Google (GOOGL.US) AI welcomes “two heavens of ice and fire”: cloud business surged 82%, and the departure of top scientists raised concerns about cutting-edge R&D

The Zhitong Finance App learned that in the past two weeks, Google's “ice and fire” pattern in the field of artificial intelligence has been fully demonstrated. On the one hand, cloud business revenue soared by 82%, and on the other hand, chief scientist Jeff Dean (Jeff Dean) announced his departure after 27 years of work.

For this tech giant, which laid the foundation for generative AI with its 2017 Transformer paper and now has a market capitalization of $4 trillion, recent events have highlighted its core strategic challenge: Where to invest the money? Building a cutting-edge model requires huge upfront costs in computing power and R&D, and future returns are difficult to determine; the cloud business has proven to be efficient and grows far faster than similar services from Amazon and Microsoft.

Alphabet CEO Sundar Pichai (Sundar Pichai) said during last month's earnings call that 90% of the Fortune 100 companies are using Gemini Enterprise, which fully demonstrates Google's ability to sell AI services to corporate customers.

Tomasz Tunguz (Tomasz Tunguz), founder of venture capital firm Theory Ventures, pointed out that there is no need for a top model to meet the needs of most companies. “I think in the AI field, especially in a large number of white-collar work scenarios, many models with decent performance are sufficient,” said Tonguz. “The next generation of models is likely to be used more in specific fields that require high-performance computers.”

Google's full-stack AI layout is an important driver of the cumulative 16% increase in its stock price this year (already surged 65% in 2025, outperforming all tech giants). However, the recent trend has been somewhat bumpy: after the release of the latest earnings report, Alphabet's stock price was under pressure due to market concerns about capital expenses; the news that Dean left his job and Demis Hassabis (Demis Hassabis) stepped down as CEO of Google DeepMind on Wednesday to become the chairman of the board, further lowered the stock price.

Despite Wall Street's overall positive attitude, not everyone within Google is happy.

According to a number of people familiar with the matter, who did not wish to be named, some researchers are increasingly dissatisfied with the acquisition of computing power — it is difficult for them to obtain the computing resources needed to advance cutting-edge projects, yet they have seen Google sell self-developed TPUs (tensor processing units, competing with Nvidia GPUs) to external customers, including Anthropic.

Furthermore, Google's internal hierarchical approval is cumbersome, and the transformation of research results into products requires multiple levels of clearance, making OpenAI, Anthropic, and even younger startups more attractive to AI developers — they prefer lab work over financial data.

Dean left his job with senior Google experts Sanjay Ghemawat (Sanjay Ghemawat), Oriol Vinyals (Oriol Vinyals), and Quoc Le to found Discovery Loop. Dean said on the X platform that the Google-funded startup will be positioned as a public service company. “The mission is to automate machine learning, science, and engineering to accelerate discovery and progress.”

Previously, many well-known researchers have left, including Noam Shazeer (Noam Shazeer), one of the authors of the 2017 landmark paper “Attention Is All You Need” — an article that laid the foundation for generative AI, and now all eight authors have left Google.

Chazel switched to OpenAI in June of this year, and it is less than two years since Google recalled it through “takeover recruitment” for nearly $3 billion. Shortly after he left his job, Nobel laureate John Jumper (John Jumper) also left DeepMind to join Anthropic.

“Be part of history”

D.A. Davidson analyst Gil Luria (Gil Luria) pointed out that there is a clear trend of top brain drain. “They're not passionate about commercializing AI, but want to be part of history,” said Luria (who recommended holding Alphabet shares), “so they see Anthropic, OpenAI, or other startups as places where history can be written.”

At Google, Dean is one of the few executives who dare to publicly criticize the Trump administration. Earlier this year, he fiercely opposed the Pentagon's decision to list Anthropic as a supply chain risk, warning that the move could damage the overall interests of the US AI industry. From a technical perspective, he also built the computing power infrastructure and neural network system that established Google's leading position in modern AI.

Hassabis co-founded DeepMind in 2010 and sold it to Google four years later. He will be re-elected as chairman of the department and assume the position of Alphabet's newly established chief scientist, focusing on long-term research and the social impact of general artificial intelligence (AGI), while also planning to spend more time at Isomorphic Labs, an AI drug discovery company incubated from DeepMind.

Koray Kavukcuoglu (Koray Kavukcuoglu), DeepMind's technical director and Alphabet's chief AI architect, will take over the department's daily management and next-generation Gemini model development. According to people close to the DeepMind team, over the past year, Kavkuolu has gradually taken on more responsibilities originally belonging to Hassabis, including guiding model development and hosting major Gemini releases; Hassabis has spent more time away from the lab to focus on the long-term impact of regulation and advanced AI.

Maximum internal friction point: distribution of computing power

Google's investment in global data centers, chips, and related infrastructure is almost unmatched, yet computing power is still in short supply. Each TPU is assigned to train models, support Google products, or fulfill cloud customer contracts, representing a choice between multiple priorities.

People familiar with the matter said that when Google announced large-scale infrastructure promises to competitive laboratories such as Anthropic (whose model directly competes with Gemini), researchers were particularly dissatisfied with the acquisition of computing power. One of the sources said that Google has long-term forecasts for demand for products and services such as research, model training, search, and Gemini, as well as cloud customer cooperation. These demands will be modeled several years in advance, but if the growth rate of a product exceeds expectations or priorities change, the allocation of computing power may also be adjusted in the short term.

In the last two earnings calls, Pichay emphasized that even if demand from cloud customers grows, Google will give priority to safeguarding DeepMind's computing power needs. When asked about TPU allocations in July of this year, he said his “top priority” was to ensure computing power to stay ahead of the AGI frontier competition, and called this work “the foundation of all of our business.”

At the same time, he added that Google will balance these needs with the computing power required for consumer products and AI models, and mitigate external demand by directly deploying TPU in third-party data centers.

Dan Niles (Dan Niles), a Google shareholder and founder of Niles Investment Management, believes that the distribution of computing power is a natural contradiction. “Google has all these other businesses, and they have to decide who to give the resources to,” Niles said. “There are always people who are not happy in this situation.”

DeepMind accelerates integration with cloud services

At the Davos World Economic Forum in January of this year, Hassabis discussed enterprise-level products and application scenarios with Google Cloud CEO Thomas Kurian (Thomas Kurian) on the same stage. According to a person familiar with the operation of Google Cloud, this scene is quite rare — the two major organizations have been separated for a long time in history, and Hassabis is even more estranged from the rest of the company's businesses.

The source believes that the appearance on the same stage marks the beginning of more active intervention in AI enterprise use cases (especially in the fields of programming, customer service, etc.), and also reflects Google's overall strategy to accelerate the deep integration of R&D and commercial operations in the face of competitive pressure from OpenAI and AnthroPic.

Over the next few months, the Google model experienced many twists and turns, the most prominent being the delay in the release of the latest flagship model, Gemini 3.5 Pro. Meanwhile, the cloud business under Curian's management recorded record explosive growth. Curian is a former Oracle executive. Since 2019, he has been in charge of Google Cloud and has built an active corporate sales organization within a company that specializes in the consumer internet.

Google Cloud independently designs AI chips, operates a global data center network, and sells models, databases, security software, and AI agent development tools.

This strategy enables Google to profit from AI requirements in multiple dimensions: selling infrastructure to labs such as OpenAI and Anthropic, providing Gemini to companies, and integrating AI into search, YouTube, Workspace, and other proprietary products.

The generative AI boom has continued for nearly four years, and now the new question that has arisen is: is it necessary for Google to develop the top AI models on its own? Or would it be wiser to let other companies bear the high costs and reap the profits themselves?

Kavkuolu told CNBC at the Google Developers Conference in May that while promoting the cutting edge, the company is also focusing on improving efficiency. He said that while providing cutting-edge capabilities, the Flash model runs four times faster and more efficiently than similar models, enabling Google to expand advanced AI to enterprise and consumer services.

Similar to Tunguz, Niles believes that most business scenarios don't require the strongest model. “The current model can meet 90% of regular needs,” he said. “We don't need a Ferrari for this; Ford is enough.” However, for scientists and researchers working to achieve the next Transformer-level breakthrough, “enough” is often far from enough.