Several sets of statistics released recently are worth reading side by side.

One overseas data site estimates that close to half of all tokens consumed worldwide now run on Chinese models.

Take weekly figures from OpenRouter, the most widely used model-routing platform outside China: in the week of October 5 to 11, 2026, the world burned through 172 trillion tokens, of which Chinese models accounted for 86.74 trillion — more than 50%.

Meanwhile, an investor relations filing published by Chinese Online in June breaks down domestic token consumption as follows: "AI short dramas and video generation" at 55%, e-commerce and marketing at 24%, and software development at 15%.

Half of China's tokens, in other words, are going into feel-good melodrama.

Just how big short dramas have become is hard to miss. As of October 2026, Hongguo Short Drama had 220 million daily active users, more than Tencent Video, iQiyi, Youku and Mango TV combined, with users spending an average of 125 minutes a day in the app.

WeChat has now entered the ring itself with "Green Bubble Short Drama," a free mini program. Its debut AI-made series, "Best Actress Yaya," racked up hundreds of millions of views within three days of launch. WeChat says average viewing time per user on Green Bubble titles runs five times the platform-wide average, with retention more than 2.5 times higher.

From Hongguo to Green Bubble, what the short-drama platforms are fighting over is not just audiences, but a steady pipeline of content.

Short Dramas Become a Token Furnace

Video is heavy industry for tokens by nature.

A text model writing a line of dialogue handles a few hundred or a few thousand text tokens. A video model has to handle continuous frames, motion, lighting, sound, and character consistency.

ByteDance's BytePlus pricing documentation for Seedance shows video token consumption rising with duration, resolution, frame rate and frame dimensions. Chinese Online's filing makes a similar point: a 1080p AI live-action-style drama typically consumes two to three times the tokens of a 720p AI animated one.

The party actually paying for all those tokens is the platforms' subsidy budget.

Short dramas already run on a mature assembly line: web-novel IP upstream, production and traffic-buying teams in the middle, and advertising, paid viewing, e-commerce and overseas distribution downstream. AI doesn't have to reinvent an industry — it just plugs in and gets to work.

Platforms, for their part, need enough content to fill their recommendation feeds, so they are competing with guaranteed payments, revenue splits, traffic and production subsidies to attract creators.

Hongguo has gone the free route: no payments, no memberships, just a dozen or so seconds of advertising to unlock the next episode — the so-called IAA model. Platform revenue comes from ad inventory, and ad inventory is tied directly to user time spent. In that model, "buying content" is "buying traffic," which makes computing power cost a customer acquisition cost.

That pushes subsidies to considerable scale: the short-drama rights arm of ByteDance's Douyin group has set aside more than 1.5 billion yuan in guaranteed support for live-action short dramas in 2026, with the average guarantee per title up roughly 60% from the previous year.

WeChat's Green Bubble Short Drama works on similar logic. Producers plug their series into the official mini program, viewers watch free, and revenue comes mainly from advertising. Under WeChat's incentive policy for the fourth quarter of 2026, strong new titles that hit 10 minutes of average viewing time per user can earn up to a 100% revenue bonus.

The result is a loop: platforms subsidize content, content producers buy traffic, users watch ads, and ad revenue funds still more content.

The tokens short dramas burn have become the fuel for that traffic loop.

At the same time, the unit price of video generation is collapsing.

TikTok CEO Shou Zi Chew said in mid-October that producing a video ad used to cost $100 to $1,000; with AI, the cost can now fall to $1 to $10. When costs drop, supply explodes.

China's National Radio and Television Administration has also disclosed that 430,000 micro-short dramas launched nationwide in the first eight months of 2026, 13 times the total for the whole of the previous year, with AI-made titles accounting for more than 90%. That works out to nearly 1,800 new titles a day.

Platform subsidies, though, are shifting from blanket cash to bets on premium titles only. Since the start of 2026, several producers have said Hongguo has tightened guarantees for some small and mid-sized projects, moving more of them to performance-based revenue sharing. Hongguo subsequently denied it had "scrapped guarantees across the board," stressing that strong scripts can still get support.

As subsidies concentrate on a small number of premium titles, the many production teams that survived on guarantees are left carrying the risk that nobody watches.

Chew also had cold water to throw at the same event. He noted that the number of creators making AI-generated content far exceeds viewership for AI content. AI content currently accounts for only single-digit percentages of total views, he said, and people still place more value on what is real and original.

A short drama that ends up with no audience still goes through script generation, storyboarding, voiceover and dozens of regeneration attempts — consuming a large volume of tokens along the way.

High token consumption, then, can signal booming demand. It can also just mean the supply side is doing an equally booming amount of trial, error and speculation.

The American Mirror: More Than Half of Tokens Go Into Code

By contrast, data suggests that more than half of US token usage goes to programming and technical R&D.

OpenRouter and a16z analyzed more than 100 trillion tokens routed through the platform. At the start of 2025, coding requests made up just 11% of the total; by year's end they had passed 50%, making it the fastest-growing use case.

OpenAI's enterprise data points the same way: as of June 2026, Codex accounted for 64% of output tokens generated by enterprise customers across Codex and ChatGPT.

At first glance, this looks like a return to a familiar storyline: Americans using gunpowder to make weapons, Chinese using it for fireworks.

But there's a problem with that comparison: the numbers don't come from the same ledger.

The claim that "55% of China's tokens go to short dramas and video generation" describes the mix of applications inside China. The claim that "coding is more than 50%" measures global traffic on overseas platforms like OpenRouter. The first mainly reflects Chinese demand; behind the second sit users from all over the world.

The directional difference is real. Domestic token consumption skews toward consumer use cases like short dramas, e-commerce and marketing, while traffic at the leading US model companies comes more from coding, agents and enterprise automation.

OpenAI, Anthropic and other US companies serve a global market. When Chinese programmers write code with Claude, or Chinese companies call OpenAI for their products, those tokens land in the coding and enterprise columns of US model companies. The reverse holds too: overseas developers calling Qwen and DeepSeek add to Chinese model usage.

So a large share of the enterprise traffic on US models comes from companies worldwide, while the consumer traffic visible in China's public data comes disproportionately from its own vast domestic market.