event-trade登录编辑
📰 DigiTimesDigiTimesDoubao · 实盘不做把握 7/10不做原因:标的无直接敞口
北京时间·🕒 发布 12:17:12·⟳ 抓取 12:21:51 +4m39s·⚡ 决策 10-27 18:26:25 +6h04m·9.3 小时前

DigiTimes复盘AI token用途之争:中文在线称55%用于短剧/视频生成遭豆包否认,文章同时指AI短剧90%厂商亏损

这是一篇行业评论和统计口径之争,不含任何上市公司新增订单、业绩指引、政策或产品商业化事实;被点名的中文在线只是争议数据披露方且引用旧数据,豆包所属字节未上市,快手仅为行业背景,无明确方向与正期望。

📄 原文DigiTimes↗ 原文
China AI token debate highlights challenge of monetizing compute
More than half of China's AI token usage is reportedly concentrated in short dramas and video generation, but ByteDance…
不做 中文在线(300364.SZ)
公司仅为争议统计的披露方,引用2026Q1旧数据且被豆包公开反驳,无新增公司级利好或实质财务利空,方向不明、无交易edge。
不做 快手(01024.HK)
文中仅以可灵V3的token消耗作为行业背景,无公司新增订单、产品或业绩事实,不构成可交易催化。
展开详情(成交明细 / 研究档案 / 原文全文 / 决策记录链)

研究档案

结论摘要

DigiTimes文章汇总「55% token用于短剧/视频」与豆包「仅占约10%算力」的口径分歧,并强调AI短剧九成厂商亏损,属行业讨论而非公司级催化;中文在线数据被质疑但无新增财务损害,字节未上市,缺乏可交易直接标的与方向,故回避。

🔎 新鲜度:最早出现 2026-10-27 · 扩散 medium · price-in 风险 low ↗

关键理由

  • 文章无任何上市公司新增可量化利好或利空:核心是token口径分歧,55%数据来自单一厂商Q1内部估算并被豆包公开否认。
  • 九成AI短剧厂商亏损是行业结构性背景,不指向特定上市标的的新增盈利下修或订单变化,无法形成做多或做空edge。

主要风险

  • 若市场将九成厂商亏损、营销占收入80%-90%解读为AI内容/短剧板块情绪利空,相关题材股可能短线波动,但该数据为行业统计、非个股新事实,方向与幅度均不可靠。
  • 中文在线作为争议数据披露方或承受舆论质疑,但其自身业务敞口未量化,贸然做空证据不足。

反向证据

  • 做多前查反向 · 来源 web · 日期 2026-10-27 · 影响 weakens
    文章未给任何上市公司新增订单、收入或指引,反而强调九成厂商亏损、变现困难,缺乏做多证据。
    检索:AI短剧 算力 受益 中文在线 快手 可灵 订单 业绩 原文 ↗
  • 做空前查反向 · 来源 web · 日期 2026-10-27 · 影响 weakens
    55%数据被豆包反驳属统计口径分歧,非中文在线自身财务暴雷或指引下修,不满足强做空催化。
    检索:中文在线 300364 AI短剧 亏损 数据被否认 业绩下修 原文 ↗

先例如何影响了这次决策

对本次决策的影响:未产生任何交易腿;确认应回避而非凭题材联想介入。

方向:无适用先例:本条为行业统计口径之争与评论性报道,不属于订单、业绩、政策或产品类有明确价格锚的事件,强行套先例反而失真。

涨幅量级:不足以校准幅度:没有以某一上市公司为主体的新增事实,也无可比的同市场同相位事件。

形态:不适用

工具执行结果 · 成功 1 · 降级 0 · 失败 0 · 取消 0
  1. search_processed_events 成功 stock_tools HTTP · 388 ms
    {"result": "search_processed_events:agent query=`ai 短剧 token 用量 中文在线 投资者交流会 | chinese online 中文在线 ai token 短剧 视频生成 55% | 豆包 字节跳动 视频生成 算力 10% 否认`;结构化索引覆盖完整 72 小时;旧记录兼容扫描近 68.2/72 小时、3826 条本臂决策(完整),返回 0 条候选。未命中只表示这些 agent query 表达式没有召回候选,不能据此证明消息首发或不是重复;按当前事件继续正常研判。"}

原文全文

展开原文 · 3,726 字
More than half of China's AI token usage is reportedly concentrated in short dramas and video generation, but ByteDance's Doubao has disputed the figure, saying video-generation models account for only about 10% of overall computing resources. The conflicting estimates highlight the difficulty of assessing AI compute consumption and its commercial value.

According to reports from Hankyung and Chinese media outlet Caixin , Chinese digital publishing company Chinese Online disclosed at an investor briefing that 55% of China's AI token usage comes from short dramas and video generation. Marketing, including e-commerce and advertising, accounted for 24%, while software development, including coding, represented just 15%.
Notably, the China Academy of Information and Communications Technology (CAICT) also cited the same figures in a report.
After the statistics surfaced, online commentators responded with self-deprecating criticism, claiming that the US uses 58% of its tokens for coding and business automation, while China spends half of its computing resources on entertainment.
Doubao, ByteDance's AI service, directly rejected the comparison in an official Weibo post, emphasizing that training and inference for video-generation models account for only about 10% of overall computing resources.
Two yardsticks, two answers
The discrepancy stems from different measurement standards. Chinese Online measured processed tokens, which are primarily reflected in output, while ByteDance cited the proportion of total computing-resource allocation.
In addition, Chinese Online's figures were estimated from internal data supplied by a single AI short-drama operator for the first quarter of 2026, combined with some external data. Critics therefore question whether they can be regarded as authoritative statistics for China as a whole.
Analysts have also pointed out that AI model pretraining consumes enormous computing power but generates little visible output in the form of tokens. Token share alone therefore cannot determine how overall computing resources are being used.
Video's heavy token appetite
There is little disagreement, however, that short-video production consumes astronomical quantities of tokens and computing power. Generating a five-second, 1080p silent video requires about 80,000 tokens per second, and creators often need to generate the footage repeatedly to improve the final result.
According to data from Tencent's Token Hub cited by Hankyung , that figure applies to Kuaishou's Kling Video V3, while generating a single image requires about 20,000 tokens.
Doubao's own usage has also continued to expand. ByteDance's Volcengine said daily token usage for Doubao models reached 180 trillion in June, more than 10 times the level of a year earlier.
Explosive growth, few winners
China's AI short-drama market is also experiencing explosive growth. Yet the high training and inference costs of advanced video models have brought the structural challenge of recovering costs and achieving profitability to the forefront.
Data from Yicai and data company DataEye showed that about 221,900 AI and animation short dramas were newly launched on Douyin in the first half of 2026. Only 1,055 of them, or 0.48%, accumulated more than 100 million views.
About 80% of the productions received fewer than 1 million views, while fewer than 1.3% surpassed the 50 million views considered the break-even point.
Although AI has reduced filming and editing costs, the savings have largely shifted to traffic-purchasing and marketing expenses. Marketing accounts for 80% to 90% of revenue; about 90% of producers are operating at a loss, and companies have begun laying off staff and exiting the market.

决策记录链 · 只追加

每次领取产生一条记录,重新决策追加一条并以 supersedes 指向被替代的记录(ADR 0014)。

  1. 867eb03560 · 2026-10-27 18:26 · doubao / doubao-seed-evolving-latest-version · 策略包 b7b4acf827b8当前