ByteDance is entering a critical new phase where its artificial intelligence (AI) research must prove it can generate revenue, not just academic breakthroughs. Gu Quanquan, a leading researcher who helped build ByteDance's Seed pre-training team, recently announced his departure from the company. His exit comes precisely as Doubao, ByteDance's flagship AI chatbot, begins charging users in June, marking a fundamental shift in how the company measures success.

For the past two years, Doubao has demonstrated that ByteDance possesses the rare ability to bring AI products to hundreds of millions of users. But now the company faces a harder question: how many of those users will actually pay for AI services? This transition from free exploration to paid products explains why Gu Quanquan's departure matters beyond typical personnel news. His exit reflects deeper organizational changes as ByteDance prioritizes commercial viability over pure research ambition.

Who Was Gu Quanquan and What Did He Accomplish?

Gu Quanquan brought an unusual combination of expertise to ByteDance. He earned his undergraduate and master's degrees from Tsinghua University's Department of Automation, then completed a Ph.D. in Computer Science from the University of Illinois at Urbana-Champaign in 2014. After postdoctoral work at Princeton University and teaching positions at the University of Virginia and UCLA, he joined ByteDance's Seed team, where he worked across two major research areas.

His work spanned AI for Science (AI4S) and large language model (LLM) pre-training, a rare combination in large technology companies. Most researchers focus on either a specific vertical field or foundational model training, but Gu Quanquan achieved significant results in both directions. In the AI4S domain, he led development of three major projects:

• SeedFold: A biomolecular structure prediction model that outperformed Google's AlphaFold3 in multiple protein-related tasks on FoldBench, a standard industry benchmark.
• SeedProteo: A protein design model capable of creating entirely new protein molecules from scratch, representing a more commercially challenging but potentially valuable application than structure prediction.
• DPLM Series: A protein language model that teaches AI to understand and generate protein sequences similarly to how large language models process human language.

These achievements carried significant academic weight. After the SeedFold paper's publication, multiple research institutions cited and reproduced the work. SeedProteo's performance in protein design tasks was considered among the strongest in the industry at that time.

Why Did ByteDance's Strategic Shift Force This Departure?

In early 2025, Gu Quanquan took on an additional responsibility that proved pivotal. He joined ByteDance's LLM pre-training work, establishing an optimization and expansion team focused on building what he called a "highly scalable pre-training technology stack." This team successfully supported the training of Seed 2.0 and subsequent frontier-scale models. The shift seemed sudden at the time, but it reflected a broader industry realization: pre-training capability depends not just on computing power but on engineering and optimization expertise.

The timing of Gu Quanquan's departure is not coincidental. A few days before he announced his exit, news emerged that ByteDance's AI4S team was undergoing organizational restructuring. Yang Zhenyuan took over leadership of the AI4S division, and several core team members, including Gu Quanquan, left or prepared to leave to start their own ventures. ByteDance clarified that it was not spinning off AI4S as an independent business, meaning the team would remain part of Seed but follow the division's overall strategic priorities.

How Does AI4S Differ From Commercial AI Products?

The fundamental challenge with AI4S research is that its value is difficult to measure using traditional product metrics. Developing a protein structure prediction model that surpasses AlphaFold3 represents a major academic breakthrough, but transforming that breakthrough into commercial revenue remains a long and uncertain journey. Additionally, achievements in AI4S tend to be closely tied to the individual researchers who created them.

This creates a critical difference between AI4S and products like Doubao. Doubao's capabilities are embedded in ByteDance's broader systems: the model platform, training infrastructure, inference architecture, and product distribution channels all belong to the company. When researchers leave, the system remains intact. AI4S, by contrast, depends heavily on the people who developed it. When key scientists depart, the institutional knowledge and momentum often leave with them.

Steps ByteDance Is Taking to Prioritize Commercial AI

ByteDance's organizational changes reflect a deliberate strategy to align research with revenue generation. The company has taken several concrete steps to reshape how it evaluates and invests in AI research:

• Charging for Doubao: Beginning in June, ByteDance will charge users for access to Doubao, shifting from a free exploration model to a revenue-generating product that must prove user demand.
• Restructuring AI4S Leadership: Yang Zhenyuan's appointment to lead AI4S signals that ByteDance is maintaining investment in this direction but with different goals and methods than before.
• Emphasizing Pre-Training Systems: ByteDance is prioritizing the engineering systems that enable continuous model iteration over individual breakthrough discoveries, as evidenced by Gu Quanquan's work building scalable pre-training infrastructure.

The departure of Gu Quanquan and other AI4S researchers represents a natural consequence of this transition. Research that produces papers and academic citations but lacks clear short-term product value faces reduced priority in a company that has entered what analysts describe as its "second stage," where AI has become the main engine driving ByteDance's growth.

ByteDance's challenge now is to prove that it can convert its massive user base and research capabilities into sustainable AI revenue. The company has demonstrated it can build AI products that reach hundreds of millions of people. Whether those users will pay for premium AI services remains the critical unanswered question that will define ByteDance's AI strategy for the next phase of its evolution.