Liang Rubo is trying to make ByteDance a leading artificial-intelligence company before its identity becomes fixed around the social platforms of the previous technology cycle. In 2026, Doubao and the Seed model family have moved to the centre of the group’s product ambition. Seedance 2.0 can generate and edit multimodal video with synchronised audio, while newer speech systems are designed for more natural interaction. The progress shows that ByteDance’s recommendation expertise, consumer distribution and engineering resources can be redirected towards generative AI. It also creates a harder problem: how to govern models that can reproduce commercial creative work, manipulate realistic media and spread through products used at immense scale.
Liang’s strategic urgency is understandable. ByteDance built Douyin, TikTok and Toutiao around the ability to understand attention and match content with users. Generative AI changes both sides of that system. It can create the content, not only recommend it, and it can replace a feed with an assistant that completes tasks. If ByteDance remains primarily a distributor, new model providers could capture the interface and economics. If it becomes a model company, it must master computing, enterprise distribution, safety and intellectual-property questions that its earlier products did not resolve.
The company’s advantage is unusually powerful. Doubao can reach consumers through a familiar domestic ecosystem, Volcano Engine can serve businesses, and CapCut or Jianying can turn model advances into creative tools. TikTok gives ByteDance experience with global product operations even as ownership and national-security disputes constrain how technology can cross borders. Liang has the assets to build an integrated AI platform. His leadership will be judged by whether integration becomes durable capability rather than a collection of heavily subsidised experiments.
Seedance turns technical progress into a rights problem
Seedance 2.0 demonstrates the strategic opportunity. Its unified architecture accepts text, images, audio and video as references, and can generate multi-shot output with synchronised sound. Better motion, consistency and editing make the model more useful for advertising, entertainment and everyday creation. They also increase the chance that users can imitate protected characters, performers, cinematography or music. A weak model produces obvious artefacts; a strong one produces disputes over ownership and substitution.
ByteDance needs a rights framework that is designed into the product. Identity verification for real-person references, tools for rights holders, provenance markers and responsive takedown processes are not peripheral compliance features. They determine whether professional creators and distributors will integrate the system. Liang should be prepared to limit some high-demand uses when consent is unclear. Market share gained through legal ambiguity can become a liability once regulation, litigation and platform policy converge.
Training transparency will remain contested because model developers protect datasets and methods. ByteDance can still disclose meaningful categories, licensing principles and exclusion processes without revealing proprietary code. Creators need a workable way to indicate that material should not be used or to negotiate compensation. The company should test collective licensing and commercial partnerships rather than assume every dispute can be handled after output appears.
Safety is broader than copyright. Realistic video and voice can be used for fraud, political manipulation and harassment. Watermarks alone are insufficient if they can be removed or are not recognised across platforms. ByteDance should combine technical provenance with account controls, risk-based access and rapid coordination with other services. The company already understands content moderation at scale; generative media requires that expertise to move upstream into model behaviour and product design.
Doubao must become useful without becoming indiscriminate
Doubao’s consumer reach gives ByteDance a route to habitual AI use in search, conversation, creation and tasks. Scale can improve products through feedback, but free or low-priced access can also conceal high inference costs. Liang needs a path from adoption to sustainable economics. Subscriptions, advertising, enterprise services and transactions each create different incentives. An assistant funded mainly by advertising may optimise engagement when users expect impartial help.
Voice interaction is especially important in China’s mobile market. Full-duplex systems that can listen while speaking make assistants feel more natural and could move AI into cars, devices and service settings. They also create privacy risk because microphones and contextual data are more intimate than a typed prompt. Clear activation, local processing where possible and strict retention controls should be part of the architecture. Convenience will not compensate for uncertainty about when a system is listening.
Enterprise adoption requires a different promise from consumer experimentation. Companies need data isolation, audit trails, predictable service and accountability when models fail. Volcano Engine can translate ByteDance’s internal infrastructure into business products, but it must give customers control over models and information. Liang should avoid making Doubao the mandatory answer to every enterprise problem. Open interfaces and support for specialised models can strengthen trust even when they reduce lock-in.
Model quality in Chinese language, regional dialects and multimodal culture can create a distinctive advantage. ByteDance should also test systems across minority languages and vulnerable groups. Safety filters trained around standard Mandarin may miss abuse or over-block legitimate expression elsewhere. A global AI ambition requires evaluation beyond the domestic contexts that produced the model.
Computing strategy is now corporate strategy
Training and serving advanced models require vast computing resources. Export controls restrict access to the most capable foreign chips, while Chinese policy favours technological self-reliance. ByteDance must balance purchases, domestic accelerators, overseas infrastructure and reported work on custom silicon. The company does not publicly disclose enough detail to treat every spending estimate as reliable. The strategic fact is simpler: compute availability and cost will shape which models can be trained, where they can operate and how cheaply Doubao can be offered.
Liang should apply portfolio discipline to research. Larger models are not automatically better businesses. Efficient architectures, inference optimisation and smaller task-specific systems may produce greater returns than a race for benchmark leadership. Teams need clear objectives tied to product usefulness, not only publication or model size. The company’s culture of rapid experimentation is an advantage when experiments can be stopped; infrastructure commitments are harder to reverse.
Supply resilience carries geopolitical and operational consequences. Custom chips may reduce dependence but require design talent, software compatibility and foundry access. Domestic alternatives can be slower or more expensive during transition. ByteDance should build portability across hardware and avoid tying critical services to one constrained route. Energy and water use will also attract scrutiny as data-centre demand expands in Asia.
Security must rise with strategic importance. Model weights, training data and chip designs are valuable targets. Internal access should be segmented, suppliers reviewed and incidents practised. At the same time, excessive secrecy can isolate researchers and slow external evaluation. Liang needs controlled collaboration with universities and independent safety experts, allowing meaningful challenge without compromising core assets.
TikTok’s geopolitical exposure remains part of the balance sheet
Competition inside China will remain intense. Alibaba, Tencent, Baidu, DeepSeek and specialised model companies approach AI from different assets and pricing models. ByteDance should not respond to every release with a larger subsidy or rushed benchmark claim. Its defensible position will come from products that combine creation, distribution and commercial tools while allowing users to move data and work elsewhere.
Financial opacity increases the need for internal discipline. As a private company, ByteDance does not publish the segment economics that public investors would use to test AI spending. The board should require comparable measures for compute commitments, user value and safety costs. Employee share programmes and secondary valuations are not substitutes for evidence that the new platform can generate durable cash.
ByteDance cannot build its AI future separately from the dispute surrounding TikTok. Governments have questioned ownership, data access and influence over recommendations. Whatever the legal and corporate outcome in individual markets, the controversy shapes how regulators will view ByteDance AI products. A company already treated as a national-security concern receives less benefit of doubt when it launches powerful generative systems.
Structural separation, local data controls and independent oversight may be necessary in some markets. Liang should evaluate them as strategic costs of access rather than temporary public-relations measures. Governance that is credible only because the parent promises restraint will not satisfy governments that distrust the parent. The company must decide which technology, personnel and data can be separated without making the service uncompetitive.
Global expansion of Seed products should therefore be selective. Local rights regimes, election rules and safety expectations differ. Delaying a launch can be better than releasing a system whose controls are not ready. ByteDance’s consumer culture prizes speed, but the economic value of AI will develop over years. Trust lost through an avoidable early incident may close markets permanently.
Organisation is Liang’s final challenge. AI talent expects research autonomy, while a consumer platform relies on product metrics and central infrastructure. The Seed team needs space to pursue uncertain work and clear responsibility for deployment consequences. Compensation should not reward only usage, and safety teams should have authority to slow releases. Founders and senior technical leaders must be challenged by governance that is more than an internal review.
The next twelve to twenty-four months will show whether ByteDance can convert distribution into an AI franchise. Doubao needs retention and credible economics. Seedance needs rights and safety systems proportional to its capability. Computing investment must produce efficient products despite supply constraints. TikTok’s geopolitical exposure must not be allowed to define every new service. Liang has put AI at the company’s centre; the harder task is proving that ByteDance can govern what it builds as quickly as it learns to build it.
