Evoken (演语科技), the parent company of image-generation platform LiblibAI, has closed a near-$300 mill...

Evoken (演语科技), the parent company of image-generation platform LiblibAI, has closed a near-$300 million Series B+ round at a valuation exceeding $2 billion, making it one of China's largest AI application-layer funding events to date. Announced June 18, the round underscores a broader pivot in AI investment logic—away from foundational model scale toward monetization and real-world deployment. Per the AI Market Watch index, Evoken is tracked with $470M in total funding, though coverage is not a census of all deals. The company, founded by former ByteDance executive Chen Mian (who led CapCut/Jianying commercialization), claims $300M in ARR as of May 2026, with revenue growing over 3000% year-over-year. Its product matrix includes LiblibAI (a C-end community with 30 million users and 500,000 LoRA models), Lovart (a B2B design tool reaching $80M ARR in five months), and LibTV (an AI short-video platform, which alone contributes over half of total ARR).

Liblib's business is essentially a token middleman thesis. It doesn't train models but aggregates APIs from upstream providers—notably ByteDance's Volcano Engine—and packages them into polished, user-friendly products, marking up token costs with service fees and community value. Founder Chen Mian compares the model to a photographer for a camera. But the strategy carries exposure risk: The firm is heavily dependent on ByteDance (it is the largest reseller of ByteDance models) and faces allegations of API reselling, though Chen denies these claims. He rebuts the "3.9% discount" narrative, saying it refers to a subscription price, not per-token API costs. The deeper challenge is that
as long as the upstream model provider is the same, price adjustments or a power-integrated app could undercut Evoken's margins and user base. Chen openly concedes that margins in AI applications should stay below 30%, viewing high margins as demand suppressors.

Evoken's rise is a case study in the shifting dynamics of the AI application layer. Its survival instincts—spotting cash-flow-negative models, pivoting from pure C-end communities to a multisystem portfolio—saved it from near-collapse in 2024 when the account balance once dipped to 4,000 yuan, after spending $3 million in burn-heavy subsidies. Today, its three-product structure (community, design, video) shares creators and assets, creating cross-subsidization. But as infra costs shrink and model vendors move down the stack, Evoken's core "token relay" moat remains unproven. One concrete implication for builders: if you’re building a distribution layer atop foundation models, prepare for a fight over pricing power and access. Evoken's model offers precedent—but also a warning—of the potential for margin compression and forced arbitrage to evaporate as the model ecosystem matures. #Evoken #AIApplications #GenAI #ByteDance #ChinaAI #TokenArbitrage