ByteDance Rethinks AI Apps as Model Revenue Outpaces Product Growth

A Quiet Merger Inside CapCut

ByteDance has begun absorbing Xiaoyunque, its artificial intelligence platform for microdrama production, into Dreamina, the video generation tool marketed in China as Jimeng AI. The consolidation preserves a separate entry point for Xiaoyunque users but unifies operations under a single leadership structure. Chen Xinran, already responsible for Dreamina's operations and marketing, now oversees both products, though Xiaoyunque's product team remains in place.

The integration took place within CapCut, ByteDance's video editing division. Dreamina has already incorporated the microdrama capabilities that defined Xiaoyunque's offering, effectively blending two tools that once competed for internal resources and user attention.

People with knowledge of the decision cited two factors: limited growth potential and a need to concentrate investment. While both products continued to add users, neither generated revenue commensurate with the capital and engineering effort required to maintain them. At the same time, the proliferation of AI-native video agents across the market made it harder to differentiate ByteDance's own suite of applications from those of rivals.

The Model Economy Reshapes Product Priorities

The merger reflects a broader recalculation of value inside ByteDance and across the technology sector. As foundation models improve, companies are reassessing the role of applications built on top of them. ByteDance has moved particularly quickly in this direction, and the Dreamina-Xiaoyunque integration is not the only consolidation under discussion.

At Opentechwire, we've tracked a similar pattern emerging at other large-scale AI developers: as model capabilities expand, the marginal value of wrapper applications shrinks, especially when those applications do not command pricing power or lock in user behaviour. ByteDance's Volcano Engine division, which licenses the Seedance video generation model to external developers, now earns revenue that far exceeds the combined intake from Dreamina and Xiaoyunque. Nearly every video agent and production workflow for short dramas and animations in China has integrated Seedance, according to people familiar with the business.

That revenue gap has altered how ByteDance evaluates its own consumer products. When outside developers using the same models can acquire paying users, internal applications must demonstrate additional value through differentiated acquisition, retention, or monetisation. If they cannot, they become redundant.

From Universal Gateway to Professional Tool

Two years ago, ByteDance held a different ambition for Dreamina. In February 2024, Kelly Zhang stepped down as chief executive of Douyin Group to lead CapCut in China and internationally. CapCut launched Dreamina in May, shortly after OpenAI introduced Sora and raised the question of whether AI video products or chatbots would become the next dominant gateway for users. ByteDance invested in both Doubao, its conversational agent, and Dreamina, betting on parallel pathways.

Zhang initially envisioned Dreamina as both a creative tool and a community for AI-generated content. The early plan mirrored the division between CapCut and Douyin: users would create on the web version and consume on mobile. After ByteDance launched Seedance, the app experimented with features that placed real people on screen and enabled interaction through digital avatars, echoing Sora's approach.

As model quality improved, however, many users found they could generate AI video content directly within Douyin or Doubao. Dreamina's role as a standalone product became less clear. The team pivoted toward professional creators, expanding beyond text-to-video and image-to-video generation to offer first-frame and last-frame controls, an intelligent canvas, and agent-assisted editing workflows.

Before China's National Day holiday in 2024, Dreamina had already reduced investment in its mobile application and begun combining app and web operations. Future resources will concentrate on the web version, targeting creators who require more granular control. The company also plans to increase investment in Dreamina's international expansion and monetisation, areas where professional tools typically command higher willingness to pay.

Intensifying Model Competition and Executive Departures

Rapid improvements in foundation models over the past two years have shifted how the market values applications. As coding agents opened opportunities in office productivity, the notion that models themselves are the applications gained traction. ByteDance's release of Seedance 2.0 in February 2025 further altered the internal dynamics.

Even before that release, Xiaoyunque, Dreamina, CapCut, and Doubao had competed to be the first to integrate new Seedance capabilities. After version 2.0 launched, competition intensified, giving model teams greater influence over product roadmaps. People at Volcano Engine described the shift as a rebalancing of power: the teams building the underlying models now held more sway than those building the interfaces.

Limited product growth has contributed indirectly to a wave of departures from CapCut over the past two years. Senior product and operations leaders who left to start their own ventures include Chen Mian, formerly responsible for global commercialisation and operations; Ming Chaoping, who led mobile editing; and product head Zhang Qizhi. In late 2024, CapCut product head Zhang Xiaoran also departed. She later joined Flova.ai, the startup founded by Guo Lie, CapCut's original founder.

The pattern is consistent with what we have observed elsewhere in the region: when internal products stall, the most entrepreneurial product leaders often leave to build on top of the same infrastructure they once managed, but with clearer monetisation models and narrower user segments.

What the Consolidation Signals

ByteDance's decision to merge Xiaoyunque into Dreamina is not simply a cost-cutting measure. It is a recognition that in a world where foundation models are widely accessible, differentiation must come from either proprietary data, unique user behaviour loops, or superior monetisation. Applications that lack these attributes become liabilities, consuming capital and attention without generating defensible value.

The company's willingness to consolidate so quickly, and to discuss further mergers, suggests it has concluded that the next phase of competition will be fought at the model layer, not the application layer. For ByteDance, that means Volcano Engine's model-as-a-service business is the strategic priority, while products such as Dreamina and Xiaoyunque serve primarily as reference implementations and distribution experiments.

This is a different calculus from the one that prevailed two years ago, when the race to build the next dominant AI application seemed open. Today, the economics point in a different direction: the companies that control the models, and can monetise them at scale through licensing, hold the advantage. Applications remain important, but only if they can prove they are more than wrappers.