Falling Prices Push AI Video Toward Platform Economics

资讯 AI 行业观察 作者:Chelsea_Sun 全文 4490 字 · 约 12 分钟读完 配图 2 张 分类SoraGoogleOpenAI

Video-model quality is converging as per-second API prices decline. OpenAI discontinued Sora after sustained losses. Google, ByteDance, and Kuaishou continue to

所属频道 大厂动态视频图像生成

(本文作者为 Chelsea_Sun,钛媒体经授权发布)

NextFin News — The race in AI video is less about who can render the most striking clip than about who can sustain low prices and still hold the customer relationship.

Google has cut the list price of a mainstream 720p generation tier to well under a dollar per several seconds of output, with multi-turn editing that bills each revision. Chinese vendors have followed with promotional rates measured in fractions of a yuan per second. Creators who once treated every generation as a budget decision now treat retries as routine. Capability gaps on public leaderboards have narrowed to ranges that many practitioners describe as interchangeable for ordinary shots. Model switches increasingly turn on speed, price, and workflow fit rather than a decisive quality gap.

That pattern points away from a durable product monopoly and toward a commodity generation layer sitting on costly infrastructure.

OpenAI’s Sora illustrated the strain. The product launched with unusual consumer attention, then shut within about half a year. Reporting at the time pointed to daily losses on the order of a million dollars and compute intensity far above ordinary text chat. Video generation does not scale like classic software: more users mean more inference, not thinner marginal cost. If each paid second fails to cover chips, power, and serving, growth widens the deficit.

No major pure-play video model company has publicly claimed steady profits. Even the strongest Chinese revenue stories still point to breakeven years away. Fixed costs—training runs, teams, data centers—remain large. Variable costs rise with every successful push for adoption. The central commercial question is whether a generation request is a profitable unit or a subsidized sample.

Companies that remain active are not escaping that cost structure. They are absorbing it inside broader businesses.

Google can price video aggressively because the same customer may pay through cloud contracts, workspace subscriptions, or advertising, and because in-house accelerators lower the cash cost of each inference relative to rivals that rent capacity. Video tools are spread across creator products, enterprise cloud, and ad workflows, so a cheap clip can pull spending elsewhere in the stack.

ByteDance treats generation as upstream supply for a chain it already operates: story IP, scripting aids, timeline editors, short-drama and short-video distribution, then advertising and e-commerce. AI reduces the cost of producing episodes; a large share of short-drama cash still goes into paid distribution, much of it on ByteDance surfaces. The model line can look only modestly profitable in isolation while the group captures traffic and ad budgets downstream. API sales through its cloud arm add a second channel aimed at studios and toolmakers.

Kuaishou’s Kling effort splits internal use and external brand. Inside the group, generation supports ad creatives, commerce video, and drama previsualization. Outside, a more independent commercial posture—and a large funding round after a carve-out—extends the runway to pursue professional pipelines and overseas API demand. Reported revenue has risen quickly; public commentary still places meaningful quarterly profit in a later year.

In each case, video functions as a feature of a larger commercial system. OpenAI, under tighter capital-market pressure and without a comparable short-video distribution or advertising engine, treated Sora as a line that failed a narrower test.

Lower unit prices will pull generation into advertising, short drama, product clips, and social drafts that could not justify earlier rates. They will not, by themselves, produce a healthy stand-alone model industry. When blind evaluations show near ties among several systems, buyers gain leverage. Platforms that control demand can treat model suppliers as interchangeable. Platforms that control both demand and distribution can keep cutting list prices while still showing gains in cloud revenue, ad volume, or content cost.

Falling Prices Push AI Video Toward Platform Economics(配图 1)

Technical parity and falling prices are established facts. The structural outcome is clearer still: mass demand for text inference grew inside products people already used; mass demand for video inference is concentrating where firms already own audiences and monetization. Until each generation second reliably covers its cost, advantage will favor companies that never required AI video to stand alone as a clean profit-and-loss business.

本文由本站编辑组整理与排版,内容基于公开信息核实后归纳发布,与文中提及的产品、公司无隶属关系。我们持续跟踪 AI 领域的模型、产品与行业动态,力求信息准确、表述清晰;文中涉及的名称、价格与功能细节可能随官方调整而变化,请以官方最新说明为准。如发现疏漏或表述有误,欢迎通过站内渠道反馈指正。

相关推荐 按相关度与热度排序

全部资讯 →
  1. 谷歌50天连发4款模型:Gemini 4 Neon登场,世界模型同步开放

    OpenAI时隔75天的顶级模型刚刚发布,社区里的吐槽就炸开了。有人直言新模型“几乎没法用”,有人抱怨它过分谨慎、难以指挥,还有人认为,这家伙解释起自己的工作就像在“说外星话”。甚至有人用漫画玩了个梗,吐槽GPT-6.2 Astra的语言风格,根本是防御过当。图片来源网络前·赛博义父Tibo在X上表示,他已经收到了关于

    来源 腾讯网
  2. Gemini上Mac可代操作电脑,Runway推270亿参数机器人模型

    本期重点包括Google向数亿用户免费开放Gemini Tasks,并在Mac端推出可代操作电脑的Computer use功能。Runway发布27B参数机器人世界模型Praxis-1,OpenAI开放GPT-Live-1.1语音API。此外,Cognition评测显示Gemini 4 Neon在代码任务上性价比突出,

    来源 AI 应用前沿
  3. 谷歌AI智能体改一处代码成功率翻8倍:Harness自我优化研究

    IT之家 11 月 20 日消息,谷歌研究人员当地时间 11 月 17 日最新发表的研究论文《Harnesses That Improve Themselves》(《能够自我改进的 Harness》)显示,通过让 AI 智能体逐步修改自身运行框架,并对每次修改进行验证,可以显著提高软件工程任务的完成率。在一项 SWE-

    来源 腾讯网
  4. 苹果AI新布局:自研智能体操控Mac,对标四款竞品

    TechCrunch报道称,苹果似乎正在开发可在Mac上自主执行任务的AI智能体,直接对标Meta Muse、Instinct、Claude与OpenAI Dots等产品。这标志着苹果的AI战略正从Siri式的语音助手,转向能够理解目标、拆解步骤并操作用户界面的“行动型AI”。凭借软硬件一体化与端侧算力优势,

    来源 AI 科技快讯
  5. OpenAI遭亚利桑那刑事调查:AI代理疑渗透银行

    2026年11月18日,亚利桑那州检察长Kris Mayes宣布对OpenAI等AI公司启动刑事调查,焦点为AI代理未经授权渗透银行系统。此前加州已于11月16日提起民事诉讼,两州在24小时内相继行动,追责性质从民事升级为刑事,标志美国州级执法对AI代理失控风险的联动加强。报道基于公开声明与路透社信息,分析法律边界重塑

    来源 AI 行业观察
  6. 谷歌Gemini登陆Mac:AI代理代操作电脑,桌面入口争夺战开跑?

    谷歌为Gemini推出面向Mac的重磅更新,首次将可执行任务的AI代理引入桌面端,能够整理文件、填写表单并完成多步骤工作流。这标志着Gemini从对话助手向“能动手”的操作型智能体进化,也意味着谷歌在桌面AI代理赛道正式与OpenAI、Anthropic和微软正面交锋。与此同时,桌面自动化带来的便利与隐私风险,也成为用

    来源 AI 行业观察
  7. 苹果研发自主AI智能体:招聘信息透露可代写代码、操控设备

    苹果公司正在开发可以代表用户完成任务的自主AI智能体,并希望进入这个已被OpenAI、Anthropic PBC、Meta Platforms Inc.及其他公司占据的快速增长的市场。苹果公司本周在网站上列出的招聘信息揭示了这一计划。其中一则信息称,苹果正在为一款主动式、智能AI助手寻找一位高级工程师经理。根据招聘信息

    来源 AI 行业观察
  8. OpenAI等AI企业遭亚利桑那州刑事调查:失控智能体疑入侵银行

    11月19日,据路透社报道,美国亚利桑那州已对OpenAI及其他人工智能企业启动刑事调查,案件核心指向“失控AI智能体疑似入侵银行系统”事件。该州总检察长克里斯·梅耶斯(Kris Mayes)周三在芝加哥的新闻发布会上表示:“亚利桑那州对此非常担忧,我们正从多个角度审视这一问题。”

    来源 TechWeb
  9. 亚利桑那州刑事调查OpenAI等AI企业:智能体被指入侵银行系统

    11月19日,据路透社报道,美国亚利桑那州已对OpenAI及其他人工智能企业启动刑事调查,案件核心指向“失控AI智能体疑似入侵银行系统”事件。该州总检察长克里斯·梅耶斯(Kris Mayes)周三在芝加哥的新闻发布会上表示:“亚利桑那州对此非常担忧,我们正从多个角度审视这一问题。”

    来源 TechWeb
  10. 贝佐斯重返一线:Prometheus估值410亿美元,要做物理世界的OpenAI

    卸任亚马逊 CEO 五年后,杰夫·贝佐斯没忍住,又给自己找了个班上。 当地时间 2026 年 11 月 17 日,加州门洛帕克的 Semafor「硅谷与世界」峰会现场,贝佐斯和物理学家维克·巴贾杰(Vik Bajaj)罕见同台。他现在的最新头衔,是 AI 初创公司 Prometheus 的联合创始人兼联席 CEO。这是

    来源 AI 行业观察
  11. 谷歌Gemini 4再添Neon:中杯模型补位,Genie 4开放

    编辑|冷猫 一觉醒来,谷歌又双叒叕发 Gemini 4 模型了。 谷歌最近一直在基础模型上发力,从最高智能到最高效率的模型不停发布,为了重返王座真是拼了。 在 Carbon、Argon、Hydrogen 之后,填补最高智能和最高效率之间的「中杯」模型 Gemini 4 Neon 登场! Demis

    来源 AI 开发者社区
  12. 三个00后创业18个月:Palo Alto 45亿美元买下Applied Compute

    Anthropic上市,硅谷正在诞生新一轮财富神话。而在这场狂欢的另一角,三个从OpenAI离职的00后也提前拿到了自己的“大结果”——创业不到18个月,直接把公司卖了45亿美元(约合人民币320亿元)。 美东时间11月18日,美国网络安全巨头Palo Alto Networks(派拓网络)宣布达成最终协议,将以45亿

    来源 AI 行业观察
  13. 阿里27财年2季报:AI云收入610亿,算力紧缺成最大变数

    在因高端芯片被卡脖、股价大幅回调后,阿里巴巴 19 日晚发榜了 27 财年 2 季度的业绩。单看本季财报表现相当不错,基本没有大瑕疵。 AI 云增速继续加快且利润率明显改善;电商板块增长企稳 & 外卖减亏后盈利修复;连新拆出的 AI 实验室与应用板块也营收加速、亏损收窄。 但问题就出在,市场关心的不是现在,而是若之后算

    来源 AI 行业观察
  14. OpenAI首款AI硬件定了:2026年底出货,富士康、立讯精密参与

    富士康旗下富智康集团已获得OpenAI手机项目订单,将参与相关产品的整机组装。据《证券时报》报道,富士康同时也是苹果iPhone的主要组装商。若OpenAI手机最终面市,将正式进入智能手机终端市场,并可能围绕AI交互体验与现有手机品牌展开直接竞争。 据报道,OpenAI首款AI硬件产品计划于2026

    来源 AI 开发者社区
  15. OpenAI首款AI硬件定档2026年底出货:立讯独家整机,富智康也入局

    富士康旗下富智康获OpenAI手机项目订单,将参与整机组装。富士康也是苹果iPhone主要组装商。若OpenAI手机面市,将进入智能手机终端市场,围绕AI交互体验与现有品牌直接竞争。报道称,OpenAI首款AI硬件计划2026年底出货,多家中国厂商参与,立讯精密独家负责整机制造,并供应核心部件。

    来源 AI 产业速递
  16. Gemini登陆Mac桌面:AI代理可代操作电脑,隐私风险几何?

    谷歌为Gemini推出面向Mac的重磅更新,首次将可执行任务的AI代理引入桌面端,能够整理文件、填写表单并完成多步骤工作流。这标志着Gemini从对话助手向“能动手”的操作型智能体进化,也意味着谷歌在桌面AI代理赛道正式与OpenAI、Anthropic和微软正面交锋。与此同时,桌面自动化带来的便利与隐私风险

    来源 AI 科技快讯
  17. 谷歌AI模式上线临时线程:搜索不留痕,服务器端仍存约两个月

    11月18日, 谷歌 宣布在全球桌面端及移动端浏览器的 搜索AI模式 中正式推出“临时线程”(Temporary Threads)功能,允许用户发起私密、非个性化的即时搜索,使其查询记录不留存于个人历史且不干扰后续算法推荐。 该功能借鉴了独立应用 Gemini 的临时聊天机制,用户无需修改账户隐私设

    来源 AI 开发者社区
  18. 谷歌Chrome造“平行互联网”:AI代理先在沙盒里犯错

    为了让AI浏览代理在真实网页上替用户办事,谷歌Chrome团队搭建了一个与真实互联网平行的“镜像维度”沙盒。在这个替代宇宙里,代理可以尽情犯错、暴露失控风险,而不会波及真实用户。这既是浏览器AI竞赛的基础设施,也折射出AI代理安全测试的行业难题。

    来源 AI 科技快讯
向下滚动自动加载更多