Diverging Paths for AI Talent at Tech Giants

Author｜Dingshan

Editor｜Xiaobai

Illustration｜AI-Generated

Produced by｜Qiangdiao Next

On August 5 (local time), Google announced its most significant AI power shift since merging Google Brain and DeepMind in 2023. Chief scientist Jeff Dean, after nearly 27 years at Google, departed alongside Sanjay Ghemawat, Oriol Vinyals, and Quoc Le to found Discovery Loop. The new venture, registered as a Delaware public benefit corporation, aims to automate complete experimental cycles in machine learning, science, and engineering using AI. Alphabet serves as a founding investor and will provide long-term cloud and computing support.

On the same day, Demis Hassabis relinquished day-to-day operational control of Google DeepMind, assuming roles as department chairman and Alphabet chief scientist while continuing to lead AI pharmaceutical firm Isomorphic Labs. Koray Kavukcuoglu, former DeepMind CTO and Google chief AI architect, was promoted to senior vice president, reporting directly to Sundar Pichai, and now oversees the Gemini model, frontier research, Gemini applications, and developer teams.

Capital markets reacted swiftly with negativity. Alphabet's Class A shares fell nearly 5.5% intraday, closing down 4.0%. However, this adjustment represents more than a talent exodus. The quartet spans large-scale computing systems, foundational models, and inference research. A more critical shift occurred in power structures: Google is ending scientist co-governance, transforming AI into a CEO-directed business line centered on model releases and product revenue. This aligns, to a certain extent, with AI organizational restructuring directions at China's Alibaba, Baidu, and other major firms over the past period.

───

01

　Taking Half of Google's AI Kingdom　■

Jeff Dean is often referred to as Google's 30th employee. He and Sanjay Ghemawat co-built search, indexing, and advertising systems before designing critical infrastructure like Google File System, MapReduce, Bigtable, and Spanner, establishing Google's global data processing foundation.

Dean co-founded Google Brain in 2011, advancing machine learning systems like DistBelief, TensorFlow, and Pathways while early advocating for custom AI acceleration chips. From distributed training frameworks to research organizations, he participated in nearly every layer of Google's modern AI capabilities.

Ghemawat, one of Google's few Senior Fellows, served as Dean's decades-long systems partner. Vinyals and Quoc Le complemented the model and inference layers. The former co-invented the seq2seq model, led AlphaStar, and serves as Gemini's technical lead; the latter, an early Google Brain core member, participated in and guided mathematical reasoning projects like AlphaGeometry.

The quartet's combined strength lies not in a single star project but in simultaneously designing models, training systems, computing resource allocation, and experimental workflows. This explains Discovery Loop's focus on "automated experimental cycles": enabling systems to propose, implement, run, and evaluate experiments while iterating based on feedback.

Machine learning experiments inherently involve code, models, and computing resources, making them easiest to close-loop first before expanding to chip design, drug development, and clean energy. The new venture has not disclosed funding amounts or valuations, nor proven that complete experimental cycles can overcome real-world laboratory, data quality, and regulatory constraints. It resembles a technical option backed by founders' credentials. However, Google has lost not ordinary management roles but a team capable of simultaneously developing algorithms and hyper-scale systems.

───

02

　Research Institutes Step Back, CEOs Step Forward　■

In 2023, Pichai merged Google Brain and DeepMind, with Hassabis as CEO and Dean as chief scientist, attempting to resolve long-standing research resource fragmentation. Three years later, Google's challenge shifted from "how to merge research forces" to "who takes responsibility for product outcomes." The new organizational arrangement provides an answer: Kavukcuoglu, eschewing the CEO title, reports directly to Pichai as senior vice president. Research, models, applications, and commercial interfaces now fall under a single accountability chain.

Hassabis's transition is not a conventional "resignation." He remains involved in model and research strategy while continuing as Isomorphic Labs CEO, but Reuters cited Alphabet stating he has few direct subordinates in his new role. Title elevation coincides with operational authority contraction. The 2023 Dean-Hassabis dual scientific core thus ends, replaced by a Pichai-Kavukcuoglu executive axis. Scientific judgment shifts to the strategic layer, while release cadence, costs, and product adoption fall to professional operational systems. This reflects Google's response to decade-long lessons. The Transformer architecture originated at Google, yet multiple core authors later departed.

While Google advanced foundational research, OpenAI commercialized large models first. As the company invests hundreds of billions in AI infrastructure, technology must promptly translate into models, traffic, and cloud revenue, making it difficult for research departments to retain full decision-making authority. Similar power shifts occurred at Chinese tech giants, with more frequent moves.

From March to June 2026, Alibaba established Alibaba Token Hub, upgraded its Tongyi large model division, and merged it with the Future Life Lab into Token Foundry, with model and application teams now under group CEO Wu Yongming's direct supervision. Former Tongyi head Zhou Jingren shifted to group chief scientist, leading the AI Future Research Institute.

This mirrors Google: long-term research space persists, but model, application, and commercial delivery fall under CEO responsibility. The difference lies in Alibaba's primary internal reorganization, while Google also offloads high-risk research to external entities via Discovery Loop and Isomorphic Labs.

Alibaba announced plans to invest at least 380 billion yuan ($53 billion) over three years in cloud and AI infrastructure, requiring model R&D to simultaneously justify reasoning costs, token consumption, cloud revenue, and user access points. Baidu's adjustments were more direct. In November 2025, Baidu established foundational model R&D and applied model R&D departments, responsible for general large models and business-specific precision models, respectively, both reporting directly to Robin Li.

Wang Haifeng retained CTO, Technical Committee Chair, and Baidu Research Institute President roles, but core large model R&D now flows directly to the CEO via new departments, dispersing permissions previously concentrated in the technical mid-platform.

A week before these adjustments, Baidu disclosed Q3 2025 AI revenue of approximately 10 billion yuan, up over 50% YoY. However, total company revenue fell 7%, with online marketing revenue dropping 18%. AI began shouldering growth responsibilities but couldn't offset legacy business declines, making coordination speed between foundational models and applications a CEO-level operational issue.

From Google to Alibaba and Baidu, a common trend emerges: as large models enter capital-intensive delivery phases, research departments no longer inherently possess complete authority over technical routes to product launches. CEOs must align budgets, iteration cadences, product access points, and revenue on a single balance sheet, bringing those capable of timely model-to-product conversion closer to daily operational centers.

───

03

　Revenue Grows, Yet Delivery Pressures Intensify　■

Alphabet reported Q2 revenue of $119.8 billion, up 24% YoY. Google Cloud revenue reached $24.8 billion, up 82% YoY, with operating profit of $8.8 billion—more than triple the prior year. Gemini apps hit 950 million monthly active users, with model APIs processing approximately 220 billion tokens per minute. By commercial metrics, Google AI shows no signs of stalling. Conversely, Gemini 3.5 Pro, initially planned for June release, remained in partner testing as of late July. Pichai acknowledged Google's need to improve in programming and agentic programming.

Prior to adjustments, Gemini technical co-leads Noam Shazeer had already joined OpenAI, while AlphaFold co-creator John Jumper moved to Anthropic. Dean, Vinyals, and Shazeer—three Gemini technical co-leads—have all left their original roles. Financial pressures also raise delivery demands. Alphabet's Q2 capital expenditures hit $44.9 billion against operating cash flow of $39.1 billion, resulting in negative free cash flow of $5.9 billion for the quarter.

The company raised its 2026 capital expenditure guidance to $195-205 billion; while Cloud's backlog reached $514 billion, faster alignment of computing resources, models, and customer demand remains essential to prove asset returns. The stock decline reflects not Google's lost AI capabilities but the compounding uncertainties of flagship model delays, concentrated researcher departures, and rising capital expenditures.

Google's complete chip, cloud, model, application, and traffic ecosystem establishes a performance floor; sustained delivery of competitive models post-personnel changes will determine its ceiling.

───

04

　Releasing Scientists, Retaining Relationships　■

Discovery Loop's relationship with Google distinguishes this departure from typical talent defections. Alphabet participated in the initial funding round, and Google will provide cloud and computing resources while collaborating on machine learning systems and infrastructure research. Though Google lost direct management control over the quartet, it retains equity stakes, computing revenue streams, research partnerships, and potential technology repatriation. This represents a lighter organizational arrangement: high-risk, long-cycle exploration gains founder equity and greater autonomy outside the company. Alphabet avoids full organizational costs while sharing upside gains. Hassabis's continued leadership of Isomorphic Labs reflects a similar approach.

The challenge lies in Discovery Loop's overlapping direction with Google's existing Co-Scientist and AlphaEvolve initiatives. The departure of four core figures to start their own venture suggests the next phase of AI scientific discovery may require not just corporate computing power but also more centralized control, higher equity incentives, and research cadences unconstrained by search, cloud, and application release schedules. This poses a long-term risk to Google: while investments and cloud partnerships can cushion a single amicable departure, they cannot be infinitely replicated. If those most capable of defining next-generation technologies increasingly seek control outside the company, Google's internal operations may gradually become machines for scaling mature research rather than preferred organizations for generating new directions.

The success of this adjustment hinges on whether Gemini 3.5 Pro can add programming capabilities and launch, whether Gemini 4 establishes a stable iteration cadence, whether Google retains the next layer of research leaders, and whether Discovery Loop demonstrates repeatable experimental efficiency gains. Google still possesses the AI industry's deepest technical assets, with revenue continuing to grow. This talent earthquake alters the old contract (xìyuē, "social contract") of research autonomy: Google reclaims product authority at its core, pushes scientists to organizational peripheries, and reconnects via capital and cloud. This may accelerate Gemini's progress but could also result in the next Transformer-level breakthrough emerging first within Google's investment portfolio rather than internally. Data in this article comes from public sources and does not constitute investment advice.