Doubao's Daily Revenue Below ¥1 Million While Seedance Achieves 70% Gross Margin
Alina Collins
ByteDance's two AI products sit at opposite extremes: Doubao has over 200 million DAU yet earns under ¥1 million daily, while video-generation model Seedance is running at $2 billion ARR with a 70% gross margin — This means → ByteDance's AI monetization gravity is shifting from free consumer apps to enterprise services.
How much does Doubao burn — and earn — each day?
Doubao has over 200 million daily active users, yet daily revenue stays below ¥1 million, mostly from e-commerce commissions.
Daily compute costs run into the tens of millions of yuan — operating expenses alone exceed Bilibili's entire running cost, yet total user time is less than one-eighth of Bilibili's.
This means → Doubao is following the classic mobile-internet playbook of "free users first, monetize later," but AI's compute bill makes that path far more expensive than it ever was for short video.
Why is Seedance actually making money?
Seedance's annualized revenue (ARR) has reached $2 billion (roughly ¥14.3 billion), with monthly revenue exceeding ¥1 billion. The vast majority comes from enterprise clients.
Gross margin sits at 70% — for every ¥10 in API revenue, server and inference costs account for about ¥3.
In plain terms = video-generation models are structurally cheaper to run: they can batch-process data and output results in parallel, depend less on chip-to-chip communication, and can run inference on a wide range of domestically made chips — all of which drives inference costs sharply lower.
Why are video models cheaper than language models?
Video models use the DiT architecture — Diffusion Transformer, a method that gradually "de-noises" static into coherent frames — enabling batch-parallel inference with low dependence on memory bandwidth and inter-chip links.
Language models generate one token at a time and must continuously reference prior context, demanding more memory and faster chip interconnects — a higher hardware bar.
The talent gap mirrors the cost gap: top multimodal researchers earn roughly 30% less than their language-model counterparts. Seedance's core algorithm team numbers only around a dozen. A full training run costs several hundred million yuan; language-model training may cost 3–5× more.
Where is Doubao's monetization stuck?
Doubao generates only about ¥10 million in daily e-commerce GMV — a thin figure for a platform with 200 million DAU.
A paid subscription tier is in preparation, but China's entrenched habit of free usage poses a significant barrier.
DeepSeek is free, open-source, and already covers a wide range of use cases, further squeezing the room for latecomers to charge a premium for being "smarter." This means → even if Doubao launches a paid version, its pricing power is squeezed from both sides.
Where is ByteDance shifting its strategic weight?
Roughly two months ago, senior ByteDance executives visited Anthropic; afterward, ByteDance began reallocating AI resources away from consumer products like Doubao and toward enterprise services.
The reference point: Anthropic's Claude Code hit $1 billion ARR within six months of launch, rising to $2.5 billion by February this year — enterprise coding tools monetize far faster than consumer chat apps.
ByteDance's large-model data-review team has expanded from roughly 1,500 to over 3,000, primarily cleaning training data for coding models. Volcengine's MaaS business has been elevated in priority, with top leadership setting a target to 10× revenue and accelerate international expansion.
Can Seedance's growth hold?
Seedance's revenue growth has reportedly begun to slow; Kuaishou's Kling saw its ARR reach $500 million by early May before its growth rate also decelerated.
Video generation's current revenue is concentrated in the short-drama industry. Whether it can expand to a broader enterprise client base hinges on continued model-performance gains and whether the supply of Nvidia processors — a scarce resource — can keep up.
The 70% gross margin itself carries an accounting caveat: ByteDance builds its own data centers and buys its own chips, and the choice of chip depreciation schedule materially affects reported costs — a debate with no settled answer across the global AI industry.
Content is for reference only, not financial advice.