Seedance Cut a Minute of Video to RMB 1,000. iQiyi Spent RMB 42 Billion

In 2006 the online rights to one episode of a Chinese sitcom sold for RMB 1,250. In 2011 a palace drama fetched RMB 300,000 an episode. By 2017 a costume epic reached RMB 9 million, or 1.5 times what the two satellite channels paying for it gave combined.

Those prices were the premium platforms paid for scarce quality under fierce bidding. In 2026 the premise is failing, after three years of AI video model iteration.

A film and television studio with a camera on a tripod and soft lighting on an empty stage
Video generation has moved from short clips to complete sequences in three years (Source: Sohu Tech)

Three years of change

In 2024 a model produced five silent seconds, and characters deformed when they moved. In 2025 mainstream models moved from UNet to the Diffusion Transformer architecture, and video generation began to follow scaling laws. More parameters, data and compute improved stability. Clips ran to fifteen or twenty seconds and character consistency improved, though scene changes and multiple people still caused drift.

In February 2026 Seedance 2.0 generated sound and picture together. The rate of usable takes rose from 20 per cent to 90 per cent. Six months later Seedance 2.5 extended a single generation from 15 seconds to 30, enough to complete setup, turn and resolution in one pass, with output ready for the editing room.

The cost structure follows. Compute is billed by the second and entry cost has fallen from yuan-scale to jiao-scale in a year. A minute of AI simulated-human short drama now costs RMB 800 to 1,000.

Feng Ji, founder of Game Science, argues that ordinary video production can no longer follow film-industry cost logic and will approach the marginal cost of compute. He predicts unprecedented inflation in content and a complete restructuring of traditional organisations and workflows.

Twenty years without a profit model

When the bottleneck moves from crews, locations and actors’ schedules to compute and prompting, the nature of content changes. It stops being a craft few teams can complete and becomes output that can be test-produced in batches.

Long-video platforms paid huge sums for exclusive rights because quality supply was scarce. That moat is being pierced by compute.

User-generated video created the first wave from 2005. It was cheap because users made and uploaded it, but fragmented and without exclusivity, so platforms could not charge. Pirated series became the main content until regulation tightened in 2009 and more than 500 BitTorrent sites closed that year. A drama episode that was worth RMB 4,000 at the start of 2009 averaged RMB 80,000 by December, and a new Romance of the Three Kingdoms asked for RMB 280,000. A seventy-fold rise in one year.

From 2015 to 2021 iQiyi’s paid members grew from 10.7 million to 119 million, while cumulative net losses reached RMB 42 billion. The company posted its first full-year profit in 2022, with non-GAAP net income of RMB 1.3 billion, earned by cutting costs rather than growing revenue.

The 2023 hit The Knockout pushed members to a record 129 million, with revenue of RMB 31.9 billion and non-GAAP profit of RMB 2.84 billion. Then the line turned down. Revenue was RMB 29.23 billion in 2024, down 8 per cent, and RMB 27.29 billion in 2025, down 7 per cent. In the first half of 2026 revenue was RMB 12.513 billion with a net loss of RMB 582 million. Tencent Video and Youku have been almost erased from their parent companies’ reports.

The rights price that rose from RMB 1,250 to RMB 9 million was a bet on scarcity. A per-second compute bill is a bet on abundance, and the two cannot coexist in the same business model.

Editor’s note: This is an adapted translation of the original Sohu Tech report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.sohu.com/a/1102566270_116132.

Leave a comment