Google presents a paradox: it holds one of the most enviable positions in artificial intelligence, yet a steady stream of top-tier talent is fleeing to leading AI labs and startups. This tension has been on full display recently.
First, Google announced an 82% surge in cloud revenue. Then, its AI organizational structure underwent a major overhaul, with Chief Scientist Jeff Dean departing after 27 years at the company. These events highlight a core dilemma for the $4 trillion tech giant: where should capital be allocated? Developing cutting-edge large language models requires massive upfront investment in computing and research, with uncertain future returns, while the cloud business boasts exceptional operational efficiency, growing much faster than similar offerings from Amazon and Microsoft.
Alphabet CEO Sundar Pichai stated on a recent earnings call that 90 of the Fortune 100 companies are using the Gemini enterprise version, demonstrating Google's powerful ability to sell AI services to cloud clients. Tomasz Tunguz, founder of venture firm Theory Ventures, believes it is increasingly clear that top-tier flagship models are not needed to meet most enterprise demands. "I think AI has reached a point where, especially for many white-collar work scenarios, moderately capable models are sufficient," he said. "Next-generation frontier models will likely only be valuable in specific, compute-intensive domains."
Where to start
Google's full-stack AI strategy is a key reason its stock has risen 16% this year, and it has gained 65% in 2025, outperforming all large tech peers. However, the road ahead is bumpy. After the latest earnings report, Alphabet's stock fell due to concerns over capital expenditure. The stock dropped further following the news of Dean's departure and Demis Hassabis stepping down as CEO of Google DeepMind to become chairman of the unit.
Despite Wall Street's generally optimistic outlook, not everyone at Google is happy with the current direction. According to sources who requested anonymity, some researchers are dissatisfied that their access to computing resources for frontier projects is limited, while Google Cloud sells its proprietary TPU chips to external clients, including rival Anthropic. Google's complex management structure also frustrates researchers, making it difficult to bring scientific results to market without layers of approval. In contrast, emerging companies like OpenAI, Anthropic, and other startups hold greater appeal for AI researchers and engineers who prefer lab work over financial statements.
Why a wave of top talent is leaving
Dean is leaving to start a new company, Discovery Loop, alongside prominent Google researchers Sanjay Ghemawat, Oriol Vinyals, and Kok Le. Dean stated on X that the startup, funded by Google, will be structured as a public benefit corporation with the mission of automating machine learning, science, and engineering to accelerate scientific discovery and technological progress. They join a growing list of notable researchers who have left, including Noam Shazeer, an author of the seminal 2017 paper "Attention Is All You Need," which is the foundation of generative AI. All eight authors of that paper have now left Google. Shazeer joined OpenAI in June, less than two years after Google rehired him through a nearly $3 billion acquisition. Shortly after, Nobel laureate John Jumper left DeepMind to join Anthropic.
D.A. Davidson analyst Gil Luria sees a clear trend behind the talent exodus. "These researchers are not enthusiastic about the commercialization of AI. They want to be witnesses to major historical breakthroughs, so they see Anthropic, OpenAI, or other startups as platforms to achieve that goal." Dean was one of the few senior Google executives willing to publicly criticize the Trump administration, opposing a Defense Department decision to label Anthropic a supply chain risk. Technically, he was instrumental in building the computing infrastructure and neural network systems that established Google as an early leader in modern AI.
Hassabis, who co-founded DeepMind in 2010 and sold it to Google four years later, will now serve as chairman of the unit while also taking on the new role of Alphabet's Chief Scientist, focusing on long-term research and the societal impact of artificial general intelligence (AGI). He also plans to devote more time to Isomorphic Labs, an AI drug discovery company spun out of DeepMind. DeepMind's technical lead and Alphabet's Chief AI Architect, Kole Kavukcuoglu, will take over day-to-day management of the unit and the development of the next-generation Gemini model. A source close to DeepMind noted that Kavukcuoglu has gradually taken over much of Hassabis's work over the past year.
Just 10 ASX 200 shares for a portfolio?
One of the biggest internal conflicts at Google is the allocation of computing resources. While Google is a leading investor in data centers, chips, and infrastructure, computing power remains scarce. Every TPU allocated for model training, powering Google's own products, or fulfilling cloud client contracts represents a trade-off between competing demands. The tension is particularly high when Google commits significant resources to competing labs like Anthropic, whose models directly compete with Gemini. Pichai has stated on earnings calls that ensuring DeepMind has sufficient compute is Google's "first priority," calling it "the foundation of all our businesses."
Google is pushing for deeper integration between DeepMind and Google Cloud. At the World Economic Forum in Davos in January, Hassabis and Google Cloud CEO Kurian appeared together to discuss enterprise AI products. Historically, the two units operated independently. This joint appearance signaled that Hassabis is beginning to focus more on enterprise AI applications, such as code generation and customer service. It also reflects Google's overall strategy to bridge its research and commercial arms in the fight against OpenAI and Anthropic.
This strategy gives Google multiple paths to monetize AI: selling computing infrastructure to labs like OpenAI and Anthropic, delivering the Gemini model to enterprises, and embedding AI into its own products like Search, YouTube, and Workspace. Nearly four years into the generative AI boom, a key question is whether Google needs to develop the world's best large language models itself or let others bear the high R&D costs. Kavukcuoglu stated at Google's developer conference in May that the company aims to push technical boundaries while constantly improving operational efficiency. The company's Flash model offers cutting-edge performance with four times the speed and energy efficiency of similar models, enabling large-scale deployment of advanced AI for enterprise services and consumer products.
Similar to Tunguz, investor Dan Niles believes the most powerful models are unnecessary for most commercial applications. "Existing models are sufficient for 90% of business needs. Many tasks don't require a 'Ferrari'; a 'family car' can handle nine out of ten use cases." However, for scientists and researchers dedicated to achieving the next breakthrough on the scale of the Transformer, "good enough" will never be enough.
Comments