According to a recent report, the Doubao application, used daily by over 200 million people, is currently generating less than one million yuan in daily revenue, primarily sourced from e-commerce commissions.
As of May this year, the daily operational costs for the Doubao app have reached tens of millions of yuan.
While text chat functions incur relatively lower costs, multimodal features such as reasoning, image recognition, voice chat, and video chat demand computing power that is several times to dozens of times more expensive.
In response to these dynamics, ByteDance is planning to increase its capital expenditure for 2026 by over 200 billion yuan, a figure representing approximately 60% of its projected 2025 profits.
Following a visit by ByteDance's senior executives to Anthropic two months ago, the company has begun adjusting its AI resource allocation.
The strategic focus is shifting from consumer-facing products like Doubao towards enterprise-oriented services.
Over the past six months, Anthropic has demonstrated that AI programming can yield returns on massive infrastructure investments.
Its Claude Code, launched in May of last year, achieved an annualized revenue run rate of $1 billion within six months, which doubled to $2.5 billion by February this year, propelling its valuation to $96.5 billion.
ByteDance's own enterprise service, Seedance, has also proven the viability of this path.
The video generation model has reached an annualized revenue (ARR) of $2 billion (approximately 14.3 billion yuan), with single-month revenue exceeding 1 billion yuan.
The vast majority of Seedance's revenue originates from corporate clients.
Currently, Seedance boasts a gross margin of 70%.
For every 10 yuan generated from API calls, server and inference costs account for only about 3 yuan.
In comparison, language models typically involve more versions, more intensive iteration cycles, and longer training periods, with overall training costs potentially being 3 to 5 times higher than for video models.
Seedance is currently concentrating its efforts on a primary video generation model, allowing its training investments to be more easily amortized by subsequent revenue.
After nearly a year of optimization, Seedance 2.0 has become the first video generation model to fully adopt a Mixture of Experts (MoE) architecture, featuring 200 billion parameters.
Since its launch in February 2026, it has rapidly established itself as the world's highest-performing video model.
Comments