Eddie Wu’s Ledger

In the era of big-model techno-romanticism, parameter counts and leaderboard rankings were once the only yardsticks. But in Eddie Wu’s ledger, the core metric for AI competitiveness was undergoing a fundamental shift: the capacity to produce and consume tokens. Even the naming of the newly established ATH business group carried the hard-nosed pragmatism of an actuary.

In an internal memo, Wu distilled the new group’s mission into three links:

“Create tokens, deliver tokens, apply tokens”

Under this structure, Tongyi Lab was no longer an independent ivory tower of research; it was positioned instead as a token “power plant.” The MaaS business line became the distribution “grid,” while Wukong and Qwen were the “appliances” plugged directly into end users. The intent behind this vertical integration was to put an end to Alibaba’s internal “warlordism” and fragmentation.

Previously, Alibaba’s AI footprint showed a kind of structural misalignment. Between the Tongyi Lab, which was responsible for foundational R&D, and the business lines responsible for front-end applications, there was no unified collaboration mechanism. During Lin Junyang’s tenure, the Qwen team functioned more like a standalone “armory”: what weapons the frontline applications could get often depended on the lab’s technical preferences.

When the applause for open-sourcing could not be directly converted into predictable profits on the financial statements, Eddie Wu chose to turn the blade inward. The “Token Hub” logic he proposed was, in essence, about pushing AI R&D from a lab stage onto an industrialized production line. Under the ATH architecture, tokens are treated as the new era’s “water, electricity, and gas”—even as the “piece-rate wage” of the compute age.

Every workflow invocation and every instruction execution is folded into a commercially closed loop that is billable and traceable.

As of fiscal 2026 Q2, Alibaba Cloud’s AI-related revenue had delivered triple-digit growth for nine consecutive quarters. But faced with the sheer scale of ByteDance’s Seed team and the efficiency storm unleashed by DeepSeek, Alibaba must prove that this “full-stack integration” kind of industrial efficiency can outperform the explosive power of individual genius.

The Shadow of ByteDance’s 2,000-Person Seed Team

Inside Alibaba, the Qwen team was once jokingly dubbed “MI6’s Q Branch.” The metaphor carried a certain tech-elitist pride: an extremely lean team that, powered by bursts of genius, exported open-source models to the world. But as the large-model war moved in early 2026 into the deep waters of transitioning from “experimentation” to “engineering,” this “secret-agent” model ran into structural pressure from ByteDance.

Two thousand people—this was the size of ByteDance’s Seed team responsible for foundation-model training, according to news reports.

By contrast, the headcount Alibaba’s Qwen put into each R&D direction was only a fraction of its peers’. 

As early as 2024, Zhou Chang, the former core lead of Qwen, left to join ByteDance's Seed team, taking more than a dozen key members with him. By March 2026, the trend had not let up. In the same week Lin publicly announced his departure, Yu Bowen, the head of post-training for Qwen, confirmed he was joining ByteDance. He took up the role of post-training lead for the Seed team’s “Multimodal Interaction and World Models” unit.

ByteDance’s logic is simple and blunt: turn the model into a fully productized offering. As of December 2025, ByteDance’s Doubao had already taken the lead by surpassing 100 million daily active users (DAU), and it had rolled out more than 20 AI apps across various niche tracks. While Alibaba’s technical prodigies were still agonizing over how to preserve the lab’s “independence,” ByteDance, backed by a massive engineering organization, had already converted the Seed family of models into an all-scenario traffic-harvesting machine spanning social, video, and workplace use cases.

This gap in scale showed up directly in the pace of infrastructure buildout. Several insiders from within Qwen had previously admitted to the media that the team had long lacked sufficient support in compute resources and Infra (infrastructure) construction. Even as Lin Junyang endorsed Qwen on X and won applause in the international open-source community, the team behind him was grappling with a kind of structural incompatibility.

The management wanted to reduce reliance on any single individual by splitting up teams and bringing in parachuted hires. The technical prodigies, however, believed this was strangling sparks of innovation. For Alibaba, the departures of Lin Junyang and Yu Bowen were not merely the loss of two P10-level talents—they also exposed a harsh reality.

As the AI race moved into the industrialized “second half” of head-to-head competition, it had already become difficult to defend a moat in the face of a standardized formation of 2,000 people on the strength of a few geniuses’ bursts alone.

The Zhou Hao Moment

In January 2026, while Lin was still sprinting full tilt to expand Qwen’s open-source footprint, a low-profile parachuted hire walked into Alibaba’s campus. His name was Zhou Hao. His first stop was not the Tongyi Lab, but a unit affiliated with Quark. Zhou earned his undergraduate degree at the University of Science and Technology of China, and he had previously been a Senior Director Research Scientist at Google DeepMind.

Zhou’s arrival signaled a foundational pivot in Qwen’s technical trajectory. In the Qwen universe led by Lin Junyang, the core pursuit was “full stack, open source, influence.” Through rapid iteration and coverage across a wide range of model sizes, Qwen established a stronghold for China-made models on leaderboards such as LMSYS.

But at its core, this path was still probabilistic prediction built on “massive data + massive compute.” When DeepSeek demonstrated through reinforcement learning (RL) that “reasoning cost” could beat “parameter piling,” Alibaba’s locus of technical anxiety shifted. What Zhou Hao brought to Alibaba was DeepMind-style “slow thinking” logic.

As the head of Gemini’s reinforcement learning and self-improvement team, Zhou Hao’s core contribution at Google was teaching AI to “think it through before it speaks.” The Gemini DeepThink mode he helped develop uses a dedicated reward function to make the model internalize constraints on factual accuracy every time it generates a token.

With Lin and Yu leaving, Zhou moved from behind the scenes to center stage, taking over the core responsibilities of post-training. Alibaba’s intent was unmistakable: Qwen needed to evolve from a language model into an agent with autonomous reasoning capabilities.

In March 2026, while restructuring ATH, Alibaba published a technical blog post on Qwen3-Max-Thinking. The model’s gains in logical reasoning and complex instruction following were positioned directly against GPT-5.4 and DeepSeek V3.2. For Alibaba, this was not just a смена of talent—it was a strategic adjustment for survival.

In the new race where algorithmic efficiency is overtaking sheer compute scale, only a scientist’s meticulous “cost accounting” can fill the combat-power gap left by departed geniuses.