钛媒体APP

wrote a column ·

ByteDance loudly pivots to B2B—why is it avoiding the main B2B stage?

Image generated by AI

ByteDance recently held two high-level company-wide meetings in quick succession.

The first took place at the end of July, when founder Zhang Yiming appeared at an all-hands meeting of the Seed team—his rare public appearance and remarks since the advent of large AI models. The second occurred in early August, led by the CEO.

These two uncharacteristic moves clearly signal that this unicorn—whose valuation ranks just behind Anthropic and OpenAI globally—is urgently seeking to align internal thinking and clarify strategic direction through more direct communication.

Executives collectively emphasized two core messages: AI is the company’s long-term strategic priority, and it is fully committed to the B2B productivity赛道, demonstrating ByteDance’s resolve to transform from a consumer-facing traffic giant into an enterprise AI service provider.

However, a puzzling disconnect has emerged between internal rhetoric and real-world external execution.

Just this past summer, at a premier national exhibition held in Shanghai, the true fragmentation of ByteDance’s AI strategy came into sharp relief.

At WAIC—the most important public platform in China for B2B and industrial AI—ByteDance once again opted against setting up an official standalone booth, maintaining only a minimal presence through third-party collaborations (such as the Dobby phone).

Just as observers assumed ByteDance was fully embracing 'cost reduction and efficiency enhancement'—scaling back PR narratives and reducing offline exhibition activities—in late July, we witnessed the company go all out at ChinaJoy.

At the BTOC exhibition hall of ChinaJoy 2026, ByteDance’s Douyin Live and its gaming division, Nuverse, each set up massive booths—their combined exhibition space exceeding that of Tencent, the world’s largest gaming company.

Douyin's livestreaming booth at ChinaJoy 2026, photo taken by the author

ByteDance’s significant investment at a gaming expo came as a surprise. Earlier this year, when ByteDance decided to sell Moonton, many assumed the company was going all-in on AI and needed to raise substantial capital to fund its AI R&D—so much so that it was willing to part with a high-quality asset that had previously given Tencent a run for its money. After all, Moonton embodied two critical strategic pillars: internationalization and gaming. Its flagship title, Mobile Legends: Bang Bang, has consistently outperformed Tencent’s international version of Honor of Kings in Southeast Asia.

The Nuverse booth at ChinaJoy 2026—the once-deprioritized and nearly divested gaming division within ByteDance—also participated, photo taken by the author

Interestingly, during ChinaJoy 2026, Hongguo Short Drama also set up an oversized exhibition space at Century Plaza on Nanjing Road Pedestrian Street—near the Bund, Shanghai’s most crowded area—as part of the 2026 G-Power Digital Entertainment Carnival (July 31–August 9), led by Huangpu District.

Hongguo Short Drama’s prime booth at the 2026 G-Power Digital Entertainment Carnival in August, photo taken by the author

Simultaneously spanning both Pudong and Puxi, ByteDance’s moves in entertainment have effectively captured the full attention of Shanghai residents.

This summer, ByteDance has loudly proclaimed its pivot to B2B—but instead of focusing on core B2B industry venues, it’s heavily investing in consumer-facing entertainment expos?

The divergence between this ground-level 'exhibition philosophy' from business units and the company’s top-down strategic narrative is genuinely puzzling—and naturally raises questions: Is this a well-considered, differentiated tactic, or merely a contradictory outcome shaped by tension among the overarching B2B strategic narrative, organizational DNA, and real-world business constraints?

The only thing that might clarify this is a deeper understanding of the urgency behind Zhang Yiming and Liang Rubo’s two recent all-hands meetings.

The last time ByteDance made a declaration akin to an 'All in'—all in on short videos—was back in September 2016, shortly after Douyin had just launched and showed no signs of becoming a market leader. At the time, the strategic pillar supporting this move was Huoshan Short Video, a product designed to compete with Kuaishou, along with the short-video channel embedded within the Toutiao app.

A decade later, Liang Rubo, who succeeded Zhang Yiming, has again issued a similar 'All in' call—this time, all in on AI for the enterprise (AI to B). It remains unclear whether ByteDance has already achieved internal alignment around this vision, but it undoubtedly faces numerous hurdles as it seeks to enter the core battleground of China’s enterprise services market.

Key challenges currently include weak brand recognition in the B2B space, shortcomings in foundational large-model capabilities, path dependency on consumer-facing traffic-driven thinking, misalignment between internal KPIs and revenue-sharing mechanisms, and a commercialization focus heavily concentrated in video content (particularly entertainment verticals).

ByteDance’s senior leadership has consistently sent clear signals.

In June this year, at the Volcano Engine FORCE conference, Liang Rubo articulated a strategic principle of 'narrowing business scope, prioritizing depth, focusing on core pillars, and optimizing for the long term,' positioning AI as the company’s central strategic pillar and explicitly identifying Model-as-a-Service (MaaS) and enterprise productivity as among AI’s most critical directions.

At the end of July, ByteDance executed its largest-ever organizational restructuring in its B2B history: the Lark product team was merged into Doubao, while all Lark sales, marketing, and customer service teams were integrated into Volcano Engine to form a Creativity Services Platform, which now centrally manages all of ByteDance’s B2B commercial operations and unifies the pipeline from 'large model products – office scenarios – cloud and MaaS sales.'

At the company-wide mid-year meeting in August, the strategic pivot toward B2B was reinforced once more. Liang Rubo stated that AI-driven productivity is advancing faster than expected, making the B2B segment increasingly vital; over 90% of new Lark customers have concurrently purchased AI products, and this integration was specifically prepared to better serve enterprise clients.

From a top-down perspective, ByteDance’s B2B vision comprises a three-tier structure: at the base, the self-developed Seed large-model foundation; in the middle, enterprise-facing products such as Doubao Enterprise Edition and Lark; and at the front end, Volcano Engine’s unified sales network delivering MaaS, SaaS, and industry-specific solutions externally.

Simultaneously, ByteDance aims to leverage its consumer-side model capabilities to penetrate the enterprise market in reverse, transitioning its identity from a consumer internet giant to an AI infrastructure provider.

Yet a contradiction persists: on one hand, senior leadership repeatedly sets the strategic tone and undertakes sweeping organizational overhauls, loudly proclaiming its intent to aggressively pursue the B2B market; on the other hand, when China’s most significant B2B industry conference kicked off—a key public forum for enterprise decision-makers—ByteDance chose to stay absent, refraining from fully revealing its B2B strategy in this critical arena.

The World Artificial Intelligence Conference (WAIC) has become the ideal stage for AI companies to comprehensively showcase their full-stack capabilities. At WAIC, ByteDance could have demonstrated its foundational general-purpose AI capabilities, Feishu and Doubao Enterprise Edition products, Volcano Engine’s industry-specific solutions, and real-world implementations with government and large enterprise clients—directly engaging B2B decision-makers to build brand recognition.

Yet for many years, ByteDance has not set up an exhibition booth at WAIC—neither at the corporate group level nor for any specific business unit—and it has never launched a comprehensive suite of offerings targeting enterprise customers. This year, glimpses of Doubao were only sporadically visible as demo units or fragmented appearances on certain partners’ booths.

Although ByteDance’s primary revenue currently comes from consumer-facing (C-end) traffic monetization, it does have B2B business activities: Volcano Engine already serves a sizable number of government and enterprise clients.

Then why has it chosen to forgo this critical public platform—an opportunity to formally assert its identity as a B2B player across the entire industry? And why is its large-scale public AI showcase heavily concentrated in C-end entertainment scenarios like short dramas and gaming?

By contrast, at WAIC, Alibaba, Tencent, Huawei, Baidu, and others have all built large-scale official exhibition booths, comprehensively showcasing their foundational AI capabilities, industry solutions, real-world deployments with government and enterprise clients, and office productivity tools.

On-site staff from major tech firms have previously told the author that all exhibited products are already officially launched; even if they currently lack compelling new offerings, attendance remains essential—because WAIC is the highest-profile public arena in China for industrial AI, government and enterprise procurement, enterprise services, and B2B large-model adoption.

The attendees gathered here represent precisely the core audience for B2B business: decision-makers from central and local state-owned enterprises, large and mid-sized corporations, system integrators, channel partners, industrial investors, and developer communities. These stakeholders also assess a company’s strategic priorities based on its public actions. If a company doesn’t show up on this industrial stage, how can B2B clients believe it is seriously committed to enterprise business?

Many observers interpret ByteDance’s absence from WAIC as a shrewd strategic retreat—a refusal to participate in the pointless benchmarking race among large models. However, internal remarks by ByteDance executives reveal another dimension: imbalances in its general-purpose foundational capabilities have led to a 'product-display dilemma' for its B2B exhibition presence.

ByteDance’s video multimodal generation model, Seedance, stands among the global leaders, with virtually no rivals worldwide in AI-generated video content. However, its general-purpose large language model still shows perceptible gaps compared to top-tier players in deep reasoning, long-context handling, code generation, and complex agent capabilities.

This is also a phenomenon acknowledged by Zhang Yiming and Liang Rubo in two internal meetings.

As a result, Zhang Yiming and Liang Rubo established a firm technical red line: they refused to boost benchmark rankings by 'distilling' third-party models and insisted on fully native, in-house development, willingly accepting a period of lag in foundational language models.

While their commitment to in-house R&D is undoubtedly admirable, in terms of commercialization, domestic B2B clients currently judge a company's AI capabilities largely by the strength of its base models. Moreover, video generation performance is not a top priority for many clients. Instead, capabilities like long-document understanding, logical reasoning, code generation, complex task decomposition, enterprise knowledge-base Q&A, and agent orchestration represent the real, practical needs for most enterprises integrating AI into their workflows.

This has placed ByteDance in an awkward position: its truly impressive and standout AI achievements are concentrated in video (entertainment) content generation, while the general-purpose foundation models that B2B industrial clients actually care about are still in a phase of catching up and refinement—insufficiently mature to serve as the centerpiece of major exhibition booths.

Internally, ByteDance can candidly acknowledge this reality and grant the Seed team a longer development cycle without chasing short-term metrics. However, this narrative cannot be directly communicated to B2B enterprise clients externally.

For B2B procurement decision-makers, it is difficult for enterprise clients to accept a vendor openly declaring, 'Our general-purpose model is temporarily behind; you can still buy our service.' Clients demand a stable, reliable, and capable commercial foundation—not an R&D project still in its catch-up phase.

In contrast, consumer entertainment expos like ChinaJoy not only safeguard ByteDance’s core commercial businesses (Douyin and Hongguo) but also offer another form of reassurance in the AI era: the audience consists of gamers, short-video creators, and content producers who care about visuals, storytelling, and gaming experiences—virtually no one there asks about a large model’s long-context reasoning or coding capabilities, allowing Seedance to showcase dazzling demos without any visible weaknesses.

But by sidestepping WAIC—a venue where vendors must fully address real B2B pain points—and retreating instead to its stronghold in C2C traffic-driven businesses, ByteDance pays a direct price: it voluntarily relinquishes its voice in the primary B2B arena.

ByteDance is a world-class expert in consumer-facing (To C) products, but many aspects of its To C mindset fundamentally contradict To B logic.

ByteDance grew and exploded during the mobile internet era. This massive traffic-processing machine built its organizational culture, performance evaluation system, and talent mindset around consumer-facing (To C) traffic-driven strategies: rapid iteration, data-driven decision-making, pursuit of short-term feedback, large-scale user validation, fast experimentation, and trading scale for efficiency…

Enterprise (To B) services prioritize long-cycle client relationships, customized delivery, channel ecosystem development, and customer success frameworks—focusing on customer retention and renewal rates. In this space, returns materialize over extended periods, making it difficult to achieve the explosive short-term data feedback typical of To C businesses.

This conflict manifests in several observable realities at the operational level.

First, ByteDance’s To B offerings naturally favor B2B models that resemble traffic-driven approaches—particularly standardized API calls. Within Volcano Engine’s MaaS (Model-as-a-Service) suite, the standout revenue generator is Seedance’s video-generation API, which primarily serves short-form drama studios, short-video MCNs, and gaming companies. Essentially, it repackages content-generation capabilities validated on the consumer side and sells them to B2B content clients.

This business model leverages ByteDance’s existing content AI capabilities, benefits from scalable API call volumes, and delivers rapid revenue visibility—aligning closely with ByteDance’s ingrained traffic-driven growth logic.

To some extent, this also reflects the original rationale behind Lark—the flagship To B product—as a commercial offering: once proven effective internally, it was then rolled out to the broader market.

However, deep customization solutions targeting industrial, government, and large traditional enterprises require intensive involvement from industry consultants, delivery teams, and customer success teams embedded within clients’ operational workflows. These projects have long cycles and slow revenue recognition, making it hard to generate impressive near-term financial metrics. Palantir’s Foundry Data Environment (FDE) has recently gained significant acclaim precisely because it had already deeply integrated AI models into specific enterprise use cases.

In a sense, Lark’s challenges encapsulate the inherent tension within ByteDance’s To B DNA.

Lark boasts a high-quality enterprise client base but has long faced profitability pressures. Today, new customers are increasingly bundling AI features into their purchases, and AI applications themselves chase user scale. Yet in the AI era, higher model usage directly translates into higher compute costs.

Mobile internet operates under an entirely different business logic. Due to network effects, mobile apps experience a tipping point in growth. During cold starts, user acquisition costs are extremely high—most apps require heavy subsidies. But once user scale surpasses a certain threshold (e.g., when WeChat established its 'real-name social graph'), the marginal cost of acquiring each additional user doesn’t rise; instead, it plummets due to word-of-mouth and social virality, eventually approaching zero as organic growth takes hold.

Moreover, the larger the user base of a mobile app, the more data it generates—data that can be used to train AI algorithms. The rise of Douyin is a prime example, and this has been the sole economic rationale for investors’ willingness to 'burn money' upfront in exchange for scale.

Deep AI adoption by enterprise clients leads to a surge in token usage, which directly drives up compute costs. Applying consumer-facing customer acquisition strategies to enterprise services no longer aligns with viable profitability models.

In fact, ByteDance’s absence from WAIC and its heavy investment in ChinaJoy reflect a calculated business decision grounded in commercial reality.

From a PR and brand marketing KPI perspective, allocating budget to showcase consumer-facing AI content at ChinaJoy yields quantifiable exposure and buzz on Douyin. In contrast, investing in B2B industry conferences for industrial outreach may not generate strong traffic metrics—and even if artificially inflated, such efforts are unlikely to drive meaningful customer conversions, making ROI difficult to assess.

The market shifts driven by AI are more dramatic than ever before. Moreover, even during its app-centric era, ByteDance did not succeed in every venture—but each time it entered or exited a market, it acted with remarkable speed, forging an organizational muscle memory within the company.

Although top management has issued directives to pivot toward B2B, frontline business units instinctively prioritize niche B2B segments that yield quicker results and lend themselves to compelling external narratives—particularly in entertainment—to ensure their survival within the organization and secure more resources. They naturally avoid capital-intensive, slow-return B2B markets in hard-tech industries requiring heavy delivery commitments.

More concretely, ByteDance’s current MaaS (Model-as-a-Service) revenue is heavily concentrated in the content and entertainment sector—including AI-generated short drama assets, short video creation, and AI-powered game content production. These businesses rely extensively on the Seedance video model, serve a vast client base, and generate high-volume API calls, making them ByteDance’s primary source of AI-related B2B revenue today.

At digital entertainment expos like ChinaJoy, ByteDance’s fellow exhibitors—game studios, short drama producers, and content agencies—are precisely the B2B clients who do not require government-enterprise industry forums. These clients already operate within ecosystems like ChinaJoy. There, ByteDance can simultaneously achieve consumer brand visibility and engage its core B2B content clients, generating tangible business leads.

Moreover, with the launch of Seedance 2.0 this year, ByteDance has achieved an industry-acclaimed commercial closed loop: leveraging Tomato Novels for IP sourcing, selling compute power via Volcano Engine (powered by Seedance), distributing content through Hongguo Short Dramas and Douyin, and monetizing ad spending through Ocean Engine.

However, the potential downsides are also evident. Enterprise and industrial clients may easily develop a stereotypical perception: ByteDance’s AI-to-B offering is fundamentally a content-generation tool tailored for the content industry. As a result, ByteDance’s B2B positioning could become confined to this vertical, rather than being recognized as a general-purpose enterprise AI foundation capable of serving countless industries.

Consequently, sales teams struggle to present sufficiently compelling, end-to-end case studies to large enterprises outside the entertainment and media sectors. Without a critical mass of cross-industry benchmark cases, ByteDance lacks hard-hitting exhibits on B2B-centric stages like WAIC.

Similarly, for the Seed foundational model team, senior leadership has set evaluation criteria that emphasize 'accepting temporary lag, disregarding short-term commercialization, focusing instead on long-term technical catch-up progress, and insisting on fully native, in-house R&D.' This framework serves a three-to-five-year strategic horizon, allowing for high investment and slow returns.

Computing power is the scarcest resource. When training general-purpose foundational models demands massive, unrewarded investments in computing capacity, while commercialization teams urgently need model APIs that can generate quick revenue, tensions over resource allocation inevitably arise.

Of course, by going all-in on B2B, ByteDance naturally holds formidable advantages: world-class multimodal engineering capabilities, vast computing power reserves, model interaction data derived from hundreds of millions of consumer users, a high-quality enterprise client base built through Lark, and the VolcEngine cloud infrastructure.

Yet it also carries significant historical baggage: deeply ingrained consumer-traffic DNA, a narrowing time window to catch up in foundational models, strong short-term KPI pressures from mature businesses, and the allure of staying within its comfortable content-industry niche.

The disconnect between executive meeting slogans and on-the-ground realities at trade shows offers a telling example of this industry tension—and perhaps ByteDance’s AI-to-B story is only just beginning. Beyond the rallying cries, the ultimate verdict will be rendered by the market, customers, and time.(This article was first published on TMT Post app. Author: Li Chengcheng; Editor: Yang Lin)

140K Views