ME News

wrote a column ·

Seed gets the meat, Flow gets the broth—as ByteDance steps back from 'all-in' on a single AI

Article author: Berkshire

Source: Xin Lichang Pro

Over the past two years, ByteDance has publicly emphasized 'All in AI,' being in the top tier of models, striving for new heights, and long-term investment. But last month, Doubao entered a less familiar phase—within fewer than ten days, it aggressively pushed forward commercialization across three fronts: conversational AI, ride-hailing, and cloud storage.

On June 22, Doubao began gray-scale testing its one-click ride-hailing feature in Beijing and Hangzhou, with Cao Cao Mobility’s driver app starting to receive dispatch alerts labeled 'Doubao service orders.' Two days later, Doubao officially launched three subscription tiers: Standard at RMB 68, Enhanced at RMB 200, and Premium at RMB 500 per month. Six days after that, a clause quietly appeared in Doubao’s user agreement regarding cloud storage expansion, capping what had previously been nearly unlimited AI cloud storage at 5TB.

Doubao is ByteDance’s largest, fastest-growing, and most mature AI product in terms of user perception. After the Lunar New Year, its daily active users surpassed 200 million, making it the only AI application in China at the time to reach the 100-million-plus DAU milestone. According to mobile internet-era playbook, such a product should keep expanding scale, solidify its position as a gateway, and then let monetization—from ads, e-commerce, subscriptions, and ecosystem revenue—unfold organically.

However, around the May Day holiday, a quietly updated payment notice on Doubao’s App Store page went viral on social media. The once high-performing product first faced criticism such as 'It’s dumb—and now it’s charging?' Monthly active users declined by an unusual 6.1 million that month.

Public sentiment has already sent a signal, yet Doubao’s commercialization timeline hasn’t paused because of it.

In ByteDance’s previous AI narrative, Doubao represented user scale while Seed symbolized model capability—both placed within the same growth loop. But starting this year, priorities within that loop have shifted: multiple media outlets reported that Seedance has achieved an annualized revenue of USD 2 billion with a gross margin of approximately 70%, with enterprise clients contributing the vast majority of revenue; Claude Code reached USD 1 billion in annual recurring revenue (ARR) within six months and hit USD 2.5 billion by February this year, serving as ByteDance’s benchmark for monetization in office and coding tools. Meanwhile, despite boasting the largest user base, Doubao has struggled to convert user engagement time and willingness to pay into actual revenue.

ByteDance has never framed its AI investment narrative as one of financial constraint. Liang Rubo set 'boldly scaling new heights' as the keyword for 2026, and the company’s commitment is directly reflected in numbers. According to the South China Morning Post, ByteDance plans to increase its 2026 capital expenditure by over RMB 200 billion—equivalent to roughly 60% of its projected 2025 profit. Separately, ByteDance is negotiating an offshore loan of approximately USD 20 billion, with a three-year term and an option to extend for two additional years. At least from a funding perspective, ByteDance shows no sign of reducing its AI investment. The real question now is where these resources should be prioritized and which businesses must demonstrate returns first.

Seed needs to prove its model capabilities, while Doubao must preserve its large-scale user value. These two goals aren’t inherently contradictory, but as computing power, talent, and management attention become increasingly costly, they can no longer always coexist. Doubao was initially told 'don’t worry about making money,' yet it has now become the first consumer-facing product in ByteDance’s AI portfolio required to demonstrate revenue-generating potential. Understanding this shift requires re-examining Seed and Flow’s positions within ByteDance’s AI landscape.

From its inception, Seed has enjoyed an unusually high priority for internal resources at ByteDance. It reports directly to the group’s executive leadership and benefits from more flexible performance metrics, stronger talent incentives, and greater organizational independence within the company. This structure was designed so ByteDance could use profits from mature businesses to fund a model team resembling a research institute—one tasked with solving foundational large-model capability challenges.

This corresponds to a specific historical phase when the global AI industry still adhered to the logic of 'spend first, ask about output later.'

In 2023, ByteDance assembled Seed by drawing personnel from departments including AI Lab, AML, and Search teams. It established strict standards—no knowledge distillation, no shortcuts—and even built an in-house data annotation team of nearly 1,000 people.

Over the following nearly three years, this structure continued to expand: the video generation team was integrated and placed under Zhou Chang, who joined from Alibaba in 2024. By the end of 2025, several directions originally under the AI Lab—including AI for Science, Embodied Intelligence, and Responsible AI—were fully transferred into Seed. Li Hang, the former head of the AI Lab, retired and was rehired on a part-time basis, now reporting to Wu Yonghui, Seed’s new leader. Even Yan Junjie, founder of MiniMax, once remarked that ByteDance has the highest concentration of AI talent.

Internally, Seed also established virtual teams such as Edge, Focus, and Base to simultaneously advance long-term research, next-generation model development, and current model delivery. This gives Seed a scale of research rarely seen in large corporations—but also brings typical challenges of research organizations: too many directions, making it difficult to allocate resources evenly forever.

The shift stemmed from growing industry scrutiny over the return on AI investments. According to LatePost, after ByteDance executives visited Anthropic, the company began adjusting its AI resource allocation, shifting focus from purely pursuing consumer-scale user growth toward placing greater emphasis on enterprise services and professional use cases.

Anthropic offered a clear benchmark: despite lacking a consumer user base comparable to ChatGPT’s, it leveraged Claude Code and enterprise services to push its valuation into the hundreds of billions of dollars. ByteDance’s own Seedance has also demonstrated that video-generation APIs can generate high-margin revenue. Model capability remains crucial, but that capability now increasingly requires validation through revenue.

In 2026, ByteDance internally distilled its AI priorities into four key initiatives:

First, ensure that the capabilities of its world model catch up with Google’s Genie 3 by year-end;

Second, maintain leadership in video models while exploring 'dynamic generation,' where users can intervene in real time to alter storylines;

Third, continue strengthening foundational capabilities in coding;

Fourth, help Doubao build commercialization capabilities in office scenarios.

AI for Science (AI4S) and robotics hardware—areas that previously received significant investment—are notably absent from this list. This reflects not only shifts in technical priorities but also reallocations of resources.

Within Seed, broad expansion is giving way to targeted reinforcement. It has almost stopped hiring mid-to-senior technical managers externally. According to ZhiNeng YongXian (Emergent Intelligence), the few external hiring slots Seed currently offers are primarily aimed at senior professionals from institutions like DeepSeek, OpenAI, DeepMind, and Meta. Candidates such as Guo Daya, a former core member of DeepSeek, and Dong Xin, a former NVIDIA researcher, come with strong credentials that eliminate the need for mass screening.

Resources are no longer spread evenly across every research direction but are concentrated on teams most likely to enhance model capabilities, generate revenue, or achieve product closure. This resource tilt is visible in the numbers: Fan Haoqi’s World Model research group received significantly more training compute this year, and its job offers include salaries 30–50% above market average. Investment in coding ranks just below the World Model; internally, more business units are now mandated to use Seed’s models to create data feedback loops.

Boundary businesses are being repriced throughout this process, with changes in the AI for Science (AI4S) direction being the first noticed externally. On June 2, Gu Quanquan, formerly co-lead of large-model pretraining and scaling at ByteDance’s Seed Lab, posted a farewell letter on his personal account, reflecting on the two parallel tracks he had advanced over the past three years—AI-driven drug discovery (including bio-computing models like SeedFold and SeedProteo) and pretraining optimization for Seed 2.0.

A few days later, Jiemian News, citing sources 'familiar with ByteDance Seed,' reported that Gu Quanquan had indeed worked on biomolecular structure prediction and LLM pre-training, but was not the head of either of these business areas. Gu promptly clarified publicly on his personal Xiaohongshu account, stating that his responsibilities and contributions in AI-driven drug discovery had 'always been widely recognized,' and explicitly noted that the title 'Head of AI Drug Discovery' had been established when he joined in July 2023.

This rare public dispute offered outsiders their first glimpse of organizational friction within Seed’s peripheral initiatives during realignment. More noteworthy than the job title itself was the fact that around the same time, Xiao Wenzhi—the head of computational biology at ByteDance who led the Protenix open-source project—and several other AI4S-related team members had already left or were preparing to leave ByteDance to start their own ventures.

By contrast, the robotics direction underwent a different kind of adjustment. Seed Robotics, which evolved from the AI Lab’s robotics team and was integrated into Seed in 2025 under Li Hang’s leadership, has now shifted its reporting line to Zhou Chang. Li Hang himself transitioned to the role of Head of Academic Collaborations, acting as an advisor to manage university partnerships. This realignment signals further consolidation of resources: multimodal capabilities, world models, and embodied intelligence are now required to deeply share algorithms and data, with these research directions merged under a single management chain.

On the flip side of this leaner approach is the fact that the ultimate decision-making power remains firmly in Seed’s hands. Whether to pursue Genie 3 in world models, invest in dynamic generation, or adopt another company’s product logic for coding—these strategic judgments originate from the research track under Wu Yonghui and Zhou Chang, and resources and talent flow accordingly.

The only item on the list that directly points to a consumer-facing product is helping Doubao build commercialization capabilities in office scenarios. This objective has been included in ByteDance AI’s overall strategic roadmap, listed alongside Seed-led technical directions such as world models, video generation, and coding. But for Flow, this means commercialization is no longer an experiment whose pace a product team can independently determine—it has become a top-down directive.

Flow’s initial mandate was to first establish Doubao as a high-frequency, widely adopted consumer entry point. At Doubao’s inception, internal messaging repeatedly emphasized it as a strategic business with no immediate need to focus on monetization. Its early design also diverged from ChatGPT’s tool-oriented approach, instead centering on anthropomorphism, companionship, proximity to users, multiple bots, and daily interactions.

These two mandates demand different product instincts: the former requires lowering barriers and expanding usage, while the latter necessitates tiered access, creating scarcity, and convincing users to pay for enhanced capabilities.

Flow plays an execution role on the product side within ByteDance’s AI division. This business unit, which operates at the same level as Douyin, oversees consumer-facing products such as Doubao and the AI hardware team Ocean, and is led by Zhu Jun. Over the past two years, Flow has fully replicated ByteDance’s proven methodology—originally validated on Toutiao and Douyin—onto Doubao: aligning with user habits, relying on data-driven feedback, and rapidly iterating. This approach, combined with relatively restrained marketing spend, has produced China’s first—and so far only—AI product to surpass 100 million daily active users.

This approach solved the challenge of going from zero to one, but it also inherently invites greater controversy when monetization begins. Users first encountered it as a free general-purpose assistant, not as a specialized office software suite. At a stage where user switching costs remain low, whoever charges first effectively gives competitors a reason to lure users away.

Consequently, rumors about Doubao introducing paid features were met with immediate backlash from the start. During the May Day holiday this year, after the news trended on social media, public opinion overwhelmingly turned critical. In response, Doubao’s internal team held meetings and focused efforts on addressing the exposed gaps in its Q&A capabilities.

Flow is not blind to reputational risk. On June 3, Doubao’s official account issued a statement denying rumors of 'downgrading basic features to push memberships,' emphasizing that the Pro version targets professional use cases like software development and data analysis, while everyday functions—including Q&A, writing, and image generation—remain free. Nevertheless, after issuing this clarification, ByteDance continued advancing its monetization plan on the original timeline.

Financial pressure on the balance sheet helps explain this 'business-as-usual' approach. As of the first half of this year, over 200 million people used Doubao daily, yet the app generated less than RMB 1 million in daily revenue, primarily from e-commerce commissions. By May, however, estimates based on Volcano Engine’s publicly listed API pricing and Doubao large model gross margins suggested daily costs in the tens of millions of RMB.

Of course, this figure does not directly equate to ByteDance’s actual internal cost. Factors such as in-house compute capacity utilization, procurement discounts, resource sharing, and inference optimization all affect the final number, and publicly listed API prices do not necessarily reflect internal transfer pricing. Still, these estimates highlight a core issue facing Doubao: for every new batch of users added, ByteDance must first pay for the corresponding inference calls; the larger the user base, the harder it becomes to indefinitely defer monetization.

Individually, subscriptions, ride-hailing integrations, and storage upgrades may not each generate significant revenue. But their sequential launch within less than ten days suggests they are part of a pre-scheduled corporate initiative.

Multiple internal financial considerations across the group ultimately converge on the Doubao interface. Office subscriptions must validate whether users are willing to pay for productivity tools, Seed needs feedback from coding and agent tasks, and ever-rising compute expenses demand a clear revenue outlet.

In this division of labor, Seed is responsible for defining capabilities, while Flow bears the consequences of how those capabilities perform once tested by users.

ByteDance’s past product teams excelled at evolving solutions based on user feedback. Recommendation algorithms could be continuously refined in real time through massive volumes of clicks, dwell time, and interactions. However, the feedback loop for large models is slower and more expensive: higher usage initially translates into greater inference costs, without necessarily yielding proportional improvements in underlying capabilities.

Liang Rubo once remarked in an all-hands meeting that Doubao has yet to demonstrate the classic internet product trait of 'the more people use it, the better it gets.' This makes it difficult for Flow to rely on ByteDance’s traditional growth playbook to justify continuously expanding investments.

Doubao’s product lead also completed a handover during the same period. In September 2025, Zhao Qi—who had spent over a decade at ByteDance and previously led Growth Middle Office and Pangle—was reassigned to Doubao to oversee its product line. Compared to building a product from scratch, Zhao’s prior experience in growth and monetization systems aligns more closely with the stable operation of a mature product. This appointment signals a stronger emphasis on growth efficiency and commercial execution.

The issue, however, is that although the cash register sits right in front of Flow, pricing power isn’t entirely in Flow’s hands. Whether Doubao delivers accurate responses, reliably completes office tasks, and further reduces per-call costs all depend on the underlying model, computing resources, and toolchain. Seed controls model capabilities, senior leadership decides resource allocation, and Flow directly faces user perception.

When Doubao delivers incorrect responses, charges high fees, offers insufficient quotas, or fails office-related tasks, public criticism initially targets the product. Yet whether the product can be sufficiently intelligent or cost-effective is not something Flow can determine alone. Flow bears responsibility for the final outcomes but cannot unilaterally control all conditions that produce those results.

The loss of 6.1 million monthly active users after May Day is a question Flow must answer; however, why subscriptions, ride-hailing, and cloud storage had to launch consecutively within ten days is not a decision Flow could make independently. ByteDance has placed the cash register in front of Flow, but the authority to determine product capabilities, costs, and supply cadence remains dispersed among its upstream units.

This year, the price of Doubao shares has been raised twice: the first repurchase in April was 30.8% above the grant price, and in June it rose another 13.5% on top of that.

Yet the incentive mechanism named after 'Doubao' primarily flows to those working on models. In 2025, Wu Yonghui announced an exclusive stock option plan called 'Doubao Shares' at an all-hands meeting for the Seed division, targeting core technical staff focused on large models. Specifically named key teams included the Large Model R&D Group, Video Generation Team, and Algorithm & Engineering Group—all operating under Seed, not Flow, which handles Doubao’s product experience and user growth.

From the flow of resources, talent, and incentives, Seed remains the internal priority center for ByteDance’s AI efforts. Doubao and its home division, Flow, occupy a position in this framework that does not fully align with their user scale. Flow serves as the largest distribution channel and the primary testing ground for model capabilities with real users, yet it does not hold the power to allocate resources.

An even bigger shift is that ByteDance has begun separating AI from a unified narrative into distinct financial ledgers. Seed’s ledger tracks model capability and talent density; Seedance’s ledger focuses on API revenue and gross margin; and Doubao’s ledger measures whether user scale can translate into subscription fees, commissions, and office-task-related income. Previously, these ledgers could all be bundled under the 'All in AI' umbrella; now, each is being evaluated independently.

This isn’t the first time ByteDance has made such strategic trade-offs. For example, in November 2023, after a business review, it decided to restructure its gaming operations. Nuverse publicly stated it would adjust its business direction, focusing on a select few innovative games and related technologies while continuing to operate already-launched titles.

In March 2026, ByteDance further agreed to sell Moonton to Savvy Games Group, a subsidiary of Saudi Arabia’s Public Investment Fund. ByteDance had acquired Moonton in 2021 at an estimated valuation of around USD 4 billion; neither party disclosed the sale price this time, though Reuters, citing informed sources, reported that Moonton’s valuation in the transaction exceeded USD 6 billion.

This timeline illustrates that once a business loses top-priority status, it doesn’t necessarily vanish immediately from ByteDance’s portfolio. More commonly, it is disaggregated: promising directions are retained for further validation, independently viable assets are divested, and resources are redirected toward new strategic centers.

Doubao and gaming are clearly different businesses. Doubao remains ByteDance AI’s largest consumer-facing entry point and plays a critical role in model distribution, scenario validation, and user feedback. However, the gaming business offers a cautionary reference: neither investment scale, user scale, nor the label of 'strategic priority' automatically guarantees continued priority in the next phase.

'All in AI' once allowed all initiatives to be accounted for on a single ledger. By the summer of 2026, ByteDance had turned the page to Doubao.

1532 Views