants to automate the discovery process itself.
Google has more clearly separated operation and search
Hassabis has not left the Alphabet ecosystem. He remains linked to Google DeepMind as Chair, becomes Chief Scientist at Alphabet and remains head of Isomorphic Labs. In his message to employees, he stated that he wanted to step away from daily operational responsibilities to have more time for broad issues related to AGI and science.
Kavukcuoglu receives a different mandate. According to Sundar Pichai, he will oversee the development of Gemini models, Frontier AI research, the Gemini app, and developer teams. Reuters confirmed that the executive will assume day-to-day responsibilities without receiving the title of CEO of DeepMind.
This distinction matters because it avoids an exaggerated interpretation of the reorganization. Hassabis was not simply replaced or removed from the AI strategy. What changed was the division of responsibilities: he stepped away from daily management, while Kavukcuoglu began to be responsible for a greater portion of technical and product execution.
There is a clear reason for this execution to gain weight. Just weeks before the reorganization, Google had released Gemini 3.6 Flash and confirmed that Gemini 3.5 Pro was still in testing, while a new round of pre-training for Gemini 4 had already begun. In the August 5 announcement, Hassabis again cited the progress of Gemini 4.
Competition hasn't stopped either. OpenAI updated GPT-5.6 Sol on August 6, a day after Google's change, while Anthropic had introduced Claude Sonnet 5 on June 30. It is not necessary to declare a winner to see the organizational problem: rapid model cycles require strong research, but also the ability to transform advances into products, APIs and distribution with little delay.
Why this change is also a story about the Search
Jeff Dean and Sanjay Ghemawat are historically associated with important parts of the infrastructure that allowed Google to scale its search and computing systems. Pichai himself highlighted in the announcement that the two participated in transitions that went from the initial Search infrastructure to the neural networks of the modern era of AI.
This doesn't mean that Dean and Ghemawat ran Google Search's current operation. The most important connection in 2026 is another: Search is increasingly dependent on models coming out of Google DeepMind.
In May, Google made Gemini 3.5 Flash the default AI Mode model globally. At I/O 2026, the company also stated that AI Overviews had surpassed 2.5 billion monthly active users and that AI Mode had surpassed 1 billion. In the August statement about the reorganization, Pichai also reported that the Gemini app had reached more than 950 million monthly users. These are numbers released by Google itself, but they help to show the scale at which a decision about models is no longer just a laboratory matter.
When Kavukcuoglu simultaneously takes over models, frontier research, application, and developer teams, he comes to command a technology layer that powers Google's various distribution surfaces. Search is one of the most important.
Therefore, the dispute over Search should not be reduced to the old comparison between search engines. OpenAI says 1 billion people use ChatGPT per week for activities including quick questions, web searches, planning, and in-depth research. These are different metrics and periods than the numbers released by Google, so they do not allow a direct comparison of participation. Still, they show how responses generated by models began to compete for activities that previously started almost automatically in a search box.
In this context, Google needs to do two things at once: continue to advance frontier models and quickly turn those advances into experiences that work in Search at the scale of billions of users. The reorganization makes this operational responsibility more explicit.
Discovery Loop brings competition into the research process
Dean's departure would be relevant in its own right after approximately 27 years at Google. But the composition of the Discovery Loop makes movement more important.
The new company brings together Dean, Ghemawat, Quoc Le and Oriol Vinyals. The group combines experience in distributed systems, large-scale infrastructure and artificial intelligence research. The company's website says its mission is to create AI systems capable of automatically solving important problems in machine learning, science and engineering.
The idea is not just to place a chatbot on a researcher's desk. In an interview with WIRED, the founders described a cycle in which AI can propose experiments, implement what is needed to run them, evaluate results, and use the findings to decide what to test next. The ambition is to execute many of these cycles in parallel.
The first planned application is especially relevant to the model competition: Discovery Loop intends to use its own technology to accelerate machine learning research. In other words, before attacking areas such as biology, chip design, medicines or materials, the company wants to apply automated cycles to improving AI systems themselves.
If this approach works, the difference will not just be in training a larger model. It will be increasing the speed with which new architectures, methods and hypotheses can be proposed, tested and discarded.
It's still a gamble. Until August 9, 2026, the public material consulted presents the mission, founding team and hiring process, but does not include a commercial product launched, its own performance benchmarks or scientific results produced by the new company. The scale of the opportunity should not be confused with results already demonstrated.
The curious point: Discovery Loop competes with a thesis that Google itself is already developing
Discovery Loop isn't entering territory unfamiliar to its former peers. Google had already been transforming “AI for science” into a concrete strategic line.
In May, the company introduced Gemini for Science, bringing together tools based on systems such as Co-Scientist, AlphaEvolve and Empirical Research Assistance. Google describes these technologies as ways to accelerate steps in the scientific method, including hypothesis generation, evaluation, and computational discovery.
AlphaEvolve, for example, combines Gemini models with automated evaluators and an iterative process to discover and optimize algorithms. Isomorphic Labs, which Hassabis will continue to lead, is trying to apply frontier AI to the drug discovery process.
This creates an unusual situation. Google loses four experts to a company whose thesis touches on an area in which Alphabet itself is investing, but simultaneously becomes that company's founding investor and infrastructure partner.
The relationship, therefore, does not fit well into the word “rival”. Discovery Loop is independent and will have its own incentives, investors and intellectual property, but it was born commercially and technically connected to the former employer of its founders.
For Google, this structure reduces the binary nature of the output. The company loses the daily work of researchers within its organization, but maintains an economic and technological relationship with the new laboratory. For Discovery Loop, the agreement offers access to a Cloud partner capable of sustaining intensive computational loads without the startup having to build this infrastructure from scratch.
The biggest risk for Google is losing talent density, not running out of research
The reorganization came after other prominent departures. In June, John Jumper, one of those responsible for AlphaFold and winner of the 2024 Nobel Prize in Chemistry alongside Hassabis, announced his move to Anthropic. During the same period, Noam Shazeer, who had once again worked on developing Gemini, left Google to join OpenAI.
Reuters framed the sequence as part of a growing scramble for top researchers and reported a 4% drop in Alphabet shares on the day of the Aug. 5 announcement. A single day's market reaction does not demonstrate future operational impact, but indicates that investors attributed relevance to the changes.
It would also be exaggerated to conclude that Google has run out of scientific capacity. The company continues to bring together teams from DeepMind, Google Research and Cloud, maintains programs such as Gemini for Science, AlphaEvolve and Co-Scientist and is training a new generation of Gemini.
The problem is more specific: exceptionally experienced researchers are not perfectly replaceable parts. The simultaneous loss of people who participated in different generations of infrastructure, models and discoveries can reduce the density of knowledge accumulated within the same organization, even when programs continue to be funded.
What to look for after reorganization
The August 5th change will be easier to judge by results than by new titles.
At Google, the test will be to see whether the division between Hassabis and Kavukcuoglu allows it to accelerate models and products without weakening long-term research. The progress of Gemini 3.5 Pro, Gemini 4, and Search integrations will be a more useful sign than speculation about internal hierarchy.
In Discovery Loop, the decisive point will be moving from the thesis to concrete evidence: working systems, reproducible results, scientific applications or advances in machine learning that show that automating experimentation cycles produces real gains.
And on Alphabet's science front, it will be important to watch how Hassabis' new role connects to Google DeepMind, Gemini for Science, and Isomorphic Labs. The company, after all, is not abandoning the idea of using AI to accelerate science. It's moving one of its top leaders even closer to that mission while former colleagues try to pursue an independent version of the same transformation.
That's what makes the reorganization bigger than executive news. It shows that the next AI fight may not just be about who has the best model. It could also be whoever builds the system is able to discover the next model — and, after that, new technologies and scientific discoveries — faster.
