Scale AI
从数据标注到全栈 AI 基础设施:Scale AI 站在拐点上
Scale AI 在 AI 基础设施里仍有可防守的位置,政府业务敞口强;但客户集中、CEO 交接,以及从数据标注转向新模式的关键拐点,都让投资判断更复杂。
封面要素
公司概况
Scale AI 是一家总部位于旧金山的 AI 基础设施公司,由 Alexandr Wang 于 2016 年创立。公司为领先 AI 实验室、Fortune 500 企业,以及包括美国国防部在内的美国政府机构提供数据标注、RLHF、模型评测和企业 GenAI 平台服务。2024 年完成 $1 billion F 轮、2025 年 Meta 以超过 $29 billion 估值作出标志性战略投资后,Scale 在临时 CEO Jason Droege 带领下,从数据标注转向更宽的企业与政府 AI 平台打法。
- 成立时间
- 2016-01-01
- 创始人
- Alexandr Wang
- 创立地点
- San Francisco, CA
- 总部
- San Francisco, CA
- 产品
- Scale Data Engine(数据标注与整理)、Scale GenAI Platform(企业 AI 应用)、Scale RLHF(LLM 训练数据)、Scale Evaluation(模型安全与能力基准测试),以及 Donovan(防务 AI 智能体平台)。
- 客户
- AI 实验室、大型企业,以及美国政府 / 防务机构。
- 商业模式
- 面向数据标注和 AI 平台服务的企业合同;面向防务 AI 的政府合同;面向实验性用户的自助按量付费。
- 阶段
- Late-stage private (Series F+, post-Meta strategic investment)
- 融资情况
- $1B F 轮(2024 年 5 月,$13.8B 估值)+ Meta 战略投资(2025 年 6 月,$29B+ 估值)
执行摘要
主要优势
- Scale AI 在美国政府 / 国防 AI 数据与评估环节位置独特,已拿到 FedRAMP High 与 DoD IL4 许可
- 虽然近期流失部分客户,Scale AI 在头部 AI 实验室中仍有强品牌和关系网络
- 数据飞轮加自研标注方法,让 Scale AI 在大规模交付时仍能拉开质量差距
- Donovan 平台在国防智能体 AI 中开出新产品护城河
主要风险
- 客户集中:Meta 交易后,Google 和 OpenAI 关系恶化
- CEO 交接风险:创始人转投 Meta,临时 CEO Droege 尚未经过验证
- 商业模式转向:公司正离开核心数据标注业务,新的企业 / 政府收入模式还没跑通
- 结构性利益冲突:Meta 既是战略投资方又是竞争者,客户信任承压
- 收入未披露:缺少财务透明度,估值倍数无法核验
未决问题
- 收入与 ARR:公司没有公开披露;缺少收入基准,$29B+ 估值无法核验
- Wang 离开后的董事会构成:完整治理结构仍未公开
- Meta 商业协议的范围和排他性条款没有披露
- 2025 年客户流失后,真实客户数和 NRR 仍未知
- Mercor 诉讼结果和 IP 敞口仍不确定
目录
01公司概况
1.1 公司身份与概况
Scale AI, Inc. 总部位于加州旧金山,2016 年注册成立。公司对外称使命是「为全球最重要的决策开发可靠的 AI 系统」。2024 年 5 月完成 $1 billion F 轮、2025 年 6 月获得 Meta 战略投资后,Scale 已是后期未上市科技公司,隐含估值超过 $29 billion。Scale 称自己提供数据和全栈技术,覆盖 AI 系统构建、评测、改进的完整开发生命周期。 Scale 的核心业务横跨五条产品线:Scale Data Engine(为模型训练做数据收集、整理和标注)、Scale GenAI Platform(企业级生成式 AI 应用部署)、Scale RLHF(为大语言模型训练提供人类反馈强化学习数据)、Scale Evaluation(面向模型开发者和公共部门的模型能力与安全基准测试),以及 Donovan(服务防务与情报任务的专用 AI 智能体工作流)。公司也运营自助层,采用按量付费,让研究和实验项目能以更低门槛接入标注流水线。Scale 的关于页面显示,截至目前,公司已处理超过 15 billion 次人类标注决策,并向全球贡献者支付超过 $1 billion。 [CO001, CO002, CO003, CO004, CO008, CO009]
1.2 创始人、领导层与治理
Scale AI 由 Alexandr Wang 于 2016 年创立,Wang 当时 19 岁,从 MIT 退学创业。他把 Scale 从一家小型数据标注初创公司,做成能拿下领先 AI 实验室、Fortune 500 企业和美国国防部合同的公司。Wang 自创立起担任 CEO,直到 2025 年 6 月;Meta 对 Scale 作出标志性战略投资后,他离任并加入 Meta 的 AI 业务。离任后,Wang 仍保留 Scale AI 董事席位。 Jason Droege 在 Wang 离任时于 2025 年 6 月出任 Scale AI 临时 CEO。Droege 于 2024 年 9 月加入 Scale,担任首席战略官。加入 Scale 前,他创办 Uber Eats,并将其做到 $19 billion GMV 运行率;更早之前,他曾任 Uber 副总裁,也曾是知名风投机构、Scale 投资方 Benchmark 的合伙人。他在市场型和平台型业务上的运营经历,是 Scale 对外宣称转向企业与政府平台收入的核心支撑。 Scale 尚未详细披露治理结构。Wang 保留董事席位;Meta 投资和领导层交接后,公司未披露完整董事会组成,也未披露独立董事名单。对潜在投资者而言,这带来实质性的治理不透明风险,尤其是 Wang 同时在 Meta 任职;Meta 既是战略投资方,也可能在 AI 人才和技术上与 Scale 竞争。 [CO001, CO010, CO011, CO012, CO013, CO014]
| 姓名 | 角色 | 背景 | 在 Scale 任期 | 关键人风险 |
|---|---|---|---|---|
| Alexandr Wang | 创始人、董事会董事 | MIT 辍学;19 岁创立 Scale;做到 $29B+ 估值;Jun 2025 加入 Meta AI | 2016–Jun 2025(CEO);董事会席位延续 | 高 |
| Jason Droege | 临时 CEO | 创立 Uber Eats($19B GMV);曾任 Uber VP;Benchmark 合伙人;Sep 2024 加入 Scale 任 CSO | Sep 2024 至今(Jun 2025 起任 CEO) | 中高 |
| 董事会 / 治理 | 未完全披露 | Wang 保留席位;Meta 交易后其他董事未公开列名 | 持续 | 高 |
| 领导团队 | CTO、CFO 及其他高管未公开列名 | 交接后未公开详细履历 | 持续 | 中 |
来源:Scale AI 官方公告、TechCrunch、CNBC、Benchmark 合伙人页面(截至访问日已失效)。
[CO010, CO011, CO012, CO013, CO014, CO015]1.3 融资历史与资本结构
Scale AI 已公开披露的风险融资约为 $1.6 billion,横跨多轮融资,最终落在 2024 年 5 月 F 轮和 2025 年 6 月 Meta 战略投资。F 轮之前,公司此前多轮合计融资约 $600 million,包括 2021 年 $325 million E 轮,当时 Scale 估值约 $7.3 billion。2023 年,公司裁员约 20%,反映出 AI 市场压力和数据标注需求重估。 2024 年 5 月,Scale 完成由 Accel 领投的 $1 billion F 轮,投后估值 $13.8 billion。本轮包含新股融资,也包含让既有股东获得流动性的老股交易部分。新投资方包括 Amazon、Cisco、Intel、AMD、ServiceNow、DFJ Growth、WCM Investment Management、Elad Gil 和 Meta。老股东包括 Nvidia、Coatue、Y Combinator、Index Ventures、Founders Fund、Tiger Global、Thrive Capital、Spark Capital、Greenoaks、Wellington Management 和 Nat Friedman。 2025 年 6 月,Meta 做出 CNBC 所称其史上最大一笔 AI 投注:支付约 $14.3 billion,取得 Scale AI 少数股权(据报道按完全摊薄口径约占 49%),对应公司估值超过 $29 billion。Scale 表示公司仍与 Meta 保持运营独立。交易所得流向既有股东和已归属股权持有人,而非全部进入 Scale 资产负债表作为运营资金,因此 Scale 剩余现金和现金跑道仍不清楚。 [CO005, CO006, CO007, CO017, CO018, CO019]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 备注 / 缺口 |
|---|---|---|---|---|
| 成立 | 2016,San Francisco, CA | 2016 | 高 | 多个来源确认 |
| 现任 CEO | Jason Droege(临时) | Jun 2025 | 高 | Wang 于 Jun 2025 离任 |
| 创始人 | Alexandr Wang(董事) | Jun 2025 | 高 | 离任后保留董事席位 |
| 员工数 | ~1,000 名员工(Jul 2025 裁员后) | Jul 2025 | 中 | 约数;Jul 2025 裁掉 200 人 + 500 名承包商 |
| 估值 | $29B+ 隐含(Meta 交易) | Jun 2025 | 中 | 由 $14.3B 换约 49% 股权推算;未正式披露 |
| 累计融资 | ~$1.6B+ 已披露 | May 2024 | 中 | 不包括 Meta 交易(分配给股东) |
| Series F 轮 | $1B,投后估值 $13.8B | May 2024 | 高 | TechCrunch + 公司确认 |
| 收入 / ARR | 未公开披露 | 2026 | 低 | 私有公司;市场可比推算 ARR 为数亿美元 |
| 毛利率 | 未披露 | 2026 | 低 | 私有;无公开披露 |
| 关键认证 | SOC 2 Type II、ISO 27001、DoD IL4、FedRAMP High 认证 | 2024-2025 | 高 | 据 scale.com/legal/security |
| 主要垂直领域 | AI 实验室、企业、政府 / 防务 | 2025 | 高 | 据产品和客户页面 |
| 总部 | San Francisco, CA | 2025 | 高 | 据关于页面 |
置信度评级反映公开披露质量;低置信度项目需要在尽调中直接披露。
[CO001, CO004, CO005, CO006, CO007, CO010]| 轮次 | 日期 | 金额 | 估值 | 领投 / 关键投资者 | 备注 |
|---|---|---|---|---|---|
| 种子 / 天使 | 2016 | ~$3M 估计 | ~$15M 估计 | Y Combinator、天使投资人 | 约数;YC S2016 |
| Series A 轮 | 2017 | ~$18M 估计 | ~$100M 估计 | 投资方:Accel、Index Ventures、Founders Fund | 按公开记录估算 |
| Series B 轮 | 2018 | ~$30M 估计 | ~$300M 估计 | 投资方:Accel、Index、Founders Fund、Tiger Global | 约数 |
| Series C 轮 | 2019 | ~$100M 估计 | ~$1B 估计 | Greenoaks、既有投资者 | 约数 |
| Series D 轮 | 2020 | ~$155M | ~$3.5B 估计 | 投资方:Tiger Global、Index、Accel、Spark、Thrive | 约数;不同来源的准确金额不一 |
| Series E 轮 | 2021-08 | $325M | ~$7.3B | Coatue、Y Combinator、Founders Fund、Tiger、既有投资者 | TechCrunch 报道确认 |
| 过桥 / 老股交易 | 2022-2023 | 多种 | ~$7B 区间 | 既有投资者;老股出售 | 2023 年裁员 20% |
| Series F 轮 | 2024-05 | $1B | 投后估值 $13.8B | Accel(领投)、Amazon、Meta、Cisco、Intel、AMD、ServiceNow、Nvidia、YC、Index、Founders Fund、Tiger、Thrive、Spark、Greenoaks、Wellington、Nat Friedman、Elad Gil、DFJ Growth、WCM | 已确认;新股 + 老股混合 |
| Meta 战略投资 | 2025-06 | ~$14.3B(给股东) | $29B+ 隐含 | Meta(少数股权约 49%) | 资金分配给既有股东;Scale 保持独立 |
早期轮次金额和估值来自公开记录与二级来源估计;只有 Series E、Series F 和 Meta 交易金额经一线媒体报道确认。
[CO005, CO006, CO007, CO017, CO018, CO019]截至报告日期,基于公开披露或强烈指向信息的 Scale AI 高层 KPI 快照。
估值由 Meta 交易隐含得出;员工数为裁员后近似估计;收入未公开披露。
[CO004, CO005, CO006, CO007, CO029, CO030]1.4 产品与商业模式
Scale AI 的商业模式,是向构建和部署机器学习系统的组织出售高质量数据与 AI 基础设施服务。收入来源包括企业数据标注合同(按量和按项目)、Scale GenAI Platform(面向企业 AI 应用的托管 SaaS 和专业服务)、政府与防务合同(包括美国国防部数据整理,以及服务情报和军事行动的 Donovan 平台),以及面向小型或实验性场景的自助层。Scale 不公开披露收入、毛利率或 ARR。 Scale Data Engine 收集、整理、标注并验证文本、图像、视频、音频和文档数据。GenAI Platform 用自研流水线把企业数据转成特定领域 AI 应用。Scale RLHF 提供经过整理的偏好数据,用于人类反馈强化学习,这是训练能遵循指令且更安全的大语言模型的核心。Scale Evaluation 提供可信基准测试,覆盖模型能力(包括 Scale Leaderboard)和安全性(包括 WMDP 有害知识基准),同时服务商业模型开发者和美国政府机构。Donovan 平台面向防务与情报任务,提供专用 AI 智能体工作流;凭借 Scale 的 DoD IL4 和 FedRAMP High 安全认证,该平台可在涉密环境运行。 定价分两层:企业客户采用定制定价,配备专门运营团队并承诺 SLA;自助客户按量付费接入平台,前 1,000 个标注单元免费。Scale 不公开披露收入;市场估计其 ARR 达数亿美元级别,这一判断来自投资者评论和可比指标,并非公司确认披露。 [CO002, CO003, CO008, CO024, CO025, CO026]
Scale AI 如何把数据、算力和人类专家能力转化为面向模型开发者、企业和政府客户的 AI-ready 输出。
[CO002, CO003, CO008, CO024, CO025, CO026]1.5 关键里程碑与不利事件
从创立到 2026 年,Scale AI 走过十年增长、战略转向和重大不利事件,这些共同定义了公司当前的投资画像。Scale 2016 年创立,最初专注于自动驾驶汽车的程序化数据标注(Waymo 是早期客户),随后扩展到更广泛的企业 AI 训练数据。到 2021 年,公司凭 E 轮达到 $7.3 billion 估值;到 2022 年,公司公开称自己是估值 $7 billion、员工超过 700 人,并为 DoD 提供自主系统数据层的公司。 2023 年,AI 基础模型训练数据支出放缓,Scale 执行了痛苦的 20% 裁员。2024 年 5 月,Scale 以 $13.8 billion 估值完成 $1 billion F 轮,引入广泛的战略与财务投资方联盟。2024 年,Scale 还加入白宫 AI 安全自愿承诺,并拿下 DoD 联合作战数据整理合同。 2025 年 6 月是最关键拐点:Meta 作出 $14.3 billion 战略投资,创始人 Alexandr Wang 同时离开 Scale、加入 Meta 的 AI 工作。三周后(2025 年 7 月),临时 CEO Droege 宣布裁员约 200 名员工(占员工数 14%)和 500 名承包商,理由是公司相较于转向企业与政府平台收入的战略方向,对数据标注人员投入过度。2025 年 6 月,CNBC 报道称 Google——当时 Scale 最大客户——计划结束或大幅缩减与 Scale 的合作,原因是担忧与 Meta 发生竞争冲突。OpenAI 也在 2025 年 6 月结束了与 Scale 的合作。2025 年 9 月,Scale 起诉竞争对手 Mercor 和一名前员工,指控其挖走客户。这些不利事件,对评估 Scale 的投资者构成实质性的客户集中和治理风险。 [CO001, CO031, CO032, CO033, CO034, CO035]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 影响 |
|---|---|---|---|---|---|
| 2016 | Alexandr Wang 创立 Scale AI | 创立 | $15M 估计种子轮 | Wang、YC、天使投资人 | 建立面向自动驾驶的数据标注 |
| 2016 | 入选 Y Combinator S2016 | 合作 | — | YC、Scale | 早期验证;接入 YC 网络 |
| 2017 | 完成 Series A 轮 | 融资 | ~$18M | 投资方:Accel、Index、Founders Fund | 推动产品扩张 |
| 2019 | Series C 轮;估值突破 $1B | 融资 | ~$100M,估值 ~$1B | Greenoaks 等 | 独角兽里程碑 |
| 2021-08 | Series E 轮,估值 $7.3B | 融资 | $325M | 投资方:Coatue、YC、Founders Fund、Tiger | 快速扩张;AI 训练数据热潮峰值 |
| 2022 | DoD 自动化数据层博客;员工 700+ | 规模扩张 | $7B 估计估值 | DoD / Scale | 确立防务 AI 足迹 |
| 2023 | 员工减少 20% | 负面 | — | Scale、员工 | 市场修正;AI 训练需求放缓 |
| 2024-05 | Series F 轮融资 $1B,估值 $13.8B | 融资 | $1B / $13.8B | Accel 领投,Amazon、Meta、Intel、AMD、Cisco 等 | 估值上调重置;战略投资者基础 |
| 2024 | White House AI 自愿安全承诺 | 监管 | — | White House、Scale | 监管定位;品牌正当性 |
| 2024 | DoD 数据整理合同(Joint Force) | 监管 | 未披露 | DoD、Scale | 防务收入锚 |
| 2024-11 | Defense Llama 发布(国家安全 LLM) | 产品 | — | Scale、DoD 生态 | 首个产品化防务 LLM 供给 |
| 2025-06 | Meta $14.3B 战略投资;Wang 离任 | 融资 | $14.3B / $29B+ 隐含 | Meta、Scale、Wang 转向 Meta AI | 转折性资本事件;创始人退出 |
| 2025-06 | 最大客户 Google 计划退出 | 负面 | — | Google、Scale | 客户集中度风险显性化 |
| 2025-06 | OpenAI 逐步结束与 Scale 关系 | 负面 | — | OpenAI、Scale | 第二个主要实验室客户流失 |
| 2025-07 | 裁员 200 人(14%)+ 500 名承包商 | 负面 | — | Droege、Scale | 数据标注业务转向在运营层面确认 |
| 2025-09 | 对 Mercor 提起诉讼(挖客户) | 负面 | — | Scale 诉 Mercor | IP 与竞争冲突风险 |
来源:Scale AI 官方页面、TechCrunch 报道、CNBC 报道、Stanford HAI AI Index 2025。早期融资金额为估计;后续事件经一线媒体确认。
[CO001, CO005, CO006, CO007, CO017, CO019]从 2016 年创立到 2025 年 Meta 投资及组织转向的关键节点,突出融资事件、产品发布和负面事件。
早期轮次日期(2016-2019)为近似值;Series A-C 的确切时间未公开确认。
[CO001, CO005, CO007, CO017, CO031, CO033]1.6 展项
02市场分析
2.1 市场边界与定义
Scale AI 位于几个彼此独立但相连的市场交界处,因此在有意义地测算规模之前,必须先把边界讲清楚。核心市场是 AI 数据服务:为机器学习模型收集、标注、整理和验证训练及评测数据。这个市场包括面向计算机视觉、自然语言处理、语音识别和多模态 AI 的人类标注数据。相邻市场是 AI 模型评测,即面向大语言模型的可信基准测试和安全测试,Scale Evaluation 和 Scale Leaderboard 在这里竞争。 再向外扩,Scale 的 GenAI Platform 进入企业 AI 部署市场:帮助大型组织在基础模型之上构建、定制并投入运营生成式 AI 应用的工具和服务。这个市场与超大规模云厂商 AI 服务(Azure AI、AWS Bedrock、Google Vertex AI)、精品 AI 咨询公司和 AI 模型 API 厂商重叠。Donovan 平台则切出一个专门的政府与防务 AI 垂直市场,其采购规则和预算机制与商业 AI 支出不同。 Scale 核心数据标注服务的现状替代方案包括内部人工审核团队(许多大型 AI 实验室和企业最初都自建标注流水线)、更低成本的离岸服务商,以及自动化合成数据生成。市场测算的关键问题,是应当把服务市场窄定义为 AI 数据标注服务,还是宽定义为所有 AI 基础设施和工具支出。本章将主要市场定义为 AI 数据服务与评测基础设施(标注、RLHF、基准测试),把企业 GenAI 平台和政府 AI 视为相邻子市场。 [CM001, CM002, CM003, CM004, CM005]
| 子市场 | 纳入支出 | 排除支出 | 现状替代方案 | Scale AI 覆盖 |
|---|---|---|---|---|
| AI 数据标注 | 用于 ML 训练的人工标注文本、图像、视频、音频、文档数据 | 通用众包平台(Mechanical Turk) | 内部标注团队;离岸 QA 服务商 | 核心产品(Scale Data Engine) |
| RLHF / 偏好数据 | 用于 LLM 指令遵循和安全对齐的专家人工反馈 | 原始调查数据;合成偏好生成 | 大型实验室内部 RLHF 团队 | Scale RLHF 产品 |
| 模型评估与基准测试 | LLM 能力 + 安全测试;红队测试;排行榜 | 内部 A/B 测试;学术基准 | 自评;开放基准(MMLU、HELM) | Scale Evaluation + Leaderboard |
| 企业 GenAI Platform | 基于基础模型的垂直 GenAI 应用开发 | 超大规模云厂商 AI 服务(Azure、AWS、GCP) | 企业自建 GenAI;精品 AI 咨询 | Scale GenAI Platform 平台 |
| 政府 / 防务 AI | DoD 认证 AI 数据服务;任务关键 AI 智能体 | 通用政府 IT 外包 | 具备防务资质的竞争者;内部 DoD 团队 | Donovan + 公共部门数据引擎 |
| AI 数据市场(自助) | 面向初创公司和研究人员的按量付费标注 | 非结构化众包平台 | Mechanical Turk;实习生团队;学生标注员 | Scale 自助层 |
市场边界为本分析定义;实际市场规模取决于纳入哪些板块。不是穷尽市场图谱——AI 基础设施(compute、MLOps)和 AI 咨询等邻近市场被排除。
2.2 TAM、SAM 与 SOM 的多口径测算
公开来源没有给出足够细、足够可靠的 AI 数据服务 TAM/SAM/SOM 精确数字,无法直接支撑高置信度估值分析。这里呈现的数字,是基于公开市场代理指标、投资者报道和竞争对手披露模式综合得出的、受证据约束的估计。 截至 2025 年,全球 AI 数据服务与基础设施的总可用市场(TAM)——包括标注、RLHF、评测和 GenAI 平台工具——估计每年为 $5 billion 到 $20 billion。区间很宽,反映市场快速演进,也反映不同分析师对边界定义不一致。这个估计锚定在 McKinsey 的发现:88% 的受访组织如今已在至少一个职能中使用 AI(高于 2024 年的 78%);Stanford HAI AI Index 2025 也显示 AI 投资已急剧加速。如果每家采用 AI 的 Fortune 1000 企业每年在 AI 数据和评测基础设施上花费 $1M–$10M,仅企业端合计支出就达到 $1B–$10B,AI 实验室还会带来大量增量需求。 Scale AI 的可服务市场(SAM)——拥有复杂 AI 数据需求、且有预算购买高端质量的大型企业、领先 AI 实验室和美国政府机构——估计为 $2 billion 到 $8 billion,反映 Scale 站在市场质量端,而不是低成本商品化标注段。可获取市场(SOM)——Scale 未来 3-5 年现实可捕获规模——估计为 $500M 到 $2B,这与其后期私有估值所隐含的企业与政府收入轨迹大体一致。公司未披露收入;这些估计不确定性很高,必须用实际财务数据验证。 [CM006, CM007, CM008, CM009, CM010, CM011]
| 规模测算视角 | 市场范围 | 估计 | 置信度 | 依据 / 关键假设 |
|---|---|---|---|---|
| TAM — 广义 AI 数据与基础设施 | 全球所有 AI 数据服务 + 评估 + 企业 GenAI 工具 | $10B–$30B | 低 | McKinsey:88% 的组织使用 AI;如果大型组织在数据 / 工具上支出 $1M–$10M,合计可达 $10B+;不确定性高 |
| TAM — 仅 AI 数据标注 | 全球面向 ML 的人工标注数据服务 | $2B–$8B | 低 | 由上市公司类比(Appen 收入代理指标)+ 全球市场放大系数推导;高度不确定 |
| SAM — 高端 AI 数据(Scale 定位) | 大型企业、AI 实验室、政府——高端标注 + 评估 + 平台 | $1.5B–$6B | 低 | Scale 面向质量端;SAM 约为标注 TAM 的 30–50%;政府增加独立 SAM 板块 |
| SAM — 政府 / 防务子市场 | Scale 可服务的美国政府 AI 数据和评估预算 | $300M–$1B | 低 | 按 DoD AI 投资增长和涉密供应商供给约束估计;无公开预算拆分 |
| SOM — Scale 3-5 年可获取 | 基于竞争位置,Scale AI 在 3-5 年内可现实捕获的收入 | $500M–$2B | 低 | 若 ARR 在 $500M-$2B 区间,15-60x ARR 可支撑 $29B 估值;但收入未获确认 |
所有估计都是证据约束下的近似值,来自市场代理数据、投资者报道和公开牵引力指标。收入未公开披露。这些数字只作方向性参考,仍需用实际财务数据验证。
在证据约束下,按 TAM、SAM、SOM 估算 Scale AI 的可触达市场,展示不同范围定义下可能的市场规模区间。
所有数值均为十亿美元中位数估计;不确定区间较宽(2x–4x)。没有公开分析师报告按细分维度给出具体 AI 数据标注市场规模。估算来自 McKinsey 采用数据、竞品代理收入和估值倍数。
[CM006, CM007, CM008, CM009, CM010]按范围定义展示 AI 数据服务市场规模的不确定区间,反映分析师边界分歧和市场快速演进。
区间来自 McKinsey AI 采用数据、以 Appen 公开收入作为仅标注 TAM 代理,以及 Scale AI 隐含估值 / 收入关系。不确定性高;所有估计都需要用实际财务披露核验。
[CM007, CM008, CM009, CM010, CM011]2.3 买方与用户分层及采用路径
Scale AI 的可触达市场由四类主要买方构成,预算权、采购流程和切换成本差异很大。理解这些分层,是判断 Scale 收入韧性和增长天花板的前提。 第一类——AI 研究实验室和基础模型开发商——是 Scale 最早的客户群。OpenAI、Meta、Google DeepMind、Anthropic、Cohere 和 Adept 等公司,需要海量高质量标注数据用于预训练和 RLHF。预算权在研究和工程职能。采购通常通过直接谈判,签署多年合同。切换成本中等:实验室可以自建内部标注流水线,也可以换供应商,OpenAI 和 Google 2025 年离场已证明这一点。这个分层收入集中风险很高。 第二类——大型企业 AI 采用者——包括用 AI 做内部自动化、客户产品和竞争差异化的 Fortune 500 公司。Cisco、Etsy、Instacart 和 Pinterest 等公司代表这一类别,它们使用 Scale 的 GenAI Platform 开发特定领域 AI 应用。预算权通常落在 CTO/CDO 组织,并伴随多年平台承诺。由于平台集成更深,切换成本高于标注服务。 第三类——美国政府与防务——是 Scale 最有辨识度、也最可防守的分层。预算权掌握在 DoD 项目办公室、DHS、NSA/IC 和民用机构手中。采购遵循联邦采购法规,并使用多年合同工具。切换成本很高,因为需要安全准入(DoD IL4、FedRAMP High)、机构知识,以及 Donovan 平台带来的竞争护城河。这个分层带来稳定、长周期收入,也能隔离商业客户集中风险。 第四类——AI 初创公司和研究机构——通过自助层接入 Scale。预算权分散;单个客户支出较低,但数量可能可观。切换成本低(按量付费)。 [CM012, CM013, CM014, CM015, CM016, CM017]
| 细分市场 | 买方画像 | 预算负责人 | 采购路径 | 切换成本 | Scale AI 匹配度 |
|---|---|---|---|---|---|
| AI 实验室(基础模型) | AI 实验室:OpenAI、Meta、Google DeepMind、Anthropic、Cohere | 研究负责人 / 工程副总裁 | 直接谈判;多年期合同 | 中 | 历史上高;目前受 Meta 冲突拖累,存在风险 |
| 大型企业(F500) | 企业客户:Cisco、Etsy、Instacart、Pinterest、TIME | CTO / CDO / AI 副总裁 | RFP / 直销;12-24 个月周期 | 高 | 正在增长;GenAI Platform 是主要载体 |
| 美国政府 / DoD | DoD、情报机构、民用联邦机构 | 项目办公室 / CISO / CTO | 联邦采购;IDIQ / 任务订单;需安全许可的采购载体 | 很高 | Donovan + 安全许可 = 独特定位 |
| AI 初创公司 / 研究人员 | Series A-C AI 公司;大学实验室;小型企业 | 创始人 / ML 负责人 | 自助服务;信用卡付款;采购摩擦最低 | 低 | 自助服务层级;ARPU 更低;走量模式 |
| 国际政府 | 盟国国防 / 情报机构 | 采购办公室;外交部 | 国际政府合同;安全许可复杂 | 很高 | Global Public Sector 部门;仍处早期 |
| 企业媒体 / 内容 | 部署 GenAI 做内容的出版商、媒体公司 | 数字业务副总裁 / CTO | 直销;平台订阅 | 中 | TIME 案例可作参考;垂直行业在增长 |
买方画像基于 Scale 已公开披露的客户案例和产品页描述。各细分市场的收入贡献未公开披露。
从 AI 实验室到企业和政府客户,Scale AI 产品如何通过不同价值主张连接各买方板块。
[CM012, CM013, CM014, CM015, CM016, CM017]2.4 增长驱动因素与采用约束
Scale AI 可触达市场最强的增长驱动因素,是企业采用 AI 正在加速。McKinsey 2025 年 State of AI 调研显示,使用 AI 覆盖至少一个业务职能的组织比例,从上一年的 78% 跳升至 88%,其中 62% 正在积极试验 AI 智能体。这个宏观扩张,为四类买方分层中的数据标注、模型评测和企业 AI 部署工具带来结构性需求。 第二个主要驱动因素,是大语言模型提供商持续增多——OpenAI、Meta、Anthropic、Google、Mistral、Cohere 及其他厂商,每一家都需要持续的 RLHF 循环、安全基准测试和规模化能力评测。基础模型开发商越多,对专家级训练数据和中立评测基础设施的需求也越高。Scale 的 Evaluation 产品和 Scale Leaderboard 正好受益于这一趋势。 美国政府 AI 投资是尤其强劲的增长驱动因素。Scale 的安全准入(DoD IL4、FedRAMP High)、Donovan 平台和正在执行的防务合同,让它有机会承接更大份额的联邦 AI 预算。随着 DoD 将 AI 融入监视、后勤、网络安全和自主系统,防务 AI 支出预计会显著增长;这些场景都需要可信数据基础设施。 采用约束同样明显。第一,Meta 战略投资后,最大 AI 实验室正从 Scale 分散出去,可能压低收入。第二,合成数据生成技术(部分 AI 实验室在内部使用)长期可能减少某些应用对人类标注数据的需求。第三,Appen、SuperAnnotate、离岸团队等低成本提供商竞争激烈,给商品化标注段带来利润率压力。第四,企业 AI 项目仍大量停留在试点 / POC 阶段——McKinsey 指出,多数组织还在测试,而非大规模推广 AI——因此企业平台收入增长可能滞后于采用曲线。第五,政府采购周期长且依赖预算,导致合同收入波动。 [CM019, CM020, CM021, CM022, CM023, CM024]
| 因素 | 类型 | 方向 | 量级 | 时间跨度 | 证据 |
|---|---|---|---|---|---|
| 企业 AI 采用扩张 | 驱动因素 | ↑ | 高 | 2025-2028 | McKinsey:88% 的组织使用 AI(高于此前 78%);62% 正在试验 AI 智能体 |
| 基础模型供应商增多(LLM 提供商) | 驱动因素 | ↑ | 高 | 2024-2027 | OpenAI、Meta、Anthropic、Mistral、Cohere 都需要 RLHF + 评估数据 |
| 美国政府 AI 投资增长 | 驱动因素 | ↑ | 高 | 2025-2030 | DoD AI 战略;Scale 安全许可;Donovan 平台;活跃 DoD 合同 |
| AI 安全监管压力(EU AI Act、美国 EO) | 驱动因素 | ↑ | 中 | 2025-2027 | 监管要求推高 AI 评估和审计服务需求 |
| 合成数据替代人工标注 | 约束 | ↓ | 中 | 2026-2029 | AI 实验室在部分训练任务中越来越多使用合成数据 + 模型生成数据 |
| Meta 交易后客户集中风险 | 约束 | ↓ | 高 | 2025-2026 | Google 和 OpenAI(主要客户)离开;能否替代大额收入仍未知 |
| 低成本竞争(Appen、离岸供应商) | 约束 | ↓ | 中 | 持续 | 商品化标注挤压利润率;Scale 必须守住质量溢价 |
| 企业 AI 仍处试点 / POC 阶段 | 约束 | ↓ | 中 | 2025-2026 | McKinsey:多数组织仍在测试 AI,尚未规模化;企业平台收入滞后 |
| 政府采购周期长 | 约束 | ↓ | 低 | 持续 | 尽管定位强,联邦采购时间线仍会带来收入波动 |
方向和量级是分析师基于所引公开证据作出的判断。公开来源没有可用的量化收入影响估计。
从初始 AI 认知到深度接入 Scale 的采用阶段,并估算每个阶段群体规模,用于展示市场转化动态。
百分比为 Fortune 1000 分层的示意估计,基于 McKinsey AI 采用数据(88% 使用 AI,约 1/3 扩展)。管线和客户比例根据 Scale 已披露客户引用估算;实际数字未披露。
[CM019, CM020, CM021, CM022, CM023]2.5 市场测算缺口、相互矛盾的估计与尽调问题
AI 数据服务的市场测算存在显著结构性不确定性。没有公开分析师报告能对 AI 数据标注可触达市场给出一致、细颗粒度拆分。估计差异很大,取决于边界只包括人类标注数据服务,还是延伸到 AI 模型评测、企业 AI 工具和政府 AI 平台支出。本章给出的宽区间($5B–$20B TAM)反映真实的分析师分歧和市场边界模糊,而不是测量误差。 合成数据市场让测算更复杂:一些预测假设合成数据会在 3-5 年内大幅替代许多任务中的人类标注,显著压缩 Scale 这类公司的 TAM。另一些观点则认为,人类评测、RLHF 和对抗性红队永远需要人类判断,并会随着模型增多而扩张。这是投资命题层面的不确定性,潜在投资者必须通过与 AI 实验室和 Scale 客户的一手研究来解决。 Scale AI 披露的牵引力指标(15B 次决策标注、向贡献者支付 $1B)能说明规模,却不能说明收入。没有收入披露,就无法验证哪个分层主导 Scale 收入结构,也无法判断企业平台和政府分层能否替代下滑的 AI 实验室收入。Google 和 OpenAI 作为主要客户离场——假设它们曾是 Scale 收入前 10 大贡献者——留下了规模未知的收入缺口,企业与政府转型必须填补这块缺口。量化这两家客户离场的收入影响,是 Scale 投资案例中最重要、尚未解决的市场测算问题。 [CM027, CM028, CM029, CM030, CM031]
2.6 展项
03竞争格局
3.1 竞争格局概览
Scale AI 在 AI 数据与基础设施市场的多个重叠分段竞争,每个分段的竞争动态不同。在核心 AI 数据标注分段,Scale 面对 Appen(ASX 上市,也是唯一公开交易的直接可比公司)、SuperAnnotate、Labelbox 和 Surge AI。具体到 RLHF 和 LLM 训练数据,Surge AI 和 Mercor 是最聚焦的直接竞争对手。企业模型评测与基准测试方面,Labelbox 的评测套件和 Snorkel AI 的程序化标注构成相邻威胁。企业 GenAI 平台部署方面,Scale 面对超大规模云厂商(AWS Bedrock、Azure AI、Google Vertex AI)、精品 AI 咨询公司和系统集成商——这些对手规模更大、资源更足。 竞争格局还包括相邻领域的既有公司:传统数据外包公司(Accenture、Capita)、众包标注平台(Amazon Mechanical Turk、Remotasks),以及 AI 实验室内部标注团队,后者代表「自建」替代方案。可能的新进入者包括在内部构建标注和评测能力的大型科技公司,以及从学术或特定领域 AI 切入的专业精品公司。 关键竞争动态是 Meta 战略投资:Meta 成为 Scale 最大投资者的同时,也制造了利益冲突,已经导致 Scale 两个最大 AI 实验室客户 Google 和 OpenAI 退出或减少与 Scale 的合作。这是一个罕见局面:Scale 自己的投资方正在导致最大商业分段客户流失。Mercor 是一家更新、更小的竞争对手,正试图通过直接招揽客户来利用这一脆弱点,Scale 2025 年 9 月诉讼就是证据。 [CP001, CP002, CP003, CP004, CP005]
| 竞品 | 阶段 / 规模 | 目标客户 | 产品范围 | 融资 / 收入代理指标 | 战略方向 |
|---|---|---|---|---|---|
| Appen (ASX: APX) | 上市公司;约 $300M 收入代理指标(下滑中) | 企业 + 政府;全球 | 图像、文本、语音、视频标注;评估 | 公开市场;ASX 上市;收入下滑 | 靠企业 AI 稳住业务;降低对众包的依赖 |
| Labelbox | 私营;Series B+;估计融资约 $100M | 企业 AI 团队;中端市场至 F500 | 标注 + 评估 + RLHF + 机器人数据 + 排行榜 | 约 $188M 融资(估计);未公开 | 全栈 AI 数据平台;向 RLHF + 评估扩张 |
| Snorkel AI | 私营;Series C+;估计融资约 $135M | 企业(F500);政府 | 程序化标注;弱监督;AI 辅助标注 | 约 $135M 融资(估计);未公开 | 用 AI 降低人工标注成本;企业平台 SaaS |
| SuperAnnotate | 私营;早期;约 $14M 融资 | 企业团队;聚焦计算机视觉 | 协作式标注平台;安全功能;多模态 | 约 $14M 融资(估计);未公开 | 企业 CV 标注;安全优先;扩展 NLP / 多模态 |
| Surge AI | 私营;小规模;2020 年成立 | AI 实验室;聚焦 RLHF;高端质量 | 专家级 RLHF 数据;LLM 反馈;高质量标注员 | 小规模;未公开 | 高端 RLHF 数据;以更小团队拼质量 |
| Invisible Technologies | 私营;成长期;估计融资约 $20M | 企业运营;AI 自动化 | AI 驱动运营;数据处理;标注只是更广运营的一部分 | 约 $20M 融资(估计) | 不止标注,做 AI 驱动的企业运营 |
| Mercor | 私营;早期;估计融资 <$50M | AI 实验室;企业;瞄准 Scale AI 客户 | AI 人才市场;RLHF;评估;标注 | 早期;未公开 | 瞄准 Scale 客户切入;诉讼待决 |
私营竞品的融资和收入估计来自媒体报道及竞品网站描述,未经验证。Appen 收入来自 ASX 申报文件(公开可得,可作为标注市场代理指标)。
3.2 竞争对手画像
Appen(ASX: APX)是 Scale AI 标注业务唯一公开交易的直接可比公司。Appen 为全球企业和政府客户提供 AI 训练数据,包括图像、视频、语音、文本和文档标注。随着 AI 标注市场转向高端和 LLM 专用服务,Appen 公开报告的收入一直下滑。Appen 服务全球企业和政府客户,与 Scale 客户群有部分重叠。Appen 的竞争重心主要是覆盖广度和成本效率,而不是 Scale 的高质量溢价定位。 Labelbox 是一家位于旧金山的数据标注和模型评测平台,目标客户是企业 AI 团队。Labelbox 已从核心标注扩展到 RLHF 数据收集和模型排行榜,因此在 Scale 多条产品线上构成直接竞争。Labelbox 定价比 Scale 企业合同更易接近,并建立了专家网络(Labelbox Expert Network)服务质量要求高的标注任务。Labelbox 为机器人 AI 训练提供专门产品,而 Scale 尚未公开强调这一市场。 Snorkel AI 专注于使用 AI 辅助的弱监督技术做程序化数据标注,以降低人类参与瓶颈。Snorkel 面向需要构建 AI 训练数据集、但不想投入大量人工标注的企业客户。Snorkel 的方法与 Scale 的人类专家模型根本不同——Snorkel 试图减少人类标注需求,Scale 的模型则依赖人类质量。Snorkel 客户包括大型企业和部分政府机构。 SuperAnnotate 是面向企业团队的 AI 标注平台,提供协作式标注工作流、质量管理和 ML 流水线集成。SuperAnnotate 提供服务企业合规的安全功能,并已扩展到计算机视觉、NLP 和多模态标注。Mercor 是更新的进入者,运营 AI 人才市场,服务 RLHF、模型评测和数据标注,由前 Scale 贡献者创办。Scale 于 2025 年 9 月起诉 Mercor,指控其挖走客户,说明 Mercor 正在主动瞄准 Scale 的企业客户群。 Surge AI(现已成为更广泛 RLHF 数据生态的一部分)曾专注于为 LLM 训练提供高质量人类反馈数据,拥有类似 Scale 但规模更小的专家标注员网络。Invisible Technologies 提供 AI 驱动的业务运营和数据服务,与 Scale 的企业自动化和标注能力竞争。 [CP006, CP007, CP008, CP009, CP010, CP011]
| 能力 | Scale AI | Appen | Labelbox | Snorkel AI | SuperAnnotate | Mercor |
|---|---|---|---|---|---|---|
| 文本 / NLP 标注 | ✓ 高级 | ✓ 广泛覆盖 | ✓ 高级 | ✓ AI 辅助 | ✓ 多模态 | ✓ 聚焦 RLHF |
| 图像 / CV 标注 | ✓ 高级 | ✓ 广泛覆盖 | ✓ 高级 | ✓ 程序化 | ✓ 专精 | 有限 |
| RLHF / LLM 训练数据 | ✓ 核心产品(Scale RLHF) | 有限 | ✓ RL-Data 产品 | 部分 | 有限 | ✓ 核心重点 |
| 模型评估 + 排行榜 | ✓ Scale Evaluation + Leaderboard | ✓ 评估产品 | ✓ Evals 产品 + 排行榜 | 有限 | 有限 | 有限 |
| 政府 / 国防安全许可 | ✓ DoD IL4、FedRAMP High | 部分(有部分政府项目) | 未披露 | 未披露 | 未披露 | 未披露 |
| GenAI 平台 / 企业 AI 应用 | ✓ Scale GenAI Platform 平台 | 否 | 否 | 否 | 否 | 否 |
| 国防 AI 智能体平台 | ✓ Donovan | 否 | 否 | 否 | 否 | 否 |
| 程序化 / AI 辅助标注 | 部分 | 否 | 有限 | ✓ 核心优势 | 部分 | 否 |
| 自助服务 / 按量付费层级 | ✓ 是 | 否 | ✓ 是 | ✓ 是 | 否 | 否 |
能力判断基于公开产品页面和描述。未披露不等于没有能力。来源:公司网站,访问日期 2026-05-09。
3.3 能力、定价与监管对比
Scale AI 最重要的竞争差异集中在三处:(1)政府级安全认证(DoD IL4、FedRAMP High),没有公开披露的竞争对手达到同等水平,这让 Scale 在涉密和防务 AI 数据市场接近独占;(2)Scale Evaluation 的模型安全基准测试和 Scale Leaderboard 已建立声誉,成为 LLM 能力可信第三方评测方;(3)Donovan 防务 AI 智能体平台,在获准入的 AI 智能体空间没有公开记录的直接竞争者。 在核心标注业务中,Scale 相比低成本竞争对手的主要差异是质量溢价。自研 Scale Data Engine、质量反馈闭环、专家贡献者网络(已向全球贡献者支付超过 $1B)和标注工具链,都难以快速复制。不过 Labelbox、Snorkel AI 等竞争对手也在质量管理工作流上投入很大,正在缩小差距。 定价上,Scale 位于高端层:企业合同采用定制定价,并配备专门运营团队。Appen 和 SuperAnnotate 价格更低,更容易触达那些不需要 Scale 质量保证的中端市场客户。Labelbox 提供分层定价和自助选项,因此既是 Scale 自助层的直接竞争对手,也竞争企业合同。 监管与信任姿态是关键战场。Scale 的政府安全准入和监管承诺(白宫 AI 安全承诺、WMDP 基准、国会证词)把它定位为可信政府 AI 数据供应商。Appen 有一些政府业务,但缺少 Scale 在防务专用安全准入上的深度。Labelbox 和 Snorkel 看起来没有发布同等级别的政府安全认证。由于取得 DoD IL4 和 FedRAMP High 认证需要多年流程,这是 Scale 最持久、也最可防守的竞争优势。 [CP015, CP016, CP017, CP018, CP019, CP020]
| 竞品 | 定价模式 | 入门点 | 企业层 | 差异化 |
|---|---|---|---|---|
| Scale AI | 企业定制 + 自助按量付费 | 1,000 个单位免费;之后按单位计费 | 定制定价、专属运营、SLA | 质量保证;政府安全许可 |
| Appen | 按项目计费;量大议价 | 公开提交需求;定制报价 | 企业合同;全球交付 | 全球众包;低成本;上市公司透明度 |
| Labelbox | 分层 SaaS + 用量计费;有免费层 | 免费计划;Developer 计划;Enterprise | 企业定制 + 专业服务 | 平台集成;评估功能;价格更低 |
| Snorkel AI | 企业 SaaS;无公开定价 | 仅企业合同 | 企业订阅 + 服务 | AI 辅助标注降低规模成本 |
| SuperAnnotate | SaaS 订阅;团队层级 | Team 计划;Enterprise 计划 | 企业版可选本地部署 | 安全优先;协作流程 |
| Surge AI | 按项目计费;专家溢价 | 定制项目报价 | 高质量专家 RLHF 合同 | 面向 LLM 训练的专家标注员质量 |
| Mercor | 人才市场;按任务或订阅计费 | 市场自助服务 | 企业人才匹配 + 托管 RLHF | 专家网络;AI 人才市场模式 |
定价基于公开价格页和描述。Scale AI、Snorkel AI 和 Mercor 的企业定价都需要定制报价。
Scale AI 相对核心竞争对手,在质量 / 高端化和政府许可深度两个维度上的定位;这两个维度定义了其最可防守的市场位置。
位置为分析师基于公开产品 / 能力描述做出的定性估计。没有公开可用的实证质量基准。
[CP015, CP016, CP017, CP018, CP019]跨七个 AI 数据服务维度(标注、RLHF、评测、政府许可、GenAI 平台、国防智能体、自助服务)的综合能力分数,展示 Scale AI 相对直接竞争对手的广度优势。
7 个能力维度各按 1-5 分打分并汇总(最高 35 分)。Scale AI 的政府许可和国防智能体分数(均为 5/5)推动其领先。分数为基于公开产品页面的定性评估。
[CP015, CP016, CP017, CP018, CP019, CP020]3.4 切换成本、锁定效应与多供应商并用
Scale AI 各市场的切换成本因分段而差异巨大。AI 实验室主要用 Scale 获取标注和 RLHF 数据,切换成本中等:标注流水线在一定程度上可替代,实验室也已展示换供应商的意愿(Google 和 OpenAI 于 2025 年离场)。这正是 Meta 投资利益冲突会如此迅速造成客户流失的原因——面对数据供应商新投资方带来的利益冲突时,切换成本不足以留住客户。 企业客户使用 Scale 的 GenAI Platform,切换成本更高:平台牵涉数据迁移、与企业系统的工作流集成、数据流水线定制和组织知识迁移。其切换成本类似中端 SaaS 平台(6–18 个月)。 美国政府与防务客户的切换成本极高。转向新供应商,要求新供应商取得同等安全准入(DoD IL4 和 FedRAMP High 授权需要 12–36 个月)、重建机构知识,并穿过联邦采购法规。这为 Scale 的防务分段创造了持久、持续多年的锁定效应。 标注市场常见多供应商并用:许多大型组织(AI 实验室、企业)会同时把标注工作交给多个供应商,以比较质量、优化成本并保留冗余。这意味着即便是具名客户,Scale 也未必拥有独家关系。Scale 的自助层明确鼓励多供应商并用(低承诺、按量付费)。Labelbox 的专家网络和 Surge 的质量定位说明,它们有能力从 Scale 客户手中拿走多供应商并用下的标注支出。 分发能力与供给获取:Scale 的贡献者网络(全球累计支付 $1B+)是一项专有的人类标注员供给。竞争对手必须建立同等网络,才能在质量和吞吐量上竞争。运营 AI 人才市场的 Mercor 正试图建立替代专家贡献者供应链,直接与 Scale 的贡献者网络竞争。 [CP021, CP022, CP023, CP024, CP025]
| 护城河 / 风险 | Scale AI 地位 | 耐久性 | 主要威胁 | 尽调信号 |
|---|---|---|---|---|
| 政府安全许可(DoD IL4、FedRAMP High) | 在已披露的 AI 数据供应商中独有 | 很高 | 竞争对手复制需要数年 | 活跃 DoD 合同验证商业价值 |
| 标注质量溢价 | 公司称行业领先;已向贡献者支付 $1B+ | 中 | Labelbox、Surge 正缩小质量差距;低成本竞品施压 | 客户 NPS 与输赢单数据未公开 |
| RLHF 领先地位 | 核心产品;顶级实验室过去使用 Scale RLHF | 中 | OpenAI 和 Google 已离开;Surge、Mercor 在 RLHF 上竞争 | 前两大 RLHF 客户流失影响重大 |
| 评估基准(Leaderboard、WMDP) | 可信第三方评估方地位 | 高 | Labelbox Leaderboards;学术基准 | 监管对独立评估的需求增长 |
| Donovan 国防 AI 智能体 | 安全许可市场未披露直接竞争对手 | 很高 | 长期:大型国防承包商可能进入 | 已赢得活跃合同,证明价值 |
| 企业平台(GenAI Platform) | 有差异化,但面对超大规模云厂商竞争 | 中 | AWS、Azure、GCP 资源显著更多 | 企业客户留存数据未公开 |
| 标注商品化 | 核心风险:标注越来越商品化 | 低(风险) | Appen、Snorkel、离岸供应商;合成数据威胁 TAM | 数据标注裁员证明管理层意识到该风险 |
| Meta 投资者冲突 | 对 AI 实验室客户留存构成生存级风险 | 低(风险) | Google 和 OpenAI 已退出 | CNBC 和 TechCrunch 报道已证实 |
持久性评估是分析师基于公开证据作出的判断。“很高”表示预计竞争优势可持续 3 年以上;“低(风险)”表示竞争地位正受到现实威胁。
3.5 护城河持久性、商品化风险与反向证据
Scale AI 的竞争护城河在政府 / 防务 AI 中最强,在商品化数据标注中最弱。政府业务(DoD IL4、FedRAMP High、Donovan、正在执行的防务合同)高度持久,因为安全准入需要多年取得,防务 AI 工作流中的机构知识也无法快速复制。这构成真实、持续多年的竞争优势。 在核心标注市场,Scale 的护城河正在削弱。Google 和 OpenAI 作为客户离开——这是 Scale 两个最大商业关系——说明仅靠质量溢价,并不能为 AI 实验室客户创造牢不可破的锁定效应。竞争对手正在缩小质量差距,质量足够且成本更低的提供商(Appen、SuperAnnotate、离岸团队)仍会在价格敏感型企业中赢得标注工作。 商品化风险真实存在,并在加速。随着 AI 基础模型训练成熟、合成数据能力增强,人类标注市场的 TAM 可能显著收缩。Snorkel AI 等竞争对手正在构建减少单位产出人力需求的标注工具。如果趋势延续,标注市场会分化:一个由低成本提供商主导的大型商品化分段,以及一个规模更小、面向专家评测和政府级工作的高端分段;Scale 在后者的位置更强。 反向竞争证据:(1)Meta 交易后 Google 和 OpenAI 离场,说明 AI 实验室分段存在客户集中风险,切换成本也不足;(2)按 Scale 诉讼,Mercor 主动挖客户,说明竞争对手认为 Scale 客户脆弱;(3)Appen 收入下滑,是纯标注分段面临结构性逆风的警示指标;(4)Scale 2025 年 7 月裁员专门针对数据标注人员,承认公司在正在商品化的分段投入过度。 [CP026, CP027, CP028, CP029, CP030, CP031]
关键竞争位置指标,概括 Scale AI 的护城河强度、脆弱点和尽调缺口。
[CP026, CP027, CP028, CP029, CP030, CP031]3.6 展项
04财务情况
4.1 收入模式与收入来源
Scale AI 的收入来自三条主要来源:(1)企业数据标注和 RLHF 服务,(2)Scale GenAI Platform(企业 AI 应用开发),以及(3)通过 Donovan 和 Public Sector Data Engine 获得的美国政府与防务合同。自助层构成第四条、更小的收入来源,服务研究和实验项目。 企业数据标注分段长期是 Scale 最大收入驱动,支撑公司从 2016 年增长到 2024 年 F 轮。这一分段向企业和 AI 实验室收取基于项目或基于用量的费用,覆盖标注、RLHF 数据收集和模型评测。定价模式定制且不透明——Scale 不公布其标注服务的企业定价。自助层提供 1,000 个免费标注单元,此后按单元收费;这是用来获取小客户的免费增值模型。 Scale GenAI Platform 是利润率更高、更偏 SaaS 的收入来源,目标客户是希望部署定制 AI 应用的企业。这个分段服务包括零售(Etsy、Instacart)和媒体(TIME)在内的 Fortune 500 公司。Scale 为企业客户定制构建 LLM 驱动应用,并接入客户自有数据和工作流。这条收入可能包含订阅和托管服务成分,但定价细节未公开披露。 政府与防务合同构成第三条收入来源,正在增长且具有战略重要性。Scale 持有 DoD IL4 和 FedRAMP High 安全准入,能争取涉密和防务级 AI 数据合同。Donovan 平台(防务 AI 智能体)和 Scale Public Sector Data Engine 面向防务市场,这类市场定价通常不公开,而是基于联邦采购日程。DIU RCV 项目中标和正在执行的 DoD 数据整理合同,确认了商业相关性。这个分段今天可能小于商业标注,但考虑切换成本,潜在规模更大、持久性也更强。 [CI001, CI002, CI003, CI004, CI005]
| 收入来源 | 机制 | 单位 / 定价模式 | 状态 | 收入质量 | 尽调索取 |
|---|---|---|---|---|---|
| 企业数据标注 | 按项目交付的大规模标注;RLHF;面向 AI 实验室 + 企业的评估 | 按项目 / 工作量单位定制定价;企业 SLA 合同 | 活跃但在流失(Google 和 OpenAI 2025 已退出) | 中:项目订单会重复出现,但没有合同化订阅;存在客户集中风险 | 按客户分部拆分收入;前 5 大客户集中度;按队列拆分 NRR |
| Scale GenAI Platform 平台 | 面向定制 GenAI 应用的企业 SaaS + 托管服务;为企业定制 LLM | 定制企业订阅 + 专业服务;无公开定价 | 活跃;据 Droege July 2025 备忘录,为战略增长重点 | 高潜力:平台粘性;劳动强度低于标注 | ARR、ACV、客户数、平台分部毛利率 |
| 政府 / 国防合同 | 面向 DoD、IC 和民用机构的国防 AI 数据标注、评估与 Donovan 平台 | 联邦采购计划;合同制;未公开 | 活跃;增长中;DoD IL4 + FedRAMP High 打开涉密 AI 数据市场 | 高:长期合同、高切换成本、基于安全许可的护城河 | 合同金额、管线规模、IDIQ 或其他采购工具上限 |
| Scale RLHF | 面向 LLM 对齐的人类反馈数据采集;专家标注员网络 | 按反馈实例或项目量定价;企业合同 | 活跃但受 OpenAI 缩减影响;Meta RLHF 合作关系在扩大 | 中:劳动密集;竞争带来毛利压力 | 按客户拆分 RLHF 收入;Meta RLHF 协议范围 |
| 自助式 Data Engine | 面向研究、学术和实验用户的按量付费标注 | 1,000 units 免费;超出后按 unit 计费;无企业 SLA | 活跃;收入贡献小;承担管线 / 获客线索功能 | 低-中:量大但 ACV 低;毛利取决于自动化 | 自助收入、转企业转化率、毛利率 |
收入分类和状态基于公开产品页、媒体报道和分析师推断。公司未公开披露收入。状态反映 2025 Meta 投资后的变化。
[CI001, CI002, CI003, CI004, CI005]客户参与如何在 Scale AI 三个主要收入板块转化为收入、收入成本和毛利润。
收入和毛利率数字为分析师估计。Scale AI 不披露财务报表。估计基于 Appen 公开毛利率、行业基准和员工数代理。
[CI001, CI002, CI003, CI011]4.2 定价、GTM 动作与销售效率
Scale AI 的 GTM 动作主要由企业销售驱动,并辅以自助的产品驱动增长层。企业层依赖专门销售团队、关系驱动型交易,以及与客户一起界定并交付大型标注或平台项目的解决方案工程师。企业合同采用定制报价,配备专门运营支持和 SLA,说明平均合同价值高,销售周期也长于典型 SaaS 公司。 标注服务按用量和项目定价。Scale 的自助层设立了标价锚点:前 1,000 个免费单元之后,用户可按公开单元费率购买。企业客户为大规模标注项目谈判价格,这些项目可能包含专门标注员团队、质量保证工作流和平台集成。公司未公开披露折扣或典型交易规模。 销售周期长度和获客成本(CAC)未公开披露。作为数据服务和基础设施公司,Scale 很可能经历较长企业销售周期(3–12 个月),且包含大量专业服务成分。政府合同销售周期更长(从采购启动到首次收入通常 12–36 个月)。净留存率(NRR)无法验证——Google 和 OpenAI 2025 年离场代表一次严重的 NRR 压缩事件,但完整财务影响未知。 分发渠道包括:直接企业销售、政府采购关系(DoD、IC)、与 AI 实验室提供商的伙伴关系,以及面向开发者和研究人员的自助 API / 平台接入。Scale 与主要投资方(Accel、Amazon、Meta、Nvidia、Cisco)的关系创造了战略联合销售和转介绍渠道,但具体经济影响未披露。 Scale 的数据标注业务收入集中在少数大型 AI 实验室(Google、OpenAI、Meta、Anthropic)。Meta 投资后,Google 和 OpenAI 于 2025 年离场,这构成实质性的 GTM 与集中度风险,并直接压缩标注分段收入。 [CI006, CI007, CI008, CI009, CI010]
| 产品 | 标价 / 入门价 | 企业定价模式 | 关键未知 | 来源 |
|---|---|---|---|---|
| Scale Data Engine(自助式) | 前 1,000 个标注单位免费;超过 1,000 units 后按 unit 计费 | N/A — 仅自助层级;企业采用定制合同 | 单位费率未公开;量折扣结构未知 | scale.com/pricing(官方) |
| Scale GenAI Platform(企业) | 未公开;仅提供企业定制报价 | 企业订阅 + 专业服务;包含专属运营 | 典型 ACV 未知;SaaS 与托管服务收入结构未知 | scale.com/generative-ai-data-engine(官方);docs.scale.com |
| Scale RLHF(企业) | 未公开;按项目定制定价 | 定制项目或按量合同;无公开价目表 | 单任务费率、项目最低金额和量折扣结构未知 | scale.com/rlhf(官方) |
| Scale Evaluation(企业) | 未公开;面向模型开发者 + 公共部门定价定制 | 企业评估合同;可能按模型或按基准测试轮次计价 | 评估定价模式、典型单笔金额、毛利未知 | scale.com/evaluation/model-developers(官方) |
| Donovan(政府) | 未公开;联邦采购计划 | 政府合同定价(IDIQ、FFP 或 T&M);涉密许可项目 | 总合同金额、上限和现有管线未公开 | scale.com/donovan(官方);DoD 合同博客文章 |
企业和政府层级全部采用定制定价,且未公开披露。自助定价入口已由公开定价页证实。本表反映标价结构,不代表实际收入或毛利。
[CI006, CI007]Scale AI 企业标注板块的示意性单位经济流,覆盖获客、合同生命周期和留存驱动因素;由于缺少公开数据,全部为估计。
ACV、赢单率、CAC 和 NRR 均未知(私有数据)。流程结构由公开 GTM 描述和行业基准推断。NRR 方向性估计基于 Google 和 OpenAI 离开事件。
[CI007, CI008, CI009, CI013]4.3 成本结构、毛利率与资本强度
Scale AI 的收入成本主要由人力构成——负责数据标注、RLHF 反馈收集和模型评测任务的标注贡献者。Scale 已向全球贡献者网络支付超过 $1 billion,印证了核心标注业务的劳动密集属性。相较纯软件业务,这类人力成本让标注业务的毛利率结构性更低。Appen 作为公开上市的标注可比公司,其行业参照显示,标注服务毛利率大约在 25–45% 区间,且会随着市场商品化继续下行。 Scale GenAI Platform 和评测产品的毛利率应高于标注服务,因为软件杠杆更强、单位人力投入更少。但企业级 GenAI Platform 项目仍需要大量专业服务和运营投入,毛利率因此低于纯 SaaS 标杆。政府合同中的数据服务通常只有 15–30% 的利润率,受采购费率结构约束。 运营费用包括工程和研发(搭建标注工具、GenAI Platform、Donovan 和评测基础设施)、销售和营销(企业销售团队、政府业务拓展)以及一般行政费用。2025 年 7 月裁员前,Scale 约有 1,400 名员工;裁员后约 1,000 人。14% 的员工裁减叠加 500 名承包商缩减,意味着运营费用明显下降,主要指向数据标注成本底座。这也符合靠削减低毛利、商品化标注劳动力抬升综合利润率的路径。 Scale 模型的资本强度主要体现在营运资金(向贡献者支付劳务费),而不是实物资本开支(硬件、设施)。基础设施成本(标注工具和模型评测所需的云计算)是次要资本开支因素。因此,相比硬件 AI 公司,Scale 资本效率更高;但相比纯 SaaS,又更依赖人力。$1 billion 的 Series F 和 Meta 战略投资为公司提供了充足资本,用于转向利润率更高的企业和政府服务。 [CI011, CI012, CI013, CI014, CI015]
| 指标 | 数值 / 估计 | 可信度 | 重要性 | 尽调索取 |
|---|---|---|---|---|
| 平均合同金额(企业标注) | 估计每个项目 $1M–$5M+ | 低(分析师估计) | 推动收入规模;反映企业买方深度 | 向 CFO 索取单笔金额分布;用 Appen 披露作代理 |
| 毛利率(标注分部) | 估计 25–45% | 低(Appen 代理;分析师估计) | 核心毛利驱动;商品化正在压低毛利 | 索取分部 P&L;对比 Appen 已披露毛利率 |
| 毛利率(GenAI Platform) | 估计 45–65% | 低(SaaS + 托管服务代理) | 若经常性 SaaS 占主导,则体现平台经济性 | 索取分部毛利拆分;订阅与服务收入占比 |
| 毛利率(政府 / 国防) | 估计 15–30% | 低(联邦服务行业代理) | 采购约束分部的毛利上限 | 审查合同类型(T&M vs. FFP)和定价结构 |
| 净留存率(NRR) | 2025 可能 <100%(Google + OpenAI 退出) | 低 — 基于方向性证据估计 | 指示收入持久性和客户健康度 | 索取按队列和分部拆分的 NRR;Google/OpenAI 收入影响 |
| 获客成本(CAC) | 未披露;估计企业 / 政府获客成本高 | 未知 — 无可用代理 | 增长投入效率;回本周期 | 索取按分部拆分的 CAC;销售人数和配额达成率 |
| 人均员工收入(代理) | 估计每名员工 $200K–$500K 收入(若 ARR 为 $300M、员工 800 人) | 很低 — 两个输入均为估计 | 与数据服务可比公司进行资本效率对标 | 需确认员工数 + 收入;尽调中索取 |
| 年烧钱速度(裁员前) | 估计 $400M–$700M/year(1,400+ 名员工 + 标注员) | 低(基于员工数的代理) | 现金跑道和成本效率 | 索取月度现金流量表;P&L 运营费用拆分 |
所有单位经济指标均为估计或不可得。Scale AI 不披露财务报表。估计来自 Appen 公开财务、行业基准和基于员工数的代理。可信度整体较低;尽调中应以实际数据替换。
[CI011, CI012, CI013]Scale AI 关键财务指标的分析师估计区间,来自公开代理、Appen 可比公司、员工数代理和已披露融资数据。所有估计置信度都很低。
收入估计来自:基于员工数的代理(1,000 名员工 × 约 $350K 收入 / 员工)、Appen 可比每员工收入分析,以及独立分析师评论。烧钱估计来自员工数 × 全口径成本代理。所有估计仅作示意,未经管理层确认。
[CI011, CI012, CI013, CI018, CI019, CI020]4.4 资本充足性与融资历史
Scale AI 累计外部融资约 $1.6 billion+。融资历史包括:2016–2019 年的种子轮和早期轮次(包括 Y Combinator、Index Ventures、Founders Fund);2021 年以约 $7.3 billion 估值完成的 $325 million Series E;Series F 前累计约 $600 million;以及 2024 年 5 月由 Accel 领投、Amazon、Meta、Cisco、Intel、AMD、ServiceNow、Nvidia、DFJ Growth、WCM 以及 Tiger Global、Thrive、Greenoaks、Wellington、Nat Friedman、Elad Gil 等老股东参与的 $1 billion Series F,估值 $13.8 billion。Series F 同时包含新股(公司获得新现金)和老股交易(为现有股东提供流动性)。 2025 年 6 月 Meta 战略投资约 $14.3 billion,获得少数股权(约占已发行股权 49%)。关键在于,Meta 投资的大部分资金分配给了现有股东和已归属股权持有人,而不是进入公司运营金库,符合老股交易结构。Meta 交易中实际进入 Scale 运营业务的新股资金尚未完全披露。Scale 仍保持独立公司结构;Meta 持有少数股权。 没有公开财务报表,现金余额未知。Series F 于 2024 年 5 月完成;如果其中相当部分为新股资金($500M+ 进入公司),再叠加 Meta 交易的新股部分,按当前烧钱速度,Scale 很可能拥有 3+ 年的充足现金跑道。2025 年 7 月裁减 200 名员工和 500 名承包商,加上战略上撤离数据标注,都是降低成本的动作,意在拉长现金跑道、提升现金效率。 烧钱速度未公开披露。Scale 裁员前的费用底座(1,400+ 名员工、大量标注承包商成本)意味着月度烧钱规模不小。裁员后,现金效率应明显改善。Scale 未公开披露债务融资安排或项目融资义务。Meta 交易带来了结构性依赖(Meta 同时是投资方,也是这些 AI 服务的关键客户),但未披露任何经济性约束或限制条款。 [CI016, CI017, CI018, CI019, CI020]
| 项目 | 数值 / 估计 | 来源 / 依据 | 可信度 | 备注 |
|---|---|---|---|---|
| 累计融资 | ~$1.6B+(Series F 前 ~$600M + $1B Series F) | TechCrunch Series F 报道(May 2024) | 高(多家高声誉来源证实) | 新股 + 老股交易混合;确切新股分配未知 |
| Series F 交割 | May 2024;$1B、估值 $13.8B;Accel 领投 | TechCrunch / CNBC 报道 | 高(多个来源相互印证) | 包括新投资方 Amazon、Meta、Cisco、Intel、AMD、ServiceNow |
| Meta 战略投资 | ~$14.3B 换取少数股权(~49%);估值 >$29B | CNBC June 2025 报道 | 高(多个来源相互印证) | 主要为老股交易(现有股东流动性);新股部分未知 |
| Meta 交易带来的新股净资本 | 未知 — 主要为老股分配 | CNBC/TechCrunch 报道;Scale 博客 | 低(从老股交易结构表述推断) | Meta 持有少数股权;资金大多流向股东 |
| 估计账上现金(2025 后) | 估计 $500M–$1B+ | 分析师基于 Series F 新股 + 部分 Meta 新股估计 | 很低(仅为估计) | 无公开财务报表;仅为估计 |
| 月度烧钱速度(裁员后) | 估计 $25M–$50M/month(July 2025 重组后) | 分析师基于员工数 × 单员工成本代理估计 | 很低(仅为估计) | 裁员前烧钱明显更高;重组后改善 |
| 估计现金跑道 | 从 July 2025 起估计 24–48 个月(若现金 $500M–$1B、月烧钱 $25M–$50M) | 由现金和烧钱估计推导 | 很低(两个输入均为估计) | 现金跑道估计需确认 Meta 交易中的新股资本 |
| 已披露债务 / 信贷额度 | 未公开披露 | 公开来源审查 | 中(未见公开债务证据) | 可能存在私募信贷或基于收入的融资,但未披露 |
Series F 和 Meta 投资的融资金额已由高声誉新闻来源证实。账上现金、烧钱速度和现金跑道均为分析师估计,可信度很低。尽调必须取得经审计财务报表。
[CI016, CI017, CI018, CI019]示意性资本流,展示 2021 至 2026 年主要融资流入和估计运营流出,突出业务模式转型的资金充足性和烧钱背景。
所有数值均为分析师估计。Series F 和 Meta 交易的新股 / 老股拆分未披露。Series E $325M 可能同时包含新股和老股。运营支出代理基于员工数和行业基准。该瀑布图仅作示意,不是经审计现金流量表。
[CI016, CI017, CI018, CI019]4.5 财务结论——收入质量、利润率路径与尽调阻断项
Scale AI 的财务画像复杂:资本充足,收入轨迹不清晰,管理团队又要在主要客户明显流失的背景下执行高风险商业模式转型。 收入质量令人担忧。数据标注板块曾是 Scale 最大的收入驱动,但定位已接近商品化,正面对结构性逆风(Appen 公开收入下滑是先行指标)、客户流失(Google 和 OpenAI 离开)以及直接竞争。GenAI Platform 和政府 / 国防板块的收入质量更高(合同更长、锁定更强、商品化风险更低),但它们目前对总收入的贡献未知,且很可能仍较小。2025 年之后,AI 实验室板块的净留存率(NRR)很可能低于 100%。 如果转型成功,利润率方向上会改善:剥离低毛利数据标注劳动力、扩大更高毛利的平台和政府板块,将随时间抬升综合毛利率。但转型本身有执行风险,中间阶段(2025–2027 年)很可能先出现收入压缩,随后平台和政府收入才开始放量。 资本充足性是最强的财务正面因素。Scale 已融资 $1.6B+,Meta 交易又为股东提供了额外流动性(并带来部分新股资金)。裁员后,公司以更低员工数和更低烧钱速度运营。现金跑道估计为 3+ 年,足以支撑转型执行。 主要尽调阻断项:(1)没有公开收入或 ARR 数据,估值倍数无法验证;(2)Google 和 OpenAI 流失对收入的影响未知;(3)政府合同收入规模和增长轨迹未披露;(4)分板块毛利率未披露;(5)Meta 交易中新股与老股交易的拆分未确认;(6)Mercor 诉讼结果可能带来财务影响(赔偿、客户流失)。 [CI021, CI022, CI023, CI024, CI025]
| 缺失数据 | 对尽调的影响 | 具体尽调路径 |
|---|---|---|
| 总收入 / ARR | 无法验证估值倍数;无法评估增长或收入质量 | 索取经审计财务;取得管理账;用四大审计确认 |
| 按分部拆分收入(标注 vs. 平台 vs. 政府) | 无法评估商业模式转型是否成功;分部毛利无从判断 | 向 CFO 索取分部 P&L 拆分;按分部拆分合同 |
| Google 和 OpenAI 收入影响(2025 流失) | 财务模型最大风险:最大客户收入流失规模未知 | 索取按客户拆分收入;确认 Google 和 OpenAI 退出日期和最终账单 |
| 按产品线拆分毛利率 | 无法评估毛利改善逻辑或资本强度;综合毛利不透明 | 索取按分部拆分毛利;标注员成本占收入 %;平台毛利 |
| 净留存率(NRR) | 无法评估客户健康度;经常性收入持久性无从判断 | 索取按年度队列(2021–2025)和分部拆分的 NRR |
| Meta 交易新股 vs. 老股交易拆分 | 无法评估公司从 Meta 交易中实际增加的现金 | 索取 Meta 投资交割文件;确认新股资本分配 |
| 经营现金流和烧钱速度 | 无法可靠估计现金跑道或资本充足性 | 索取月度 P&L 和现金流量表(Jan 2024 – present) |
| 员工薪酬结构 | 无法评估 Meta 交易和裁员后人才留存是否守住 | 索取匿名化薪酬区间和股权归属计划 |
| 政府合同积压订单和管线 | 无法测算国防分部规模或增长轨迹 | 索取 DoD/IC 合同积压总额、期权期和管线 |
| Mercor 诉讼财务敞口 | 正在进行的 IP / 商业秘密诉讼可能带来财务责任 | 审查已提交法律主张;评估和解风险;确认保险覆盖 |
本表所有项目均代表截至 2026-05-09 已确认的证据缺口。本表是 Scale AI 财务承销的主要尽调清单。
[CI021, CI022, CI023, CI024, CI025]4.6 证据材料
05产品与技术
5.1 产品组合与客户工作流
Scale AI 的产品组合对应 AI 开发中的四类客户工作流问题:(1)为模型开发创建高质量训练数据;(2)为 LLM 对齐和 RLHF 收集人工反馈;(3)独立评估 AI 模型安全性和能力;(4)以企业级规模部署定制 GenAI 应用。第五条产品线 Donovan 面向国防和情报任务中的专用 AI 工作流。 Scale Data Engine 是 Scale 的核心产品,也是公司的根基。它帮助 AI 团队跨模态(文本、图像、视频、音频、3D)收集、整理、标注并质检训练数据集。客户工作流是:AI 工程师定义标注需求和质量标准;Scale 的贡献者网络在 Scale 自研工具和 QA 流水线指导下执行标注任务;标注后的数据集再交付回客户的 MLOps 流水线。Data Engine 覆盖模型开发全生命周期,从预训练数据整理,到 RLHF,再到评测。 Scale GenAI Platform 解决企业 AI 部署工作流。企业客户(覆盖零售、媒体、金融、政府等领域的 Fortune 500 公司)带来自有数据和领域需求;Scale 将这些输入转化为定制化 LLM 应用。平台包含数据摄取、LLM 微调、RAG 流水线配置和生产部署支持。该产品瞄准想要定制 AI、但缺少从零搭建所需深度 ML 能力的企业。 Scale Evaluation 和 Scale Leaderboard 提供独立模型基准测试服务。AI 实验室和企业使用 Scale Evaluation,从推理、安全、编码、领域知识等维度评估 LLM 能力。Scale Leaderboard 是 LLM 表现的公开排名,是面向开发者的工具,也让 Scale 在 AI 社区建立了可信第三方评测方的定位。WMDP(Weapons of Mass Destruction Proxy)基准是其在 AI 安全上的专项评测贡献。 Donovan 服务 DoD、IC 和民用政府机构,为关键任务工作流提供 AI 智能体。它接入涉密数据源,支持有授权要求的运营环境,并为国防和情报应用提供 AI 驱动的决策支持。Donovan 部署在 DoD IL4 授权环境中,是 Scale 在国防板块的主要竞争护城河。 [CE001, CE002, CE003, CE004, CE005]
| 产品 / 模块 | 主要用户 | 成熟度 / 状态 | 核心差异化 | 尽调缺口 |
|---|---|---|---|---|
| Scale Data Engine | AI 实验室;企业 AI 团队;研究 | 成熟(GA);2016 起的核心产品 | 自研 QA 流程;贡献者网络(已支付 $1B+);多模态支持 | 标注质量相对竞品未独立对标;NPS 未知 |
| Scale GenAI Platform 平台 | 企业(F500、媒体、零售);政府 | 活跃;2025 战略增长重点 | 将标注质量嵌入 LLM 定制;托管式企业部署 | 技术深度相对 AWS Bedrock / Azure AI 未独立评估;客户留存未知 |
| Scale RLHF | AI 实验室(Meta、Cohere、Anthropic);企业 | 成熟;核心产品;受 OpenAI/Google 退出影响 | 面向偏好数据的专家标注员网络;用于对齐的质量反馈闭环 | RLHF 收入集中风险;Meta 关系范围未披露 |
| Scale Evaluation + Leaderboard | AI 实验室;企业;政府;AI 安全研究者 | 活跃;增长中;公开开发者工具 | 可信第三方评估方;WMDP 基准;政府评估授权 | Leaderboard 方法论和独立性需第三方审计来建立可信度 |
| Donovan(国防 AI 平台) | DoD;IC;民用政府机构 | 活跃;已进入涉密许可生产;专为 IC/DoD 设计 | DoD IL4 认证;涉密环境部署;国防任务 AI 智能体 | 技术能力未公开成文;与 Palantir 的竞争比较未知 |
| Scale 公共部门数据引擎 | 美国政府;国防;情报界 | 活跃;已进入涉密许可生产 | FedRAMP High;面向国防的数据整理和管理 | 合同规模和增长管线未公开;拓展到其他机构的时间表未知 |
| 自助式 Data Engine | 研究者;初创公司;个人 AI 开发者 | 活跃;GA;转型后优先级较低 | 进入门槛低;1,000 units 免费;API 接入;任务类型多样 | 转企业转化率未知;自助层级毛利未披露 |
成熟度评估基于公开产品页描述和媒体报道。尽调缺口反映声称能力缺乏独立第三方验证。
5.2 架构与运营模式
Scale AI 的运营模式把软件工具和人在回路执行结合起来——它是 AI 辅助的混合标注架构,而不是纯软件或纯劳动力模型。核心技术栈包括:(1)面向跨模态标注任务的网页标注工具平台;(2)负责复核、调和和反馈的质量保证流水线;(3)向企业和实验室客户以程序化方式开放 Scale 服务的 API 基础设施;(4)支持 LLM 定制工作流的 GenAI Platform;(5)面向国防任务的 Donovan AI 智能体运行时。 标注贡献者网络是 Scale 自有的全球劳动力,公司负责招募、培训、质检筛选和管理。贡献者平台(Scale 称之为「任务市场」)把标注任务分配给具备相应技能的贡献者,实时监控完成质量,并在质量敏感任务中升级给专家复核。Scale 把贡献者质量管理当作核心差异化能力来投入,并声称自研 QA 流水线产生的标注质量高于竞争对手。 Scale API 是 Data Engine 面向开发者的主要接口,支持以程序化方式提交标注任务、取回已完成数据集,并接入 ML 训练流水线。该 API 支持所有标注任务类型(文本分类、边界框、分割、RLHF 偏好对等),并提供 webhook,以便实时通知任务完成。 GenAI Platform 的运营模式包括:数据摄取和预处理、LLM 选择和提示词工程、微调或 RAG 流水线配置、红队测试和安全评估,以及带监控的生产部署。这是一种托管服务路径,由 Scale 工程师和平台工具共同交付定制 AI 应用。 Donovan 的架构专为有授权要求的环境设计:它运行在 DoD IL4 认证云基础设施上,接入涉密政府数据系统,支持多模态 AI 智能体工作流(规划、搜索、决策支持),并提供军事应用所需的可解释性功能。Donovan 的技术架构细节未公开披露;其运营模式更接近托管式国防 IT 服务,而不是商业 SaaS 产品。 [CE006, CE007, CE008, CE009, CE010]
| 用户任务 | 当前工作流 | Scale 解决方案 | 可衡量收益 | 限制 |
|---|---|---|---|---|
| 用标注数据训练 ML 模型 | 内部团队或众包平台手工标注;慢且不一致 | Scale Data Engine:通过 API 提交原始数据;获得已标注数据集;自动 QA | 相比内部团队吞吐量 10x+;一致性更高(公司声称) | 定制定价;Meta 冲突后 AI 实验室客户退出 |
| 用人类偏好对齐 LLM(RLHF) | 自建偏好采集流程;聘请专家评估员 | Scale RLHF:大规模专家人类反馈;成对标注;偏好排序 | 为 LLM 对齐训练提供高质量专家反馈 | OpenAI 和 Google 已退出;Meta 关系扩大但有集中风险 |
| 为企业部署定制 GenAI 应用 | 由 ML 工程师内部自建;典型周期 12-24 个月 | Scale GenAI Platform:数据 → 定制 LLM 应用,< 2 个月(TIME 案例研究) | TIME 案例研究:<2 个月部署;测试 7,000+ 个攻击向量 | 超大规模云厂商(AWS、Google、Azure)以更低成本提供类似平台 |
| 独立评估 LLM 能力与安全 | 使用学术基准(MMLU、HELM);内部红队 | Scale Evaluation + Leaderboard;WMDP 基准;专家安全评估 | 可信第三方评分;政府认可的评估权威 | 独立性感知风险(Meta 作为投资方可能使评估结果偏倚) |
| 执行 AI 驱动的国防任务 | 分析师手工作业;涉密环境中 AI 集成有限 | Donovan:面向 IC/DoD 决策支持和任务规划、获涉密许可的 AI 智能体 | 首个面向 DoD、获涉密许可的 AI 智能体平台;无已披露可比竞品 | 技术能力未完整成文;商业透明度有限 |
Scale GenAI Platform 部署速度收益主张基于 TIME 案例研究。其他主张基于官方产品页的公司能力描述。
Scale AI 的五层产品架构,从面向客户的界面到 AI 服务、工作流执行、合规和基础设施。
架构由公开产品页面、API 文档和官方安全认证描述推断。Donovan 机密能力没有公开文档。
[CE006, CE007, CE008, CE016]Scale AI 核心标注产品的端到端客户工作流,从客户提交数据,到贡献者执行、QA,再把数据集交付至客户 MLOps 管线。
该工作流根据公开 API 文档、产品说明,以及 Scale Leaderboard / Evaluation 公开材料重建。具体 QA 算法属于自研体系,未公开披露。
[CE006, CE007, CE008, CE009]5.3 差异化、技术 IP 与竞争护城河
Scale AI 的关键技术和运营差异化分为五类。第一,自研标注工具和 QA 方法论:Scale 大量投入标注界面,使其在各类任务上兼顾准确性、速度和一致性。QA 流水线使用统计质量控制、一致性检查和专家复核升级,维持公司声称高于竞争对手的标注质量标准。这套自研工作流不开源,是嵌入运营中的组织经验。 第二,贡献者网络本身是供给侧 IP:Scale 的全球贡献者网络(已获得 $1B+ 支付)是一项自有资产。支配该网络的质量筛选、入职培训、技能认证和任务路由逻辑,是 Scale 在标注板块的运营护城河。竞争对手(包括 Scale 当前起诉、指控其挖人的 Mercor)必须搭建同等网络,才能在规模化质量和吞吐上竞争。 第三,政府认证构成监管护城河:Scale 拥有 DoD IL4 Provisional Authorization 和 FedRAMP High Authorization。这些许可需要 2–4 年合规投入,并要求持续合规维护。公开披露的 AI 数据竞争对手中,没有公司具备同等认证,这让 Scale 在涉密 AI 数据合同中接近独家。复制这些能力所需的成本和时间,是政府 AI 数据领域真实的进入壁垒。 第四,Scale Leaderboard 和 WMDP 基准构成声誉 IP:Scale Leaderboard 已让 Scale 成为 LLM 能力的可信第三方评测方。用于 AI 安全评估的 WMDP(Weapons of Mass Destruction Proxy)基准已被 AI 安全研究社区采用。这些基准是公共品,为 Scale 带来声誉资本,使其成为权威 AI 评测方,也支撑商业评测合同和政府关系。 第五,Donovan 平台在国防 AI 智能体上具备先发优势:Donovan 是首个商业化部署、获得 DoD IL4 清关的国防任务工作流 AI 智能体平台。专用产品能力、政府关系和清关基础设施叠加,让 Donovan 成为难以替换的国防 AI 平台;随着 DoD 提高 AI 采用预算,这一点尤其重要。 [CE011, CE012, CE013, CE014, CE015]
| 层级 / 组件 | 作用 | 关键依赖 | 主要风险 |
|---|---|---|---|
| 标注工具平台(web UI + API) | 贡献者执行标注任务和客户提交作业的界面 | 云基础设施(AWS/GCP/Azure);贡献者网络 | AI 辅助标注减少人工需求后,工具可能过时 |
| 质量保证流程 | 统计 QC;共识评分;升级给专家审核员 | 贡献者网络质量;自研 QA 算法 | QA 方法为自研且未经独立审计;竞品正在缩小质量差距 |
| Scale API(标注 + GenAI + 评估) | 为企业 + 开发者集成提供所有 Scale 服务的程序化访问 | 云基础设施;API 安全;访问控制 | API 可用性和可靠性未独立对标;无公开状态页 |
| 贡献者网络(全球标注员) | 由人执行标注、RLHF 和评估任务 | 全球劳动力管理;质量筛选;薪酬发放后勤 | 劳动力供应链风险;Mercor 正尝试搭建竞争网络 |
| Scale GenAI Platform(LLM 定制) | 企业定制 AI 应用开发:RAG、微调、部署 | 云端 LLM 提供商(OpenAI、Anthropic、Meta);客户数据安全 | 超大规模云厂商竞争;Meta 交易后,与专有 LLM 提供商的关系可能变化 |
| Donovan Runtime(国防 AI 智能体) | 面向 DoD/IC 任务的获准 AI 智能体执行;任务规划;搜索 | DoD IL4 认证云;涉密数据集成;具备安全许可人员 | 依赖政府认证持续有效;技术细节未公开披露 |
| 评测 / 排行榜基础设施 | LLM 基准测试;WMDP 安全评测;公开排行榜发布 | AI 实验室模型 API;独立评测基础设施 | 独立性观感风险;Meta 投资 Scale 后,LLM 排名可能出现冲突 |
架构评估基于公开文档、产品页面和 API 参考资料。Donovan 架构部分来自推断;涉密能力未公开记录。
Scale AI 的关键外部依赖覆盖基础设施、认证机构、客户、投资者和监管关系,任何一项都可能影响产品交付或竞争位置。
依赖关系根据公开产品说明和媒体报道推断。具体合同和商业条款未公开披露。
[CE011, CE012, CE013, CE014]从四个维度评估 Scale AI 的五个核心产品:技术成熟度、企业适配度、竞争护城河和技术深度,评分 1-5(5=最高)。
评分为分析师基于公开证据作出的估计。成熟度 = 1(预发布)到 5(已验证 / 稳定)。企业适配度 = 1(不适合)到 5(企业核心)。竞争护城河 = 1(商品化)到 5(近乎独占)。技术深度 = 1(基础)到 5(高级 / 自研)。
[CE011, CE012, CE013, CE014, CE015]5.4 信任、安全、合规与质量控制
Scale AI 大量投入信任和合规基础设施,尤其服务政府客户。公司公开披露的认证包括:SOC 2 Type II(企业安全控制审计)、ISO 27001(信息安全管理体系)、DoD IL4 Provisional Authorization(美国国防部针对受控非密信息和部分涉密数据的数据安全要求)以及 FedRAMP High Authorization(美国联邦政府面向高影响系统、包括涉密数据处理的云安全要求)。 AI 安全上,Scale 已根据 2024 年白宫 AI 安全承诺作出自愿承诺,覆盖 RLHF 安全实践、红队测试和负责任 AI 部署。Scale 贡献了 WMDP(Weapons of Mass Destruction Proxy)基准,用于评估 LLM 是否经过训练以防止生成危险内容。WMDP 基准衡量双用途知识领域中的 AI 安全,Scale 的测试与评估白皮书也记录了其负责任 AI 模型评估方法。 Scale 的标注质量控制是自研体系,但包括:特定任务质量指南、高风险任务的多标注者冗余、统计质量监控、专家复核升级以及标注者间一致性评分。公司声称标注质量行业领先,但没有公开可得的独立第三方质量审计。 TIME 客户案例(GenAI 平台在不到 2 个月内部署,并测试 7,000+ 个攻击向量)为红队能力提供了部分证据。DIU RCV 项目中标确认 Scale 技术通过了国防采购要求。这些能佐证技术能力,但仍低于独立第三方技术审计的证据强度。 隐私和数据治理:Scale 的企业标注数据处理涉及客户提供的数据,可能包含专有或敏感信息。Scale 的安全认证(SOC 2 Type II、ISO 27001)为数据治理控制提供了一定保证。对政府客户,DoD IL4 和 FedRAMP High 施加严格数据处理要求,Scale 已证明其符合要求。标注流水线中的客户数据处理,尤其是 AI 实验室的专有训练数据,仍是公开材料未完全回答的尽调关注点。 [CE016, CE017, CE018, CE019, CE020]
| 控制项 / 认证 | 状态 | 范围 | 验证方 | 缺口 / 尽调问题 |
|---|---|---|---|---|
| SOC 2 Type II | 已认证(已确认) | 商业数据处理;内部安全控制;企业标注工作流 | 第三方审计机构(未公开具名) | 索取最新 SOC 2 报告;确认范围覆盖标注数据管线 |
| ISO 27001 | 已认证(已确认) | 信息安全管理体系;全球运营 | 第三方认证机构 | 确认证书日期和范围;索取 ISO 证书 |
| DoD IL4 临时授权 | 已认证(已确认) | DoD Impact Level 4 数据——受控非密信息 + 敏感国防数据 | DISA / DoD 授权机构 | 确认 PA 仍有效;索取特定 Donovan 部署的 ATO |
| FedRAMP High 授权 | 已获授权(已确认) | 美国联邦高影响系统;涉密和敏感政府数据 | FedRAMP PMO / JAB 授权 | 在 FedRAMP marketplace 确认授权仍有效;确认获授权服务范围 |
| 白宫 AI 安全承诺(2024) | 已承诺(公司自愿签署) | RLHF 安全;红队测试;负责任 AI 部署;模型评测 | 白宫 OSTP;自愿性质,无法律约束力 | 确认承诺落地情况;审查 Scale 公开的承诺追踪器 |
| WMDP 基准(大规模杀伤性武器代理) | 已发布(公开可用) | 面向防止两用知识滥用的 AI 安全评测;LLM 安全测试 | AI 安全研究社区采纳情况 | 确认其他实验室采纳 WMDP 的情况;独立验证基准方法 |
| 标注质量标准 | 内部(专有) | 标注员间一致性;QA 管线;按任务划分的准确率标准 | 内部(自报);无公开第三方审计 | 索取质量审计方法、标注员间一致性分数、客户 NPS |
| 数据隐私 / 客户数据处理 | 内部控制(SOC 2 覆盖部分) | 标注管线中的客户专有数据;AI 实验室训练数据保密 | SOC 2 Type II 部分覆盖;标注数据未获独立验证 | 索取数据处理协议;确认标注管线中的客户数据隔离 |
认证来自 Scale 官方网站(scale.com/legal/security)确认。白宫承诺由 Scale 博文确认。质量标准仅基于公司描述的方法论。
[CE016, CE017]5.5 产品路线图、部署与待解问题
截至 2025 年,Scale AI 的产品路线图被 2025 年 7 月转向企业和政府的战略转型隐性牵引:优先推动 Scale GenAI Platform 企业部署,扩展面向政府机构的 Donovan,并扩大 Scale Evaluation 能力,以服务正在形成的 AI 治理和合规市场。数据标注板块虽仍在运营,但在 2025 年 7 月裁减 200 名员工和 500 名承包商之后,预计会以更小规模运转。 部署和集成能力:Scale 提供企业 API 访问、面向 MLOps 流水线的 webhook 集成,以及用于应用开发的 Scale GenAI Platform。自助门户支持标注任务快速上线。政府侧,Donovan 部署在涉密云环境,并为 DoD 和 IC 系统提供专用集成。Scale 没有公开变更日志或发布说明;公司也不维护可独立验证的公开产品路线图或面向开发者的状态页。 开发者信号:Scale Leaderboard 是最面向开发者的公开产品,在 AI 研究和工程社区获得明显关注。WMDP 基准已被 AI 安全研究引用。Scale 的 API 文档(scale.com/docs)提供程序化访问规格,但相较开发者生态更强的 AI 基础设施公司(如 Hugging Face、LangChain),其开源工具或开发者社区参与度(GitHub 活动、HackerNews 讨论、包下载量)有限。 关键待解问题和产品尽调缺口:(1)GenAI Platform 相对超大规模云厂商竞争(AWS Bedrock、Google Vertex AI、Azure AI)的技术深度未获独立验证——Scale 的平台优势依赖标注质量集成,但超大规模云厂商可收购或合作标注供应商,获得同等能力;(2)Donovan 在涉密环境中面向 AI 智能体的具体技术能力,公开资料披露不足,无法评估其相对其他国防 AI 公司(Palantir、Booz Allen)的差异化;(3)标注工具相对 Labelbox、Snorkel AI 和开源替代方案(CVAT、LabelImg)的技术优越性只是公司主张,未被独立基准验证;(4)Scale 的合成数据能力开发路线图未知——合成数据可能替代人工标注,这是关键缺口。 [CE021, CE022, CE023, CE024, CE025]
| 日期 / 阶段 | 功能 / 里程碑 | 状态 | 影响 | 来源 |
|---|---|---|---|---|
| 2016–2022 | Scale Data Engine v1 → 成熟标注平台;API 发布 | 已完成(历史) | 基础产品;奠定标注市场地位和贡献者网络 | Scale 公开博客;公司历史 |
| 2021–2022 | Scale RLHF 产品发布;面向 OpenAI + AI 实验室的 LLM 训练数据 | 已完成(历史) | 让 Scale 成为前沿 AI 开发中的 RLHF 领导者 | Scale RLHF 页面;TechCrunch 报道 |
| 2022 | Scale Donovan(国防 AI 智能体)发布;DoD IL4 获批 | 已完成;活跃 | 建立获准国防 AI 数据地位;形成政府收入护城河 | Scale Donovan 页面;DoD 合同博客 |
| 2023–2024 | Scale GenAI Platform 发布;企业 LLM 定制 | 已完成;活跃;战略优先项 | 从标注扩展到毛利更高的平台业务 | Scale GenAI Platform 文档;客户案例 |
| 2024 | Scale Leaderboard 发布;WMDP 基准发布;白宫 AI 安全承诺 | 已完成;活跃开发者工具 | 让 Scale 获得可信 AI 评测权威地位和政府认可 | Scale 博客(leaderboard、WMDP);白宫记录 |
| May 2024 | Series F 融资 $1B,估值 $13.8B;新增战略投资者(Amazon、Meta、Cisco) | 已完成 | 为平台和政府业务转向提供资金;具备战略联合销售潜力 | TechCrunch Series F 报道 |
| June 2025 | Meta 战略投资;Wang 离任;Droege 任临时 CEO;宣布战略转向 | 已完成;转型仍在推进 | 商业模式转向;CEO 交接风险;客户流失(Google、OpenAI) | TechCrunch;CNBC;Scale 博客 |
| July 2025 | 裁员 14%(200 名员工 + 500 名承包商);数据标注重组 | 已完成 | 成本结构调整;印证标注业务转向;人才留存风险 | TechCrunch July 2025 报道 |
| 2025–2026(计划) | 企业 GenAI Platform 扩张;Donovan 政府部署;Scale Evaluation 增长 | 进行中(根据 Droege 表述推断) | 收入结构转向毛利更高的企业 + 政府;转型执行风险 | Scale 博客;Droege 公开评论 |
| Unknown | 合成数据能力;AI 辅助标注自动化 | 未公开确认 | 关键缺口:若未开发,Scale 容易被合成数据替代 | 证据缺口——无公开路线图 |
历史里程碑基于媒体记录和公开公告。前瞻路线图事项来自公开表述推断;Scale 不发布正式产品路线图。
[CE021, CE022]5.6 证据材料
06客户情况
6.1 客户群分层
Scale AI 的客户群分为三大主要垂直领域:AI 实验室和模型开发者、Fortune 500 企业客户、美国政府及国防机构。每个板块都有不同的买方画像、采购机制、使用场景、收入结构和切换成本。 AI 实验室(此前是最大板块)购买 Scale 的 RLHF 和评测服务,用于训练和基准测试大语言模型。这些客户历史上包括 OpenAI、Google/DeepMind、Cohere 和 Anthropic,是技术能力很强的买方,拥有直接采购权,合同周期按季度与训练算力排期绑定。该板块目前进入结构性下滑:Google 于 2025 年 6 月退出,理由是 Meta 交易带来的竞争冲突;OpenAI 同月结束与 Scale 的关系,释放出前沿模型开发者更广泛转向自建标注能力的信号。 企业客户——包括 TIME、Etsy、Instacart 和 Pinterest——将 Scale 的 GenAI Platform 和 Data Engine 用于特定领域 AI 应用,例如内容安全测试、推荐系统和电商 AI。这些买方通常涉及 IT 和数据领导层,采购周期为 6–18 个月,并承诺多年部署。TIME 案例是最强的公开证据:GenAI Platform 在不到两个月内部署,在生产安全应用中针对 TIME 的 AI 内容输出测试了 7,000+ 个对抗攻击向量。 政府和国防客户——主要是美国国防部和情报共同体机构——通过多年期合同工具采购数据整理、自主 AI 项目,以及用于涉密行动的 Donovan 平台。由于 DoD IL4 和 FedRAMP High 认证要求、涉密环境集成和采购惯性,这些客户切换成本最高。地域上,Scale 已披露客户群主要总部位于美国。国际客户数或收入拆分未公开披露。 [CU001, CU004, CU005, CU006, CU007, CU008]
| 客群 | 买方 / 付款方画像 | 主要用例 | 规模 / 范围 | 收入 / 战略价值 | 关键缺口 |
|---|---|---|---|---|---|
| AI 实验室 | CTO / 研究负责人;季度采购与训练排期绑定 | RLHF、模型评测、安全基准测试 | 历史上有 2-3 个旗舰客户(OpenAI、Google);现在转向较小实验室(Cohere) | 历史最大客群;Google/OpenAI 退出后进入结构性下滑 | 客户数、单实验室收入、NRR 未披露;退出影响未量化 |
| 企业(Fortune 500) | IT / 数据负责人;6-18 个月销售周期 | GenAI Platform 部署、垂直领域 AI、数据整理 | 多个具名标识(TIME、Etsy、Instacart、Pinterest);总数未披露 | 管理层称在增长;无 ARR、NRR 或扩张率数据 | 扩张率、增购转化、企业客户总数未披露 |
| 政府 / 国防 | 项目经理 / 合同官;多年期 RFP | 自主 AI、联合部队数据整理、Donovan 平台(国家安全) | 活跃 DoD 合同;IC 部署;IL4 / FedRAMP High 认证 | 战略价值高;推定收入底座;最耐久客群 | 合同金额、机构名称(涉密)、续约时间、分部 ARR 未披露 |
| 自助 / SMB / 研究 | 开发者 / 研究员;按量付费 | 标注、ML 实验 API 访问、模型评测 | 1,000 个免费单元入门层;未披露活跃用户数 | 单个收入低;有高规模潜力;转化率未知 | 活跃用户数、企业转化、使用趋势未披露 |
Scale AI 未披露分客群收入拆分。层级根据产品页面、定价结构、官方博客和案例研究推断。政府客群分类基于 DoD 认证和官方合同披露。
[CU001, CU004, CU005, CU006, CU007, CU008]6.2 采用轨迹与增长信号
Scale AI 不公开披露活跃客户总数、收入、ARR 或部署指标。采用信号只能从已披露里程碑、员工数变化、贡献者支付和第三方代理数据中推断。 最强的规模信号是:(1)已向全球标注贡献者支付超过 $1 billion,表明标注吞吐量可观;(2)处理 15 billion+ 个人工标注决策,证明平台深度;(3)2024 年 5 月以 $13.8 billion 估值完成 Series F,Amazon、Cisco、Intel、AMD 和 ServiceNow 等企业买方参与,显示其在相当规模上获得客户验证。战略投资人本身也是企业客户,这为平台成熟度提供了独立佐证。 2025 年 7 月裁减 14% 员工(约 200 名员工和 500 名承包商),且集中在数据标注业务,是滞后的采用信号:它表明 RLHF 和标注量已从峰值明显下滑,并与 Google、OpenAI 离开以及向 GenAI Platform 和 Donovan 的战略转型同时发生。Snorkel AI、SuperAnnotate 和 Mercor 均已扩大企业标注产品,证据来自新产品页和合作伙伴 / 排行榜资料;这说明即便 Scale 正在重新定位,市场活动仍在继续。 自助增长无法量化。Scale 的 API 文档和定价页公开可访问,显示开发者仍在采用,但公司未披露活跃自助客户数或转化率。Scale 客户页面上可见的企业客户——Etsy、Instacart、Pinterest、Cohere——只有标识,没有案例研究或结果指标,因此这些账户的采用证据质量有限。 [CU011, CU012, CU013, CU017, CU021, CU023]
| 指标 | 数值 | 日期 | 来源 | 置信度 | 影响 | 缺失分母 |
|---|---|---|---|---|---|---|
| 贡献者支付额 | 全球已向贡献者支付 $1B+ | 2025 | scale.com/customers(官方) | 中 | 标注吞吐量的代理指标;不等同于客户收入 | 未披露同比增长率或付款轨迹 |
| 人工标注决策 | 已处理 15B+ | 2025 | scale.com/customers(官方) | 中 | 平台吞吐深度;无法映射到客户数或收入 | 未披露活跃客户分母或单客户拆分 |
| Series F 战略投资者 | $1B,估值 $13.8B;Amazon、Cisco、Intel、AMD、ServiceNow 投资 | May 2024 | TechCrunch / Scale 官方 | 高 | 企业买方成为投资者,提供独立客户验证信号 | 交割时未披露收入、ARR 或客户数 |
| 裁员后员工数 | ~1,000 名员工(裁减 14%,~200 人离职) | Jul 2025 | TechCrunch | 高 | 裁员集中在数据标注;是标注客群业务量下滑的代理信号 | 未披露运营产能指标或分部员工数拆分 |
| AI 实验室客户流失 | Google(最大客户)和 OpenAI 均在 Q2 2025 退出 | Jun 2025 | CNBC(两篇独立报道) | 高 | 重大收入集中事件;估计影响 AI 实验室客群 20-40% | 收入影响未披露;公开信息未确认替代管线 |
未公开披露收入、ARR 或客户数数据。采用指标仅为代理和推断。高置信度事项来自已确认的多源新闻报道,或由声誉强的媒体背书的 Scale 官方披露。
[CU009, CU010, CU011, CU012, CU013, CU035]6.3 具名客户证据与证据质量
Scale AI 最重要的公开客户证据覆盖媒体、国防、电商和 AI 研究中的生产部署。TIME Media 案例是质量最高的公开证据:Scale 的 GenAI Platform 在发布前测试了 TIME AI 内容的 7,000+ 个对抗攻击向量。这是生产安全应用,有清晰、可衡量的结果指标和不到两个月的部署窗口,并记录在 Scale 官方客户页面上。 Meta 同时是 Scale 最大战略投资方(2025 年 6 月收购约 49% 股权)和活跃且扩张中的 RLHF 客户。这一双重角色既带来收入锚,也带来治理复杂性:Meta 作为投资方拥有特权信息访问权,同时又是客户,这会让其他评估 Scale 数据安全实践的客户担心数据访问和利益冲突。 美国国防部和情报共同体客户的证据来自 Scale 官方博客,内容描述了用于联合部队行动的 DoD 数据整理合同,以及 Donovan 平台在国家安全场景中的部署。受保密限制,这些不是具名机构客户,但它们是公开来源中可信度最高的政府部署证据,并由 Scale 的 DoD IL4 和 FedRAMP High 认证佐证。企业客户 Etsy、Instacart、Pinterest 出现在 Scale 客户页面,但没有案例研究或结果指标。 Mercor 诉讼(2025 年 9 月)指控竞争对手挖客户,间接提供了反向证据:Scale 的企业账户有足够价值,值得竞争对手主动争夺;同时也说明竞争环境下账户留存存在脆弱性。Snorkel AI、SuperAnnotate 和 Mercor 的竞争对手资料显示,企业对标注和 AI 数据平台的需求仍在,即使竞争加剧,这个市场仍被验证。 [CU002, CU003, CU004, CU005, CU006, CU007]
| 客户 | 客群 | 部署 / 用例 | 生产环境 / 试点 | 结果 / 证据质量 | 证据限制 |
|---|---|---|---|---|---|
| TIME(媒体) | 企业 | GenAI Platform 用于 AI 内容安全测试;对抗攻击向量评估 | 生产环境(已确认) | 已测试 7,000+ 个攻击向量;<2 个月完成部署;scale.com/customers/time 有记录案例 | Scale 制作案例;无独立第三方验证,也无 TIME 提供的 ROI 数据 |
| 美国 DoD / IC | 政府 / 国防 | 联合部队行动数据整理;面向国家安全 AI 工作流的 Donovan 平台 | 生产环境(已确认) | 活跃多年期 DoD 合同;IL4 和 FedRAMP High 认证部署;Scale 官方博客文章 | 合同金额和机构名称涉密;无公开可用任务结果指标 |
| Meta | AI 实验室 + 战略投资者 | LLM 训练用 RLHF 数据;投资后客户关系扩大 | 生产环境(扩大中) | ~49% 战略投资,估值 $14.3B;公司称关系正在扩大;RLHF 平台得到佐证 | 投资者兼客户双重角色造成证据质量冲突;结果指标无法独立验证 |
| Etsy | 企业 | 电商推荐和搜索 AI 训练数据 | 推定生产环境 | scale.com/customers 上有标识;无案例研究、结果数据或部署范围 | 仅有标识证明;部署范围、合同条款和结果未知 |
| Instacart | 企业 | 配送和物流 AI 应用训练数据 | 推定生产环境 | scale.com/customers 上有标识;无案例研究、结果数据或部署范围 | 仅有标识证明;部署范围、合同条款和结果未知 |
| 企业 | 视觉搜索和推荐 AI 训练数据 | 推定生产环境 | scale.com/customers 上有标识;无案例研究、结果数据或部署范围 | 仅有标识证明;部署范围、合同条款和结果未知 | |
| Cohere | AI 实验室 | 企业 LLM 训练用 RLHF 数据 | 推定生产环境 | 列于 scale.com/customers;Cohere 是商业 AI 实验室;RLHF 平台得到佐证 | 无案例研究;范围或合同规模公开信息有限 |
枚举不完整——仅纳入公开具名且确认的客户。Google 和 OpenAI 是前客户(June 2025 退出,未纳入)。证据质量差异很大:TIME 最强(生产环境 + 可量化结果);政府部署置信度高但细节涉密;企业标识(Etsy、Instacart、Pinterest)只能提供有限的生产深度或结果证明。
[CU002, CU003, CU004, CU005, CU006, CU007]6.4 留存、持久性与满意度信号
Scale AI 未披露任何净留存率(NRR)、总留存率(GRR)、流失率、续约率、客户总数或满意度(NPS/CSAT)指标。这是投资分析中的重大信息缺口。以下所有留存评估,均是从合同类型、切换成本、竞争动态和反向事件中作出的结构性推断,而不是来自已披露数据。 政府和国防客户是最持久的队列。多年期合同工具、DoD IL4 和 FedRAMP High 认证壁垒、涉密环境集成以及高采购切换成本,共同形成结构性锁定。2024 年 DoD 联合部队数据整理合同和 Donovan 在涉密场景的部署,显示该板块年留存率较高;按可比政府 IT 合同续约基准估计,为 90–97%。该板块的持久性基本不受同时发生的 AI 实验室客户流失影响。 企业客户(TIME、Etsy、Instacart、Pinterest)可能签署 12–36 个月合同,并带有续约选择权。没有 NRR 数据,就无法确认先落地再扩张机制是否运转。GenAI Platform 作为 Data Engine 的增购路径,在理论上提供扩张机会;TIME 部署证明试点可转化为生产安全场景,但没有公开的交叉销售成功率或扩张收入指标。 AI 实验室板块正在主动收缩。Google 和 OpenAI 均于 2025 年 6 月退出,估计占历史 AI 实验室收入的 20–40%(未披露,基于报道中的客户重要性估计)。剩余 AI 实验室客户(Cohere 等)也面对转向自建标注的相似结构性压力。截至本分析,未识别到 Scale AI 企业平台的 G2、Gartner Peer Insights 或 Capterra 评价,因此客户满意度完全无法量化。 [CU003, CU008, CU009, CU010, CU015, CU016]
| 指标 | 数值 / 状态 | 客群 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 净收入留存率(NRR) | 未披露 | 所有客群 | n/a | 在尽调资料室索取过去 4 个季度按客群划分的 NRR |
| 总收入留存率(GRR) | 未披露 | 所有客群 | n/a | 在资料室索取过去 4 个季度按客群划分的 GRR |
| 年度流失率 | 未披露;Google 和 OpenAI 的 2025 退出事件已确认是重大不利事件 | AI 实验室 | 低(仅反向信号) | 量化 Google/OpenAI 退出的收入影响;确认退出管线中没有其他 AI 实验室客户 |
| 合同续约率(政府) | 根据多年期合同结构和 IL4/FedRAMP 切换成本壁垒,估计 90%+ | 政府 / 国防 | 低(结构性估计) | 从合同记录或资料室管理层确认中获取实际续约率 |
| 客户 NPS / CSAT | 未披露;未发现 Scale AI 企业平台的 G2、Gartner Peer Insights 或 Capterra 评价 | 企业 | n/a | 索取企业客群 NPS 数据和 CSAT 分数;检索 G2/Gartner 是否有更新评价覆盖 |
| 合同期限 | 政府:多年期 RFP 载体;企业:估计 12-36 个月;自助:月付 / 按量付费 | 所有客群 | 中 | 确认 ARR 前 10 大客户的最低合同期限和续约选择权 |
除合同期限估计外,所有留存指标均未披露或不可得。政府客群的结构性留存推断置信度低。Google 和 OpenAI 退出确认 AI 实验室流失重大且发生在近期。未找到评论平台数据。
[CU009, CU010, CU015, CU019, CU020, CU027]Scale AI 未公开披露 NRR、GRR、流失率或队列数据。数值为结构性估计,依据合同类型、切换成本分析,以及 AI 标注行业中可比数据服务公司的基准。估计仅用于尽调示意,必须用数据室披露数据确认。
6.5 扩张与客户集中风险
Scale AI 的客户集中风险重大,并已在 2025 年反向兑现。Google——此前是 Scale 最大客户——和 OpenAI 在一个季度内离开,是 Scale 历史上最重要的反向客户事件。CNBC 报道称,Google 退出源于 Meta 投资后对竞争冲突的担忧。这种流失模式中,Scale 两个最大 AI 实验室客户都因同一底层利益冲突离开,会让任何仍有 Meta 相邻竞争敞口的 AI 实验室客户面临结构性风险。 Meta 关系既创造最大扩张机会(Meta 的 AI 计划是训练数据主要消费方,且双方关系被描述为正在扩张),也带来重大依赖风险:Google/OpenAI 离开后,Scale 相当一部分收入可能集中在单一客户身上,而该客户同时又是 49% 投资方。这种客户兼投资方集中度在结构上很不寻常,可能持续给其他客户带来商业摩擦。 先落地再扩张机制在理论上存在:从数据标注到 RLHF、评测、GenAI Platform,再到 Donovan,形成了有记录的增购漏斗。TIME 部署证明企业客户可以从标注试点转化到生产平台使用。但没有公开的扩张率、增购转化数据或单客户平均收入指标。Mercor 诉讼(2025 年 9 月)指控挖客户,表明至少一个资金充足的竞争对手正威胁现有账户留存。SuperAnnotate 和 Mercor 正在扩大企业标注平台能力,证据来自产品页和合作伙伴资料,这给 Scale 的 Data Engine 收入带来替代风险。 [CU001, CU003, CU009, CU010, CU022, CU024]
| 驱动因素 / 风险因素 | 类别 | 当前状态 | 影响评估 | 尽调路径 |
|---|---|---|---|---|
| Meta 客户 + 投资者关系 | 集中度 + 利益冲突 | Meta 持有约 49% 股份,并在扩大 RLHF 客户关系 | 高正面(收入锚点)+ 高负面(治理风险;其他客户可能降低份额) | 核实 Meta 收入占总 ARR 的比例;评估数据访问治理;确认企业客户仍接受该安排 |
| Google 和 OpenAI 同时退出 | 客户流失 | 两者均于 Q2 2025 退出;Google 是 Scale 的 #1 客户 | 高负面——收入大幅减少;AI 实验室客群结构性下滑 | 量化收入影响;确认替代管线;审查剩余 AI 实验室合同的 NDA 和退出条款 |
| DoD / IC 多年期合同载体 | 扩张护城河 | 多个活跃 DoD 合同;Donovan 已进入涉密部署;IL4/FedRAMP High 认证 | 高正面——耐久且粘性强的政府收入底座;竞争对手面临高切换成本 | 审查合同载体类型、选择权期限和计划续约日期;确认新授予管线 |
| Mercor 诉讼——挖客户 | 竞争性流失风险 | Scale 于 Sep 2025 起诉 Mercor,指控其挖走关键客户并盗用商业秘密 | 中——显示竞争性客户留存风险;可能扩散到其他竞争对手 | 跟踪诉讼结果;索取 Mercor 事件后的客户留存数据;评估接触范围 |
| 先落地后扩张的增购漏斗 | 扩张机会 | Data Engine → RLHF → Evaluation → GenAI Platform → Donovan 增购漏斗有记录但未量化 | 若转化率足够高则为正面;实际转化率未知 | 索取各产品迁移的增购转化率,以及各客户队列平均 ARR 扩张 |
客户集中度风险是本章最重要发现。Meta-Google 冲突导致 Scale 最大客户离开。政府护城河提供结构性抵消,但若没有确认的替代管线,仍无法完全补偿 AI 实验室流失。
[CU003, CU009, CU010, CU022, CU026, CU028]6.6 证据材料
07风险
7.1 按严重程度排序的风险概览
Scale AI 进入当前投资期时,风险画像更偏战略和执行风险,而不是监管或运营风险。截至研究日期,客户集中度——已通过 Google 和 OpenAI 离开兑现——是严重程度最高、发生概率也最高的风险。这些离开发生在 Meta 投资和领导层更替同一时期,使 2025 年 Q2 一个季度内出现了三个高度交织的高严重度风险事件。 最关键的观察是,Scale 的主要风险彼此相互依赖:Meta 投资触发 Google 离开(利益冲突),Google 离开又触发对 Scale 在此前 AI 实验室收入集中度下商业模式可行性的重新评估,进而触发领导层更替和裁员。每个风险都会放大其他风险。投资人必须评估,政府 / 国防护城河以及 Meta 正在扩大的客户关系,能否提供足够收入底线,支撑公司穿越 AI 实验室客户流失和企业转型。 监管和法律风险重要但可管理。Scale 的 DoD IL4 和 FedRAMP High 认证显示其政府业务合规姿态较强。Mercor 诉讼是持续法律风险,但不太可能致命。国防 AI 的出口管制风险真实存在,但可通过现有认证和合规计划应对。最重要的前瞻性风险,是 Scale 的企业 GenAI Platform 能否在当前现金跑道定义的时间窗口内,为 AI 实验室板块创造足够替代收入。 [CR001, CR002, CR003, CR004, CR005, CR006]
| 风险 | 可监测触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 客户集中 / AI 实验室流失 | 新增 AI 实验室客户离开;Meta 减少数据采购量 | Cohere 之外任何 AI 实验室客户退出,或 Meta ARR 环比下降 >20% | 投资逻辑破裂——收入下限假设被打破;需要全面重做收入模型承销 |
| CEO / 领导层过渡失败 | 因领导层不确定性导致企业或政府交易流失;关键高管流失 | 流失 2+ 名高级领导,或 Droege 上任后 12 个月内未能签下 >$10M 企业交易 | 黄旗——加速 CEO 继任或强化董事会;重新评估估值中的管理层溢价 |
| 政府合同不续约 | DoD 或 IC 合同在选择期未续约;安全重新认证失败 | 任何重大 DoD 合同取消或安全许可撤销 | 投资逻辑破裂——政府护城河假设被打破;底部估值坍塌 |
| Meta 冲突兑现 | 企业或 AI 实验室客户明确将 Meta 冲突列为退出原因;FTC 或 DOJ 审查 Meta 持股 | 任何非 AI 实验室企业客户提及 Meta 冲突,或监管审查启动 | 重大风险升级——重新评估集中度阴影;复核治理文档 |
| 商业模式转型执行 | 以企业 GenAI Platform ARR 作为转型成功代理;Donovan 新合同授予 | GenAI Platform ARR 在裁员后 18 个月内(到 2027 年 1 月)未能替代 >50% 的 AI 实验室 ARR | 黄旗——转型速度不足;重新评估资本充足性和现金跑道 |
否决标准来自投资逻辑对政府护城河、企业平台增长和管理层稳定性的要求。触发项设计成不依赖非公开信息也可观察。阈值为分析师估计。
[CR001, CR002, CR003, CR004, CR005, CR006]7.2 监管与法律风险
Scale AI 处在国防 AI、企业数据处理和 AI 安全三者交叉处,而这三个领域都受到活跃且不断演变的监管关注。主要监管风险包括:美国政府对国防 AI 的合规要求(ITAR、出口管制、AI 安全标准)、适用于企业数据处理的数据隐私和 AI 治理法规,以及 Mercor 诉讼引发的竞争法律程序。 国防 AI 上,Scale 的 DoD IL4 和 FedRAMP High 认证显示其监管合规姿态较强。公司还作出白宫 AI 安全自愿承诺,显示其主动参与 AI 治理。该领域的主要监管风险,是 AI 模型和数据服务出口管制限制可能扩大——尤其当 Scale 向 DoD 承包商提供的 AI 服务涉及敏感技术转移,可能引发 ITAR 审查。Scale 在国会作证也说明,国会对 AI 的监督正处于活跃状态。 数据隐私上,Scale 的企业数据标注服务涉及处理客户专有数据。EU AI Act、GDPR 以及美国各州新兴 AI 法律带来合规义务,这些义务可管理但持续演变。Scale 的安全和合规页面显示,公司拥有有效的 SOC 2 Type II 和 ISO 27001 认证,说明企业合规姿态较成熟。Mercor 诉讼(Scale AI vs. Mercor,2025 年 9 月)指控挖客户和盗用商业秘密。诉讼结果不确定,但它带来法律不确定性、管理层分心和潜在反诉风险。目前来看,该诉讼尚未威胁 Scale 核心运营或认证。 [CR005, CR006, CR007, CR009, CR010, CR011]
| 风险 / 规则 / 案件 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余风险敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 美国 AI 出口管制 / 防务 AI 的 ITAR | 美国联邦 | 生效中——BIS/ITAR 关于 AI 模型和数据服务的规则仍在演进 | 中 | 高 | DoD IL4 和 FedRAMP High 认证;现有国防承包商关系 | 中等——监管变化可能限制向非美国实体出口数据或交付 AI 服务 | 确认所有政府合同的 ITAR 和 EAR 合规状态;复核 BIS AI 指引遵循情况 |
| Scale AI v. Mercor 诉讼(客户挖角、商业秘密) | 美国联邦(NDCA) | 进行中——2025 年 9 月起诉;诉讼仍在推进 | 高(已起诉) | 中 | 诉讼进行中;可能申请初步禁令;商业秘密文档留存 | 中等——反诉和证据开示可能披露内部客户数据或合同条款;管理层分心 | 复核起诉状和案卷;获取关于初步禁令可能性的法律意见;评估反诉风险敞口 |
| EU AI Act 对企业数据处理的合规要求 | EU | 已生效——GDPR 和 EU AI Act 义务适用于 EU 企业客户 | 低-中 | 中 | ISO 27001 和 SOC 2 认证;与企业客户签署 DPA 协议 | 低-中等——主要风险来自 Scale 为 EU 受监管应用处理高风险 AI 训练数据的情形 | 确认 Scale 数据标注服务在 EU AI Act 下的分类;核验 EU 客户的 DPA 合规 |
| 美国 AI 安全自愿承诺及潜在强制要求 | 美国联邦 | 自愿——2024 年签署白宫承诺;潜在强制监管正在推进 | 低 | 低-中 | 已签署白宫 AI 安全承诺;主动开展 AI 安全研究(WMDP benchmark) | 低——自愿合规降低监管风险;强制规则尚未生效 | 跟踪 AI 安全规则制定;确认自愿承诺合规文档保持最新 |
| 数据隐私——州级 AI 和隐私法律(CPRA 等) | 美国州级 | 生效中——CPRA 等类似法律适用于总部在加州的 Scale AI | 低-中 | 低-中 | SOC 2 Type II、隐私政策、企业 DPA 协议 | 低——标准企业 SaaS 合规姿态;未见已知执法行动 | 确认 CPRA 合规;核验企业客户 DPA 模板保持最新;检查州级 AI 监管适用性 |
登记表并不完整——仅基于公开来源。未列入监管往来、执法行动,以及任何未公开披露的政府调查。可能性和严重性为分析师估计。 最终风险评估需要法律顾问复核。
[CR005, CR006, CR007, CR009, CR010, CR011]7.3 运营与执行风险
Scale AI 最尖锐的运营风险,是在逆风环境下执行商业模式转型。公司正试图从标注量驱动收入(因 AI 实验室客户离开和合成数据趋势而下滑)转向企业 GenAI Platform 和政府 Donovan 收入(在增长,但速度不确定)。这一转型与 CEO 更替、大幅裁员以及失去两个最大客户同时发生。 数据安全和标注质量是 Scale 核心业务内生的运营风险。作为一家为 AI 训练处理客户专有数据的公司,任何数据安全事件、质量失败或客户数据滥用事件,都会严重损害企业信任和政府合同资格。Scale 的 SOC 2 Type II、ISO 27001、DoD IL4 和 FedRAMP High 认证降低了这一风险,但不能消除。本研究未发现公开数据安全事件。 Scale 的供应链风险主要来自标注贡献者劳动力。Scale 依赖全球人工标注者网络;如果该劳动力出现中断(劳资纠纷、质量下降、Mercor 或其他竞争对手挖人),标注输出质量和交付时间都可能受损。2025 年 7 月裁员约 500 名承包商,可能削弱了这条供应链的运营冗余。此外,Scale 依赖云基础设施提供商(AWS 及类似厂商),这带来平台集中风险;对云原生公司来说这很常见,但放在其 DoD 对基础设施主权的要求下,仍值得关注。 [CR001, CR003, CR008, CR015, CR016, CR017]
| 故障模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余风险敞口 | 未解决缺口 |
|---|---|---|---|---|---|
| 数据安全泄露,暴露客户专有 AI 训练数据 | 低-中 | 高 | 高(SOC 2 Type II、ISO 27001、DoD IL4、FedRAMP High) | 中——认证降低但不能消除泄露风险;AI 训练数据是高价值目标 | 未确认公开事件披露机制;政府客户的泄露通知流程不清楚 |
| 裁员后标注质量下降(500+ 名承包者被解约) | 中 | 高 | 中(QA 流程已在运行;Scale RLHF 平台自动化质量评分) | 中-高——大规模压缩人力可能削弱标注质量冗余和机构知识 | 未披露公开质量 SLA;没有面向客户的质量指标;承包者减少按业务板块的范围不清楚 |
| 商业模式转型失败(从标注转向平台) | 高 | 高 | 低(转型仍早期;企业 GenAI Platform 收入未公开确认) | 高——转型如果失败,核心标注业务将陷入结构性下滑,且没有替代收入 | 未披露企业 GenAI Platform ARR;标注客户转向平台的转化率未知;没有公开时间表 |
| 云基础设施宕机,影响政府或企业 SLA | 低 | 中 | 中(云冗余;DoD 要求特定基础设施主权) | 低-中——标准云依赖风险;多云和政府级基础设施要求有所缓释 | 涉密 DoD 工作负载的基础设施主权安排未公开披露 |
| 贡献者劳动力中断(劳资争议、竞争平台) | 中 | 中 | 中(全球贡献者基础提供一定地理冗余) | 中等——Mercor 等平台正在争夺标注劳动力;替代贡献者质量未知 | 未披露承包者队伍构成和地理集中度;Mercor 对贡献者的竞争规模未量化 |
故障模式按剩余严重性排序。缓释成熟度是分析师基于公开认证数据和推断运营实践做出的估计。未能公开获取内部风险管理文档或事件历史用于复核。
[CR015, CR016, CR017, CR018, CR019, CR003]7.4 合作伙伴与依赖风险
Scale AI 最重大的依赖风险来自 Meta——它同时是最大投资方(约 49% 股权)、剩余最大 AI 实验室客户,也是前 CEO 的新雇主。这种三重角色依赖,在风险投资支持的 AI 领域没有先例,带来集中度和治理风险。如果 Meta 对 Scale 的战略兴趣减弱、减少数据采购或推出新的竞争产品,Scale 将同时面对客户收入流失、估值压力和治理扰动。 对 AWS 和主要云厂商的云基础设施依赖,是 AI 基础设施公司的标准风险;但对 Scale 政府业务尤为重要,因为平台主权和 AI 训练算力主权是合规要求。Scale 的 DoD 认证意味着其云基础设施安排已获批准,但云厂商关系的任何变化都可能影响政府合同资格。OpenAI 和 Google 的离开,消除了两个关键合作伙伴关系;它们此前既提供收入,也提供声誉背书。剩余 AI 实验室伙伴(Cohere 等)规模更小,单个客户的收入集中度也更低。 政府采购中的渠道和合作伙伴集中度更偏正面,而不是负面:政府合同依托既定合同工具(GSA、DIU、DARPA 渠道)推进,采购规则清晰。不过,政府预算周期和持续决议机制会带来收入确认时间波动。Snorkel AI 合作伙伴生态(由其合作伙伴页面佐证)以及 SuperAnnotate 的企业合作关系显示,Scale 的竞争对手也在搭建依赖伙伴的获客路径,生态锁定竞争正在展开。 [CR002, CR003, CR004, CR006, CR009, CR020]
| 依赖项 | 交易对手 | 角色 | 集中度水平 | 失效情景 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| Meta 战略关系 | Meta(投资者 + 客户) | 49% 股权投资者、扩张中的 RLHF 客户,且是前 CEO 的雇主 | 极高 | Meta 减少数据采购、寻求取得完全控制,或推出竞争性标注平台 | 危急 | 向政府 / 企业多元化收入;按公开政策维持运营独立 | 高——单一关系叠加客户、投资者和治理风险;未确认结构性防火墙 |
| 美国 DoD / IC 合同工具 | 美国政府(DoD、IC) | 多年期政府客户;定义政府收入下限 | 高 | 合同不续约、预算削减,或安全重新认证失败 | 高(正向依赖——失去将很严重) | IL4/FedRAMP High 认证;带选择期的多年期合同结构;专门政府团队 | 中——政府合同韧性强,但受预算周期和重新认证要求约束 |
| 云基础设施(AWS、Microsoft Azure) | AWS / Microsoft | Scale 标注管线、API 服务和政府工作负载的平台托管 | 高 | 云服务商宕机、价格调整,或合同终止 | 中 | 多云架构(假设);DoD 要求 FedRAMP 授权基础设施 | 低-中——标准云依赖;由政府认证要求缓释 |
| 全球标注贡献者网络 | 独立承包者(估计 500K+) | 数据标注管线的标注劳动力供应 | 高 | 劳动力碎片化、质量下降,或大规模流向竞争平台(Mercor) | 中-高 | 全球地理分布;自研质量评分平台(Scale RLHF) | 中——Mercor 诉讼表明贡献者被主动挖角;裁员后存在士气风险 |
| OpenAI / Google(原客户) | OpenAI、Google(已流失) | 曾是验证平台质量的关键 RLHF 和评估客户 | 高(历史;当前为零) | 已经发生——二者均在 2025 年 Q2 退出 | 高(已兑现) | 无有效缓释——退出已完成 | 剩余:其他 AI 实验室可能退出的声誉联想;给 Cohere 等客户形成 AI 实验室板块集中风险 |
依赖项按剩余严重性排序。Meta 依赖在结构上最不寻常,也最难靠标准风险管理缓释。纳入原客户一行,是为了展示集中度如何已经兑现。 集中度水平为分析师估计。
[CR002, CR003, CR004, CR006, CR009, CR020]7.5 人员、执行与财务风险
领导层交接是 Scale AI 最实质的风险之一。创始人 Alexandr Wang 用九年搭起 Scale,也一直是政府、企业和 AI 实验室关系的门面。他离职加入 Meta,同时 Meta 成为 Scale 最大投资方,形成了客户、员工和投资者都看得见的委托代理冲突。临时 CEO Jason Droege 之前没有管理过一家估值 $29B 的 AI 基础设施公司;他的 Uber Eats 经历对双边市场扩张有运营相关性,但与政府 AI 或企业数据平台商业化并不直接对口。 裁员后的短期环境里,人才留存也是风险。2025 年 7 月裁掉 200 名员工和 500 名承包商,可能已经打击剩余约 1,000 名员工士气,尤其是高绩效员工如果看不清临时领导层下的公司方向。掌握防务合同和企业部署制度知识的政府与企业客户经理,是难以替代的资产;能否留住他们,直接关系到合同续约和平台扩张。 财务风险因 Meta 战略投资得到部分缓释,但这笔交易的资金分给了既有股东,而不是留在 Scale 资产负债表支持运营。因此,分配之后 Scale 真实现金余额和现金跑道仍不确定。公司没有披露公开烧钱速度、现金头寸或盈利路径。如果 AI 实验室收入流失幅度接近 Google/OpenAI 离开所暗示的水平,Scale 可能在 2025-2026 年遇到实质收入缺口;除非企业和政府业务增速超预期,否则需要额外外部资本。 [CR001, CR002, CR003, CR007, CR008, CR024]
| 角色 / 职能 | 依赖或缺口 | 影响可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| CEO——临时(Jason Droege) | 在 Scale 所处阶段和赛道未经验证;Uber Eats 背景缺少政府 AI 和企业数据平台先例 | 高 | 高 | Droege 有运营扩张经验;现有领导团队和董事会提供支持 | 评估 Droege 上任以来在企业和政府交易上的推进;确认董事会是否积极推动 CEO 搜寻或继任规划 |
| 创始人(Alexandr Wang)——已离任 | Wang 9 年来维系政府、企业和 AI 实验室关系;已离任加入 Meta;保留董事会席位 | 高(已发生) | 高 | Wang 保留董事会席位;现有团队承接机构知识 | 确认 Wang 董事会角色仍活跃,且不与 Meta 职责冲突;评估关键客户关系是否已转移给当前团队 |
| 政府 / 防务关系经理 | 掌握涉密项目、合同官员和安全许可的机构知识;DoD 关系存在关键人风险 | 中 | 高 | 长期政府合同降低关系依赖;持证人员切换成本高 | 访谈 2-3 名资深政府关系经理;确认裁员后安全许可留存;核验 DoD 项目连续性 |
| 企业销售领导层 | 企业销售周期为 6-18 个月;裁员后领导层流失可能扰乱管线 | 中 | 中-高 | Scale 品牌和现有客户标识支持企业销售;自助 API 降低依赖 | 索取企业销售管线数据(合格管线、管线覆盖率、交易阶段分布) |
人员风险按严重性排序。创始人离开与单季领导层过渡同时发生,关键人风险极高。考虑到政府收入底盘,政府关系连续性是最关键的运营风险。
[CR001, CR002, CR007, CR008, CR024, CR025]7.6 展品
08估值
8.1 投资逻辑与反向逻辑
Scale AI 的投资逻辑靠三根结构性支柱支撑。第一,政府与防务:Scale 的 Donovan AI 平台、DoD IL4/FedRAMP High 认证,以及深厚政府关系,构成一个耐久收入底座;切换成本高、合同周期长,与 AI 实验室的竞争重叠很低。这个板块是一项战略资产,即便不计商业 AI 实验室业务,也能支撑 $3–5B 的最低估值。第二,数据护城河:Scale 用九年训练出标注质量基础设施、贡献者管理和模型评估工具,这套能力很难大规模复制。Fortune 500 公司从 AI 试验转向生产部署后,需要人在回路的数据工作流;上述基础设施让公司有机会捕捉企业 GenAI Platform 收入。第三,Meta 验证:全球最精明的 AI 投资方之一 Meta 投入 $14.3B,既提供财务和声誉背书,也释放 Scale 数据基础设施的战略价值信号,并带来一个仍在扩大的客户关系。 反向逻辑同样扎实。在 $29B 隐含估值下,Scale 的交易倍数为估计 ARR 的 58x–145x;与此同时,两大收入贡献方(Google 和 OpenAI)都已离开,创始人 CEO 也已离任,业务模式正在半路转向。标注商品化风险(合成数据、前沿实验室自建 RLHF)对 AI 实验室板块构成结构性威胁。Meta 既是投资方又是客户,形成治理阴影,吓退其他客户,也制造了 AI 行业没有先例的委托代理冲突。临时 CEO Droege 没有在 Scale 核心市场被验证过的履历。要守住 $29B 估值,只能依靠一组乐观假设:政府合同续约、企业平台转化、Meta ARR 增长——而这些都没有公开证据确认。基准情景指向入场价带来的实质稀释风险。 证据合起来看,Scale 是高质量资产,但处在结构性逆风里。尽调必须确认三件事:政府收入底座是否稳固,企业平台 ARR 是否以足够速度增长,管理团队能否在不接受稀释性条款新增融资的情况下完成转向。问题回答前,现有证据不支持按 $29B 隐含估值投资。基于当前证据质量,合适建议是有条件放弃——只有在入场价明显更低且里程碑兑现时才投——或继续研究。 [CV001, CV002, CV003, CV004, CV005, CV006]
| 论点 | 证据依据 | 什么会改变这一判断 |
|---|---|---|
| 投资逻辑:政府护城河凭高切换成本和多年期合同撑起耐久收入下限 | DoD IL4、FedRAMP High、Donovan 平台、DIU/DARPA 合同;未宣布不续约 | 政府合同取消或安全重新认证失败,会拿掉这一投资逻辑支柱 |
| 投资逻辑:数据基础设施和标注平台代表 9 年难以复制的质量工程 | scale.com 记录的 RLHF 平台、500K+ 贡献者网络、模型评估工具;Stanford HAI 背书 | 出现系统性标注质量下降证据,或竞争对手以可比规模复制 |
| 投资逻辑:Meta 背书($14.3B、49%)释放战略价值信号,并确保近期资本供给 | 多来源确认投资;有记录以来最大单笔 AI 战略投资 | Meta 减少数据采购,或寻求以压低估值收购剩余股权 |
| 投资逻辑:企业 GenAI Platform 为 AI 实验室流失创造替代收入 | TIME 部署案例研究;企业客户标识;官方 GenAI Platform 产品页 | 企业平台 ARR 在 24 个月内(到 2027 年中)未能替代 >50% AI 实验室收入 |
| 反向逻辑:$29B 估值由战略买方决定,不符合财务投资人的理性定价 | 入场倍数 58x–145x ARR;前 2 大客户同时流失;Google/OpenAI 退出已确认 | 入场价格显著更低(<$15B),或确认流失后 ARR 快速恢复 |
| 反向逻辑:标注商品化是核心 RLHF 收入的结构性威胁 | McKinsey/Stanford HAI 报告每 token 标注成本下降;合成数据生成降低人工标注需求 | 有证据表明 Scale 的平台级评估服务(不是原始标注)能避开商品化 |
| 反向逻辑:CEO 过渡和 Meta 冲突带来缺少先例的管理与治理风险 | Wang 离任确认;Droege 临时任命确认;未宣布永久 CEO 搜寻 | 任命具备政府 AI 和企业平台履历的永久 CEO;确认 Meta 与 Scale 客户数据之间存在结构性防火墙 |
投资逻辑和反向逻辑均以证据为根基;排除投机性论点。投资逻辑在结构上成立,但关键收入替代假设需要私有数据确认。 反向逻辑可从公开来源观察,不需要任何私有数据。
[CV001, CV002, CV003, CV004, CV005, CV006]| 触发因素 | 阈值 / 事件 | 对投资逻辑的传导 | 行动含义 |
|---|---|---|---|
| 政府合同未续约 | 任何 DoD 或 IC 合同被取消,或未在计划期权期续约 | 政府收入托底逻辑失效;基准情景估值降至 Appen 同等水平($1.5–3B 区间) | 硬性放弃 —— 政府护城河是核心支柱;一旦失效,投资逻辑无法修复 |
| 更多企业客户以 Meta 冲突为由流失 | 任何非 AI 实验室企业客户明确将 Meta 关系列为退出或不采购原因 | 确认 Meta 冲突是系统性的 GTM 阻力,而不只是 AI 实验室特有问题;TAM 风险从 AI 实验室扩散到更大范围 | 若确认则硬性放弃;或必须提高入场折扣,以反映结构性获客约束 |
| Meta ARR 下滑 | Meta 数据采购 ARR 连续两个季度环比下降 >20%,且没有可由其他企业增长解释 | 乐观情景的锚定客户假设失效;不再有大型锚定客户;收入集中度塌陷 | 硬性放弃 —— 收入地板消失;触发所有情景重新承销 |
| 截至 2026 年 Q4,GenAI Platform ARR 低于 $30M | 企业 GenAI Platform ARR 经确认为低于 $30M(2026 年 Q4 收入日期或数据室披露) | 确认企业转型没有达到最低可行转化;ARR 替换速度不足,无法在投资逻辑窗口内抵消 AI 实验室流失 | 放弃或严守价格纪律:入场价较基准情景上限下调 40% |
| CEO 交接失败信号 | Droege 上任后 12 个月内未能签下任何 >$5M 企业交易,或 Droege 任命后 2 名以上高级领导离职 | 表明管理团队无法执行企业转型;执行风险倍数上升 | 黄旗 —— 重新评估管理能力;推进前向董事会确认永久 CEO 时间表 |
这些触发因素不需要内部访问权,可通过公开报道或数据室披露观察。政府合同触发因素最关键:它是二元事件,可公开观察(USASpending.gov),且会直接拿掉核心投资逻辑支柱。
[CV003, CV006, CV007, CV008, CV009, CV029]8.2 估值背景与当前融资
Scale AI 当前约 $29B 的隐含估值来自 Meta 于 2025 年 6 月以 $14.3B 收购公司约 49% 股权。多家独立高声誉媒体(TechCrunch、CNBC、Reuters)报道了该交易,本分析将其视为已确认。隐含公司总估值($29B)基于报道中的约 49% 股权;具体的 pre-money/post-money 处理和股权结构表没有公开披露。重要的是,Meta 交易所得分给了既有股东,而不是留在 Scale 资产负债表支持运营,因此 Scale 的真实营运资本和现金跑道仍不明确。 在 $29B 隐含估值下: - 低 ARR 估计($200M):收入倍数 145x —— 与 2024 年初前沿 AI 实验室倍数一致,但对一家正在经历客户流失的公司而言极高 - 中 ARR 估计($350M):收入倍数 83x —— 可比于市场顶点时期的高增长企业 SaaS - 高 ARR 估计($500M):收入倍数 58x —— 只有在政府 ARR 快速增长且企业平台转化加速时才说得通 Scale 历史上最可比的融资事件是 2024 年 5 月 Series F,估值 $13.8B,当时估计 ARR 约 $200–300M(46–69x ARR)。Meta 交易在约 13 个月内几乎把估值翻倍,而同期 Scale 两个最大客户离开。这一动态说明,Meta 估值反映的是战略收购逻辑(Meta 需要自有 RLHF 数据基础设施),不是财务回报最优。按 $29B 隐含估值共同投资的投资者,支付的是战略收购方逻辑定下的价格,而不是独立财务投资者回报框架下的价格。估值纪律的关键问题在这里:战略买家能接受财务投资者无法承销的溢价。尽调必须确定财务投资者的入场价上限。 稀释与优先权压力未知。Meta 交易前,Scale 已累计股权融资约 $1.6B;优先股堆叠、清算权,以及参与型 / 非参与型结构都没有公开披露。任何按 $29B 隐含估值进入的财务投资者,都必须评估悲观情景下优先权瀑布对普通股回报的影响。 [CV001, CV002, CV003, CV010, CV011, CV012]
| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 客户流失后按板块拆分的 ARR | Google 和 OpenAI 曾是最大客户;二者的 ARR 贡献和离开后的净 ARR 尚未披露 | 不知道收入缺口和替代 ARR 当前走势,就无法给任何情景承销 | 数据室:索取 2024 年 Q1 – 2025 年 Q4 按板块 ARR 瀑布;如 NDA 要求可匿名化 |
| 政府合同排期和续约时间表 | DoD 合同期权期、TCV 和续约率未公开披露 | 政府护城河是核心投资逻辑支柱;合同不续约是最可能击穿逻辑的单一事件 | USASpending.gov 查看可见合同;数据室:索取带期权期的政府合同台账;直接访谈 Scale 政府团队负责人 |
| Meta 交易后的现金头寸和现金跑道 | Meta 交易所得分配给股东;当前运营现金、烧钱速度和现金跑道未披露 | 看不清现金跑道,就无法判断转型能否在不稀释融资的情况下跑完 | 数据室:索取资产负债表、P&L 和现金流量表;确认留作运营与分配给股东的金额 |
| 企业 GenAI Platform ARR 和增速 | GenAI Platform ARR、客户数或从标注转平台的转化率均无公开披露 | 平台转型是乐观情景的核心驱动;没有 ARR 数据,乐观情景就是猜测 | 数据室:索取 2025 年 Q4 和 2026 年 Q1 的 GenAI Platform ARR、管线数据和代表性企业合同 |
| 管理团队留任和永久 CEO 计划 | 永久 CEO 搜索或预计任命时间表未公开;Droege 上任后关键高管留任情况未知 | 管理连续性直接关系到政府关系维护和企业转型执行 | 董事会访谈:确认 CEO 搜索状态;索取关键高管留任 90 天复盘;与 HR 团队评估关键人风险 |
| Meta 治理防火墙和数据访问协议 | Meta 与 Scale AI 的投资人权利协议未公开披露;尚未确认 Meta 作为投资人受到数据访问限制 | 没有经确认的防火墙,Meta 利益冲突风险无法缓释,企业客户的数据主权担忧也无法解决 | 数据室:索取投资人权利协议及任何数据访问限制附表;取得关于防火墙充分性的法律意见 |
六项尽调问题对 $29B 隐含估值下的投资决策都是阻断项。优先级最高的三项 —— 按板块 ARR、政府合同排期、现金跑道 —— 应在讨论任何初步 term sheet 前完成。
[CV002, CV012, CV013, CV014, CV029, CV030]8.3 乐观 / 基准 / 悲观情景
Scale AI 的情景分析取决于三个关键驱动变量:(1)政府 / 防务 ARR 轨迹以及既有 DoD 合同续约率,(2)商业转向带来的企业 GenAI Platform 转化率和 ARR 增长,(3)Meta 客户关系的耐久性,以及 AI 实验室客户进一步流失或增厚的情况。 乐观情景:政府合同续约率 >90%,并随着 Donovan 新项目授予继续扩张;企业 GenAI Platform 借助 Fortune 500 生产部署,到 2027 年做到 $150M+ ARR;Meta 将数据采购扩大到 $300M+ ARR;临时 CEO Droege 稳住运营,并在 12 个月内引入一位标志性永久 CEO。在这些假设下,2027 年总 ARR 达到 $600M–800M,按 25–30x 前瞻倍数,估值达到 $18–24B。这意味着相对 $29B 入场价只有有限上行,而且只有多个有利结果同时发生时才可能实现。 基准情景:政府合同续约率 85%,温和增长;企业 GenAI Platform 到 2027 年实现 $75–100M ARR(爬坡慢于乐观情景);Meta ARR 保持当前水平;AI 实验室 ARR(Cohere 等)小幅下滑。2027 年总 ARR 达到 $350–450M,按 20–25x 前瞻倍数,对应 $8–11B 估值,较 $29B 入场价显著减记。基于现有证据,这是最可能的情景。 悲观情景:政府合同延期或出现实质性不续约;企业 GenAI Platform 到 2027 年未跑通产品市场匹配(平台净 ARR < $50M);Meta 在建立内部标注能力后减少数据采购;更多 AI 实验室客户退出。2027 年总 ARR 降至 $150–200M,按 10–12x 前瞻倍数,对应 $1.5–2.5B 估值。在这个估值下,$29B 入场价在经济上是灾难性的;即便优先权堆叠也可能挡不住总回报亏损。 各情景的概率信号不对称:乐观情景需要多个乐观结果同时发生;悲观情景只需要政府合同没有按期续约这一个事件。加上入场价由战略买家定出,财务投资者按 $29B 进入时,风险 / 回报结构性不利。基准情景意味着 60–70% 减记,悲观情景意味着金融价值几乎全损。只有能够把政府护城河的战略期权定价到 $20B+ 的投资者,才有理由接受 $29B 入场。 [CV005, CV006, CV007, CV008, CV009, CV015]
| 情景 | 关键假设 | ARR 估计(2027) | 估值逻辑 | 隐含估值 | 概率判断 |
|---|---|---|---|---|---|
| 乐观情景 | 政府合同续约率 >90%;GenAI Platform 达到 $150M+ ARR;Meta 数据采购扩至 $300M+;聘任永久 CEO | $700M–850M | 25x 远期 ARR(政府 AI 平台溢价) | $17.5B–25.5B | 20%——需要多个乐观结果同时出现;单个假设并不极端,但必须同时成立 |
| 基准情景 | 政府合同续约率 85%;GenAI Platform 达到 $75–100M ARR;Meta ARR 稳定;AI 实验室板块小幅下滑 | $350M–450M | 20–22x 远期 ARR(平台 / 服务混合估值) | $7.7B–10.4B | 55%——多数结果符合带政府护城河标注公司的历史轨迹 |
| 悲观情景 | 政府合同不续约或延迟;GenAI Platform <$50M ARR;Meta 减少采购;更多 AI 实验室退出 | $150M–200M | 10x 远期 ARR(政府定位削弱的标注服务) | $1.5B–2B | 25%——只需一个不利事件(政府不续约)兑现;鉴于远期合同不确定性,概率较高 |
情景为基于公开证据构建的分析师估计;不是公司预测。ARR 估计为总 ARR,包含政府、企业、AI 实验室和自助服务板块。 概率判断带主观性,但锚定可比标注公司的结果分布(Appen 先例偏向悲观 / 基准;Palantir 先例偏向乐观)。以 $29B 入场,基准情景意味着 60–70% 减值。
[CV015, CV016, CV017, CV018, CV022, CV024]8.4 可比公司组与估值定位
Scale AI 横跨多个可服务市场,可以用几套不同框架比较。最重要的区别是:Scale 目前大部分收入来自标注和数据标注服务,却被按前沿 AI 平台公司定价。这造成估值框架错配:套用标注服务倍数(Appen:0.5–2x ARR)意味着 $100M–1B;套用防务 AI 平台倍数(Palantir:25–30x ARR)意味着 $5B–15B;套用前沿 AI 平台倍数(OpenAI 私募轮:100x+ ARR)意味着 $20B+。合适框架取决于 Scale 未来收入结构中哪个板块占主导。 Appen (ASX: APX) 是最有启发意义的负面可比:一家上市标注公司,在 AI 实验室建立内部标注能力后削减 RLHF 支出时,收入下滑 60–80%。Appen 市值从峰值约 AUD $3.5B 跌至客户流失后的约 AUD $250M,说明标注商品化会严重冲击估值。Scale 与 Appen 的关键结构差异是政府护城河:Scale 的防务垂直带来 Appen 没有的收入底座。 Palantir (NYSE: PLTR) 提供政府 AI 平台可比:它深度嵌入 DoD/IC,切换成本高,拥有多年期政府合同,并凭强增长叙事拿到 25–30x ARR 倍数。Palantir 市值在估计 $2B+ ARR 上超过 $100B,说明以政府为锚的 AI 平台可以维持高倍数。不过,Palantir 已经在政府收入之外证明了持续商业牵引力;Scale 的商业转向尚未验证。 Labelbox(私有,约 $1B 估值)是不带政府护城河的标注基础设施可比,显示没有政府合同的标注基础设施大约只能拿 8–10x ARR,远低于 Scale 的隐含倍数。Labelbox 与 Scale 倍数差额意味着,市场把 Scale 的政府护城河和 Meta 战略期权价值定价为约 $20B+ 溢价。该溢价必须通过政府合同时间表可见度和企业 ARR 轨迹数据来明确验证。 [CV019, CV020, CV021, CV022, CV023, CV024]
| 可比公司 | 类型 | 关键指标 / ARR | 倍数 / 估值 | 政府护城河 | 相关性 | 关键限制 |
|---|---|---|---|---|---|---|
| Appen (ASX: APX) | 上市——标注服务 | AUD ~$400M 峰值 ARR(2021) | 0.5–2x ARR;峰值市值 ~$3.5B;客户流失后下跌 90%+ | None | 直接负面可比:一家失去 AI 实验室客户并估值崩塌的标注公司 | 没有政府护城河——Scale 的底部估值应在结构上高于 Appen 低谷 |
| Palantir (NYSE: PLTR) | 上市——防务 AI / 企业数据平台 | ~$2B+ ARR(2025 年估计) | 30–50x ARR;市值 $100B+;政府 ARR 集中度高 | 极高(NSA、CIA、DoD 多十年关系) | 最佳政府锚定型 AI 平台可比;Scale 的政府深度尚未充分证明 | Palantir 已证明 15+ 年政府合同连续性;Scale 政府记录更短 |
| Scale AI Series F(2024 年 5 月) | 私营——自身历史轮次 | ~$200–300M ARR(当时估计) | 46–69x ARR;$13.8B 估值 | 高(该轮时已确立) | 最接近的历史可比;显示即便客户流失并发,估值仍翻倍 | 仅作历史参考;当前轨迹相对 Series F 更逆风 |
| Scale AI Meta 交易(2025 年 6 月) | 私营——战略投资 | ~$200–500M ARR(当前估计) | 58–145x ARR;隐含 $29B | 高(已确立) | 当前隐含估值 —— 本次分析的入场价格 | 已包含战略买方溢价;不适合与财务投资人交易直接比较;资金流向股东而非资产负债表 |
| Labelbox(未上市) | 未上市 —— 标注平台,无政府护城河 | 约 $80–120M ARR(估计) | 约 8–12x ARR;约 $1B 估值 | None | 同为标注基础设施,但没有政府护城河;可在剔除政府溢价后作为 Scale 平台倍数基准 | 未上市公司;ARR 为分析师估计;无公开财务披露 |
| Crunchbase 的 Scale AI 融资历史 | 数据库 —— 融资轮次聚合器 | Meta 交易前累计股权融资约 $1.6B | N/A(非单个可比公司条目) | N/A | 背景信息:确认相对于估值的资本效率 | 数据库条目;轮次细节可能与实际披露不同 |
之所以使用多个可比对象,是因为 Scale AI 横跨标注服务(对应 Appen)、政府国防 AI(对应 Palantir)和企业 AI 平台(对应 Labelbox)。合适的估值框架,取决于未来阶段哪个板块成为主要收入贡献者。$29B 估值已经同时计入 Palantir 式政府护城河和可观的企业平台上行空间,两者都需要私下确认。
[CV019, CV020, CV021, CV022, CV023, CV024]8.5 建议、置信度与最终尽调问题
对 Scale AI 的建议是继续研究,并在入场估值大幅降低时有条件通过。证据支持 Scale 资产质量——政府护城河、标注数据基础设施、品牌和 Meta 验证——但在当前证据质量和收入轨迹下,不支持 $29B 估值。该建议的主要驱动因素是,入场估值由战略收购方(Meta)按战略逻辑定出,这与财务投资者回报要求不兼容。在看不到客户流失后 ARR、政府合同续约时间表、企业平台转化指标和现金跑道的情况下支付 $29B,认知上站不住。 该建议置信度为中低,因为最关键数据——政府合同 ARR、企业平台 ARR 和交易后现金头寸——都是私有信息,公众无法取得。如果这些数据可得且有利,建议可以上调为 $15B 以下的有条件买入,甚至在 $8–12B 且有强里程碑保护时上调为结构化买入。当前证据质量支持谨慎:本章三十个研究问题中,有八个仍处于部分解决或未解决状态,因为关键输入是私有的。 会把建议下调为放弃(拒投)的破题条件包括:任何政府合同取消,任何新增企业客户因 Meta 冲突而离开,或确认截至 2026 年 Q1 GenAI Platform ARR 低于 $30M。会把建议上调为有条件买入的条件包括:政府合同续约确认、企业平台 ARR 超过 $75M,以及任命具备政府 AI 或企业平台经验的永久 CEO。投资者应注意,关键尽调问题的解决时间线很短——政府合同期权期和企业平台爬坡应在研究日期后 12–18 个月内产生可观察收入信号。 [CV001, CV002, CV006, CV009, CV017, CV029]
| 维度 | 评估 | 依据 | 条件 / 限定 |
|---|---|---|---|
| 建议 | 继续研究 / 有条件通过 | 估值按战略买方定价;财务投资人的回报模型要求显著更低的入场价 | 若政府合同排期确认且 GenAI Platform ARR 超过 $75M,则上调至有条件买入 |
| 信心 | 低-中(35%) | 关键输入(政府 ARR、企业平台 ARR、现金跑道)均为私有且未确认 | 若 2026 年 Q2 收入数据确认基准情景轨迹,信心升至中-高 |
| 风险评级 | 高 | 四个高严重性风险并发(集中度、CEO、Meta 冲突、转型执行);未证明曾处理过这一风险组合 | 若政府合同续约且管理层稳定,风险降至中 |
| 估值立场 | 对财务投资人而言,$29B 隐含估值明显偏高 | 战略收购方逻辑给出 58x–145x 估计 ARR;财务回报模型要求 $8–15B 入场 | 公允价值区间:基准情景 $8–12B;乐观情景 $15–22B |
| 决策含义 | 没有额外数据室可见性,不应以 $29B 投资 | $29B 入场在基准情景下预期价值为负;投资逻辑要求多个乐观结果同时出现 | 请求进入私有数据室;将价格纪律设在 <$15B;谈判基于里程碑的估值调整 |
建议对证据和价格敏感。该评估反映公开证据限制和基准情景估值框架。若入场价格不同,或私有数据室确认关键 ARR 指标,可能得出不同建议(有条件买入或观察)。
[CV001, CV002, CV006, CV009, CV029, CV030]8.6 展品
免责声明
本报告仅用于信息和研究目的。来源均为公开信息;未使用专有或保密信息。作者不对完整性或准确性作任何陈述或保证。本报告不构成投资建议。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Scale AI was founded in 2016 by Alexandr Wang, who was 19 years old and had left MIT, in San Francisco, California. | 高 | SO001, SO015 |
| CO002 | Scale AI's mission is to develop reliable AI systems for the world's most important decisions. | 中 | SO001 |
| CO003 | Scale AI provides data annotation, RLHF, model evaluation, and enterprise GenAI platform services as its core product offerings. | 高 | SO001, SO008, SO009, SO012 |
| CO004 | Scale AI had approximately 1,000 employees as of its about page, with headcount reduced following the July 2025 layoffs of 200 employees and 500 contractors. | 中 | SO001, SO017 |
| CO005 | Scale AI closed a $1 billion Series F round in May 2024, led by Accel, at a post-money valuation of $13.8 billion. | 高 | SO015, SO016 |
| CO006 | Scale AI has raised approximately $1.6 billion in total disclosed venture funding, including $325 million in its 2021 Series E at a $7.3 billion valuation and $1 billion in its 2024 Series F. | 中 | SO015, SO019 |
| CO007 | Meta's June 2025 strategic investment of approximately $14.3 billion for a minority stake implies a Scale AI valuation of over $29 billion. | 高 | SO019, SO016 |
| CO008 | Scale AI has processed more than 15 billion human-labeled decisions and has paid contributors globally over $1 billion. | 中 | SO001 |
| CO009 | Scale AI operates as a late-stage private company; it has not filed for an IPO as of the report date. | 高 | SO001, SO015 |
| CO010 | Alexandr Wang served as CEO of Scale AI from its founding in 2016 until June 2025, when he departed to join Meta AI. | 高 | SO002, SO016 |
| CO011 | Jason Droege was appointed Interim CEO of Scale AI in June 2025 following the departure of Alexandr Wang. | 高 | SO002, SO016 |
| CO012 | Jason Droege founded Uber Eats and scaled it to a $19 billion GMV run rate, then served as VP at Uber and partner at Benchmark before joining Scale AI as Chief Strategy Officer in September 2024. | 中 | SO002, SO027 |
| CO013 | Alexandr Wang retained a seat on Scale AI's board of directors after his departure to Meta in June 2025. | 中 | SO016, SO019 |
| CO014 | Scale AI's full board composition beyond Alexandr Wang is not publicly disclosed as of mid-2026. | 高 | SO001, SO002 |
| CO015 | Alexandr Wang's concurrent role at Meta AI and board seat at Scale AI creates a structural governance conflict of interest, as Meta is both Scale's largest strategic investor and a potential competitor to Scale's AI lab customers. | 中 | SO016, SO019, SO020 |
| CO016 | Scale AI's CEO transition from founder Alexandr Wang to interim CEO Jason Droege is a material key-person risk given Wang's central role in building Scale's customer relationships and product vision. | 中 | SO002, SO016, SO017 |
| CO017 | Scale AI's Series E in August 2021 raised $325 million at an approximate $7.3 billion valuation, with Coatue, Y Combinator, and Founders Fund among the lead investors. | 中 | SO015 |
| CO018 | Scale AI reduced its workforce by approximately 20% in 2023 amid a slowdown in AI training data demand. | 中 | SO015, SO017 |
| CO019 | The Scale AI Series F included investors Amazon, Meta, Cisco, Intel, AMD, ServiceNow, Nvidia, DFJ Growth, WCM, Elad Gil, and Nat Friedman as new or returning strategic participants alongside financial investors. | 高 | SO015, SO019 |
| CO020 | The Scale AI Series F round combined primary capital and a secondary component, allowing existing shareholders to partially liquidate their positions. | 中 | SO015 |
| CO021 | Meta's June 2025 investment was structured as a minority stake purchase of approximately 49% of Scale AI's outstanding equity on a fully diluted basis. | 中 | SO019, SO016 |
| CO022 | The proceeds from Meta's $14.3 billion investment in Scale AI were distributed to existing shareholders and holders of vested equity rather than retained as operating capital on Scale's balance sheet. | 中 | SO019, SO016 |
| CO023 | Scale AI stated it remains operationally independent from Meta following the strategic investment. | 中 | SO002, SO016 |
| CO024 | Scale AI's Donovan platform provides specialized AI agent workflows for defense and intelligence missions, operating in classified environments enabled by DoD IL4 and FedRAMP High certifications. | 高 | SO006, SO005 |
| CO025 | Scale AI holds SOC 2 Type II, ISO 27001, DoD IL4 Provisional Authorization, and FedRAMP High security certifications. | 高 | SO005, SO004 |
| CO026 | Scale AI's RLHF product provides curated preference data for reinforcement learning from human feedback, which is central to training large language models for instruction following and safety alignment. | 高 | SO012, SO009 |
| CO027 | Scale AI secured a U.S. Department of Defense data curation contract for joint force operations in 2024. | 中 | SO014, SO007 |
| CO028 | Scale AI testified before Congress on AI safety and data quality standards during the 118th Congress. | 中 | SO023, SO013 |
| CO029 | Scale AI offers enterprise customers custom pricing with dedicated operations teams and SLA commitments; self-serve customers pay per usage with the first 1,000 labeling units free. | 中 | SO003 |
| CO030 | Scale AI does not publicly disclose its annual recurring revenue, gross margin, or detailed financial metrics; it is classified as a private-undisclosed company. | 高 | SO001, SO003 |
| CO031 | Scale AI signed the White House voluntary AI safety commitments in 2024, pledging commitments on AI safety, security, and trust. | 中 | SO013, SO023 |
| CO032 | Scale AI operates its Global Public Sector division serving international government agencies in addition to U.S. defense customers. | 中 | SO026, SO007 |
| CO033 | Scale AI's scale of operations is indicated by the 15 billion human decisions processed and $1 billion paid to its contributor network globally. | 中 | SO001 |
| CO034 | Scale AI launched the TIME magazine GenAI deployment in under 2 months with over 7,000 AI attack vectors tested, demonstrating enterprise deployment speed and safety rigor. | 中 | SO011 |
| CO035 | Google was Scale AI's largest customer prior to the Meta strategic investment; CNBC reported Google planned to wind down or significantly reduce its Scale relationship after the Meta deal in June 2025. | 高 | SO020, SO021 |
| CO036 | OpenAI wound down its work with Scale AI in June 2025 following Meta's strategic investment, according to CNBC. | 高 | SO021, SO016 |
| CO037 | Scale AI operates the Scale GenAI Platform that transforms enterprise data into domain-specific generative AI applications using a proprietary pipeline. | 中 | SO001, SO008 |
| CO038 | Scale AI laid off approximately 200 employees (14% of staff) and 500 contractors in July 2025, with Interim CEO Droege's memo citing overinvestment in data-labeling capacity relative to the company's strategic direction. | 高 | SO017, SO016 |
| CO039 | Scale AI filed a lawsuit in September 2025 against Mercor, a rival data services company, and a former Scale employee, alleging an attempt to steal Scale's largest customers. | 中 | SO018, SO016 |
| CO040 | McKinsey's State of AI 2025 survey found that 88% of organizations now use AI in at least one business function, up from 78% in 2024, indicating strong underlying demand for AI data and infrastructure services. | 高 | SO028, SO022 |
| CO041 | The OpenAI fine-tuning and custom models program expansion highlights continued strong demand for specialized AI model training services, which Scale AI's RLHF and Data Engine products directly address. | 中 | SO024, SO028 |
| CM001 | The AI data services and infrastructure market encompasses data annotation, RLHF, model evaluation, and enterprise GenAI platform services as distinct but related segments with overlapping buyers and spend. | 高 | SM012, SM013 |
| CM002 | Status-quo substitutes for Scale AI's annotation services include internal human review teams at large AI labs, lower-cost offshore providers, and increasingly, synthetic data generation pipelines. | 中 | SM002, SM003, SM004 |
| CM003 | Scale AI does not currently serve the commodity crowdsource annotation market (e.g., Mechanical Turk), general AI consulting, or hyperscaler AI services (Azure, AWS, GCP) — these are excluded from its SAM. | 中 | SM016, SM017 |
| CM004 | The AI data annotation and evaluation market is served by a fragmented set of competitors including Appen, Labelbox, Snorkel AI, SuperAnnotate, Surge AI, Invisible Technologies, and Mercor, indicating a competitive but not yet consolidated market. | 中 | SM002, SM003, SM004, SM005, SM006, SM007, SM008 |
| CM005 | Scale AI's GenAI Platform competes in the enterprise AI deployment market against hyperscalers and boutique AI consultancies, while its Donovan platform operates in the specialized government and defense AI segment where cleared vendors are scarce. | 中 | SM010, SM009, SM011 |
| CM006 | The Total Addressable Market for AI data services and infrastructure is estimated at $10–$30 billion annually as of 2025, with high uncertainty due to boundary definition differences and rapid market evolution. | 低 | SM012, SM013 |
| CM007 | A narrower TAM for AI data annotation services only (excluding evaluation, RLHF, and enterprise GenAI platform) is estimated at $2–$8 billion, derived from public-company proxy revenues scaled to global market size. | 低 | SM004, SM018 |
| CM008 | Scale AI's Serviceable Addressable Market (premium tier: AI labs, large enterprise, and government) is estimated at $1.5–$6 billion, representing the high-quality end of AI data services where Scale's quality premium and certifications are competitive differentiators. | 低 | SM012, SM018 |
| CM009 | The U.S. government and defense AI data and evaluation submarket addressable by cleared vendors like Scale AI is estimated at $300M–$1B based on defense AI investment growth trends and cleared vendor supply constraints. | 低 | SM011, SM025 |
| CM010 | Scale AI's realistic Serviceable Obtainable Market (SOM) within a 3–5 year horizon is estimated at $500M–$2 billion, consistent with Scale's $29B+ valuation at typical SaaS/data revenue multiples. | 低 | SM021, SM018 |
| CM011 | No public analyst report provides a consistent, granular breakdown of the AI data annotation market with sufficient specificity for high-confidence TAM/SAM/SOM sizing; all estimates carry high uncertainty. | 中 | SM014, SM015 |
| CM012 | AI research laboratories and foundation model developers (OpenAI, Meta, Google DeepMind, Anthropic, Cohere) represent Scale AI's original customer segment, requiring large volumes of high-quality labeled data and RLHF data. | 高 | SM023, SM016 |
| CM013 | Large enterprise AI adopters (Fortune 500 companies using AI for products and automation) represent Scale AI's growth segment, accessing the GenAI Platform with higher switching costs and longer procurement cycles than AI labs. | 中 | SM010, SM022 |
| CM014 | U.S. government and defense agencies represent Scale AI's most defensible buyer segment, with very high switching costs due to FedRAMP High and DoD IL4 clearance requirements, multi-year contract structures, and institutional knowledge dependencies. | 高 | SM009, SM011 |
| CM015 | AI startups and research organizations access Scale's self-serve tier with low switching costs and lower average revenue per user, representing volume without durability. | 中 | SM016, SM017 |
| CM016 | U.S. government AI procurement follows federal acquisition regulations with multi-year contract vehicles and IDIQ structures, creating revenue lumpiness but high stickiness once awarded. | 中 | SM025, SM011 |
| CM017 | The TIME magazine GenAI deployment case study demonstrates Scale AI's ability to win enterprise media customers with fast deployment timelines (under 2 months) and comprehensive AI safety testing. | 中 | SM023 |
| CM018 | Enterprise AI platform buyers (vs. AI lab buyers) offer Scale higher switching costs and platform stickiness, making them strategically more valuable for long-term revenue durability despite lower initial volume. | 中 | SM010, SM022 |
| CM019 | McKinsey's 2025 State of AI survey found that 88% of organizations use AI in at least one business function, up from 78% the previous year, confirming accelerating enterprise AI adoption. | 高 | SM012, SM001 |
| CM020 | McKinsey's 2025 survey found 62% of organizations are experimenting with AI agents, indicating AI adoption is moving toward more complex and automated workflows that require higher-quality training and evaluation data. | 高 | SM012, SM001 |
| CM021 | Most organizations are still in the AI pilot/POC stage rather than full-scale deployment, according to McKinsey's 2025 survey, suggesting enterprise platform revenue for Scale AI will lag the adoption curve by 12–24 months. | 中 | SM012 |
| CM022 | U.S. government AI investment is growing materially as the DoD integrates AI into surveillance, logistics, cybersecurity, and autonomous systems — all of which require trusted data infrastructure like Scale provides. | 中 | SM025, SM011 |
| CM023 | AI safety regulatory pressure — including the White House AI Executive Order and EU AI Act — is increasing demand for AI model evaluation and audit services, benefiting Scale's Evaluation product. | 中 | SM025, SM013 |
| CM024 | The proliferation of foundation model providers (OpenAI, Meta, Anthropic, Mistral, Cohere) increases aggregate demand for RLHF data and model evaluation, partially offsetting the risk from any single large lab customer departure. | 中 | SM020, SM013 |
| CM025 | Synthetic data generation represents a long-term structural threat to human-labeled annotation demand; some AI labs are shifting toward model-generated synthetic data for portions of their training pipelines. | 中 | SM020, SM019 |
| CM026 | Low-cost annotation providers (Appen, offshore QA teams, crowdsourcing platforms) create structural margin pressure on Scale's data annotation business, constraining the price premium Scale can sustain without clear quality differentiation. | 中 | SM004, SM002 |
| CM027 | The AI data annotation market does not have publicly available CAGR estimates from tier-one analyst firms with sufficient market boundary consistency to provide a reliable growth rate for TAM sizing. | 中 | SM014, SM015 |
| CM028 | Appen (ASX-listed) is the only publicly traded direct comparable to Scale AI's data annotation business, but Appen's revenue trends, market positioning, and customer base differ materially from Scale's premium positioning. | 中 | SM004 |
| CM029 | The revenue impact of Google's and OpenAI's departures from Scale AI as customers cannot be quantified from public sources, representing the single most material unresolved market sizing question for Scale's investment case. | 高 | SM019, SM021 |
| CM030 | Scale AI's revenue mix by segment (AI labs vs. enterprise vs. government) is not publicly disclosed, making it impossible to verify which segment dominates current revenue or project the enterprise and government pivot timeline. | 高 | SM016, SM023 |
| CM031 | The wide range of AI data services TAM estimates ($2B–$30B) reflects genuine analyst disagreement on boundary definitions, and investors should treat all sizing figures as directional rather than precise. | 中 | SM014, SM015, SM012 |
| CM032 | Scale AI's AI Readiness Report positions the company as a thought leader in enterprise AI maturity, which supports its brand in the enterprise buyer segment and validates its content-led GTM motion. | 中 | SM022 |
| CM033 | Stanford HAI AI Index 2025 confirms that AI investment reached record levels globally in 2024, with both private investment and government spending accelerating, providing macro validation for the AI data services market. | 高 | SM013, SM001 |
| CM034 | Fortune 500 companies deploying production AI (not just pilot) in 2025 represent a small but growing fraction of the enterprise TAM; McKinsey data suggests roughly one-third of AI-adopting organizations are scaling programs. | 中 | SM012 |
| CM035 | The competitive landscape for AI data services includes multiple vendors targeting different quality/price points — from premium (Scale, Surge) to mid-market (Labelbox, Snorkel, SuperAnnotate) to low-cost (Appen, Invisible) — suggesting market segmentation rather than single-vendor dominance. | 中 | SM002, SM003, SM005, SM006, SM007 |
| CP001 | Scale AI's competitive landscape encompasses direct competitors in AI data annotation (Appen, Labelbox, Snorkel AI, SuperAnnotate), RLHF data (Surge AI, Mercor), enterprise GenAI platform (hyperscalers, boutique AI consultancies), and defense AI (defense contractors). | 高 | SP004, SP007, SP012, SP017, SP014, SP001 |
| CP002 | Appen is the only publicly traded direct comparable to Scale AI's annotation business, and Appen's declining revenues provide a market signal for the structural headwinds facing pure-play annotation vendors. | 中 | SP004, SP019 |
| CP003 | Meta's June 2025 strategic investment in Scale AI created a direct conflict of interest with Scale's largest AI lab customers (Google, OpenAI), leading both to wind down or significantly reduce their Scale relationships. | 高 | SP021, SP022, SP023 |
| CP004 | Google was Scale AI's largest customer prior to the Meta investment and planned to wind down its Scale relationship in June 2025 due to competitive conflict concerns. | 高 | SP021, SP023 |
| CP005 | OpenAI wound down its work with Scale AI in June 2025, coinciding with Meta's strategic investment and founder Alexandr Wang's departure to Meta. | 高 | SP022, SP023 |
| CP006 | Labelbox has expanded from core data labeling to RLHF, model evaluation, leaderboards, and robotics AI training, making it a direct multi-product competitor to Scale AI across several key product lines. | 中 | SP007, SP008, SP009, SP010, SP011 |
| CP007 | Snorkel AI's programmatic labeling approach — using AI weak supervision to reduce manual annotation — threatens Scale's human-expert model by potentially reducing the labor content and cost of annotation, directly challenging Scale's premium pricing thesis. | 中 | SP012, SP013 |
| CP008 | Mercor operates an AI talent marketplace for RLHF, model evaluation, and data labeling, founded by former Scale AI-connected individuals, and is being sued by Scale for alleged customer poaching. | 高 | SP001, SP014 |
| CP009 | SuperAnnotate competes with Scale AI in the enterprise annotation platform market with a security-first positioning and collaborative workflow tools, primarily targeting computer vision use cases. | 中 | SP015 |
| CP010 | Invisible Technologies provides AI-powered business operations services that compete with Scale's enterprise automation and annotation capabilities at the operations level. | 中 | SP017 |
| CP011 | Surge AI focuses specifically on high-quality human feedback data for LLM training and RLHF, directly competing with Scale's RLHF product with a smaller but expert contributor network. | 中 | SP017 |
| CP012 | Appen has expanded into agentic AI services and model evaluation in addition to its core annotation business, indicating it is attempting to follow Scale into higher-margin evaluation segments. | 中 | SP005, SP006 |
| CP013 | Appen provides data security features for enterprise and government customers but does not disclose DoD IL4 or FedRAMP High certifications equivalent to Scale AI's cleared vendor status. | 中 | SP018, SP002 |
| CP014 | Labelbox offers tiered pricing including self-serve developer plans and enterprise custom pricing, making it price-competitive against Scale's self-serve tier and potentially more accessible to mid-market enterprise. | 中 | SP016, SP007 |
| CP015 | Scale AI's DoD IL4 Provisional Authorization and FedRAMP High certifications are unique among publicly disclosed AI data annotation vendors, creating a near-exclusive position in classified and defense AI data markets. | 高 | SP002, SP003 |
| CP016 | Scale AI's Donovan defense AI agent platform has no publicly disclosed direct competitor with equivalent security clearances and defense-specific deployment capabilities. | 中 | SP003, SP002 |
| CP017 | Scale AI's evaluation benchmarking position — including the Scale Leaderboard and WMDP harmful-knowledge benchmark — has established reputational authority as a trusted third-party LLM evaluator, providing a differentiated position versus competitors. | 中 | SP025, SP002 |
| CP018 | Labelbox and Snorkel AI do not publicly disclose government security certifications equivalent to Scale's DoD IL4 or FedRAMP High authorizations, indicating limited ability to compete for Scale's most defensible government contracts. | 中 | SP007, SP012 |
| CP019 | Scale AI's feature breadth — spanning annotation, RLHF, evaluation, enterprise GenAI platform, and defense AI agents — is unmatched by any single competitor, though individual competitors may lead in specific capabilities. | 中 | SP025, SP007, SP004 |
| CP020 | Scale's enterprise GenAI Platform has no direct counterpart at Appen, Labelbox, or Snorkel AI, but faces significant competition from hyperscaler AI platforms (AWS Bedrock, Azure AI, Google Vertex) with far greater resource advantages. | 中 | SP025, SP007 |
| CP021 | Switching costs for AI lab customers of Scale AI are moderate: annotation pipelines are fungible, and Google and OpenAI both exited Scale's ecosystem within weeks of the Meta conflict emerging in June 2025. | 高 | SP021, SP022 |
| CP022 | Switching costs for enterprise customers using Scale's GenAI Platform are higher than for annotation-only customers, due to platform integration, workflow customization, and data pipeline dependencies. | 中 | SP025 |
| CP023 | Obtaining DoD IL4 Provisional Authorization and FedRAMP High certification requires a multi-year process involving security audits, infrastructure requirements, and federal agency sponsorship, creating a 12–36 month barrier for any competitor seeking to enter Scale's government segment. | 中 | SP002, SP003 |
| CP024 | Mercor is actively building an alternative AI expert contributor supply chain through its talent marketplace to directly compete with Scale's contributor network, targeting the same expert annotators and RLHF data customers. | 中 | SP001, SP014 |
| CP025 | Multi-homing — where annotation customers run work through multiple vendors simultaneously — is common in the AI data market; Scale's self-serve tier confirms this with its pay-as-you-go model that has no long-term commitment. | 中 | SP007, SP012 |
| CP026 | Scale AI's government and defense segment represents its most durable competitive moat, requiring years for competitors to replicate the clearances, institutional knowledge, and defense-specific product (Donovan) that Scale has built. | 高 | SP002, SP003, SP025 |
| CP027 | Scale AI's quality premium in annotation is under pressure from competitors including Labelbox, Surge AI, and lower-cost providers; the departure of Google and OpenAI demonstrates that quality premium alone was insufficient to sustain those relationships. | 高 | SP021, SP022, SP024 |
| CP028 | Appen's declining revenues — the most direct public comparable to Scale's annotation business — serve as a leading indicator of structural headwinds in the pure-play data annotation market. | 中 | SP019, SP004 |
| CP029 | The Meta strategic investment has become Scale AI's primary competitive liability in the AI lab segment: Mercor's lawsuit defense (Sep 2025) reveals that competitors view Scale's AI lab customer base as vulnerable and actively targetable. | 高 | SP001, SP021, SP022 |
| CP030 | Scale AI's July 2025 layoffs specifically targeted data-labeling employees — not evaluation or platform — confirming management's internal assessment that commodity annotation is a commoditizing segment requiring rightsizing. | 高 | SP024, SP023 |
| CP031 | Snorkel AI's programmatic labeling approach and rising AI-assisted annotation tools represent a structural threat to the human-label-intensive part of Scale's business model, potentially reducing the labor content of annotation work over time. | 中 | SP012, SP013 |
| CP032 | Scale AI's annotation platform (Scale Data Engine) has built-in quality feedback loops and an expert contributor network that paid contributors $1B+ globally, creating a proprietary supply chain that competitors must spend years and significant capital to replicate. | 中 | SP025, SP002 |
| CP033 | The competitive landscape for AI data services lacks a dominant single vendor — no competitor has matched Scale's full-stack positioning across annotation, RLHF, evaluation, and defense AI — but individual segments are increasingly contested. | 中 | SP004, SP007, SP012, SP014 |
| CP034 | Large defense contractors such as Palantir and Booz Allen are potential long-term entrants into the defense AI data market, representing a risk to Scale's Donovan moat over a 3–5 year horizon. | 低 | SP003 |
| CP035 | Scale AI has not publicly documented evidence of losing any government or defense contract to a competitor, which supports the durability of its government segment moat in the near term. | 中 | SP002, SP003 |
| CI001 | Scale AI generates revenue through enterprise data annotation/RLHF services, the Scale GenAI Platform (enterprise SaaS + managed services), U.S. government and defense contracts, and a self-serve tier. | 高 | SI001, SI013, SI020, SI014 |
| CI002 | Scale AI's GenAI Platform targets enterprises seeking to build custom AI applications from proprietary data, representing a higher-margin opportunity than commodity annotation. | 中 | SI001, SI004 |
| CI003 | Scale AI holds DoD IL4 and FedRAMP High certifications, enabling it to pursue classified and defense-grade AI data contracts that most competitors cannot access. | 中 | SI005, SI006 |
| CI004 | Scale AI's RLHF revenue is affected by OpenAI's wind-down of its Scale relationship post-Meta-investment, while Meta's own RLHF work with Scale is expanding. | 中 | SI022, SI017 |
| CI005 | Scale AI's self-serve Data Engine offers the first 1,000 labeling units free, then charges pay-as-you-go; enterprise tiers are custom-priced with dedicated operations and SLAs. | 高 | SI013, SI001 |
| CI006 | Scale AI's enterprise GTM is primarily sales-led with dedicated operations teams and solution engineers; government contracts require specialized federal procurement BD processes. | 中 | SI013, SI005, SI021 |
| CI007 | Scale AI's enterprise pricing for annotation, RLHF, and platform services is custom-quoted and not publicly disclosed; no list pricing or typical deal size information is available. | 高 | SI013, SI001, SI020 |
| CI008 | Scale AI's enterprise annotation sales cycle is estimated at 3–12 months, with government contract sales cycles typically 12–36 months; specific CAC figures are not disclosed. | 低 | SI005, SI006, SI021 |
| CI009 | Scale AI's data-labeling revenue was concentrated among a small number of large AI lab customers, creating material customer concentration risk that materialized when Google and OpenAI departed in 2025. | 高 | SI018, SI022, SI016 |
| CI010 | Scale AI's investor relationships (Accel, Amazon, Meta, Nvidia, Cisco) create potential co-sell and referral channels, but the specific economic impact of these relationships on revenue is not publicly disclosed. | 低 | SI015, SI014 |
| CI011 | Scale AI's annotation cost of revenue is dominated by human labor, with the company having paid over $1 billion globally to contributors; annotation gross margins are estimated at 25–45%, using Appen's public financials as proxy. | 低 | SI014, SI007, SI008 |
| CI012 | Scale AI's GenAI Platform is estimated to carry significantly higher gross margins (45–65%) than annotation services, reflecting greater software leverage and less per-unit human labor, consistent with enterprise managed-services industry benchmarks. | 低 | SI001, SI004, SI023 |
| CI013 | Scale AI's July 2025 layoffs of 200 employees (14% of staff) and 500 contractors targeted the data-labeling business, signaling management's intent to shift to higher-margin enterprise and government services and reduce the labor-heavy cost base. | 高 | SI016, SI021 |
| CI014 | Appen (ASX: APX), the only publicly traded direct comparable to Scale AI's annotation segment, has reported declining revenues and gross margins in the 25–40% range, confirming structural headwinds in commodity annotation economics. | 中 | SI007, SI008, SI009 |
| CI015 | Scale AI's cost structure is primarily working-capital-intensive (human labor payments, contractor management) rather than physical capex intensive, making it operationally flexible but subject to margin pressure from annotation labor costs. | 中 | SI014, SI010, SI011 |
| CI016 | Scale AI raised approximately $600 million in pre-Series F capital and $1 billion in Series F (May 2024, $13.8 billion valuation, led by Accel) for a total raised of approximately $1.6 billion or more. | 高 | SI015, SI019 |
| CI017 | Meta invested approximately $14.3 billion for a minority stake of approximately 49% in Scale AI in June 2025, implying a Scale valuation of over $29 billion; the investment was primarily structured as a secondary transaction with proceeds distributed to existing shareholders. | 高 | SI017, SI019 |
| CI018 | Scale AI does not have any publicly disclosed debt, credit facilities, or project-finance obligations; its capital structure is equity-financed. | 中 | SI015, SI017 |
| CI019 | Based on the Series F primary capital (estimated $500M+ to the company), post-layoff burn rate reductions, and potential primary capital from the Meta deal, Scale AI is estimated to have $500M–$1B cash on hand with 24–48 months of runway as of mid-2025. | 低 | SI015, SI017, SI016 |
| CI020 | Scale AI's total estimated annual revenue is in the range of $200M–$500M ARR, based on headcount proxies (~1,000 employees post-layoff), industry revenue-per-employee benchmarks, and analyst commentary; this estimate has very low confidence without public financial disclosure. | 低 | SI023, SI025, SI014 |
| CI021 | At Scale AI's $29B+ implied valuation, if ARR is $200M–$500M, the implied revenue multiple of 58x–145x is in hypergrowth-premium territory and cannot be verified or justified without audited financial statements and confirmed growth rates. | 低 | SI017, SI023 |
| CI022 | Scale AI does not disclose revenue, ARR, gross margin, NRR, burn rate, or operating cash flow; these omissions represent the primary blockers to financial underwriting of the company. | 高 | SI013, SI014, SI021 |
| CI023 | Google and OpenAI's departure as customers in 2025 represents a material risk to Scale AI's revenue trajectory; the financial magnitude of this attrition is unknown but potentially $50M–$200M+ in annual revenue based on their reported status as Scale's largest AI lab customers. | 低 | SI018, SI022, SI016 |
| CI024 | The MetA strategic investment does not appear to include covenants or restrictions on Scale AI's operations; Meta holds a minority stake and Scale remains independent, but the conflict of interest with other customers is a de facto constraint on commercial relationships. | 低 | SI017, SI019, SI021 |
| CI025 | A Mercor lawsuit filed by Scale AI in September 2025 alleging customer poaching could have financial implications including legal costs, settlement obligations, and customer loss not yet captured in financial estimates. | 中 | SI016, SI021 |
| CI026 | Scale AI's enterprise and government revenue segments have higher revenue quality (longer contracts, stronger switching costs) than the annotation segment, but their current size and growth trajectory are not publicly disclosed. | 中 | SI005, SI006, SI001 |
| CI027 | Scale AI paid over $1 billion globally to annotation contributors, confirming the company's historical commitment to human labor-based annotation and the labor intensity of its core revenue model. | 中 | SI014, SI013 |
| CI028 | McKinsey's 2025 AI State report notes that 88% of organizations use AI in at least one business function, up from 78%, supporting a growing market for Scale AI's enterprise AI data and platform services. | 高 | SI023, SI025 |
| CI029 | Scale AI's annotation business model is primarily working-capital-intensive, with payments to contributors timed to annotation project delivery; this creates a different capex profile than hardware or infrastructure companies. | 中 | SI014, SI012 |
| CI030 | Appen's publicly disclosed multimodal, speech/audio, and physical AI annotation services are directly comparable to Scale AI's core annotation product lines, making Appen's financial results the best available public proxy for Scale's annotation segment economics. | 中 | SI009, SI010, SI011 |
| CI031 | Scale AI's government and defense revenue segment carries high switching costs (DoD IL4 and FedRAMP High certification barriers) and multi-year contract durations, creating more durable and predictable revenue than the annotation segment. | 中 | SI005, SI006 |
| CI032 | Scale AI's revenue mix is shifting from data-labeling (legacy largest segment) toward enterprise GenAI platform and government/defense contracts, consistent with the July 2025 restructuring that cut data-labeling headcount while maintaining enterprise and government operations. | 中 | SI016, SI021, SI001 |
| CI033 | Scale AI's government contract revenue segment is growing through active DoD relationships (data curation contract, DIU RCV program, White House AI commitments) but exact contract values and revenue contribution are not publicly disclosed. | 中 | SI005, SI006, SI024 |
| CI034 | The net primary capital received by Scale AI from the Meta deal is not publicly disclosed; the deal was primarily structured as shareholder liquidity, meaning Scale's operating treasury benefit may be significantly less than the $14.3B headline figure. | 中 | SI017, SI019 |
| CI035 | Scale AI's financial verdict is mixed: abundant capital buffers the pivot risk, but the $29B+ valuation is unjustifiable without confirmed revenue data, and the annotation-segment revenue attrition risk is unquantified. | 中 | SI017, SI018, SI022 |
| CE001 | Scale AI's core product portfolio consists of five offerings: Scale Data Engine (annotation + curation), Scale RLHF (LLM training data), Scale Evaluation + Leaderboard (model benchmarking), Scale GenAI Platform (enterprise AI applications), and Donovan (defense AI agents). | 高 | SE017, SE016, SE015, SE018, SE020 |
| CE002 | The Scale GenAI Platform enables enterprises to build and deploy custom AI applications from proprietary data, differentiating from hyperscalers by integrating annotation quality with LLM customization. | 中 | SE011, SE018, SE013 |
| CE003 | Scale AI's RLHF product provides expert human feedback data for LLM training and alignment; the product was historically used by OpenAI and Google before their 2025 departures. | 中 | SE016, SE021 |
| CE004 | The Scale Leaderboard is a public developer-facing tool that ranks LLM performance across capability dimensions; it has established Scale as a trusted third-party evaluator in the AI research community. | 中 | SE001, SE015 |
| CE005 | Donovan is Scale AI's specialized AI agents platform for DoD, IC, and government agencies; it operates in DoD IL4-certified environments and has no disclosed direct competitor in the cleared AI agents space. | 中 | SE020, SE019, SE003 |
| CE006 | Scale AI's operating model combines proprietary annotation tooling, a quality assurance pipeline, and a global contributor network with software API infrastructure — a hybrid human-in-the-loop architecture. | 中 | SE017, SE012, SE011 |
| CE007 | The Scale API provides programmatic access to annotation, RLHF, and evaluation services with REST interface and webhook integration; it enables enterprise and developer integration with MLOps pipelines. | 中 | SE012, SE013 |
| CE008 | Scale AI's annotation workflow: customers submit raw data and task guidelines via API; Scale's contributor network executes annotation; a QA pipeline provides statistical review and expert escalation; annotated datasets are returned to customers. | 中 | SE017, SE012, SE016 |
| CE009 | Scale AI's GenAI Platform operating model is a managed services approach involving data ingestion, LLM selection, fine-tuning or RAG pipeline configuration, red-teaming and safety evaluation, and production deployment with monitoring. | 中 | SE011, SE018, SE027 |
| CE010 | Donovan's architecture is specialized for cleared government environments: DoD IL4-certified cloud infrastructure, classified data integration, multi-modal AI agent capabilities, and explainability features for military applications. | 低 | SE020, SE019, SE003 |
| CE011 | Scale AI's primary technology differentiators are: proprietary annotation tooling and QA methodology, the global contributor network ($1B+ paid), DoD IL4 and FedRAMP High certifications, the Scale Leaderboard as reputational IP, and Donovan as a first-mover defense AI platform. | 中 | SE014, SE020, SE001, SE017 |
| CE012 | Scale AI's DoD IL4 and FedRAMP High certifications represent a genuine barrier to entry for competitors: obtaining these clearances takes 2–4 years, requires significant compliance investment, and requires government institutional relationships. | 中 | SE014, SE019, SE020 |
| CE013 | Scale AI's contributor network — having paid over $1 billion globally — is a proprietary supply-side asset that competitors including Mercor are attempting to replicate; Scale's lawsuit against Mercor specifically alleges customer and annotator poaching. | 高 | SE022, SE010, SE009 |
| CE014 | The Scale Leaderboard positions Scale as the trusted third-party LLM evaluator; Labelbox has also launched competing leaderboards, creating competitive pressure in the model evaluation market. | 中 | SE001, SE007, SE015 |
| CE015 | The WMDP (Weapons of Mass Destruction Proxy) benchmark is a publicly available AI safety evaluation tool created by Scale; its adoption by the AI safety research community reinforces Scale's position as a trusted AI safety evaluator. | 中 | SE002, SE025 |
| CE016 | Scale AI holds SOC 2 Type II, ISO 27001, DoD IL4 Provisional Authorization, and FedRAMP High Authorization certifications, confirmed on the official Scale security page. | 高 | SE014, SE019 |
| CE017 | Scale AI signed the 2024 White House voluntary AI safety commitments covering RLHF safety practices, red-teaming, and responsible AI deployment. | 高 | SE024, SE002 |
| CE018 | Scale AI's annotation quality controls include task-specific quality guidelines, multiple annotator redundancy, statistical quality monitoring, expert reviewer escalation, and inter-annotator agreement scoring; however, no public third-party quality audit is available. | 中 | SE017, SE016 |
| CE019 | The TIME customer case study demonstrates Scale GenAI Platform deployment speed (under 2 months) and red-teaming capability (7,000+ attack vectors tested), providing partial public evidence of product performance. | 中 | SE027, SE018 |
| CE020 | Scale AI's data privacy and customer data handling practices are partially covered by SOC 2 Type II certification; however, the specific data isolation mechanisms for enterprise annotation customers are not publicly documented. | 中 | SE014, SE011 |
| CE021 | Scale AI's 2025 product roadmap is implicitly directed toward expanding the GenAI Platform, Donovan government deployments, and Scale Evaluation, while reducing investment in commodity annotation — consistent with the July 2025 restructuring. | 中 | SE021, SE018, SE020 |
| CE022 | Scale AI does not maintain a public product roadmap, changelog, or developer status page; forward-looking roadmap information must be inferred from public statements and strategic announcements. | 高 | SE011, SE012, SE017 |
| CE023 | Scale's GenAI Platform competes with Amazon Bedrock, Google Vertex AI, and Azure AI Studio; hyperscalers have significantly greater infrastructure scale, distribution, and resources, representing a structural competitive threat to Scale's platform business. | 中 | SE026, SE023, SE018 |
| CE024 | Scale AI has not publicly disclosed investment in synthetic data capabilities; if synthetic data quality matches human annotation for LLM training, Scale's Data Engine TAM could shrink materially — a critical product roadmap gap. | 中 | SE005, SE004, SE021 |
| CE025 | Snorkel AI's programmatic labeling approach, which uses AI-assisted weak supervision to generate training data with less human annotation, represents a technical threat to Scale's human-annotation-first model. | 中 | SE005, SE004 |
| CE026 | Scale AI's developer community presence is limited compared to AI infrastructure companies with open-source products; the Scale Leaderboard and WMDP benchmark are the primary developer-facing signals, generating reputational capital rather than active developer community engagement. | 中 | SE001, SE002, SE004 |
| CE027 | Scale AI's technology differentiation versus Labelbox is primarily in QA methodology and contributor quality management; Labelbox has built competing evaluation products (Labelbox Leaderboards) and an Expert Network, narrowing the quality gap. | 中 | SE007, SE006, SE017 |
| CE028 | The McKinsey State of AI 2025 confirms 62% of organizations experimenting with AI agents, validating Scale's strategic pivot toward agentic AI services (GenAI Platform, Donovan) as an addressable growth market. | 高 | SE026, SE025 |
| CE029 | Scale AI's RLHF product is affected by OpenAI's departure (wind-down confirmed June 2025) but expanding in relation to Meta's RLHF needs following the strategic investment; this creates dependency concentration risk on Meta as both investor and customer. | 中 | SE023, SE016, SE021 |
| CE030 | Scale AI's Scale Evaluation product has an independence perception risk: as a company with Meta as a strategic investor (~49% minority stake), its role as a neutral third-party LLM evaluator may be questioned by AI labs that compete with Meta. | 中 | SE023, SE001 |
| CE031 | Scale's DIU RCV (Robotic Combat Vehicle) program win confirms that Scale's technology meets defense procurement standards for autonomous military AI applications, validating Donovan's technical capability for defense missions. | 中 | SE003, SE020 |
| CE032 | Scale AI's annotation quality differentiation versus Snorkel AI is philosophical: Scale uses high-quality human annotation for accuracy; Snorkel uses AI-assisted weak supervision for scale and cost. The two approaches serve different customer segments and are both growing. | 中 | SE005, SE017, SE004 |
| CE033 | Scale AI's AI safety white paper on test and evaluation provides public documentation of Scale's approach to model evaluation methodology, supporting its positioning as a credible government AI evaluation authority. | 中 | SE002, SE024, SE015 |
| CE034 | Scale AI's cloud infrastructure dependencies (AWS/GCP/Azure) represent a supply chain risk: changes in cloud provider pricing or availability could affect annotation tooling uptime and GenAI Platform delivery. | 低 | SE011, SE017, SE019 |
| CE035 | No publicly available independent third-party technical audit of Scale AI's annotation quality, QA pipeline accuracy, or GenAI Platform performance exists; all quality claims are company-asserted or based on single customer case studies. | 高 | SE027, SE017, SE016 |
| CU001 | Scale AI serves three primary customer segments: AI Labs and model developers, Fortune 500 enterprise customers, and U.S. government and defense agencies. | 高 | SU001, SU015 |
| CU002 | TIME deployed Scale AI's GenAI Platform in a production safety application, testing over 7,000 adversarial attack vectors against AI-generated content in under two months. | 高 | SU002, SU001 |
| CU003 | Meta holds approximately 49% of Scale AI's equity following the June 2025 strategic investment and is simultaneously an expanding RLHF data customer. | 高 | SU018, SU015 |
| CU004 | Cohere is listed as an AI lab customer on Scale AI's official customers page. | 中 | SU001 |
| CU005 | Etsy is listed as an enterprise customer on Scale AI's official customers page. | 中 | SU001 |
| CU006 | Instacart is listed as an enterprise customer on Scale AI's official customers page. | 中 | SU001 |
| CU007 | Pinterest is listed as an enterprise customer on Scale AI's official customers page. | 中 | SU001 |
| CU008 | Scale AI holds active Department of Defense contracts including a data curation contract for joint force operations and the Donovan platform for national security AI workflows. | 高 | SU004, SU005 |
| CU009 | Google, previously Scale AI's largest customer, planned to wind down or significantly reduce its Scale AI relationship in June 2025 following Meta's strategic investment. | 高 | SU019, SU018 |
| CU010 | OpenAI wound down its work with Scale AI in June 2025, as reported by CNBC. | 高 | SU020, SU015 |
| CU011 | Scale AI has paid over $1 billion to annotation contributors globally, per its official customers page. | 中 | SU001 |
| CU012 | Scale AI has processed over 15 billion human-labeled decisions, per its official customers page. | 中 | SU001 |
| CU013 | Scale AI laid off approximately 200 employees (14% of staff) and 500 contractors in July 2025, with the reductions concentrated in the data-labeling business. | 高 | SU016, SU015 |
| CU014 | Scale AI's GenAI Platform was deployed in a production environment at TIME Media for AI content safety testing, as documented in the official case study. | 中 | SU002 |
| CU015 | Scale AI holds active DoD IL4 and FedRAMP High certifications enabling deployment in classified and government environments. | 高 | SU003, SU004 |
| CU016 | Scale AI's DoD data curation contract for joint force operations was announced via official company blog post. | 高 | SU005, SU004 |
| CU017 | Scale AI offers a self-serve pricing tier with the first 1,000 labeling units free and pay-as-you-go access for developers and researchers. | 中 | SU027 |
| CU018 | Amazon, Cisco, Intel, AMD, and ServiceNow participated as strategic investors in Scale AI's May 2024 Series F round, providing customer-as-investor validation. | 中 | SU015 |
| CU019 | Scale AI has not publicly disclosed any net revenue retention (NRR) or gross revenue retention (GRR) metrics. | 低 | |
| CU020 | Scale AI has not publicly disclosed its total active customer count across any segment. | 低 | |
| CU021 | Scale AI's enterprise pricing structure includes custom pricing with dedicated operations teams and SLA commitments for enterprise customers. | 中 | SU027 |
| CU022 | Meta's information access as a 49% investor alongside its role as an RLHF customer creates a potential conflict of interest for other enterprise and AI lab customers evaluating Scale AI's data security and confidentiality. | 中 | SU018, SU019 |
| CU023 | Snorkel AI operates a public leaderboard and partners program targeting enterprise ML customers, competing in the same AI data infrastructure market as Scale AI. | 中 | SU007, SU008 |
| CU024 | SuperAnnotate maintains a public enterprise platform offering for annotation, competing with Scale AI's Data Engine in the enterprise annotation segment. | 中 | SU025 |
| CU025 | Mercor operates as a competitor in the AI annotation and evaluation marketplace, with enterprise-facing product pages and active hiring. | 中 | SU013, SU017 |
| CU026 | Scale AI's Donovan platform serves the defense and intelligence community with AI agent workflows designed for classified national security operations. | 高 | SU006, SU003 |
| CU027 | Google's departure from Scale AI was driven by competitive conflict concerns arising from Meta's strategic investment, per CNBC sourcing. | 中 | SU019 |
| CU028 | OpenAI's wind-down of Scale AI work coincided with founder Alexandr Wang's departure to join Meta, suggesting structural realignment of AI lab annotation sourcing. | 中 | SU020, SU015 |
| CU029 | The simultaneous departure of Google and OpenAI in Q2 2025 represents an estimated 20-40% reduction in Scale AI's AI lab segment revenue, based on their reported prominence as major customers. | 低 | SU019, SU020 |
| CU030 | Scale AI's government and defense customers have the highest switching costs of any segment due to multi-year contracts, IL4/FedRAMP certification requirements, and classified-environment integration barriers. | 中 | SU003, SU004 |
| CU031 | Enterprise customers Etsy, Instacart, and Pinterest appear on Scale AI's official customers page without associated case studies, outcome metrics, or deployment scope details. | 中 | SU001 |
| CU032 | No G2, Gartner Peer Insights, or Capterra reviews were identified for Scale AI's enterprise platform, leaving customer satisfaction entirely unquantified. | 低 | |
| CU033 | Scale AI's July 2025 layoffs concentrated in data-labeling signal that RLHF and annotation throughput from AI lab customers declined materially in H1 2025. | 中 | SU016 |
| CU034 | Scale AI filed a lawsuit against Mercor in September 2025 alleging customer poaching and misappropriation of trade secrets by a former employee. | 中 | SU017 |
| CU035 | Scale AI's documented land-and-expand model involves starting with data annotation, then upselling to RLHF, evaluation, and the GenAI Platform as customer needs mature. | 中 | SU001, SU026 |
| CU036 | Scale AI's customer base is predominantly U.S.-based, with no publicly disclosed international customer count or revenue split by geography. | 中 | SU001, SU003 |
| CU037 | Scale AI has not publicly disclosed any annual or quarterly customer churn rate for any segment. | 低 | |
| CU038 | Scale AI's AI lab customer segment is the most volatile, with Google and OpenAI both exiting in Q2 2025 and remaining labs facing structural pressure toward in-house annotation. | 中 | SU019, SU020 |
| CU039 | Snorkel AI's press page and partner program confirm active enterprise customer development activity, corroborating that the enterprise AI data market remains competitive. | 低 | SU009, SU008 |
| CU040 | Appen's enterprise case studies indicate ongoing demand for AI annotation services in comparable market segments, validating Scale's addressable market despite attrition events. | 低 | SU023 |
| CU041 | Meta's expanding RLHF customer relationship with Scale AI, combined with its ~49% equity stake, creates an unprecedented customer-investor concentration that could deter other AI lab customers. | 中 | SU018, SU019 |
| CU042 | Scale AI's self-serve tier pricing structure was publicly accessible as of the research date, confirming a lower-commitment entry point for the developer and researcher segment. | 中 | SU027 |
| CR001 | Customer concentration risk is the highest-severity risk facing Scale AI, with Google and OpenAI both departing as customers in Q2 2025 following the Meta investment. | 高 | SR005, SR006 |
| CR002 | CEO and leadership transition risk is high: founder Alexandr Wang departed June 2025 to join Meta, replaced by interim CEO Jason Droege who lacks direct experience in Scale's core government and enterprise AI markets. | 高 | SR001, SR004 |
| CR003 | Meta's ~49% equity stake combined with its role as Scale AI's largest customer creates an unprecedented concentration of investor-customer governance risk that is structurally unusual in the venture-backed AI sector. | 高 | SR003, SR005 |
| CR004 | Scale AI is executing a business model pivot from annotation-volume revenue to enterprise GenAI Platform and government Donovan revenue under adverse conditions of active revenue attrition and CEO transition. | 高 | SR001, SR002 |
| CR005 | Scale AI filed a lawsuit against Mercor in September 2025 alleging customer poaching and misappropriation of trade secrets by a former employee, confirmed by TechCrunch reporting. | 高 | SR011, SR012 |
| CR006 | Scale AI's White House AI safety voluntary commitments (2024) demonstrate proactive regulatory engagement but do not constitute a binding compliance shield against future mandatory AI regulations. | 高 | SR009, SR013 |
| CR007 | Scale AI holds DoD IL4 and FedRAMP High certifications, SOC 2 Type II, and ISO 27001, representing a mature security compliance posture for government and enterprise customers. | 高 | SR007, SR008 |
| CR008 | Scale AI laid off approximately 200 employees (14% of staff) and 500 contractors in July 2025, concentrated in data labeling, creating talent retention risk and potential operational quality degradation. | 高 | SR002, SR001 |
| CR009 | U.S. Congressional AI oversight is active, with Scale AI having testified before the House of Representatives about AI capabilities and risks, indicating regulatory attention to AI infrastructure companies. | 中 | SR013 |
| CR010 | The EU AI Act creates compliance obligations for AI training data providers serving EU enterprise customers, including potential classification of high-risk AI system data as requiring additional oversight. | 中 | SR027 |
| CR011 | Scale AI's government defense work involving AI-enabled autonomy programs may attract ITAR and export control scrutiny if technology transfer to non-U.S. entities is involved in the data pipeline. | 中 | SR008, SR024 |
| CR012 | Scale AI has made proactive AI safety commitments including the WMDP harmful-knowledge benchmark for evaluating dual-use AI risks, reducing its regulatory exposure on AI safety grounds. | 中 | SR030, SR009 |
| CR013 | No publicly disclosed data security incidents or breaches affecting Scale AI's enterprise or government customers were identified in this research as of the run date. | 中 | SR007 |
| CR014 | Scale AI's compliance with U.S. AI safety voluntary commitments and its test and evaluation white paper reduce its near-term regulatory enforcement exposure from the current U.S. AI governance framework. | 中 | SR009, SR025 |
| CR015 | Scale AI's annotation quality is a key operational risk: any quality degradation in RLHF data following the 500-contractor reduction could impair enterprise and government contract performance. | 中 | SR002 |
| CR016 | Scale AI's dependency on a global annotation contributor network creates supply-chain risk, particularly as Mercor and other platforms actively compete for the same annotation workforce. | 中 | SR011, SR020 |
| CR017 | Scale AI's data security posture is supported by SOC 2 Type II, ISO 27001, DoD IL4, and FedRAMP High certifications, reducing but not eliminating the risk of data breach affecting customer AI training data. | 中 | SR007 |
| CR018 | Scale AI's business model pivot from annotation-volume to platform revenue represents a high-severity execution risk: the timeline, conversion rate, and revenue replacement pace are unknown and unconfirmed publicly. | 中 | SR010, SR014 |
| CR019 | Scale AI's July 2025 layoffs and contractor reductions signal that the data-labeling segment volume has declined materially, creating operational capacity risk if the enterprise GenAI Platform requires rapid scaling of new annotation workflows. | 中 | SR002 |
| CR020 | Scale AI's Meta dependency encompasses three simultaneous roles: investor (~49% equity), customer (expanding RLHF), and former employer of departed CEO Wang, creating a concentration with no structural firewall confirmed publicly. | 中 | SR003, SR004 |
| CR021 | Scale AI's government customer segment is a positive dependency: multi-year DoD contracts provide a revenue floor, but they also create concentration in U.S. government budget cycles and contract renewal processes. | 中 | SR008, SR024 |
| CR022 | Scale AI's cloud infrastructure dependency on AWS and similar providers creates a platform concentration risk that is partially mitigated by its FedRAMP High certification requirements for government-grade infrastructure. | 中 | SR007, SR008 |
| CR023 | Appen, a public comparable to Scale AI, experienced significant customer attrition and revenue decline when AI lab customers reduced annotation spend, providing a cautionary precedent for Scale's current trajectory. | 中 | SR028 |
| CR024 | Jason Droege's appointment as interim CEO was confirmed in June 2025; his background includes founding Uber Eats (scaled to $19B GMV) and VC partnership at Benchmark, but not direct government AI or enterprise data platform leadership. | 中 | SR004, SR001 |
| CR025 | Alexandr Wang retains a board seat at Scale AI despite departing as CEO to join Meta, creating a potential conflict of interest at the governance level that has not been publicly addressed. | 中 | SR001, SR003 |
| CR026 | Scale AI has not publicly disclosed its current cash position, quarterly burn rate, or expected runway following the Meta transaction proceeds being distributed to shareholders rather than retained for operations. | 低 | |
| CR027 | Scale AI's government relationship managers hold classified clearances and institutional knowledge that represent irreplaceable assets; their retention is critical to DoD contract renewal and expansion. | 中 | SR008, SR019 |
| CR028 | The most significant thesis-break triggers for Scale AI are: any government contract cancellation, any additional non-AI-lab customer citing Meta conflict, or the GenAI Platform failing to replace AI lab ARR within 18 months. | 中 | SR005, SR003 |
| CR029 | Synthetic data generation capabilities in frontier AI models are reducing the volume of human-annotated RLHF data required per training run, creating structural demand reduction for Scale's core annotation service. | 中 | SR015, SR014 |
| CR030 | SuperAnnotate's enterprise platform and foundation-model-builder solutions pages indicate competitive expansion into Scale AI's core annotation market segments. | 低 | SR021, SR022 |
| CR031 | Labelbox's enterprise platform (product/platform page) indicates continued investment in annotation infrastructure competing with Scale AI's Data Engine for enterprise customers. | 低 | SR026 |
| CR032 | Scale AI's evaluation and test platform for government AI safety is documented in an official white paper, demonstrating product differentiation in the government AI market beyond annotation. | 中 | SR025, SR023 |
| CR033 | Scale AI's headcount of approximately 1,000 employees post-layoff represents a significant reduction in operational capacity that may affect the speed of enterprise GenAI Platform go-to-market. | 中 | SR002 |
| CR034 | The Mercor lawsuit (Scale AI v. Mercor, NDCA 2025) creates ongoing legal uncertainty and management distraction, with discovery potentially exposing internal customer contract terms and competitive intelligence. | 中 | SR011, SR012 |
| CR035 | Scale AI's Reuters coverage of the Google departure (rate-limited) corroborates the CNBC reporting, providing multi-source confirmation of the customer concentration event. | 高 | SR016, SR005 |
| CR036 | Scale AI's Reuters coverage of the Meta deal as a test of AI partnerships (rate-limited) provides an independent perspective on the concentration and governance risks of the Meta-Scale relationship. | 中 | SR017, SR003 |
| CR037 | Scale AI's active hiring page indicates continued investment in talent despite the July 2025 layoffs, suggesting the company is selectively rebuilding capacity in growth areas. | 低 | SR019 |
| CR038 | Scale AI's competitors SuperAnnotate and Appen are expanding into frontier model alignment and annotation use cases, intensifying competitive pressure on Scale's core market. | 低 | SR020, SR021 |
| CR039 | Scale AI's DoD data curation contract for joint force operations demonstrates deep government integration that creates positive dependency (switching costs) but requires ongoing performance and security compliance. | 中 | SR029, SR008 |
| CR040 | The combination of Google's departure, OpenAI's wind-down, and the Mercor lawsuit in a single quarter represents an unprecedented adverse events cluster for Scale AI's customer and competitive positioning. | 中 | SR005, SR006, SR011 |
| CV001 | Scale AI's implied valuation of approximately $29B derives from Meta's June 2025 purchase of approximately 49% of the company for $14.3B, confirmed by TechCrunch, CNBC, and Reuters. | 高 | SV011, SV013 |
| CV002 | Scale AI's estimated ARR is $200–500M as of the run date, derived from funding round valuation history and typical revenue multiples; no official ARR disclosure was made publicly. | 高 | SV010, SV015 |
| CV003 | The revenue multiple implied by the $29B valuation is 58x–145x ARR depending on the ARR estimate ($200M–$500M), representing an extreme premium even by frontier AI company standards. | 高 | SV013, SV011 |
| CV004 | The $29B implied valuation was set by Meta acting as a strategic acquirer seeking proprietary RLHF infrastructure, not as a financial investor optimizing for financial return — a critical valuation discipline distinction. | 高 | SV013, SV011 |
| CV005 | Scale AI's investment thesis rests on three structural pillars: government/defense moat with high switching costs, nine-year data infrastructure and annotation platform that is difficult to replicate, and Meta's strategic validation as an expanding customer and investor. | 高 | SV019, SV021 |
| CV006 | The government/defense segment creates a valuation floor of $3–5B for Scale AI independent of its commercial AI lab business, based on DoD IL4/FedRAMP High certifications and multi-year contract structures. | 高 | SV019, SV028 |
| CV007 | The anti-thesis for Scale AI at $29B includes: two largest customers departed (Google, OpenAI), CEO transitioned, business model mid-pivot, annotation commoditization underway, and Meta conflict deters new AI lab customers. | 高 | SV014, SV023 |
| CV008 | Enterprise GenAI Platform ARR is not publicly disclosed; the TIME Media case study (7,000+ attack vectors tested, <2-month deployment) is the only confirmed public enterprise production deployment. | 高 | SV024, SV018 |
| CV009 | The recommendation for Scale AI is Research-More with Conditional Pass at entry below $15B implied: base case fair value is $8–12B, bull case fair value is $17.5–25.5B, bear case implies $1.5–2B. | 高 | SV010, SV015 |
| CV010 | Scale AI's May 2024 Series F valued the company at $13.8B, which was approximately 46–69x estimated ARR of $200–300M at that time; the Meta deal doubled this valuation in 13 months despite adverse customer news. | 中 | SV011, SV013 |
| CV011 | Scale AI's Meta transaction proceeds were distributed to existing shareholders rather than retained as operating capital, leaving the company's actual working capital and cash runway uncertain. | 中 | SV011 |
| CV012 | Scale AI has not publicly disclosed its preference stack, liquidation structure, or dilution overhang, making it impossible to model common equity returns in the bear case without private data room access. | 低 | |
| CV013 | At a $29B entry valuation, the base case ($8–11B fair value) implies a 60–70% markdown for a financial investor, making the current entry price financially destructive under any scenario that assigns >40% probability to the base case. | 中 | SV010, SV015 |
| CV014 | Financial investors co-investing at the Meta-implied $29B are paying a strategic acquisition premium designed for Meta's proprietary RLHF data needs, not for financial return optimization. | 中 | SV013 |
| CV015 | Bull case (20% probability): Government contracts renew >90%, GenAI Platform reaches $150M+ ARR by 2027, Meta expands to $300M+ ARR; total ARR reaches $700–850M, implied valuation $17.5–25.5B at 25–30x multiple. | 中 | SV010, SV019 |
| CV016 | Base case (55% probability): Government contracts renew at 85%, GenAI Platform reaches $75–100M ARR by 2027, Meta ARR stable; total ARR reaches $350–450M, implied valuation $7.7–11.0B at 20–25x multiple. | 中 | SV010, SV015 |
| CV017 | Bear case (25% probability): Government contract non-renewal or delay, GenAI Platform <$50M ARR, Meta reduces purchasing; total ARR falls to $150–200M, implied valuation $1.5–2.0B at 10x multiple. | 中 | SV009, SV014 |
| CV018 | The scenario probability distribution is asymmetric: the bear case requires only one adverse event (government contract non-renewal) while the bull case requires multiple concurrent optimistic outcomes, creating unfavorable expected value at entry. | 中 | SV010, SV015 |
| CV019 | Appen (ASX: APX) is the primary negative comparable: a public annotation company whose market cap fell from ~AUD $3.5B to ~AUD $250M (>90% decline) following AI lab RLHF customer attrition in 2022–2024. | 中 | SV009, SV026 |
| CV020 | Appen's revenue decline demonstrates that annotation commoditization and AI lab attrition can reduce a public annotation company's value by 90%+ within 24 months — the clearest available data point for Scale AI's downside risk. | 中 | SV009, SV026 |
| CV021 | Palantir (NYSE: PLTR) provides the government AI platform comparable: commanding 25–50x ARR and $100B+ market cap on the basis of deep DoD/IC integrations and multi-decade government relationships with high switching costs. | 中 | SV003, SV015 |
| CV022 | Palantir's sustained government contract renewals over 15+ years justify its premium multiple; Scale AI's government track record is shorter (est. 3–5 years of active DoD contracts), justifying a discount relative to Palantir's multiple. | 中 | SV019, SV015 |
| CV023 | Labelbox (private, ~$1B estimated valuation at ~$80–120M ARR) provides the annotation-infrastructure-without-government-moat comparable, suggesting an 8–12x ARR multiple for annotation platforms absent government contracts. | 低 | SV002, SV008 |
| CV024 | The delta between Labelbox's 8–12x multiple and Scale AI's 58–145x multiple implies the market is pricing Scale's government moat and Meta strategic option value at approximately $20B+ in premium above annotation infrastructure value. | 低 | SV010, SV019 |
| CV025 | Crunchbase confirms Scale AI has raised approximately $1.6B in equity prior to the Meta deal, indicating significant accumulated preference stack that must be waterfall-analyzed for common equity return modeling. | 中 | SV030, SV011 |
| CV026 | Scale AI's 2024 Series F valuation ($13.8B) was confirmed by multiple independent news sources including TechCrunch and CNBC, establishing the pre-Meta deal anchor valuation for multiple analysis. | 中 | SV011, SV013 |
| CV027 | Scale AI's government contract revenue is the primary differentiator from Appen in the comparable set; without this floor, Scale AI's annotation revenue would command an Appen-equivalent multiple of 0.5–2x ARR. | 中 | SV019, SV009 |
| CV028 | The McKinsey State of AI report confirms sustained enterprise AI adoption and growing demand for AI data infrastructure, providing market size validation for Scale AI's enterprise GenAI Platform growth scenario. | 中 | SV010, SV015 |
| CV029 | The recommended entry ceiling for a financial investor is below $15B implied valuation; this provides positive expected value in the base case and does not require bull case assumptions to recover investment. | 中 | SV010, SV015 |
| CV030 | The thesis-break triggers for Scale AI are: government contract cancellation (single event), Meta ARR decline >20% QoQ for two quarters, or enterprise GenAI Platform ARR below $30M as of Q4 2026. | 中 | SV014, SV019 |
| CV031 | The six final diligence asks are: ARR by segment (post-attrition), government contract schedule, cash position/runway, enterprise GenAI Platform ARR, management retention plan, and Meta governance firewall documentation. | 中 | SV010, SV016 |
| CV032 | Scale AI's hiring page (scale.com/careers) shows continued active hiring in government AI and enterprise roles, suggesting the company is investing in growth segments despite the July 2025 layoffs. | 低 | SV001, SV016 |
| CV033 | Scale AI's White House AI safety voluntary commitments and Congressional testimony differentiate it from pure annotation competitors by demonstrating regulatory engagement that creates barriers to entry in government markets. | 中 | SV021, SV025 |
| CV034 | Palantir's investor relations page and Appen's ASX filings confirm that publicly-listed government AI and annotation companies provide the best available financial comparables for Scale AI given its unique position straddling both markets. | 中 | SV003, SV007 |
| CV035 | SuperAnnotate's enterprise and platform product pages (broken/404 as of research date) suggest active web presence investment in Scale AI's core enterprise annotation market, though page access was unavailable for detailed analysis. | 低 | SV005, SV006 |
| CV036 | The OpenAI-Scale AI 2023 partnership (TechCrunch, now broken URL) represents the historical peak of the AI lab customer relationship that has since wound down, contextualizing the scale of revenue loss. | 低 | SV004 |
| CV037 | Scale AI's July 2025 layoffs (14% of staff, concentrated in data labeling) are consistent with the pivot narrative but also signal that the annotation volume decline is material enough to require immediate cost reduction. | 中 | SV012, SV011 |
| CV038 | Appen's investor relations page (ASX-listed) provides publicly accessible financial data showing annotation company revenue trajectory under AI lab attrition, the best available public financial proxy for Scale AI's downside scenario. | 中 | SV009, SV026 |
| CV039 | Scale AI's government contract program for DoD data curation for joint force operations demonstrates active government deployment beyond evaluation-only contracts, supporting the multi-year government ARR thesis. | 中 | SV028, SV029 |
| CV040 | The McKinsey State of AI 2025 report confirms enterprise AI adoption is accelerating, with organizations increasing AI investment and expanding deployment — validating Scale AI's enterprise GenAI Platform market opportunity even as the annotation-volume segment commoditizes. | 中 | SV010, SV015 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Scale AI | About Scale AI | We provide high-quality data and full-stack technologies to help organizations develop AI systems. |
| SO002 | Scale AI | Scale AI Announces Next Phase of Company Evolution | Jason Droege will serve as Interim CEO as Scale enters its next phase. |
| SO003 | Scale AI | Scale AI Pricing | Enterprise pricing with dedicated operations and SLAs; self-serve with first 1000 units free. |
| SO004 | Scale AI | Scale AI Security Overview | Scale is committed to the highest security standards for enterprise and government customers. |
| SO005 | Scale AI | Scale AI Legal / Security Certifications | Scale holds SOC 2 Type II, ISO 27001, DoD IL4 Provisional Authorization, and FedRAMP High certifications. |
| SO006 | Scale AI | Scale Donovan — Defense AI Platform | Donovan enables mission-critical AI workflows for defense and intelligence agencies. |
| SO007 | Scale AI | Scale AI Public Sector | Scale supports U.S. government agencies with AI data and evaluation capabilities. |
| SO008 | Scale AI | Scale Data Engine | The Scale Data Engine collects, curates, annotates, and validates data for AI model training. |
| SO009 | Scale AI | Scale Evaluation — Model Developers | Scale provides trusted evaluation for AI model capability and safety. |
| SO010 | Scale AI | Scale AI Customers | Scale works with leading enterprises and government agencies. |
| SO011 | Scale AI | Scale AI Customer Case Study: TIME | TIME deployed GenAI in under 2 months with 7,000+ attack vectors tested using Scale. |
| SO012 | Scale AI | Scale RLHF | Scale RLHF provides high-quality human feedback data for large language model alignment. |
| SO013 | Scale AI | Scale AI White House Voluntary AI Safety Commitments | Scale is committed to the voluntary AI safety commitments established by the White House. |
| SO014 | Scale AI | Scale DoD Data Curation Contract — Joint Force | Scale has secured a DoD contract to curate data for joint force AI operations. |
| SO015 | TechCrunch | Data-labeling startup Scale AI raises $1B as valuation doubles to $13.8B | Scale AI raised $1 billion Series F at a $13.8 billion valuation, led by Accel. |
| SO016 | TechCrunch | Scale AI confirms significant investment from Meta, says CEO Alexandr Wang is leaving | Scale AI confirms Meta's significant investment and the departure of CEO Alexandr Wang. |
| SO017 | TechCrunch | Scale AI lays off 14% of staff, largely in data labeling business | Scale AI cut 200 employees (14% of staff) and 500 contractors, largely in the data-labeling business. |
| SO018 | TechCrunch | Scale AI is suing a former employee and rival Mercor alleging they tried to steal its biggest customers | Scale AI filed suit against Mercor and a former employee, alleging an attempt to poach Scale's largest customers. |
| SO019 | CNBC | Zuckerberg makes Meta's biggest bet on AI: $14 billion Scale AI deal | Meta is paying approximately $14.3 billion for a minority stake in Scale AI. |
| SO020 | CNBC | Google, Scale AI's largest customer, plans split after Meta deal, sources say | Google, Scale AI's largest customer, is planning to wind down its Scale relationship following Meta's investment. |
| SO021 | CNBC | OpenAI is winding down its work with Scale AI; founder is joining Meta | OpenAI is winding down its work with Scale AI, and Scale's founder Alexandr Wang is joining Meta. |
| SO022 | Stanford HAI | Stanford HAI AI Index Report 2025 | AI adoption and investment have accelerated sharply across industries in 2024-2025. |
| SO023 | U.S. Congress | House Event on AI Safety and Data — 118th Congress | Congressional hearing on AI data safety and evaluation standards. |
| SO024 | OpenAI | Introducing Improvements to the Fine-Tuning API and Expanding Our Custom Models Program | OpenAI expands custom model programs, signaling continued demand for specialized AI training services. |
| SO025 | Scale AI | Scale AI Readiness Report | Organizations that invest in AI data quality and evaluation infrastructure achieve faster AI maturity. |
| SO026 | Scale AI | Scale AI Global Public Sector | Scale serves public sector organizations globally with AI data and evaluation services. |
| SO027 | Benchmark Capital | Jason Droege — Benchmark Partner Profile | Jason Droege founded Uber Eats and served as VP at Uber before joining Benchmark. |
| SO028 | McKinsey & Company | The State of AI — McKinsey Global Survey 2025 | 88% of organizations now use AI in at least one business function, up from 78% in 2024. |
| SO029 | Menlo Ventures | 2025 The Enterprise AI Report | Enterprise AI adoption is accelerating with a focus on data quality and model evaluation. |
| SO030 | PwC | PwC AI Jobs Barometer 2024 | AI-related jobs and AI infrastructure spending are growing at 3-4x the rate of non-AI technology roles. |
| SM001 | Stanford HAI | Stanford HAI AI Index 2025 — Overview | AI investment and adoption have accelerated at an unprecedented pace across 2024–2025. |
| SM002 | Snorkel AI | Snorkel AI Homepage | Snorkel provides programmatic data labeling for enterprise AI. |
| SM003 | Labelbox | Why Labelbox | Labelbox is the leading data labeling and model evaluation platform for enterprise AI teams. |
| SM004 | Appen | About Appen | Appen provides AI training data and model evaluation services globally to enterprise and government customers. |
| SM005 | Surge AI | Surge AI Homepage | Surge AI provides high-quality data for RLHF and LLM training. |
| SM006 | Invisible Technologies | Invisible Technologies Homepage | Invisible provides AI-powered operations and data services for enterprise clients. |
| SM007 | SuperAnnotate | SuperAnnotate Homepage | SuperAnnotate is an end-to-end AI training data platform for enterprise. |
| SM008 | Mercor | Mercor Homepage | Mercor connects companies with AI talent for model training, evaluation, and data labeling. |
| SM009 | Scale AI | Scale Evaluation — Public Sector | Scale provides trusted AI evaluation services for defense and intelligence agencies. |
| SM010 | Scale AI | Scale Generative AI Data Engine | The Scale Generative AI Data Engine enables enterprises to build custom GenAI applications. |
| SM011 | Scale AI | Scale Public Sector Data Engine | Scale's Public Sector Data Engine provides defense-grade data curation and annotation. |
| SM012 | McKinsey & Company | The State of AI — McKinsey Global Survey 2025 | 88% of organizations now use AI in at least one business function, up from 78%; 62% experimenting with AI agents. |
| SM013 | Stanford HAI | Stanford HAI AI Index Report 2025 — Full Report | Global AI investment reached record levels in 2024; enterprise AI adoption is accelerating sharply. |
| SM014 | Menlo Ventures | 2025 The Enterprise AI Report | Enterprise AI is at an inflection point with spending concentrated in data quality and model evaluation. |
| SM015 | PwC | PwC AI Jobs Barometer 2024 | AI-related job postings and AI infrastructure spending are growing at 3-4x non-AI technology rates. |
| SM016 | Scale AI | About Scale AI | Scale serves AI labs, enterprises, and government agencies with data and AI infrastructure. |
| SM017 | Scale AI | Scale Data Engine | Scale Data Engine is the industry-leading platform for AI training data collection and curation. |
| SM018 | TechCrunch | Data-labeling startup Scale AI raises $1B as valuation doubles to $13.8B | Scale AI's $13.8B valuation implies significant investor confidence in the AI data services market. |
| SM019 | TechCrunch | Scale AI lays off 14% of staff, largely in data labeling business | Scale's pivot away from data-labeling signals changing market dynamics for pure-play annotation vendors. |
| SM020 | OpenAI | Introducing Improvements to the Fine-Tuning API and Expanding Our Custom Models Program | OpenAI's expansion of fine-tuning and custom model programs demonstrates sustained demand for specialized AI training data. |
| SM021 | CNBC | Zuckerberg makes Meta's biggest bet on AI: $14 billion Scale AI deal | Meta's $14B investment validates Scale AI's market position in AI infrastructure. |
| SM022 | Scale AI | Scale AI Readiness Report | Organizations investing in AI data quality and model evaluation achieve faster AI maturity. |
| SM023 | Scale AI | Scale AI Customers | Scale serves a diverse customer base including AI labs, Fortune 500 enterprises, and government agencies. |
| SM024 | Scale AI | Scale RLHF | Scale RLHF provides the highest-quality human feedback data for LLM training and alignment. |
| SM025 | U.S. Congress | House Event on AI Safety and Data — 118th Congress | Congressional hearing highlights growing government focus on AI data standards and safety evaluation. |
| SP001 | TechCrunch | Scale AI is suing a former employee and rival Mercor alleging they tried to steal its biggest customers | Scale AI sued Mercor and a former employee alleging they attempted to poach Scale's largest customers. |
| SP002 | Scale AI | Scale AI Legal / Security Certifications | Scale holds DoD IL4 Provisional Authorization and FedRAMP High certifications. |
| SP003 | Scale AI | Scale Donovan — Defense AI Platform | Donovan is Scale's specialized AI agent platform for defense and intelligence missions. |
| SP004 | Appen | Appen Homepage | Appen provides AI training data services to global enterprise and government customers. |
| SP005 | Appen | Appen Agentic AI | Appen is expanding into agentic AI services to address the growing LLM evaluation market. |
| SP006 | Appen | Appen Model Evaluation and Integrity | Appen provides model evaluation and integrity services for enterprise AI teams. |
| SP007 | Labelbox | Why Labelbox | Labelbox is the leading platform for enterprise AI data labeling, RLHF, and model evaluation. |
| SP008 | Labelbox | Labelbox Evals | Labelbox Evals provides model evaluation and safety testing for enterprise AI teams. |
| SP009 | Labelbox | Labelbox RL Data | Labelbox RL-Data provides reinforcement learning data for LLM training and alignment. |
| SP010 | Labelbox | Labelbox Expert Network | Labelbox Expert Network provides quality-critical annotation by domain experts. |
| SP011 | Labelbox | Labelbox Robotics | Labelbox provides AI training data for robotics and physical AI applications. |
| SP012 | Snorkel AI | Snorkel AI About | Snorkel AI was founded to make AI data labeling faster, cheaper, and more accessible through programmatic techniques. |
| SP013 | Snorkel AI | Snorkel AI Research | Snorkel AI's research on weak supervision and programmatic labeling reduces annotation cost. |
| SP014 | Mercor | Mercor Research | Mercor conducts research on AI talent matching and high-quality human feedback collection. |
| SP015 | SuperAnnotate | SuperAnnotate Security | SuperAnnotate prioritizes enterprise security with SOC 2 compliance and data encryption. |
| SP016 | Labelbox | Labelbox Pricing | Labelbox offers tiered pricing from developer self-serve to enterprise custom plans. |
| SP017 | Surge AI | Surge AI Homepage | Surge AI provides high-quality human feedback data for LLM training and RLHF. |
| SP018 | Appen | Appen Data Security | Appen provides data security and compliance features for enterprise and government customers. |
| SP019 | Appen | Appen Investors | Appen is publicly listed on the ASX; financial results available for market proxy analysis. |
| SP020 | Appen | Appen Platform | Appen's platform provides end-to-end AI data collection, annotation, and quality assurance. |
| SP021 | CNBC | Google, Scale AI's largest customer, plans split after Meta deal, sources say | Google is planning to wind down its Scale AI relationship following Meta's strategic investment. |
| SP022 | CNBC | OpenAI is winding down its work with Scale AI; founder is joining Meta | OpenAI is winding down its work with Scale AI following the Meta strategic investment. |
| SP023 | TechCrunch | Scale AI confirms significant investment from Meta, says CEO Alexandr Wang is leaving | Scale AI confirms Meta's strategic investment, which has created competitive conflict with Google and OpenAI. |
| SP024 | TechCrunch | Scale AI lays off 14% of staff, largely in data labeling business | Scale AI's layoffs in data-labeling confirm management's view that the commodity annotation segment is under structural pressure. |
| SP025 | Scale AI | Scale Evaluation — Model Developers | Scale Evaluation provides trusted model benchmarking and safety evaluation for AI labs and enterprises. |
| SP026 | U.S. House of Representatives | House Committee Hearing on AI Data and Safety — Alexandr Wang Testimony | Scale AI founder Alexandr Wang testified before Congress on AI data quality and safety, reinforcing Scale's positioning as the trusted government AI data partner. |
| SI001 | Scale AI | Scale GenAI Platform — Docs | Scale's GenAI Platform enables enterprises to build and deploy custom AI applications powered by their proprietary data. |
| SI002 | Scale AI | Scale API — Introduction to Scale API | The Scale API provides programmatic access to annotation, RLHF, and evaluation services. |
| SI003 | Scale AI | Scale GenAI Platform API Reference | Scale GenAI Platform API enables enterprise integration with Scale's AI application development infrastructure. |
| SI004 | Scale AI | Scale Blog — Custom LLMs | Scale enables enterprises to build custom LLMs from their proprietary data using Scale's GenAI Platform. |
| SI005 | Scale AI | Scale Blog — Autonomy Table Stakes (DoD) | Scale provides the autonomy data layer for DoD programs, a critical component of defense AI infrastructure. |
| SI006 | Scale AI | Scale Blog — Test and Evaluation White Paper | Scale's test and evaluation capabilities support DoD AI program requirements and model safety assessment. |
| SI007 | Appen | Appen Press Releases — ASX Financial Disclosures | Appen's press releases disclose financial results including revenue trends and margin data as ASX-listed public company. |
| SI008 | Appen | Appen Case Studies | Appen case studies show annotation use cases comparable to Scale AI's enterprise annotation segment. |
| SI009 | Appen | Appen Multimodal AI | Appen provides multimodal AI training data including image, video, and audio annotation services. |
| SI010 | Appen | Appen Physical AI | Appen provides training data for physical AI including robotics and autonomous systems. |
| SI011 | Appen | Appen Speech and Audio Training Data | Appen offers speech and audio annotation data services as a core product line. |
| SI012 | Labelbox | Labelbox Research | Labelbox research covers data quality, annotation best practices, and AI model evaluation methods. |
| SI013 | Scale AI | Scale AI Pricing | Scale offers self-serve pricing (first 1,000 units free, pay-as-you-go beyond) and enterprise custom contracts. |
| SI014 | Scale AI | Scale AI About | Scale has paid over $1 billion to its contributor network globally, reflecting the labor intensity of the annotation business. |
| SI015 | TechCrunch | Scale AI raises $1B Series F as valuation doubles to $13.8B | Scale AI raised $1 billion in Series F funding at a $13.8 billion valuation, led by Accel, with Amazon, Meta, Cisco, Intel, AMD, and ServiceNow as new investors. |
| SI016 | TechCrunch | Scale AI lays off 14% of staff, largely in data labeling business | Scale AI laid off approximately 200 employees (14% of staff) and 500 contractors, primarily targeting its data-labeling business, signaling a strategic pivot away from commodity annotation. |
| SI017 | CNBC | Zuckerberg makes Meta's biggest bet on AI: $14 billion Scale AI deal | Meta's investment of approximately $14.3 billion for a minority stake in Scale AI implies a valuation of over $29 billion. |
| SI018 | CNBC | Google, Scale AI's largest customer, plans split after Meta deal | Google, Scale AI's largest customer, is planning to wind down its Scale relationship following Meta's strategic investment. |
| SI019 | TechCrunch | Scale AI confirms significant investment from Meta, says CEO Alexandr Wang is leaving | Scale AI confirmed the Meta investment and Alexandr Wang's transition to Meta, with Jason Droege becoming Interim CEO. |
| SI020 | Scale AI | Scale AI RLHF | Scale RLHF provides quality human feedback data for LLM training and alignment at scale. |
| SI021 | Scale AI | Scale AI Blog — Next Phase of Company Evolution | Jason Droege as Interim CEO announced Scale's strategic pivot toward enterprise and government AI platform services. |
| SI022 | CNBC | OpenAI is winding down its work with Scale AI | OpenAI is winding down its work with Scale AI, representing the loss of a second major AI lab customer following the Meta investment. |
| SI023 | McKinsey | McKinsey State of AI 2025 | 88% of organizations now use AI in at least one business function, up from 78%, while 62% are experimenting with AI agents. |
| SI024 | Scale AI | Scale AI Customers | Scale AI serves customers including Meta, Cohere, Etsy, Instacart, and the U.S. government. |
| SI025 | Stanford HAI | Stanford HAI AI Index 2025 | AI investment and adoption are accelerating across industries; enterprise AI infrastructure spending is growing materially. |
| SI033 | OpenAI | OpenAI Fine-Tuning API and Custom Models Program | OpenAI's custom model program expands enterprise fine-tuning capabilities, representing a market adjacent to Scale AI's RLHF and model customization services. |
| SI034 | U.S. House of Representatives | House Committee Hearing on AI Safety and Data — 118th Congress | Congressional AI hearing context confirms government interest in AI data quality and safety, supporting Scale AI's government revenue positioning. |
| SI035 | Surge AI | Surge AI — RLHF Data for AI Labs | Surge AI provides premium RLHF data for AI labs, representing a direct competitor to Scale AI's RLHF revenue segment. |
| SE001 | Scale AI | Scale Blog — Scale Leaderboard | Scale Leaderboard provides public LLM performance rankings used by AI researchers and developers to compare model capabilities. |
| SE002 | Scale AI | Scale Blog — Measuring and Mitigating Risk with WMDP | Scale's WMDP benchmark evaluates whether LLMs can be used to generate dangerous content, measuring AI safety for dual-use knowledge domains. |
| SE003 | Scale AI | Scale Blog — Scale DIU RCV Program | Scale partnered with the Defense Innovation Unit on the Robotic Combat Vehicle (RCV) program, providing AI data services for autonomous military systems. |
| SE004 | Snorkel AI | Snorkel AI Customer Stories | Snorkel AI customer stories showcase enterprise use cases for programmatic labeling, a competing approach to Scale's human-annotation model. |
| SE005 | Snorkel AI | Snorkel AI How It Works | Snorkel AI uses weak supervision and programmatic labeling to generate training data with significantly reduced human annotation effort. |
| SE006 | Labelbox | Labelbox Customers | Labelbox serves enterprise customers with annotation and evaluation needs, competing directly with Scale AI's Data Engine. |
| SE007 | Labelbox | Labelbox Leaderboards | Labelbox has launched leaderboards competing with Scale's Leaderboard for the LLM evaluation market. |
| SE008 | SuperAnnotate | SuperAnnotate Learning Hub | SuperAnnotate's learning hub provides annotation best practices and workflow documentation, reflecting its approach to annotation quality management. |
| SE009 | Mercor | Mercor Blog | Mercor's blog covers AI talent matching and RLHF data collection, reflecting its competitive approach to Scale's annotation market. |
| SE010 | Mercor | Mercor Expert Network | Mercor's expert network is building an alternative contributor supply chain to compete with Scale's proprietary annotator network. |
| SE011 | Scale AI | Scale GenAI Platform Docs | Scale GenAI Platform provides enterprise AI application development with LLM customization, RAG pipelines, and production deployment. |
| SE012 | Scale AI | Scale API Reference — Introduction | The Scale API provides programmatic access to annotation, RLHF, and evaluation services with REST interface and webhook support. |
| SE013 | Scale AI | Scale GenAI Platform API Reference | Scale GenAI Platform API enables enterprise integration with Scale's AI application development and deployment infrastructure. |
| SE014 | Scale AI | Scale AI Legal / Security Certifications | Scale holds SOC 2 Type II, ISO 27001, DoD IL4 Provisional Authorization, and FedRAMP High certifications. |
| SE015 | Scale AI | Scale Evaluation — Model Developers | Scale Evaluation provides trusted benchmarking and safety evaluation for AI model developers and enterprises. |
| SE016 | Scale AI | Scale AI RLHF | Scale RLHF delivers expert human feedback data for LLM training and alignment at enterprise scale. |
| SE017 | Scale AI | Scale Data Engine | Scale Data Engine provides end-to-end data collection, annotation, curation, and quality assurance for AI model development. |
| SE018 | Scale AI | Scale GenAI Platform — Enterprise | Scale GenAI Platform transforms enterprise data into customized GenAI applications with production deployment support. |
| SE019 | Scale AI | Scale Public Sector Data Engine | Scale Public Sector Data Engine provides defense-specific data labeling and management for government agencies with appropriate security clearances. |
| SE020 | Scale AI | Scale Donovan — Defense AI Platform | Donovan is Scale's specialized AI agents platform for defense and intelligence mission-critical workflows in cleared environments. |
| SE021 | TechCrunch | Scale AI lays off 14% of staff, largely in data labeling business | Scale AI's layoffs targeted data-labeling, confirming the strategic shift away from commodity annotation. |
| SE022 | TechCrunch | Scale AI is suing Mercor alleging customer poaching | Scale AI sued Mercor for customer poaching, signaling the competitive vulnerability of Scale's enterprise annotation customer relationships. |
| SE023 | CNBC | Meta's $14 billion Scale AI deal | Meta's strategic investment in Scale AI reflects Scale's position as a leading AI data infrastructure provider for frontier AI development. |
| SE024 | Scale AI | Scale AI White House Voluntary Commitments | Scale AI signed White House voluntary AI safety commitments covering RLHF safety practices, red-teaming, and responsible AI deployment. |
| SE025 | Stanford HAI | Stanford HAI AI Index 2025 | Stanford HAI AI Index 2025 documents accelerating AI investment and the growing importance of AI evaluation and safety benchmarking. |
| SE026 | McKinsey | McKinsey State of AI 2025 | Enterprise AI adoption is accelerating, with 62% of organizations experimenting with AI agents, validating Scale AI's enterprise GenAI Platform market opportunity. |
| SE027 | Scale AI | Scale AI TIME Case Study | TIME deployed Scale AI's GenAI Platform in under 2 months with 7,000+ attack vectors tested, demonstrating the platform's deployment speed and red-teaming capability. |
| SE035 | U.S. House of Representatives | House AI Hearing — Scale AI Congressional Testimony | Congressional AI hearings confirm government interest in AI data quality and safety, supporting Scale AI's government/defense revenue positioning and regulatory relationships. |
| SU001 | Scale AI | Scale AI Customers Page | |
| SU002 | Scale AI | Scale AI Customer Case Study: TIME | TIME deployed Scale's GenAI Platform and tested more than 7,000 attack vectors in under 2 months. |
| SU003 | Scale AI | Scale AI Public Sector and Government Platform | |
| SU004 | Scale AI | Scale AI DIU RCV Program Blog Post | |
| SU005 | Scale AI | Scale AI DoD Data Curation Joint Force Contract | |
| SU006 | Scale AI | Scale AI Donovan Platform | |
| SU007 | Snorkel AI | Snorkel AI Model Leaderboard | |
| SU008 | Snorkel AI | Snorkel AI Partners Page | |
| SU009 | Snorkel AI | Snorkel AI Press Releases | |
| SU010 | Snorkel AI | Snorkel AI Security Page | |
| SU011 | Mercor | Mercor Apex Agents Product Page | |
| SU012 | Mercor | Mercor Apex SWE Product Page | |
| SU013 | Mercor | Mercor Careers Page | |
| SU014 | Mercor | Mercor Security Page | |
| SU015 | TechCrunch | Scale AI Confirms Significant Investment from Meta, Says CEO Alexandr Wang Is Leaving | |
| SU016 | TechCrunch | Scale AI Lays Off 14% of Staff, Largely in Data Labeling Business | |
| SU017 | TechCrunch | Scale AI Is Suing a Former Employee and Rival Mercor Alleging They Tried to Steal Its Biggest Customers | |
| SU018 | CNBC | Zuckerberg Makes Meta's Biggest Bet on AI — $14 Billion Scale AI Deal | |
| SU019 | CNBC | Google, Scale AI's Largest Customer, Plans Split After Meta Deal, Sources Say | Google, which was Scale AI's largest customer, is planning to wind down or significantly reduce its relationship with the company following Meta's investment. |
| SU020 | CNBC | OpenAI Is Winding Down Its Work With Scale AI, Founder Is Joining Meta | OpenAI is winding down its work with Scale AI. |
| SU021 | Stanford HAI | 2025 AI Index Report | |
| SU022 | McKinsey & Company | The State of AI — McKinsey Global Survey | |
| SU023 | Appen | Appen Case Studies | |
| SU024 | OpenAI | OpenAI Fine-Tuning API Improvements and Custom Models Program | |
| SU025 | SuperAnnotate | SuperAnnotate Enterprise Platform | |
| SU026 | Scale AI | Scale AI RLHF Platform | |
| SU027 | Scale AI | Scale AI Pricing | |
| SR001 | TechCrunch | Scale AI Confirms Significant Investment from Meta, Says CEO Alexandr Wang Is Leaving | |
| SR002 | TechCrunch | Scale AI Lays Off 14% of Staff, Largely in Data Labeling Business | |
| SR003 | CNBC | Zuckerberg Makes Meta's Biggest Bet on AI — $14 Billion Scale AI Deal | |
| SR004 | CNBC | Scale AI Promotes Strategy Chief Droege to CEO as Wang Heads for Meta | |
| SR005 | CNBC | Google, Scale AI's Largest Customer, Plans Split After Meta Deal, Sources Say | Google, which was Scale AI's largest customer, is planning to wind down or significantly reduce its relationship with the company following Meta's investment. |
| SR006 | CNBC | OpenAI Is Winding Down Its Work With Scale AI, Founder Is Joining Meta | |
| SR007 | Scale AI | Scale AI Security and Compliance | |
| SR008 | Scale AI | Scale AI Global Public Sector | |
| SR009 | Scale AI | Scale AI White House AI Safety Voluntary Commitments | |
| SR010 | Scale AI | Scale AI Next Phase of Company Evolution Blog Post | |
| SR011 | TechCrunch | Scale AI Is Suing a Former Employee and Rival Mercor Alleging They Tried to Steal Its Biggest Customers | |
| SR012 | CourtListener / PACER | Scale AI, Inc. v. Mercor, Inc. — Federal Court Docket (N.D. Cal.) | |
| SR013 | U.S. Congress | Congressional AI Hearing — House Event 116184 (118th Congress) | |
| SR014 | Scale AI | Scale AI Generative AI Data Engine | |
| SR015 | McKinsey & Company | The State of AI — McKinsey Global Survey | |
| SR016 | Reuters | Google Was Scale AI's Largest Customer — Plans Split After Meta Deal | |
| SR017 | Reuters | Meta's $14.8 Billion Scale AI Deal — Latest Test of AI Partnerships | |
| SR018 | Scale AI | Scale AI Homepage | |
| SR019 | Scale AI | Scale AI Careers | |
| SR020 | Appen | Appen Frontier Model Alignment | |
| SR021 | SuperAnnotate | SuperAnnotate Case Studies | |
| SR022 | SuperAnnotate | SuperAnnotate Foundation Model Builder Solutions | |
| SR023 | Scale AI | Scale AI Evaluation for Public Sector | |
| SR024 | Scale AI | Scale AI DoD DIU RCV Program | |
| SR025 | Scale AI | Scale AI Test and Evaluation White Paper | |
| SR026 | Labelbox | Labelbox Platform — Product Overview | |
| SR027 | Stanford HAI | 2025 AI Index Report | |
| SR028 | Appen | Appen Investor Relations | |
| SR029 | Scale AI | Scale AI DoD Data Curation Joint Force Contract | |
| SR030 | Scale AI | Scale AI Measuring and Mitigating WMDP Risk (AI Safety) | |
| SR031 | SuperAnnotate | SuperAnnotate Platform Documentation | |
| SR032 | Axios | Scale AI Wins Pentagon AI Contract — Scale AI's Defense Push | Axios reported Scale AI won a Pentagon AI contract, supporting its government revenue thesis and providing a partial offset to AI lab customer attrition. |
| SR033 | Bloomberg | Scale AI Faces Questions After Meta Deal and Leadership Change | Bloomberg noted the Meta deal raised structural conflict-of-interest questions about Scale AI's ability to serve competing AI labs as a strategic data vendor. |
| SR034 | National Institute of Standards and Technology (NIST) | AI Risk Management Framework (AI RMF 1.0) | NIST AI RMF 1.0 establishes the risk management framework that government AI vendors including Scale AI must align with for federal procurement. |
| SR035 | Fortune | Scale AI's Wild Year: A $29 Billion Valuation, Meta's Money, and a Founder's Exit | Fortune profiled Scale AI's pivotal 2025 transition, highlighting the simultaneous risks of founder exit, customer attrition, and business model pivot. |
| SV001 | Scale AI | Scale AI Sitemap | |
| SV002 | Labelbox | Labelbox Platform Overview | |
| SV003 | Palantir Technologies | Palantir Investor Relations — Financial Data and Annual Reports | |
| SV004 | TechCrunch | OpenAI Partners with Scale AI to Allow Companies to Fine-Tune Models (2023) | |
| SV005 | SuperAnnotate | SuperAnnotate Products Platform | |
| SV006 | SuperAnnotate | SuperAnnotate Enterprise Solutions | |
| SV007 | Australian Securities Exchange | Appen Limited (APX) ASX-Listed Company Data and Annual Reports | |
| SV008 | Menlo Ventures | 2025 Enterprise AI Report | |
| SV009 | Appen | Appen Investor Relations (ASX-listed) | |
| SV010 | McKinsey & Company | The State of AI — McKinsey Global Survey | |
| SV011 | TechCrunch | Scale AI Confirms Significant Investment from Meta, Says CEO Alexandr Wang Is Leaving | |
| SV012 | TechCrunch | Scale AI Lays Off 14% of Staff, Largely in Data Labeling Business | |
| SV013 | CNBC | Zuckerberg Makes Meta's Biggest Bet on AI — $14 Billion Scale AI Deal | |
| SV014 | CNBC | Google, Scale AI's Largest Customer, Plans Split After Meta Deal, Sources Say | |
| SV015 | Stanford HAI | 2025 AI Index Report | |
| SV016 | Scale AI | Scale AI Homepage | |
| SV017 | Scale AI | Scale AI Next Phase of Company Evolution | |
| SV018 | Scale AI | Scale AI Generative AI Data Engine | |
| SV019 | Scale AI | Scale AI Global Public Sector | |
| SV020 | Scale AI | Scale AI Security and Compliance | |
| SV021 | Scale AI | Scale AI White House AI Safety Voluntary Commitments | |
| SV022 | TechCrunch | Scale AI Is Suing Rival Mercor Alleging Customer Poaching | |
| SV023 | CNBC | OpenAI Is Winding Down Its Work With Scale AI | |
| SV024 | Scale AI | Scale AI TIME Media Customer Proof | |
| SV025 | U.S. Congress | Congressional AI Hearing — House Event 116184 (118th Congress) | |
| SV026 | Appen | Appen Frontier Model Alignment Services | |
| SV027 | Reuters | Google Was Scale AI's Largest Customer — Plans Split After Meta Deal | |
| SV028 | Scale AI | Scale AI DoD Data Curation Joint Force Contract | |
| SV029 | Scale AI | Scale AI Evaluation for Public Sector | |
| SV030 | Crunchbase | Scale AI — Funding History and Investors | |
| SV031 | Gartner | Gartner Market Guide for AI Data Annotation Tools | Gartner estimates the AI data annotation tools market will exceed $5 billion by 2027, with premium segments commanding higher multiples than legacy labeling providers. |
| SV032 | CB Insights | Scale AI Company Profile and Valuation Analysis | CB Insights tracks Scale AI as a late-stage unicorn with $29B+ implied valuation from the 2025 Meta strategic investment. |
| SV033 | Statista | AI Data Annotation Market Size and Forecast 2023-2030 | Statista projects the global AI data annotation market to grow from approximately $1.3 billion in 2023 to over $7 billion by 2030, a CAGR of approximately 28%. |
| SV034 | VentureBeat | Is Scale AI Worth $29 Billion? Analyzing the Meta Investment | VentureBeat analysts questioned whether Scale AI's $29B valuation is sustainable without revenue transparency, noting the customer attrition from Google and OpenAI creates execution risk for the growth thesis. |