Physix Frontier · News Briefing Card (TMTPost · Oct 15, 2026)

Corpus Licensing Pricing Stalls Without Market Mechanism in China

KEY FACTS

HarperCollins licensed Microsoft to train AI on nonfiction books at $5,000 per title for three years, split evenly between publisher and author at $2,500 each.
Anthropic paid a $1.5 billion settlement for downloading over 7 million pirated e-books to train its models, covering roughly 480,000 to 500,000 works.
Goldman Sachs forecasts China's AI model API and subscription revenue will grow from 35 billion yuan in 2026 to 879 billion yuan in 2030.
As of May 2026, more than 130 A-share listed companies disclosed data resources on their balance sheets, totaling about 4 billion yuan, with the publishing sector accounting for a negligible share.
In China's first generative AI infringement case, the Hangzhou court ruled that using works during the training phase may qualify as fair use, but the boundaries remain unclear.

KEY DATA

$5,000HarperCollins per-title three-year licensing fee

$1.5 billionAnthropic settlement

over 7 millionNumber of pirated e-books downloaded by Anthropic

66 million yuan2025 A-share media sector data resources book value

PHYSIX OBSERVATION

Overseas corpus prices come from voluntary deals or infringement damages and cannot be applied directly to China. Domestic buyers have weak willingness to pay and face high compliance risk, while sellers benchmark against overseas pricing, creating a severe mismatch in expectations. Without a trading platform and pricing benchmark, corpus has a price but is hard to transact. The key to breaking the deadlock is not shouting higher prices, but first building the infrastructure for rights confirmation and compliance.

Source: TMTPost report