我的征尘是星辰大海。。。
The dirt and dust from my pilgrimage forms oceans of stars...
-------当记忆的篇章变得零碎,当追忆的图片变得模糊,我们只能求助于数字存储的永恒的回忆
作者:黄教授
手机视频列表
AI时代的小公司诅咒
视频
音频
原始脚本
AI 时代的小公司诅咒,为何互联网的成功经验在今天全成了死路?当互联网创业者还在怀念一台服务器、几行代码就能撬动百万用户的黄金时代时,AI 赛道的小公司正陷入一场无声的绞杀。 三个月烧光融资,核心人才被巨头挖走,好不容易落地的客户拒绝数据复用。 这种做一单亏一单,做一个死一个的困境,不是能力问题,而是 AI 时代给小公司设下的系统性诅咒。 从成本、人才、数据到客户,每一个环节都布满了互联网时代从未有过的死结。 一、成本诅咒,从边际成本趋近于0,到每多一个用户就多一分亏损。 互联网时代的小公司靠免费拉流量、广告变现的模式就能活下来。 一台 Web 服务器成本几千元,能服务上万用户,用户越多,单用户成本越低。 但 AI 时代的成本逻辑彻底颠覆了这种规模红利。 大模型的成本是双重刚性的。 一方面是算力硬件成本,A100H100显卡的显存成本是普通内存的100倍,单卡时租20~50美元,一次复杂推理需要占用8张显卡。 成本直接突破160美元。 存储成本更吓人,一个千亿参数模型的权重文件就有数百 GB,长期托管需要专用存储服务器,年成本超10万元。 另一方面是人力维护成本,模型微调需要算法工程师,年薪50万加。 标注专家月薪3万加,硬件运维人员月薪2.5万加。 一个3人小团队的月人力成本就超10万,还不算数据采购、合规审核的费用。 更致命的是,AI 的边际成本根本降不下来。 互联网时代,一台服务器服务1万人,用户从1万增长到10万,服务器成本可能只增两倍。 但 AI 时代,用户从100增长到1000,GPU 服务器成本几乎线性增长。 因为模型推理是计算密集型工作,每多一个并发用户,就必须额外投入一份显卡资源,不存在闲置资源复用的空间。 有创业者算过一笔账,做一个 ToC 的 AI 绘画工具,单次生成成本0.5元。 若免费开放,每天1万用户调用,月亏损就达15万。 用户破10万时,月亏损超150万,根本撑不到变现那天。 这种成本结构直接堵死了小公司用小钱撬动大 大市场的路。 互联网时代可以先亏后赚,但 AI 时代只能先亏后死。 光是维持模型运行的算力成本,就足以让小公司在用户破万前烧光所有资金。 二,人才诅咒,懂 AI 加懂行业的人要么稀缺。 要么请不起 AI 时代的创业,缺的不是写代码的程序员,而是既懂 AI 技术又懂行业规则的复合型人才。 但这类人才对小公司来说几乎是奢侈品。 首先是 AI 人才荒,国内算法工程师缺口超50万。 能独立完成模型微调,解决推理优化的资深工程师,年薪普遍80万加,还要求股票期权、弹性办公。 小公司哪怕开出百万年薪,也很难吸引到合适的人。 巨头的 AI 团队能提供千亿参数模型训练经验,全球顶会发表机会。 而小公司只能给几个人几台显卡的简陋条件,人才用脚投票的结果显而易见。 更难的是行业加 AI 的跨界人才。 做医疗 AI 需要懂影像诊断的 AI 工程师,既考过执业医师资格证,又会用 PyTorch 调参。 这种人全国不超过1000人,大多在医院或头部药企,根本不屑于去小公司。 做法律 AI 需要懂合同法的 AI 产品经理,既知道合同审核的风险点,又能设计模型的推理逻辑。 这种人要么在顶尖律所拿百万年薪,要么被大厂抢走,小公司连面试的机会都没有。 更残酷的是,行业内的高收入人群没有动力帮小公司做 AI 律师、医生、芯片工程师这些高人力成本岗位,本身拿着高薪。 AI 替代的是他们的重复劳动,比如律师检索法条、医生看常规 CT 片。 他们凭什么花时间教小公司行业规则,给数据标注?除非公司层面推动,但大公司又倾向于自己做 AI 不会把机会留给小公司。 小公司陷入缺人才,做不出产品,没人付费,更招不到人才的死循环,根本无解。 三,数据诅咒。 行业核心数据是禁区,拿到了也不会用,AI 的核心是数据。 但小公司面临的现实是,高价值行业的数据,要么拿不到,要么拿到了也用不了。 高人力成本领域的核心数据从来不是公开可得的。 律师行业的胜诉案例库、合同审核标准,只在精度、中伦等顶尖律所内部流转。 医疗行业的影像数据、病例资料,受数据安全法限制,只有医院内部团队能接触。 金融行业的风控模型数据、交易决策逻辑,是公司的核心机密,绝不会对外泄露。 小公司想拿数据,要么通过灰色渠道,面临法律风险,要么花高价采购,单份医疗影像 数据成本超10元,百万级数据采购费就超千万,根本承担不起。 哪怕侥幸拿到数据,小公司也不懂怎么用。 数据标注不是贴标签那么简单,工业质检的缺陷数据,外行人看就是一个斑点。 但内行人知道这个斑点在哪个位置,多大尺寸才是致命缺陷。 律师合同的标注需要区分效力性条款和任意性条款,没考过法考的标注员根本分不清。 这种隐性知识不身在行业内部根本无法理解,而 AI 模型的微调恰恰需要这种懂行的标注。 有小公司花50万采购了一批工业质检数据,结果因为标注标准错误,模型训练出来后准确率不足50%,钱全打了水漂。 更糟的是数据复用难题,小公司好不容易给一家工厂做了 AI 质检模型。 想把经验复用到另一家工厂时,却发现两家的产品缺陷标准、生产流程完全不同。 之前的数据和模型参数全没用,每个行业、每个客户的需求都是定制化的。 小公司只能做一单,重新训一次模型,根本无法规模化。 这种数据孤岛让小公司永远停留在小作坊阶段,成不了气候。 四,客户诅咒,高价值客户只认巨头,小公司连入场资格都没有。 AI 时代的小公司,想拿到高人力成本领域的客户,比登天还难。 不是产品不好,而是客户的信任门槛和合规门槛太高。 高价值客户,比如医院、律所、金融机构的采购逻辑。 首先看信任背书,他们不关心你的模型准确率多高,只关心有没有成功案例,会不会有风险。 医院不会用小公司的 AI 辅助诊断工具,万一误诊,责任谁担?律所不会用小公司的合同审核 AI 万一漏看条款,损失谁赔?金融机构不会用小公司的风控 AI 万一出现坏账,谁来负责?相比之下,他们更信任行业巨头加 AI 巨头的组合。 比如医院会优先选西门子医疗加阿里云的 AI 影像工具,律所会选金杜律所加通义千问的合同工具。 小公司哪怕产品更好,也没新人背书,连投标的资格都没有。 合规门槛更是小公司的死穴。 医疗 AI 需要 NMPA 认证,周期1~2年,成本超百万。 金融 AI 需要银保监会备案,需要专业法务团队。 工业 AI 需要 ISO 体系认证,流程 成复杂。 小公司既没有足够的资金,也没有专业的合规团队,往往卡在资质审核环节,眼睁睁看着订单被巨头抢走。 有一家做 AI 税务筹划的小公司,好不容易谈下一家上市公司客户。 却因为没有税务师事务所资质,最终被客户放弃,前期投入的研发成本全打了水漂。 更残酷的是客户锁定,高价值客户一旦合作,会要求模型专属化。 比如一家律所会要求小公司把训练好的合同审核模型私有化部署,不准给其他律所使用。 一家工厂会要求小公司删除训练数据中的自家产品信息,防止泄露给竞争对手。 这种定制化加独占性的要求,让小公司根本无法复用经验,只能做一单赚一单,没下一单就饿死,永远成不了规模化企业。 五,破局无望,小公司的唯一出路是沦为巨头的附庸。 AI 时代的小公司,想复制互联网时代从小到大独立上市的路径,几乎不可能。 能活下来的只有两种选择,一种是依附巨头做分包。 比如跟着阿里云做行业 AI 的细分模块,给阿里云的医疗 AI 做影像预处理。 跟着华为做工业 AI 的端侧适配,靠巨头的资源拿数据、接客户,赚点辛苦的分包费。 这种模式下,小公司没有自主权,巨头说换技术路线,小公司就得跟着改。 巨头说降报价,小公司就得跟着降,利润被压得薄如纸。 另一种是聚焦超细分痛点,做小而美。 不做全流程替代,只做某个极细分的环节。 比如律师合同里的图文数据提取工具、工业质检里的某个特定缺陷识别工具。 靠精准解决一个痛点打动小部分客户,赚点服务费。 这种模式下,小公司永远做不大,只能有一单做一单,遇到巨头进入赛道,分分钟被碾压。 说到底,AI 时代的小公司已经失去了互联网时代从0到1建生态的可能。 他们不是独立的创业者,更像是巨头生态里的临时工,靠某一项细分技能谋生。 随时可能被替代。 这种附庸式生存就是 AI 时代给小公司最现实的解药,也是最无奈的诅咒。 结语,AI 不是小公司的风口,而是筛选器。 互联网时代的创业是人人皆可参与的游戏,因为基础设施普惠、壁垒后发。 AI 时代的创业是少数玩家的游戏,因为基础设施稀缺,壁垒前置。 那些还在试图用互联网经验做 AI 的小公司,本质是在用农耕时代的工具应对工业时代的战争,不是努力不够。 而是赛道的底层逻辑已经变了。 AI 不是小公司的风口,而是筛选器。 它筛选掉那些幻想靠轻资产流量变现的创业者,只留下那些愿意依附巨头做垂直细分的务实者。 AI 时代的小公司诅咒,本质是技术进步带来的资源集中,算力、人才、数据、客户全向巨头聚集,小公司只能在缝隙中求生。 这不是暂时的现象,而是 AI 技术重资产属性和移动互联网闭环属性共同决定的,未来只会越来越明显。 对小公司来说,接受附庸式生存,放弃规模化幻想,或许才是 AI 时代最现实的选择。
修正脚本
AI 时代的小公司诅咒,为何互联网的成功经验在今天全成了死路?当互联网创业者还在怀念一台服务器、几行代码就能撬动百万用户的黄金时代时,AI 赛道的小公司正陷入一场无声的绞杀。 三个月烧光融资,核心人才被巨头挖走,好不容易落地的客户拒绝数据复用。 这种做一单亏一单,做一个死一个的困境,不是能力问题,而是 AI 时代给小公司设下的系统性诅咒。 从成本、人才、数据到客户,每一个环节都布满了互联网时代从未有过的死结。 一、成本诅咒,从边际成本趋近于0,到每多一个用户就多一分亏损。 互联网时代的小公司靠免费拉流量、广告变现的模式就能活下来。 一台 Web 服务器成本几千元,能服务上万用户,用户越多,单用户成本越低。 但 AI 时代的成本逻辑彻底颠覆了这种规模红利。 大模型的成本是双重刚性的。 一方面是算力硬件成本,A100H100显卡的显存成本是普通内存的100倍,单卡时租20~50美元,一次复杂推理需要占用8张显卡。 成本直接突破160美元。 存储成本更吓人,一个千亿参数模型的权重文件就有数百 GB,长期托管需要专用存储服务器,年成本超10万元。 另一方面是人力维护成本,模型微调需要算法工程师,年薪50万加。 标注专家月薪3万加,硬件运维人员月薪2.5万加。 一个3人小团队的月人力成本就超10万,还不算数据采购、合规审核的费用。 更致命的是,AI 的边际成本根本降不下来。 互联网时代,一台服务器服务1万人,用户从1万增长到10万,服务器成本可能只增两倍。 但 AI 时代,用户从100增长到1000,GPU 服务器成本几乎线性增长。 因为模型推理是计算密集型工作,每多一个并发用户,就必须额外投入一份显卡资源,不存在闲置资源复用的空间。 有创业者算过一笔账,做一个 ToC 的 AI 绘画工具,单次生成成本0.5元。 若免费开放,每天1万用户调用,月亏损就达15万。 用户破10万时,月亏损超150万,根本撑不到变现那天。 这种成本结构直接堵死了小公司用小钱撬动大市场的路。 互联网时代可以先亏后赚,但 AI 时代只能先亏后死。 光是维持模型运行的算力成本,就足以让小公司在用户破万前烧光所有资金。 二、人才诅咒,懂 AI 加懂行业的人要么稀缺,要么请不起。AI 时代的创业,缺的不是写代码的程序员,而是既懂 AI 技术又懂行业规则的复合型人才。 但这类人才对小公司来说几乎是奢侈品。 首先是 AI 人才荒,国内算法工程师缺口超50万。 能独立完成模型微调,解决推理优化的资深工程师,年薪普遍80万加,还要求股票期权、弹性办公。 小公司哪怕开出百万年薪,也很难吸引到合适的人。 巨头的 AI 团队能提供千亿参数模型训练经验,全球顶会发表机会。 而小公司只能给几个人几台显卡的简陋条件,人才用脚投票的结果显而易见。 更难的是行业加 AI 的跨界人才。 做医疗 AI 需要懂影像诊断的 AI 工程师,既考过执业医师资格证,又会用 PyTorch 调参。 这种人全国不超过1000人,大多在医院或头部药企,根本不屑于去小公司。 做法律 AI 需要懂合同法的 AI 产品经理,既知道合同审核的风险点,又能设计模型的推理逻辑。 这种人要么在顶尖律所拿百万年薪,要么被大厂抢走,小公司连面试的机会都没有。 更残酷的是,行业内的高收入人群没有动力帮小公司做 AI 律师、医生、芯片工程师这些高人力成本岗位,本身拿着高薪。 AI 替代的是他们的重复劳动,比如律师检索法条、医生看常规 CT 片。 他们凭什么花时间教小公司行业规则,给数据标注?除非公司层面推动,但大公司又倾向于自己做 AI,不会把机会留给小公司。 小公司陷入缺人才,做不出产品,没人付费,更招不到人才的死循环,根本无解。 三、数据诅咒,行业核心数据是禁区,拿到了也不会用,AI 的核心是数据。 但小公司面临的现实是,高价值行业的数据,要么拿不到,要么拿到了也用不了。 高人力成本领域的核心数据从来不是公开可得的。 律师行业的胜诉案例库、合同审核标准,只在金杜、中伦等顶尖律所内部流转。 医疗行业的影像数据、病历资料,受数据安全法限制,只有医院内部团队能接触。 金融行业的风控模型数据、交易决策逻辑,是公司的核心机密,绝不会对外泄露。 小公司想拿数据,要么通过灰色渠道,面临法律风险,要么花高价采购,单份医疗影像数据成本超10元,百万级数据采购费就超千万,根本承担不起。 哪怕侥幸拿到数据,小公司也不懂怎么用。 数据标注不是贴标签那么简单,工业质检的缺陷数据,外行人看就是一个斑点。 但内行人知道这个斑点在哪个位置,多大尺寸才是致命缺陷。 律师合同的标注需要区分效力性条款和任意性条款,没考过法考的标注员根本分不清。 这种隐性知识不身在行业内部根本无法理解,而 AI 模型的微调恰恰需要这种懂行的标注。 有小公司花50万采购了一批工业质检数据,结果因为标注标准错误,模型训练出来后准确率不足50%,钱全打了水漂。 更糟的是数据复用难题,小公司好不容易给一家工厂做了 AI 质检模型。 想把经验复用到另一家工厂时,却发现两家的产品缺陷标准、生产流程完全不同。 之前的数据和模型参数全没用,每个行业、每个客户的需求都是定制化的。 小公司只能做一单,重新训一次模型,根本无法规模化。 这种数据孤岛让小公司永远停留在小作坊阶段,成不了气候。 四、客户诅咒,高价值客户只认巨头,小公司连入场资格都没有。 AI 时代的小公司,想拿到高人力成本领域的客户,比登天还难。 不是产品不好,而是客户的信任门槛和合规门槛太高。 高价值客户,比如医院、律所、金融机构的采购逻辑。 首先看信任背书,他们不关心你的模型准确率多高,只关心有没有成功案例,会不会有风险。 医院不会用小公司的 AI 辅助诊断工具,万一误诊,责任谁担?律所不会用小公司的合同审核 AI,万一漏看条款,损失谁赔?金融机构不会用小公司的风控 AI,万一出现坏账,谁来负责?相比之下,他们更信任行业巨头加 AI 巨头的组合。 比如医院会优先选西门子医疗加阿里云的 AI 影像工具,律所会选金杜律所加通义千问的合同工具。 小公司哪怕产品更好,也没信任背书,连投标的资格都没有。 合规门槛更是小公司的死穴。 医疗 AI 需要 NMPA 认证,周期1~2年,成本超百万。 金融 AI 需要银保监会备案,需要专业法务团队。 工业 AI 需要 ISO 体系认证,流程复杂。 小公司既没有足够的资金,也没有专业的合规团队,往往卡在资质审核环节,眼睁睁看着订单被巨头抢走。 有一家做 AI 税务筹划的小公司,好不容易谈下一家上市公司客户。 却因为没有税务师事务所资质,最终被客户放弃,前期投入的研发成本全打了水漂。 更残酷的是客户锁定,高价值客户一旦合作,会要求模型专属化。 比如一家律所会要求小公司把训练好的合同审核模型私有化部署,不准给其他律所使用。 一家工厂会要求小公司删除训练数据中的自家产品信息,防止泄露给竞争对手。 这种定制化加独占性的要求,让小公司根本无法复用经验,只能做一单赚一单,没下一单就饿死,永远成不了规模化企业。 五、破局无望,小公司的唯一出路是沦为巨头的附庸。 AI 时代的小公司,想复制互联网时代从小到大独立上市的路径,几乎不可能。 能活下来的只有两种选择,一种是依附巨头做分包。 比如跟着阿里云做行业 AI 的细分模块,给阿里云的医疗 AI 做影像预处理。 跟着华为做工业 AI 的端侧适配,靠巨头的资源拿数据、接客户,赚点辛苦的分包费。 这种模式下,小公司没有自主权,巨头说换技术路线,小公司就得跟着改。 巨头说降报价,小公司就得跟着降,利润被压得薄如纸。 另一种是聚焦超细分痛点,做小而美。 不做全流程替代,只做某个极细分的环节。 比如律师合同里的图文数据提取工具、工业质检里的某个特定缺陷识别工具。 靠精准解决一个痛点打动小部分客户,赚点服务费。 这种模式下,小公司永远做不大,只能有一单做一单,遇到巨头进入赛道,分分钟被碾压。 说到底,AI 时代的小公司已经失去了互联网时代从0到1建生态的可能。 他们不是独立的创业者,更像是巨头生态里的临时工,靠某一项细分技能谋生。 随时可能被替代。 这种附庸式生存就是 AI 时代给小公司最现实的解药,也是最无奈的诅咒。 结语、AI 不是小公司的风口,而是筛选器。 互联网时代的创业是人人皆可参与的游戏,因为基础设施普惠、壁垒后发。 AI 时代的创业是少数玩家的游戏,因为基础设施稀缺,壁垒前置。 那些还在试图用互联网经验做 AI 的小公司,本质是在用农耕时代的工具应对工业时代的战争,不是努力不够。 而是赛道的底层逻辑已经变了。 AI 不是小公司的风口,而是筛选器。 它筛选掉那些幻想靠轻资产流量变现的创业者,只留下那些愿意依附巨头做垂直细分的务实者。 AI 时代的小公司诅咒,本质是技术进步带来的资源集中,算力、人才、数据、客户全向巨头聚集,小公司只能在缝隙中求生。 这不是暂时的现象,而是 AI 技术重资产属性和移动互联网闭环属性共同决定的,未来只会越来越明显。 对小公司来说,接受附庸式生存,放弃规模化幻想,或许才是 AI 时代最现实的选择。
英文翻译
**The Curse of Small Companies in the AI Era: Why All the Success Formulas from the Internet Age Have Become Dead Ends** When Internet entrepreneurs still fondly recall the golden age when a single server and a few lines of code could attract millions of users, small companies in the AI track are trapped in a silent massacre. In just three months, they burn through their funding, their core talent is poached by giants, and the hard-won customers refuse to allow data reuse. This predicament—where every project loses money and every attempt fails—is not a matter of capability but a systemic curse imposed on small companies by the AI era. From costs, talent, data to customers, every link is fraught with dead ends that never existed in the Internet age. **1. The Cost Curse: From Marginal Cost Approaching Zero to Losing Money with Every Additional User** In the Internet age, small companies survived by offering free services to attract traffic and monetizing through advertising. A web server costing a few thousand yuan could serve tens of thousands of users; the more users, the lower the per-user cost. But the cost logic of the AI era has completely overturned this scale dividend. The costs of large models are doubly rigid. On one hand, there is the hardware cost for computing power. The memory cost of A100/H100 GPUs is 100 times that of regular memory; renting a single GPU costs $20–$50 per hour, and a single complex inference requires 8 GPUs, directly pushing the cost beyond $160. Storage costs are even more staggering: a model with hundreds of billions of parameters has weight files totaling hundreds of GB, requiring dedicated storage servers for long-term hosting, costing over 100,000 yuan annually. On the other hand, there is the cost of human maintenance. Fine-tuning models requires algorithm engineers (annual salary 500,000 yuan+), annotation experts (monthly salary 30,000 yuan+), and hardware operations personnel (monthly salary 25,000 yuan+). A small team of three incurs monthly labor costs exceeding 100,000 yuan, not including data procurement and compliance auditing fees. More deadly is that the marginal cost of AI simply cannot be reduced. In the Internet age, one server serving 10,000 users; when users increase from 10,000 to 100,000, server costs might only double. But in the AI era, when users increase from 100 to 1,000, GPU server costs grow almost linearly. This is because model inference is compute-intensive; each additional concurrent user requires an extra allocation of GPU resources, leaving no room for idle resource reuse. One entrepreneur calculated that for a ToC AI drawing tool, the cost per generation is 0.5 yuan. If offered for free, with 10,000 daily calls, the monthly loss would be 150,000 yuan. When users reach 100,000, the monthly loss exceeds 1.5 million yuan, making it impossible to survive until monetization. Such a cost structure directly blocks the path for small companies to leverage small capital to capture large markets. In the Internet age, one could lose money first and profit later, but in the AI era, losing money first means dying first. Just the computing cost to keep the model running is enough to burn through all funds before the user count reaches 10,000. **2. The Talent Curse: People Who Understand Both AI and the Industry Are Either Scarce or Unaffordable** Entrepreneurship in the AI era does not lack coders but multi-talented individuals who understand both AI technology and industry rules. But such talent is almost a luxury for small companies. First, there is a shortage of AI talent. The gap for algorithm engineers in China exceeds 500,000. Senior engineers who can independently complete model fine-tuning and solve inference optimization command an annual salary of 800,000 yuan+, along with stock options and flexible working conditions. Even if a small company offers a million-yuan salary, it is difficult to attract suitable people. Giants’ AI teams can provide experience in training models with hundreds of billions of parameters and opportunities to publish at top global conferences, while small companies can only offer a few people and a few GPUs under rudimentary conditions. The result of talent voting with their feet is obvious. Even harder to find are cross-disciplinary talents combining industry expertise with AI. To build medical AI, you need an AI engineer who understands imaging diagnostics, has passed the medical practitioner exam, and can tune parameters in PyTorch. There are fewer than 1,000 such people in the country, most working in hospitals or top pharmaceutical companies, with no interest in joining a small company. To build legal AI, you need an AI product manager who understands contract law, knows the risk points in contract review, and can design the model’s reasoning logic. Such people either earn millions in top-tier law firms or are snatched up by big tech firms; small companies can’t even get an interview. Even more cruel is that high-income professionals in the industry have no incentive to help small companies with AI. Lawyers, doctors, chip engineers—these high-cost positions come with high salaries themselves. AI replaces their repetitive work, such as lawyers searching legal provisions or doctors reading routine CT scans. Why would they spend time teaching small companies industry rules or annotating data? Unless it is driven by the company level, but big companies prefer to do AI in-house and won’t leave opportunities for small companies. Small companies fall into a death spiral: lack of talent → can’t build products → no paying customers → can’t hire talent. There is no escape. **3. The Data Curse: Core Industry Data Is Forbidden Territory; Even If Obtained, It Can’t Be Used** The core of AI is data. But small companies face the reality that high-value industry data is either inaccessible or unusable even if obtained. Core data in high-labor-cost fields is never publicly available. Case-winning databases and contract review standards in the legal industry circulate only within top firms like King & Wood and Zhong Lun. Medical imaging data and patient records are restricted by the Data Security Law, accessible only to hospital internal teams. Financial risk control model data and trading decision logic are core company secrets never disclosed. Small companies wanting data must either use gray channels (facing legal risks) or pay high prices for procurement. A single medical imaging dataset costs over 10 yuan, and a million-level data procurement would cost over 10 million yuan—unaffordable. Even if they get the data, small companies don’t know how to use it. Data annotation is not just about pasting labels. For industrial quality inspection defect data, an outsider sees a spot, but an insider knows its location and size to determine if it is a fatal defect. Contract annotation requires distinguishing validity clauses from arbitrary clauses; annotators who haven’t passed the bar exam cannot tell the difference. This tacit knowledge is only understandable within the industry, yet fine-tuning AI models precisely requires such informed annotation. One small company spent 500,000 yuan buying a batch of industrial quality inspection data, but due to incorrect annotation standards, the trained model had an accuracy below 50%, wasting all the money. Worse is the data reuse problem. A small company may build an AI quality inspection model for one factory, but when trying to replicate the experience for another factory, they find completely different product defect standards and production processes. The previous data and model parameters are useless. Every industry and every client’s needs are customized. Small companies can only do one project at a time, retraining the model each time, making scaling impossible. This data silo keeps small companies stuck in a small workshop stage, never reaching scale. **4. The Customer Curse: High-Value Customers Only Trust Giants; Small Companies Can’t Even Get a Seat at the Table** For small companies in the AI era, obtaining customers in high-labor-cost fields is extraordinarily difficult. It’s not that their products are bad, but the trust and compliance barriers are too high. High-value customers—hospitals, law firms, financial institutions—have procurement logic that first looks at trust endorsement. They don’t care how high your model’s accuracy is; they care whether you have successful cases and whether there are risks. Hospitals won’t use a small company’s AI diagnostic tool—if a misdiagnosis occurs, who bears responsibility? Law firms won’t use a small company’s AI contract review—if a clause is missed, who compensates for the loss? Financial institutions won’t use a small company’s AI risk control—if bad debts arise, who is accountable? In contrast, they trust combinations of industry giants plus AI giants. For example, hospitals prefer Siemens Healthineers plus Alibaba Cloud’s AI imaging tools; law firms choose King & Wood plus Tongyi Qianwen’s contract tools. Even if a small company has a better product, it lacks trust endorsement and cannot even qualify for bidding. Compliance thresholds are the death knell for small companies. Medical AI requires NMPA certification, a process lasting 1–2 years and costing over a million yuan. Financial AI requires filing with the China Banking and Insurance Regulatory Commission and a professional legal team. Industrial AI requires ISO system certification, involving complex procedures. Small companies lack both sufficient funds and professional compliance teams, often getting stuck at the qualification review stage while watching orders being snatched by giants. A small company doing AI tax planning once secured a listed company client, but because it lacked a tax firm qualification, the client ultimately abandoned it, wasting all the upfront R&D costs. Even more brutal is customer lock-in. Once a high-value customer collaborates, they demand model exclusivity. For example, a law firm may require the small company to deploy the trained contract review model privately, prohibiting other law firms from using it. A factory may require the small company to delete its product information from the training data to prevent leakage to competitors. Such customization plus exclusivity means small companies can never reuse experience; they can only earn from one project at a time, starving if there is no next project, forever unable to become a scaled enterprise. **5. No Way Out: The Only Escape for Small Companies Is to Become Appendages of Giants** Small companies in the AI era find it almost impossible to replicate the path of growing from small to independent IPOs as in the Internet age. Only two survival options exist. One is to attach to giants as subcontractors. For example, following Alibaba Cloud to develop sub-modules of industry AI, doing image preprocessing for Alibaba Cloud’s medical AI; or following Huawei to do edge-side adaptation for industrial AI, relying on the giant’s resources to get data and customers, earning a meager subcontracting fee. In this model, small companies have no autonomy; if the giant changes its technical route, the small company must follow; if the giant lowers its price, the small company must follow, squeezing profits paper-thin. The other option is to focus on ultra-niche pain points and be small but beautiful. Instead of full-process replacement, they only do an extremely specific link. For instance, a tool for extracting text and images from lawyer contracts, or a tool for identifying a particular defect in industrial quality inspection. By precisely solving one pain point to win over a small group of customers, they earn service fees. In this model, small companies never grow big; they take one project at a time, and when a giant enters the track, they are crushed instantly. In the end, small companies in the AI era have lost the possibility of building an ecosystem from 0 to 1 as in the Internet age. They are not independent entrepreneurs but rather temporary workers in the giants’ ecosystem, making a living through one niche skill, replaceable at any time. This dependent existence is the most realistic antidote the AI era offers to small companies, and also the most helpless curse. **Conclusion: AI Is Not a Windfall for Small Companies; It Is a Filter** Entrepreneurship in the Internet age was a game anyone could join because infrastructure was universal and barriers were late. Entrepreneurship in the AI era is a game for a few players because infrastructure is scarce and barriers are upfront. Those small companies still trying to use Internet-era experience to do AI are essentially using farming-era tools to fight an industrial-era war. It’s not a lack of effort; the underlying logic of the track has changed. AI is not a windfall for small companies; it is a filter. It filters out those entrepreneurs who fantasize about leveraging light-asset traffic monetization, leaving only pragmatic ones willing to attach to giants and focus on vertical niches. The curse of small companies in the AI era is essentially the concentration of resources brought by technological progress—computing power, talent, data, and customers all gather toward giants, leaving small companies to survive in the cracks. This is not a temporary phenomenon but a result jointly determined by the heavy-asset nature of AI technology and the closed-loop characteristics of mobile Internet. It will only become more pronounced in the future. For small companies, accepting dependent survival and abandoning the illusion of scaling may be the most realistic choice in the AI era.
back to top