01 logo

ByteDance Just Bet $29.6 Billion on One Idea. There’s No Going Back.

Zhang Yiming folded Feishu into Doubao, cut reporting lines, and went to war with Tencent and Alibaba over the AI office. Here’s what the gamble actually means.

By JinPublished 8 days ago • 6 min read

Zhang Yiming's No-Return Bet: Feishu Moves Under Doubao

On the morning of July 30, 2026, an internal email circulated among about 20,000 ByteDance employees. It had three paragraphs. Feishu's product team would merge into Doubao to form the Doubao product team. Doubao head Zhao Qi would lead it. Feishu CEO Xie Xin would report to Zhao Qi. The same day, Feishu's GTM team, covering marketing, sales, and customer service, was folded into Volcano Engine under Tan Dai. Xie Xin's reporting line went from Liang Rubo plus one to Liang Rubo minus three.

Many people felt sorry for Xie Xin. Feishu had run for ten years. It had once been an independent business unit alongside Douyin. Now its leader had become a product manager. But the organization exists to serve the strategy. Every personnel change and structural adjustment pointed to one goal: control the super entry point of the AI era, a general Agent.

A decision made with numbers

Feishu was not merged because it had failed. By outside measures, Xie Xin's record looked strong. Feishu revenue passed 3 billion yuan in 2025. In the second quarter of 2026, revenue grew more than 100% year over year. Some industry observers expected Feishu to become the largest domestic enterprise collaboration vendor by revenue in 2026.

ByteDance made the move when Feishu had momentum. The two businesses had different cost structures. Feishu kept about 4,000 people. Industry estimates put its annual labor cost between 3.2 billion and 4 billion yuan. Xie Xin had pushed Feishu's profit target to 2030. Doubao's compute bill was larger. By June 2026, Doubao's large model handled more than 180 trillion tokens per day. A high-growth SaaS business with no near-term profit and an AI assistant that burned hundreds of millions of yuan in compute each day were both financial holes on their own. Zhang Yiming welded them together.

Three parts, one loop

The new structure was blunt. Feishu split in two. The product team went to Doubao and would put AI into documents, spreadsheets, meetings, and other office settings. The GTM team went to Volcano Engine under a new To B sales group called the Creativity Service Platform. That group would sell ByteDance's MaaS and SaaS products and handle customer service.

On August 25, 2026, the integrated Doubao Work launched. It is an Agent product for productivity. It can break down tasks around a user's goal, call tools, and push complex workflows forward. It connects deeply with Feishu. After a user signs in with a Feishu account, Doubao Work inherits the enterprise knowledge and work context allowed by that account. It can call chat records, documents, meeting notes, and calendars stored in Feishu to finish tasks. The output returns to Feishu and becomes reusable enterprise knowledge.

The loop works like this. Doubao supplies intelligence and Agent capability. Volcano Engine supplies models, compute, and enterprise services. Feishu supplies the collaboration environment and the work site. Liang Rubo told employees that climbing the AI peak was ByteDance's most important task. Volcano Engine's MaaS business was becoming ByteDance's basic business. By June 2026, the Doubao app had 382 million monthly active users. On a July average consumption basis, ByteDance's large model ARR reached $4 billion.

Rivals moved first

ByteDance was not the first company to see this direction. It was the last of the three big Chinese internet companies to act.

Tencent's WorkBuddy got out first. On January 17, 2026, a team of about ten people built an initial prototype in two days. By June, WorkBuddy had 20.97 million visits in a single month. It ranked first in China's PC AI native office agent market, with more visits than the second and third products combined. Pony Ma called it the most widely used efficiency AI agent in the domestic market during an earnings call. On September 2, WorkBuddy launched an open platform with more than 100 partners. Liu Yi, a Tencent Cloud vice president, said the goal was to make WorkBuddy the operating system for the Agent era.

Alibaba moved quickly too. On July 27, 2026, it merged three agent products into Qwen Office and put DingTalk's new CEO in charge. Within one month, Qwen Office had more than 30 million registered users. More than half were enterprise employees. DingTalk's 26 million enterprise organizations gave Qwen Office a ready customer base. Guo Ziya, a quality engineer at Changan Automobile, had worked 13 years without writing a line of code. He now uses Qwen Office to check certificates. The time for one vehicle dropped from more than ten minutes to one or two minutes.

The three giants consolidated their AI office products within ten days. On July 20, Tencent merged QClaw into WorkBuddy. On July 27, Alibaba completed the Qwen Office integration. On July 30, ByteDance sent the internal email about Feishu and Doubao. The competition had moved from whose model was smarter to whose product could sit inside a user's workflow. Tencent's WorkBuddy had 20.97 million monthly visits. Alibaba's Qwen Office had 30 million users. ByteDance's Doubao Work first appeared in public in September 2026, after the merger.

$29.6 billion and one sentence

If the organizational merger was the software side of Zhang Yiming's strategy, the money was the hardware side. ByteDance was moving from an asset-light model built on algorithms and advertising to a heavy-asset war.

On September 3, Yicai reported that ByteDance would receive a $29.6 billion syndicated loan. The initial plan was $20 billion. Bank orders exceeded $30 billion, and the final size rose to $29.6 billion. It was the second-largest USD-denominated loan in Asia in 2026, after SoftBank's $40 billion bridge loan in March. ByteDance last entered the global loan market in 2024, raising about $10.8 billion through more than 20 banks. This time the size was nearly three times larger.

Where did the money go? Media reports said ByteDance was discussing raising its 2026 capital expenditure to as much as $70 billion, more than double the previous year. The money would go to data centers and other AI infrastructure. If economic and business conditions allowed, the company might raise it to $100 billion in 2027. About 85 billion yuan was earmarked for AI chip purchases.

In late July, at a Seed team all-hands meeting, Zhang Yiming said one sentence: ByteDance would not use distillation as a shortcut to improve AI model capability, even if Seed's models temporarily lagged major domestic competitors. He added that model work needed long-termism and delayed gratification, not other companies' output in exchange for a temporary leaderboard ranking. The company should be willing to sacrifice some short-term gains for long-term goals.

He did not explain why. He did not add a caveat. The sentence was the conclusion. Internal restrictions on API detection for open-source model distillation were tightened afterward.

That statement and the Feishu merger followed the same logic. Both pointed to one judgment: AI competition would not be decided by current product forms or short-term revenue. It would be decided by whether the underlying model could build a barrier over a long enough time. If model capability was the deciding variable, then every business had to organize around that variable. Feishu's independent BU status, Xie Xin's reporting level, and even the Feishu brand were variables that could be adjusted.

A road with no return

By putting Feishu under Doubao, Zhang Yiming bet on one judgment: the super entry point of the AI era would not be a chatbot. It would be an intelligent agent that thinks and acts for the user. Whoever controlled that agent would control the operating system of the next era. In that logic, Feishu's value was not how much money it could make on its own. Its value was the work scenarios, enterprise data, and workflows it gave the Agent.

The risk was just as clear. Tencent's WorkBuddy had spent eight months building more than 20 million monthly visits. Alibaba's Qwen Office had DingTalk's large enterprise customer base. ByteDance's Doubao Work had just finished its integration and public debut. A $29.6 billion loan, hundreds of billions of yuan in annual capital expenditure, and a full shift from light assets to heavy assets. Zhang Yiming, 43, had started the largest bet of his life. The asset-light king built on algorithms and advertising was fighting a heavy-asset war it did not know well.

He chose a road with no return.

how totech newsfact or fictionfuturethought leaders

About the Creator

Jin

Writer of reamstories

https://reamstories.com/jin

Enjoyed the story? Support the Creator.

Subscribe for free to receive all their stories in your feed. You could also become a paid subscriber, letting them know you appreciate their work.

Subscribe For Free

Reader insights

Comments

There are no comments for this story

Be the first to respond and start the conversation.

Sign in to comment
    Written by Jin