⚡ Key Highlights
With a major wave of core developers departing, including Jeff Dean, market concerns are escalating regarding Google's AI model development competitiveness and the performance of its frontier models.
Despite explosive growth in Google Cloud and a surge in Gemini users, Google's technological leadership is being challenged by delays in releasing high-performance models for developers.
To overcome talent leakage and performance stagnation, Google is pursuing a market turnaround through a full-stack AI strategy that combines its proprietary infrastructure with a vast service ecosystem.
Lately, Google's situation appears unusual. Legendary developers are departing in succession, while users continue to voice concerns that model performance is not what it used to be. Some are even questioning whether Google might be heading toward a decline. Yet, looking at its stock price and financial position, that hardly seems to be the case. What is really going on at Google to prompt talk of a downturn? In today's OhGraph, we look into Google's current standing through various data points and charts.
Has an 'Exodus' Begun? Wave of Resignations by Legendary Developers
On August 5, the global developer community was shaken by the news of Google's legendary developer Jeff Dean's resignation. Dean joined Google in 1999 as employee number 30 and became an iconic figure at the company. Under Google's internal leveling system, the standard ceiling is Level 10 (L10), but Dean was famously promoted to L11.
Jeff Dean is the pivotal architect who built Google's modern engineering foundation. Early iterations of Google's primary services—including Google Search, Google Translate, and Google Ads—were born from his hands. He also spearheaded critical internal initiatives such as the machine learning framework TensorFlow, the distributed storage system Bigtable, and the large-scale data processing system MapReduce.
In 2011, he joined Google's secretive research facility, Google X, to conduct deep learning research. Recognizing that deep learning could be a true game-changer, he recruited top AI talent worldwide and founded Google Brain, Google's early AI research team.
Furthermore, Google's custom AI accelerator, the TPU, originated from Dean's conceptual work. Dean is well known for "napkin calculations"—approximating key figures on a napkin rather than running exhaustive simulations. In 2013, when Google's deep learning-based speech recognition made significant strides, Dean began contemplating the next phase: If voice recognition improved drastically and saw widespread use, could Google's existing infrastructure handle the load?
With that question in mind, Dean carried out a napkin calculation. Assuming every Google user utilized voice search for just three minutes a day, he calculated that Google would need to double its entire existing fleet of CPUs. Realizing that general-purpose CPUs alone could not keep up with the computational demands of the AI era, Google developed custom hardware—which became today's TPU. Achieving feats of this magnitude is why he earned the L11 title.
This L11 rank, known internally as Senior Fellow, has only been granted to two people in Google's history. The other recipient is Sanjay Ghemawat, who joined around the same time and is often described as Dean's intellectual counterpart. In Silicon Valley, the two were legendary for pair programming so seamlessly that they were said to share a single brain. These two individuals, who guarded Google's engineering core with unmatched achievements and expertise, have now left the company together.
In addition, Oriol Vinyals, who led the development of Gemini, and Quoc Le, a founding member of Google Brain, also parted ways with Google, bringing the total to four prominent departures. Together, they teamed up to launch a new startup called Discovery Loop.
Demis Hassabis, the long-time leader of DeepMind, was also reported to have expressed a desire to step down. However, Google actively persuaded him to stay, fearing that his departure right after Jeff Dean's could trigger a sharp sell-off in the company's stock. In the end, Hassabis did not leave entirely; he stepped back from frontline operations to serve as Chair of DeepMind and transitioned to Chief Science Officer at Google's parent company, Alphabet.
These are not the only high-profile figures leaving. Prominent developers who once led Google's AI efforts are departing in rapid succession.
While Hassabis transitioned to a higher-level advisory role and the quartet founded Discovery Loop, John Jumper—who co-developed AlphaFold with Hassabis and won the Nobel Prize in Chemistry—moved to Anthropic. Furthermore, Noam Shazeer, an author of the landmark paper "Attention Is All You Need," which introduced the Transformer architecture foundational to modern AI, made his second exit from Google.
Shazeer had previously left Google after publishing the paper to co-found Character.ai, which became a major success. Google subsequently acquired Character.ai for $2.7 billion in an effort to bring Shazeer and his team back. However, within less than two years of returning, Shazeer departed once again, this time joining OpenAI.
Objectively speaking, talent mobility between AI firms is not uncommon in this rapidly evolving market. However, this particular exodus stands out because heavyweight leaders who shaped Google's AI identity left simultaneously. More importantly, data indicates that Google DeepMind has been steadily experiencing a talent drain over the past few years.
According to data released last year by venture capital firm SignalFire tracking talent flow between DeepMind and Anthropic, for every single person who moved from Anthropic to DeepMind, ten times as many engineers and researchers moved from DeepMind to Anthropic. Aside from Hugging Face, DeepMind has seen a net outflow of talent to competitors such as Cohere and OpenAI.
Strong Financials, but Lagging AI Performance?
Despite the talent drain, Google's balance sheet tells a different story. In its second-quarter earnings report released in late July, Google comfortably beat market forecasts, driven by strong demand for AI services. Total revenue, operating profit, and operating margin all increased compared to the same period last year.
The explosive growth in its cloud business was particularly notable. In the first quarter of 2026, Google Cloud revenue surged 63% year-over-year. That momentum carried into the second quarter, where revenue jumped 82% year-over-year to $24.8 billion. Google Cloud's market share in the global cloud industry has also shown steady gains.
Traffic metrics also appear favorable for Google. Looking at generative AI web traffic, a significant number of users are adopting Google Gemini.
Looking at the web traffic share among generative AI chatbot platforms from 2024 to 2026, ChatGPT's share has gradually contracted, reaching around 53% in May 2026. In contrast, Gemini's traffic share expanded significantly, rising to 28%. Based on these metrics, Google Gemini has made the most notable dent in ChatGPT's consumer-facing monopoly.
Monthly active users (MAUs) for the Gemini app have also been climbing rapidly. From 650 million in October of last year, MAUs grew to approximately 750 million in February this year. More recently, Gemini surpassed 1 billion monthly active users, making it the fastest-growing service in Google's history.
While Google is performing well in the consumer market, the developer landscape presents a different picture. In high-performance frontier models designed to satisfy developers, Google has struggled to assert dominance. Google recently unveiled its Gemini 3.6 model with extensive marketing, but its benchmark specifications still lagged behind rival frontier models. Although Gemini 3.7 was released just three weeks later, expectations have largely centered on Gemini 3.5 Pro. At Google's annual developer conference last May, Google CEO Sundar Pichai asked developers for their patience.
"We are also very excited about 3.5 Pro. We are currently using it internally and have seen significant improvements. Please bear with us until next month, and we will make it available to you."
However, the promised June timeline has passed, as have July and August, yet Gemini 3.5 Pro remains unreleased. According to some reports, the launch was delayed because internal testing revealed unsatisfactory performance gains. While Anthropic has introduced higher-performing models like Opus 5 and Fable 5, and OpenAI has unveiled GPT-5.6 Sol, internal feedback suggests that Gemini's coding performance improvements have progressed too slowly.
Some reports even suggest that Google scrapped the model under development and restarted pre-training from scratch. If these rumors hold true, an official release will take considerable time, during which Google's ranking in the frontier model race could slip further.
A comparison of current frontier model performance illustrates the landscape. The companies behind the top 10 models include U.S. firms Anthropic, OpenAI, Meta, and xAI, alongside Chinese companies Moonshot and Alibaba. The U.S. players are core members of the AI Big Tech cohort known as MANGOS. Excluding Nvidia, whose primary business is not foundational model development, Google stands as the only MANGOS member absent from the top 10 frontier models.
Top-Tier in Research, Lagging in Commercialization? DeepMind's Strategic Shift
This does not mean Google DeepMind has been idle. DeepMind's official announcements have centered more on scientific breakthroughs than incremental gains in Gemini's conversational capabilities. DeepMind's founding mission is to build intelligence to understand the fundamental nature of the universe, and it has consistently conducted scientific research to tackle major scientific challenges using AI.
A prime example is the AlphaFold series, which solved protein folding, one of biochemistry's grand challenges. Using AlphaFold, DeepMind established Isomorphic Labs, an AI-driven drug discovery startup. DeepMind also officially unveiled AlphaGenome, an AI model for genetic variant effect prediction, in Nature this year. Beyond biology and chemistry, the lab is developing models for mathematics, earth sciences, and various scientific fields, recently announcing Co-Scientist, a general-purpose AI agent designed to assist across diverse scientific research disciplines.
DeepMind's research-first orientation has largely been shaped by Demis Hassabis. Hassabis believes AI technology should drive scientific progress and ultimately benefit humanity. As the pace of AI advancement accelerates, concerns regarding potential unintended consequences have grown. Recent public statements by Hassabis reflect deep contemplation on preparing for the post-AGI era, emphasizing model safety and its societal impact over racing to ship commercial frontier models. Consequently, those who prioritize Google's immediate competitive standing in commercial AI models view Hassabis's transition as a necessary strategic pivot.
With Hassabis stepping back from direct day-to-day management, expectations are rising that DeepMind will restructure into a more business- and product-driven organization.
Google has always excelled at foundational research.
However, it has often struggled to commercialize its own breakthroughs. The Transformer paper, "Attention Is All You Need," originated at Google, yet OpenAI was the first to successfully commercialize it on a massive scale.
DeepMind will now report directly to Google CEO Sundar Pichai. The organization is expected to transition from an autonomous research lab into one directly accountable for revenue and commercial performance. How, then, does Google intend to monetize AI moving forward?
Some industry experts argue that Google does not necessarily need to hold the number-one spot in the standalone frontier model race. By focusing on its existing strengths, revenue will follow naturally. Google controls the world's most widely distributed mobile operating system, web browser, email platform, and map service. By leveraging user data across these services and integrating AI capabilities directly into daily consumer workflows, Google may not need to rely solely on having the absolute highest-ranking benchmark model.
Data from South Korea already highlights this advantage: the Google app recently surpassed the Naver app in monthly active users for the first time in history.
Google's monthly active users reached 47.02 million last month, surging from the prior year and surpassing Naver's figure for the first time since records began.
At the same time, Google is not pulling out of the frontier model race altogether. Generating significant enterprise revenue from model APIs requires attracting developers who consume large volumes of tokens. On July 21, Google announced that training had commenced for its next-generation model, Gemini 4, describing it as the most ambitious training run in Google's history and raising industry expectations.
For enterprise clients, Google plans to leverage its full-stack capabilities. Google is one of the few companies capable of offering a vertically integrated AI stack—spanning hardware infrastructure, foundational AI models, agent platforms, software services, and hardware devices. Google plans to utilize this end-to-end control to embed security across every layer from the design stage, enabling enterprises to deploy AI services securely.
Although Google faces talent departures and competitive pressure in frontier model benchmarks, it still commands the world's most widely used consumer platforms. If it successfully embeds AI across this vast ecosystem, it can maintain a formidable presence in the evolving AI landscape. How Google ultimately navigates this AI competition remains to be seen. That concludes today's OhGraph. Thank you for reading.
References
- Jeff Dean (@JeffDean) | X
- Sundar Pichai (@sundarpichai) | X
- TPU Training Day for I/O '26 | YouTube
- Hinton et al., "Deep Neural Networks for Acoustic Modeling in Speech Recognition," 2012
- Vaswani et al., "Attention Is All You Need," 2017
- The future of intelligence | Demis Hassabis (Co-founder and CEO of DeepMind) | YouTube
- Alphabet Q1 and Q2 2026 Earnings Releases
- Google I/O '26 Keynote | YouTube
- AI Model & API Providers Analysis | Artificial Analysis
- The State of Tech Talent Report 2025 | SignalFire
- The 2026 Generative AI Landscape Report | Similarweb
- Avsec, Ž et al., "Advancing regulatory variant effect prediction with AlphaGenome," 2026
- Gottweis et al., "Accelerating Scientific Discovery with Co-Scientist," 2026
- Demis Hassabis, Nobel Prize in Chemistry 2024: Official interview | YouTube
Written by
: An Hyemin
Designed by
: Ahn Jun-seok
Intern
: Shin Yeon-seong
※
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.