After 27 years at Google, Jeff Dean has made a move that would have seemed unlikely just a few years ago: he left one of the world’s most powerful technology companies to build an AI startup of his own.
Dean, a longtime Google executive and former chief scientist, is now focused on Discovery Loop, a company aimed at using artificial intelligence to automate parts of the scientific research process.
Speaking at Stanford University earlier this month, Dean explained that his decision was driven less by dissatisfaction with Google and more by a changing AI landscape. Advances in cloud computing, he argued, have made it possible for relatively small teams to pursue projects that once required the resources of a technology giant.
His move also highlights a broader shift in AI: the ability to access enormous computing infrastructure without actually owning it is lowering the barrier to building ambitious technology companies.
Why did Jeff Dean leave Google after 27 years?
Dean described leaving Google as an emotional decision, particularly after spending nearly three decades at the company.
He remains optimistic about Google’s AI ambitions, including its Gemini family of models. But the opportunity to build a focused company around science, engineering and AI-driven research ultimately proved more compelling.
Rather than competing directly with Google across consumer technology, advertising and general-purpose AI, Dean is betting on a narrower problem: whether AI agents can perform meaningful portions of the scientific discovery process.
That includes tasks that traditionally require teams of researchers, engineers and laboratory specialists.
Dean’s decision reflects how quickly the economics of AI development have changed. A startup no longer necessarily needs to construct its own massive computing infrastructure before it can begin developing sophisticated AI systems.
How has cloud computing changed AI startups?
One of Dean’s central arguments is that cloud computing has dramatically reduced the infrastructure burden facing new AI companies.
In the past, an ambitious AI project could require substantial investments in:
- Specialized computing hardware
- Data centers
- Networking infrastructure
- Storage systems
- Engineering teams to operate the infrastructure
- Long-term capital commitments
Today, startups can rent computing capacity from major cloud providers and scale their usage as their needs change.
That does not make AI development cheap. Training and operating advanced models can still cost enormous amounts of money. But companies can increasingly treat computing as a service rather than building every layer of infrastructure themselves.
For a founder like Dean, who has spent decades working on large-scale computing systems, that shift is particularly significant.
What is Discovery Loop trying to build?
Discovery Loop is focused on automating scientific research.
The company’s ambition extends beyond using AI to summarize academic papers or help researchers write code. Its stated vision involves AI systems participating in multiple stages of the research process.
That could include:
- Proposing experiments based on existing scientific knowledge.
- Designing and implementing experiments using available tools and systems.
- Analyzing experimental results.
- Generating new hypotheses from those results.
- Repeating the process to accelerate scientific discovery.
The underlying idea is that AI could eventually function as a research partner capable of continuously testing hypotheses rather than simply responding to questions.
If successful, such systems could potentially compress research cycles that currently take months or years.
Which scientific fields could Discovery Loop target?
Discovery Loop has indicated that it wants to tackle major scientific and engineering problems rather than focus exclusively on conventional software applications.
Potential areas include neuroscience and nuclear security, among other fields.
These are particularly demanding areas because useful AI systems would need to do more than generate plausible text. They would have to reason about scientific evidence, work with experimental data and potentially interact with real-world research environments.
That creates a much higher technical bar.
An AI system that produces a convincing explanation is one thing. An AI system that proposes an experiment, correctly predicts what should happen, interprets the results and uses them to design the next experiment is a considerably more ambitious proposition.
Why is Discovery Loop a public benefit corporation?
Discovery Loop has been established as a public benefit corporation, a corporate structure designed to allow a company to pursue specified public benefits alongside traditional business objectives.
That structure is particularly notable given the company’s mission.
Scientific research can produce benefits that extend far beyond the commercial interests of the company conducting it. A breakthrough in neuroscience, for example, could have implications for medicine and public health.
Similarly, work involving nuclear security could have consequences that go beyond the company’s eventual commercial returns.
By adopting the public benefit corporation structure, Discovery Loop can formally incorporate broader societal considerations into its corporate decision-making.
That does not mean the company is a nonprofit. It remains a commercial business and can raise private capital.
Who is backing Jeff Dean’s new AI company?
Discovery Loop has attracted backing from several prominent venture capital firms and technology investors.
Reported investors include:
- Radical Ventures
- Khosla Ventures
- Lightspeed
- Kleiner Perkins
- Doerr Capital
- Alphabet, Google’s parent company
Alphabet’s involvement is particularly interesting because Dean spent nearly three decades at Google before launching the startup.
Alphabet is reportedly not only a founding investor but also a cloud partner for Discovery Loop.
That arrangement could give the startup access to significant computing resources while allowing Dean to pursue a more focused strategy outside Google.
Is Jeff Dean raising $1 billion for Discovery Loop?
Reports indicate that Dean is in discussions to raise approximately $1 billion, potentially valuing Discovery Loop at around $10 billion.
If completed at those terms, it would place the company among the more heavily funded AI startups despite being at a relatively early stage.
The scale of the reported fundraising also illustrates the unusual economics of frontier AI.
Cloud computing has lowered the infrastructure barrier for startups, but advanced AI research can still require enormous amounts of computing power, specialized talent and experimentation.
In other words, cloud computing has made it easier to start an AI company. It has not necessarily made cutting-edge AI inexpensive.
Why does Jeff Dean’s move matter for Google?
Dean’s departure is significant because of his extraordinary history at Google.
He joined the company in 1999 and became one of its most influential technical leaders. His work has been closely associated with Google’s large-scale computing infrastructure and later with its AI efforts.
His decision does not necessarily signal a loss of confidence in Google.
In fact, Dean has continued to express optimism about the company’s AI direction.
Instead, his departure illustrates the growing competition between established technology companies and startups for AI talent.
Large companies have enormous advantages: computing resources, data, research teams and distribution. Startups, however, can offer something different — a narrow mission, fewer layers of decision-making and the ability to build an organization around one specific technological bet.
For Dean, that appears to be the attraction.
What makes Discovery Loop different from a typical AI startup?
Many AI startups are building applications around existing foundation models.
Discovery Loop’s ambition appears to be considerably broader.
Its goal is to apply AI to the process of generating scientific knowledge itself.
That creates several potential advantages if the technology works:
- AI could help researchers explore far more hypotheses.
- Experiments could potentially be designed and analyzed faster.
- Researchers could spend more time on high-level scientific questions.
- AI systems could identify patterns across enormous scientific datasets.
- Research organizations could potentially operate more continuously rather than relying entirely on human working hours.
But there are equally significant challenges.
Scientific research is full of uncertainty. Experiments can fail for reasons that are difficult to model, data can be incomplete, and correlations do not necessarily establish causation.
An AI system that confidently proposes a scientifically plausible but fundamentally flawed experiment could waste significant resources.
For Discovery Loop, therefore, the challenge is not simply building a smarter AI model. It is building systems that can operate reliably within the messy, evidence-driven world of science.
What does Jeff Dean’s move say about the future of AI?
Dean’s decision points to a broader evolution in the AI industry.
The first phase of the AI boom was dominated by the companies capable of training enormous foundation models. The next phase may increasingly focus on what those models can actually do.
That includes AI agents capable of writing software, operating tools, conducting analysis and, potentially, carrying out scientific research.
If that transition happens, companies may compete less on simply having the largest model and more on building AI systems that can complete complicated, multi-step tasks in the real world.
Discovery Loop is betting that scientific discovery will be one of those areas.
The bigger bet behind Discovery Loop
Jeff Dean is not simply betting that AI can make researchers more productive.
He is betting that AI can become part of the research engine itself.
That is a much bigger proposition.
The technology could eventually help scientists generate hypotheses, conduct experiments and learn from results at a pace that humans alone cannot match. If it works, the implications could stretch across medicine, neuroscience, engineering, energy and national security.
But the uncertainty is equally substantial.
Dean has acknowledged that building a new company comes with risks. After 27 years at one of the world’s most successful technology companies, he could have remained at Google and continued working on its enormous AI operation.
Instead, he chose to build something much smaller — with an equally enormous ambition.
The bet behind Discovery Loop is straightforward: if AI can accelerate the process of scientific discovery, the next major AI breakthrough may not be another chatbot or search engine. It could be a machine that helps discover what humans have not yet found.