Twelve Labs
July 2026 $100M Series B 轮后的视频理解基础模型独角兽——战略可信,但财务不透明
Twelve Labs 的视频 AI 战略可信,但财务透明度不足;若按约 $1B 入场,更适合观察 / 继续研究,而不是买入。
封面要素
公司概况
Twelve Labs (TwelveLabs) 是一家总部位于旧金山的视频智能公司,成立于 2021,专注构建视频理解基础模型。其平台通过 Marengo 模型家族(多模态视频嵌入与检索)和 Pegasus 视频语言模型(视频转文本描述、摘要和结构化元数据)提供 Search、Analyze 和 Embed 工作流。公司通过自有 API、Amazon Bedrock 和 AWS Marketplace 分发模型,服务媒体与娱乐、广告、体育、安全、政府和汽车等场景。July 2026,公司完成由 NEA 和 NAVER Ventures 共同领投的 $100M Series B 轮,累计融资约 $207M。
- 成立时间
- 2021-03-31
- 创始人
- Jae Lee, Aiden Lee
- 创立地点
- San Francisco, California, USA (with founding-team roots in Korea)
- 总部
- San Francisco, California, USA (additional operations in Seoul; New York and London expansion in 2026)
- 产品
- 视频理解基础模型和 API——Marengo 3.0 用于多模态搜索和 512 维嵌入,Pegasus 1.5 用于 schema 优先的视频转文本和基于时间的元数据,另有 Embed API;通过自助 API、Amazon Bedrock 和 AWS Marketplace 销售。
- 客户
- 需要搜索、分析大规模视频档案并让其进入业务流程的媒体与娱乐、广告、体育、安全 / 监控、政府和汽车企业及开发者。
- 商业模式
- 按用量变现 API(索引 / 推理)叠加企业计划,通过直销以及 AWS Marketplace / Amazon Bedrock 分发。
- 阶段
- Series B
- 融资情况
- $100M Series B 轮于 July 1, 2026 宣布(NEA 和 NAVER Ventures 共同领投);从种子轮、Series A($50M,2024)到 Series B,累计融资约 $207M。
执行摘要
主要优势
- 聚焦视频专用基础模型,定位有差异:Marengo 3.0 做搜索 / embeddings,Pegasus 1.5 做视频转文本;多数对手仍把视频当作通用多模态能力的附加项。
- 战略和财务股东质量高,包括 NEA、NAVER Ventures、Amazon、Radical Ventures、Index Ventures、Korea Investment Partners、Quadrille Capital 和 Red Bull Ventures。
- AWS 在多年 Trainium 承诺下成为首选云服务商,再叠加 Amazon Bedrock 和 AWS Marketplace 分发,有机会改善推理经济性并触达企业客户。
- 所处市场规模大且增长快,覆盖视频分析、生成式和多模态 AI、视频监控、体育分析;多个测算口径都能支撑空间。
主要风险
- ARR、收入运行率、毛利率、烧钱速度、现金跑道和投后估值都未公开;假定约 $1B 的价格,无法靠公开证据完成承销。
- 大厂多模态基础模型会带来商品化压力:Google Gemini、OpenAI、Microsoft / Amazon 原生视频服务都能把视频理解打包进去。
- 监管和法律暴露不低:EU AI Act 高风险 / 生物识别监控分类、GDPR 和生物识别隐私法,以及生成式 AI 训练数据 / 版权诉讼都会压上来。
- 集中度和关键人风险并存:公司依赖 CEO Jae Lee,也依赖 AWS / NVIDIA / NAVER 的伙伴和云关系。
未决问题
- 收入、ARR、毛利率、净烧钱速度和现金跑道均未披露,阻断承销判断。
- 投后估值、股权结构表、清算优先权和期权池条款均未公开。
- 客户集中度,以及已点名部署中生产环境与试点的比例仍不清楚。
- 净收入留存、总留存和流失指标均不可得。
目录
01公司概况
1.1 身份、运营足迹与产品模式
应把 TwelveLabs 视为一家总部位于旧金山的视频智能基础模型公司,而不是普通媒体搜索厂商。目前最可靠的一组来源显示,公司让机器通过统一感知、知识和推理的架构来感知、理解并推理视频。产品边界很具体:Search 和 Embed 以 Marengo 为核心,Analyze 和结构化视频转文本工作流以 Pegasus 为核心。公司还通过自有 API、AWS Marketplace 和 Amazon Bedrock 提供接入路径。地理证据异常新:招聘页列出 San Francisco、Seoul、New York、London 和 Pangyo 办公室,Series B 通稿则称总部在 San Francisco,运营覆盖 Seoul、New York、Los Angeles 和 London。这为后续章节提供了可用的身份基线,但仍需谨慎看待具体办公室成熟度和岗位分布。[CO001, CO002, CO003, CO009, CO010, CO011]
| 指标 | 数值 / 状态 | 时间背景 | 置信度 | 缺口 |
|---|---|---|---|---|
| 成立 | 2021;Tracxn 显示 2021 年 3 月 31 日注册 | 历史 | 高 | |
| 总部 / 基地 | 总部在旧金山,并在首尔、纽约、洛杉矶、伦敦运营 | 2026-07 | 高 | 各办公室按职能分工未公开 |
| 当前阶段 | 未上市 Series B 轮阶段 | 2026-07-01 | 高 | |
| 最新融资 | $100M Series B 轮,由 NEA 和 NAVER Ventures 共同领投 | 2026-07-01 | 高 | |
| 累计融资 | 当前市场数据来源显示约 $207M 至 $207.1M | 2026-07 | 高 | Crunchbase 档案仍滞后在 $107.1M |
| 估值 | 已抓取官方 / 通讯社 Series B 轮来源未披露 | 2026-07 | 中 | 需要原始股权结构表、投资条款清单或可靠估值报道 |
| 员工数 | 约 200 人,分布在首尔和旧金山 | 2026-07-06 | 高 | 按岗位拆分的组织图未公开 |
| 收入 / ARR / 毛利率 | 已保留来源未公开披露 | 2026-07-21 | 中 | 需要管理层数据室或客户 / 收入尽调 |
| 客户 / 用户证据 | 2024 年 30,000 名用户;点名客户包括 MLSE、AMC Global Media、UNICEF | 2024-2026 | 中 | 当前付费客户数未披露 |
| 办公室版图 | 招聘页列出旧金山、首尔、纽约、伦敦、Pangyo | 2026-07-21 | 高 | 各办公室开设日期未单独记录 |
各行混合使用官方页面、融资新闻稿和市场数据来源;私有财务指标和估值仍是明确尽调缺口。
[CO001, CO002, CO003, CO004, CO017, CO020]Twelve Labs 如何把身份、模型、基础设施、资本、客户和尽调缺口串成一张公司快照。
流程图是概念性的;除来源支持的公司逻辑外,不暗示技术依赖顺序。
[CO001, CO009, CO010, CO011, CO014, CO015]来源支撑的成熟度和风险指标,用于公司概览。
KPI 值保留公开证据缺口,不推测私下财务或估值。
[CO017, CO020, CO021, CO028, CO029, CO030]1.2 创始人、管理梯队与治理可见度
领导层画像是创始人主导,并高度围绕 Jae Lee 展开。多家独立和合作伙伴来源都将 Jae Lee 认定为联合创始人兼 CEO;NEA 的访谈尤其有用,因为 Lee 把创业起点连接到 2021 与四位 Korean Cyber Command 密友的合作。CB Insights 给出最完整的创始人枚举,列名 Aiden Lee、Dave Chung、Jae Lee、SJ Kim 和 Soyoung Lee,但没有逐一给出角色细节。因此,公开治理证据足以支撑概览,却不足以评估董事会控制权。公司公布的顾问阵容包括 Fei-Fei Li、Silvio Savarese、Jeffrey Katzenberg、Alex Wang、Lukas Biewald、Nicolas Dessaigne 和 Jay Simons;WEF 还补充了 Jae 在 Korea Foundation Model Association 的董事会角色。核心尽调风险是关键人依赖:融资、投资人引述和产品叙事都持续经由 Jae Lee。[CO005, CO006, CO007, CO008, CO037, CO038]
| 人物 | 职务 / 公开状态 | 背景证据 | 职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Jae Lee | 联合创始人兼 CEO | TechCrunch 将其定位为数据科学家;WEF 档案称其拥有 UC Berkeley EECS 背景,并在 Korea Foundation Model Association 任董事 | 创始人愿景、融资叙事、产品逻辑 | 高:投资人、产品和公司引述反复以他为中心 |
| Yoon Kim | 总裁兼首席战略官,2024 年报道 | 据 TechCrunch,曾任 SK Telecom CTO 和 Siri 架构师 | 战略与扩张领导 | 中:公开来源是一篇 2024 年任命报道 |
| Aiden Lee | CB Insights 点名的创始人 | 第三方创始人名单点名;已保留来源未公开角色细节 | 创始人群体 / 早期技术覆盖 | 低:只有创始人身份有充分来源 |
| Dave Chung | CB Insights 点名的创始人 | 第三方创始人名单点名;已保留来源未公开角色细节 | 创始人群体 / 早期技术覆盖 | 低:只有创始人身份有充分来源 |
| SJ Kim | CB Insights 点名的创始人 | 第三方创始人名单点名;已保留来源未公开角色细节 | 创始人群体 / 早期技术覆盖 | 低:只有创始人身份有充分来源 |
| Soyoung Lee | CB Insights 点名的创始人,Tracxn 显示为董事会成员 | 第三方创始人和董事会相关来源点名 | 创始人群体 / 治理信号 | 中:公开角色细节仍不完整 |
公开领导层列举不完整:官方页面没有发布完整高管名单,因此创始人 / 群体行依赖第三方档案,并保留明确证据缺口。
[CO005, CO006, CO007, CO008, CO028, CO037]1.3 融资历程与利益相关方图谱
融资记录很强,但公开数据库之间并未完全对齐。当前最清晰的事件是 July 1, 2026 的 $100 million Series B,由 NEA 和 NAVER Ventures 共同领投,Amazon、Radical Ventures、Korea Investment Partners、Index Ventures、Quadrille Capital 和 Red Bull Ventures 参与。此前融资包括 2024 由 NEA 和 NVIDIA 的 NVentures 共同领投的 $50 million Series A、December 2022 由 Radical Ventures 领投的 $12 million 种子轮延展,以及 2022 由 Index Ventures 领投的 $5 million 种子轮。当前市场数据来源支持累计融资约 $207 million 至 $207.1 million,但抓取快照中的 Crunchbase 仍停在 $107.1 million。利益相关方含义很重要:TwelveLabs 同时拥有传统 VC 背书、韩国战略连接、Amazon/AWS 基础设施分发和 NVIDIA 算力验证。未解项是估值、持股、董事会权利以及任何老股成分。[CO017, CO018, CO020, CO021, CO022, CO023]
| 利益相关方 | 角色 | 控制权或经济重要性 | 尽调要求 |
|---|---|---|---|
| NEA | Series B 轮共同领投;Series A 轮共同领投 | 反复担任领投方,并通过 Tiffany Luck 发出董事 / 合伙人声音 | 确认持股、董事席位、按比例跟投权和保护性条款 |
| NAVER Ventures | Series B 轮共同领投 | 据引述为首位投资人;韩国 / APAC 战略信号 | 确认与 NAVER 的商业关系和战略权利 |
| Amazon / AWS | Series B 轮参投方兼首选云服务商 | 资本加多年 Trainium / AWS 基础设施承诺 | 审查合同最低消费、模型首发义务和云集中风险 |
| Radical Ventures | 种子延期轮领投;Series B 轮参投方 | 从种子轮到后续轮次的长期 AI 专业投资人 | 确认后续跟投承诺和 AI 治理支持 |
| Index Ventures | 种子轮领投 / 既有及 Series B 轮参投方 | 早期机构投资人延续到后续轮次 | 确认持股历史和老股交易活动 |
| Korea Investment Partners | Series A 轮和 Series B 轮参投方 | 韩国相关机构支持 | 确认本地市场准入和控制权 |
| NVIDIA / NVentures | Series A 轮共同领投方和 GPU 合作伙伴 | 战略算力 / 基础设施验证 | 澄清任何供应、共同开发或首选平台义务 |
| Quadrille Capital / Red Bull Ventures 等战略投资人 | Series B 轮参投方 | 2026 年辛迪加中的新财务 / 战略验证 | 确认票额、战略权利和客户接入承诺 |
投资人图谱按利益相关方梳理,并非完整股权结构表;准确持股、董事权利和老股交易活动未公开披露。
[CO017, CO018, CO022, CO023, CO026, CO027]1.4 里程碑、产品成熟度与规模信号
里程碑序列显示,公司正从 API 驱动的视频搜索走向更大的视频认知栈。2022,公开叙事是上下文视频搜索:一个由种子轮支持的产品,可以在视频内部找到精确片段,而不是只查文件周边元数据。到 2024 Series A,公司宣布 Marengo 2.6、Pegasus-1 beta、Embeddings API、30,000 名 API 用户,并计划大举招聘。到本次运行,来源将 Marengo 3.0 描述为支撑 Search 和 Embed 的嵌入模型,将 Pegasus 1.5 描述为结构化、基于时间的元数据引擎,并把 Rodeo 描述为应用层推进。合作里程碑与产品里程碑同样重要。AWS 现在是首选云,带有 Trainium 优化和模型首发承诺;NVIDIA 仍是 GPU 加速和战略融资信号。已报道的客户证明包括 MLSE、AMC Global Media、UNICEF、创作者、体育俱乐部和 Hollywood 制片厂,但当前付费客户数量未公开。[CO012, CO013, CO015, CO016, CO019, CO024]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2021-03-31 | Tracxn 和 Crunchbase 显示法律注册 / 成立年份 | 创立 | 2021 年注册 / 成立 | 包括 Jae Lee 在内的创始人群体 | 为后续章节设定标准成立日期 |
| 2022-03-16 | 种子轮融资支持开放服务产品建设 | 融资 | $5M 种子轮 | Index Ventures、Radical Ventures、Expa、Techstars Seattle、天使投资人 | 视频搜索 API 逻辑获得早期机构验证 |
| 2022-12-05 | 产品仍处封闭测试时完成种子延期轮 | 融资 | $12M 延期轮;当时引用累计 $17M | Radical Ventures、Index Ventures、WndrCo、Spring Ventures、天使投资人 | 延长基础模型 / API 开发现金跑道 |
| 2024-06-04 | Series A 轮公布,同时更新多模态产品 | 融资 | $50M Series A 轮 | NEA、NVentures、Index、Radical、WndrCo 与 Korea Investment Partners | 战略算力和风投验证 |
| 2024-06-04 | 重点发布 Marengo 2.6、Pegasus-1 beta 和 Embeddings API | 产品 | 产品套件扩展 | TwelveLabs、NVIDIA 基础设施 | 从搜索 API 走向多模态基础模型平台 |
| 2024-12-12 | Yoon Kim 据报道加入,出任总裁兼首席战略官 | 治理 | 领导层补强 | Yoon Kim;TwelveLabs | 为创始人主导公司补入资深战略、电信和 Siri 背景 |
| 2025-12-01 | Marengo 3.0 定位为生产级向量嵌入模型 | 产品 | 512 维向量嵌入;声称基准领先 | TwelveLabs 研究 / 产品团队 | 强化 Embed / Search API 和检索经济性 |
| 2026-06-01 | Pegasus 1.5 转向基于时间的元数据提取 | 产品 | 模式优先的结构化视频输出 | TwelveLabs 研究 / 产品团队 | 为分析和智能体加入结构化数据层 |
| 2026-07-01 | Series B 轮完成 / 公布 | 融资 | $100M;已抓取发布未披露估值 | NEA、NAVER Ventures、Amazon、Radical、KIP、Index、Quadrille 与 Red Bull | 支持研发、全球扩张和 Video Cognition System 推进 |
| 2026-07-01 | AWS 首选云和 Trainium 关系加深 | 合作 | 多年承诺;新模型优先在 AWS 上线 | AWS / Amazon;TwelveLabs | 带来云集中尽调项,同时形成战略分销渠道 |
| 2026-07-21 | 公开法律 / 监管检查未发现被引用诉讼或执法行动 | 监管 | 已保留来源未发现诉讼或程序 | Tracxn、Crunchbase、CB Insights、TechCrunch 搜索集 | 监管行记录的是未发现事项,但仍需持续监控 |
| 2026-07-21 | Series B 轮后融资数据库相互冲突 | 反向 | 当前来源约 $207M;Crunchbase 快照为 $107.1M | Tracxn、CB Insights、Crunchbase | 使用当前多来源数字,并标记档案滞后风险 |
时间线是截至 runDate 的概览记录;估值、当前收入和完整治理 / 股权结构数据仍不在公开证据内。
[CO004, CO017, CO020, CO021, CO022, CO026]带日期的里程碑勾勒 TwelveLabs 的成立、融资、产品演进、AWS 伙伴关系和不利数据冲突。
日期采用发布日期或来源标注日期;产品博客日期是抓取页面里最好的公开时间锚点。
[CO004, CO017, CO020, CO021, CO022, CO026]1.5 公开指标、反向信号与尽调缺口
概览部分有一个反向来源,以及多个不构成红旗但重要的不确定点。第一,融资数据库不同步:Tracxn、CB Insights、Digital Today 和 2026 融资报道都在 Series B 后收敛到 $207 million 或更高,而抓取到的 Crunchbase 快照仍显示 $107.1 million,且最后一轮融资早于 July 2026 公告。第二,没有已保留的公开来源披露收入、ARR、毛利率、烧钱速度、现金跑道、准确付费客户数、估值、股权结构表所有权或详细董事会权利。第三,合作伙伴叙事带来集中度问题,因为 AWS 既接近投资方,又在多年 Trainium 承诺下成为首选云。第四,抓取来源没有浮现诉讼、制裁或执法行动;这是负面检索结果,不是不存在的证明。因此,后续章节应把本概览当作已验证的身份和融资基线,同时在管理层或可靠来源给出前,把估值和私人运营指标排除在封面事实之外。[CO015, CO020, CO021, CO030, CO031, CO036]
1.6 展示项
02市场分析
2.1 市场边界、相邻领域与现状
不能把 Twelve Labs 当作泛基础模型公司或摄像头 / VMS 硬件厂商来投资评估。其公开产品界面定义了更窄的任务:通过 API、SDK、视频原生模型,以及在知识库上运行的智能体,让视频可搜索、可分析、可嵌入、可推理。因此,纳入的支出覆盖视频理解 API、视频搜索与检索、视频转文本和摘要、多模态嵌入、语料库级推理智能体,以及建立在视频档案上的工作流自动化。相邻支出包括企业视频平台、媒体资产管理、视频监控分析、体育分析、广告内容分析和汽车感知。明确排除基础摄像头、显示器、存储设备、商品化视频会议、不处理视频的通用 LLM 订阅,以及除作为现状替代之外的人工打标劳动力。这个边界很重要,因为公开市场页面常把软件、硬件、服务、监控和广义 AI 混在一起;本章因此保留多重视角,而不是强行给出一个 TAM 标题。[CM001, CM002, CM022, CM023, CM024, CM025]
| 细分 / 品类 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 Twelve Labs 的相关性 |
|---|---|---|---|---|
| 视频理解 API | 用 API 或 SDK 工作流对视频做搜索、分析、总结、嵌入和推理 | 手工打标人力;通用纯文本 LLM 席位 | 产品、数据、工程、媒体运营或创新预算 | Twelve Labs 模型和智能体的核心可变现层 |
| 视频分析软件 | 物体 / 活动检测、告警、检索、事件复盘、人群或交通分析 | 摄像头、监视器、存储设备和通用 VMS 硬件 | 安全运营、公共安全、零售运营、设施、政府 | 最接近的成熟分析师品类,但偏监控的定义会高估契合度 |
| 多模态 AI | 结合视频、图像、音频和文本,用于搜索、生成和推理的模型 | 纯文本生成或非视频分析 | AI 平台团队、应用开发者、企业创新团队 | 视频原生模型需求和 API 买方的最佳窄口径代理 |
| 企业视频 / MAM | 视频内容管理、档案发现、合规审查、元数据丰富化 | 视频会议硬件,以及不含理解能力的通用内容分发 | 媒体运营、品牌团队、合规、企业传播 | 档案规模大且元数据不完整时契合度高 |
| 体育分析 | 球员追踪、战术分析、球探片段、转播洞察、球迷工作流 | 票务、场馆运营和非视频球迷 CRM | 球队、联盟、转播方、竞技表现部门 | 视频密集工作流带来吸引力,但绝对支出较小的垂直领域 |
| 汽车与物理世界感知 | 面向 ADAS / 自动驾驶数据集的片段检索、标注、场景搜索和视频推理 | 车辆硬件、传感器和非 AI 汽车软件 | ADAS / 自动驾驶工程和数据平台团队 | 视频搜索能降低数据工程负担的邻近扩张方向 |
边界是分析口径,刻意窄于宽泛 AI 或摄像头市场估算。各行标出市场报告支出可能完全映射、部分映射,或应从 Twelve Labs 口径中剔除的位置。
[CM001, CM002, CM022, CM023, CM024, CM030]2.2 规模测算视角与相互矛盾的估计
公开市场估计方向有利,但差异很大。核心视频分析池中,2026 估计从 The Business Research Company 的 $11.59B,到 Mordor Intelligence 的 $15.04B,再到 Precedence Research 的 $18.53B;Polaris 则报告 $17.62B。CAGR 估计集中在中双位数到二十出头,但供应商对品类定义不同,常纳入监控、零售、政府和 VMS 层,而 Twelve Labs 可能只部分覆盖。更宽的 TAM 视角是生成式 AI,Precedence 估计 2026 为 $55.51B,Fortune 估计 2026 为 $161B;这个差距提醒我们,宽口径 AI TAM 可能夸大视频理解 API 的近期收入池。更窄的切入口是多模态 AI,Precedence 估计 2026 为 $3.43B,MarketsandMarkets 将 Twelve Labs 列为供应商之一,并预计该市场到 2028 达到 $4.5B。投资结论是:市场大且在增长,但精度低。[CM003, CM004, CM005, CM006, CM007, CM008]
| 发布方 | 年份 | 地理范围 | 数值 | CAGR | 方法 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| Precedence Research,生成式 AI | 2026 | 全球 | 2026 年 USD 55.51B;2035 年达 USD 1,206.24B | 2026 至 2035 年 36.97% | 生成式和多模态 AI 应用的宽口径 TAM 代理 | 中 | 远宽于视频理解;可能高估 Twelve Labs 直接收入池 |
| Precedence Research,多模态 AI | 2026 | 全球 | 2026 年 USD 3.43B;2035 年达 USD 51.76B | 2026 至 2035 年 35.34% | 多模态工作负载的更窄模型 / API 滩头代理 | 中 | 包含非视频模态和许多垂直领域;可能低估视频分析买方 |
| MarketsandMarkets,多模态 AI | 2028 | 全球 | 2028 年达 USD 4.5B | 预测期内 35.0% | 供应商 / 品类扫描;将 Twelve Labs 列为供应商之一 | 中 | 搜索页摘录缺少完整方法和终年基准细节 |
| Mordor Intelligence,视频分析 | 2026 | 全球 | 2026 年 USD 15.04B;2030 年达 USD 33.74B | 2026 至 2030 年 22.18% | 视频分析软件需求的核心 SAM 代理 | 中 | 监控、政府和边界防护组合只部分映射到 Twelve Labs |
| Precedence Research,视频分析 | 2026 | 全球 | 2026 年 USD 18.53B;2035 年达 USD 109.85B | 2026 至 2035 年 21.94% | 带长期预测的更高核心 SAM 代理 | 中 | 品类口径宽、预测周期长,会带来上行偏差风险 |
| The Business Research Company,视频分析 | 2026 | 全球 | 2026 年 USD 11.59B;2030 年 USD 24.73B | 2026 年基准后到 2030 年 20.8% | 保守的 2026 年 SAM 下限代理 | 中 | 仍混合安全、交通、零售和其他用例 |
| IMARC,视频分析 | 2025 | 全球 | 2025 年 USD 9.8B;2034 年 USD 35.3B | 2026 至 2034 年 14.81% | 增速较低的佐证性类别估计 | 中 | 未单独拆出 AI 原生 API 收入 |
| Mordor Intelligence,企业视频 | 2026 | 全球 | 2026 年 USD 28.98B;2031 年 USD 46.93B | 2026 至 2031 年 10.12% | 相近的企业视频基础设施与工作流预算池 | 中 | 包含会议和视频平台,不限于理解 / 搜索 |
| Grand View Research,体育分析 | 2026 | 全球 | 2026 年 USD 7.0B;2033 年 USD 23.1B | 2026 至 2033 年 18.5% | 体育视频、竞技表现和战术分析的垂直 SAM 视角 | 中 | 体育分析涵盖视频智能之外的数据类型和产品 |
这些数值不能相加,因为类别边界重叠。表格保留相互矛盾的估计,把它们作为市场规模视角,而不是断言一个单一 TAM。
[CM003, CM004, CM005, CM006, CM007, CM008]Twelve Labs 的宽口径 TAM、核心 SAM 和窄滩头市场视角;能用 2026 年市场数值时优先采用。
各层是视角,不是可相加市场。生成式 AI 刻意取宽;视频分析偏重安防监控;多模态 AI 更窄,但包含非视频模态。
[CM026, CM027, CM028, CM047]同一 2026 年全球视频分析市场规模的低 / 基准 / 高估算,单位为十亿美元。
低、中、高分别来自 The Business Research Company、Mordor Intelligence 和 Precedence Research;品类定义不同,因此这是尽调区间,不是统计置信区间。
[CM029]2.3 买方、用户、付款方分层与采纳路径
经济买方随细分市场变化,但共同采纳路径是:某个工作流负责人有太多视频,集成团队必须连接档案或视频流,预算负责人需要可衡量的节省时间、搜索质量、合规改善或收入影响。在媒体与娱乐领域,买方通常是媒体运营、档案、产品或后期制作负责人;编辑、合规审核员、制片人和研究团队是用户。在安全与监控领域,买方是公共安全、安全运营或设施团队;分析员和调度团队使用系统;隐私 / 法务团队可能阻断部署。体育队、联盟和广播商购买表现、战术、球探和球迷体验分析;广告和营销团队把视频智能用于内容分类、个性化和活动工作流。汽车买方是需要视觉感知和片段检索的 ADAS/autonomy 工程与数据团队。这个买方图谱支持 API 优先,但更大规模部署仍需要数据治理、模型评估、采购和变更管理。[CM030, CM031, CM032, CM033, CM034, CM035]
| 细分市场 | 买方 | 用户 | 付款方 | 工作流 | 预算归口 | 采用触发点 |
|---|---|---|---|---|---|---|
| 媒体娱乐 / 归档库 | 媒体运营、产品、后期制作、档案、合规 | 剪辑师、制片人、研究员、合规审核员 | 工作室、广播电视台、流媒体平台、版权方或平台 | 搜索档案、总结素材、分类场景、审核内容、补全 MAM 元数据 | 内容运营、产品工程、AI 创新 | 大型视频库元数据差,或人工审核太慢 |
| 安防与监控分析 | 安全运营、公共安全、设施管理、智慧城市团队 | 分析师、调度员、调查员、防损人员 | 企业、市政部门、机构或基础设施运营方 | 检测事件、搜索事故、分诊告警、回看摄像头画面 | 安防、设施、公共安全、运营 | 需要更快的事件响应、更少人工监控工时,或智慧城市分析 |
| 体育分析与转播 | 球队表现、联盟媒体、广播电视台、球探负责人 | 教练、分析师、球探、转播制作团队 | 球队、联盟、广播电视台、协会或赞助商 | 球员追踪、战术剪辑、球探检索、转播集锦生成 | 竞技表现、分析、媒体或球迷互动预算 | 从比赛和训练视频中获取竞争洞察,或更快制作剪辑 |
| 广告 / 品牌视频工作流 | 营销运营、创意技术、广告技术产品团队 | 创意人员、媒介策划、品牌安全、效果营销人员 | 品牌、代理商、平台、零售媒体网络 | 分类素材、分析创意、个性化视频、加快内容合规 | 营销技术、数字媒体、AI 转型 | 视频创意量上升,需要可搜索的活动素材 |
| 汽车 / 出行感知数据 | ADAS / 自动驾驶工程、数据平台、仿真负责人 | ML 工程师、数据整理人员、验证团队 | OEM、AV 开发商、Tier 1、车队运营商 | 检索边缘场景、理解驾驶片段、构建训练 / 评估数据集 | 工程、自动驾驶、AI 基础设施 | 数据标注卡住,需要找到罕见驾驶场景 |
| 企业开发者与 AI 平台 | 产品工程、数据科学、应用平台负责人 | 开发者、分析师、内部应用构建者 | 业务单元或中央 AI / 平台团队 | 把视频搜索 / 推理嵌入客户或内部应用 | 云、平台工程、创新、产品 | 需要 API 能力,但不想从零训练视频模型 |
这些行是细分市场图谱,不是完整客户清单。买方、用户和付款方经常分散在技术团队和工作流团队,因此采用取决于集成和 ROI 证明。
[CM030, CM031, CM032, CM033, CM034, CM035]按细分市场拆解视频理解采用中的买方、用户、付款方和工作流。
矩阵基于公开市场品类证据和 Twelve Labs 产品任务;不识别未披露的 Twelve Labs 客户。
[CM030, CM031, CM032, CM033, CM034, CM023]企业采用视频理解 AI 时,从视频痛点到续约决策的路径。
数值是示意性漏斗指数,不是实测转化率。公开来源支持这些阶段,但不支持经过基准验证的转化曲线。
[CM035, CM042, CM043]2.4 增长驱动、采纳约束与尽调缺口
增长逻辑很强,因为视频量、智能摄像头、边缘计算、云原生企业视频和多模态生成式 AI 都把组织推向自动化视频理解。Twelve Labs 自身产品主张也匹配这种拉力:用文本或图像查询搜索片段、分析视频、生成嵌入,并跨整个知识库推理。约束逻辑同样重要。Polaris 提醒 VMS 集成复杂度、算力和存储需求、数据隐私以及误报。RAND 的 AI 项目研究警告称,超过 80% 的 AI 项目失败;引用调查中只有 14% 的组织已完全准备好集成 AI,根因在问题定义、数据、基础设施、工作流匹配和技术可行性。EU AI Act 禁止或限制若干生物识别和 CCTV 用途,NIST 强调 AI 风险管理,Stanford 记录了加速中的监管和事故。最大尽调缺口不是视频 AI 是否在增长,而是 Twelve Labs 在试点、治理审查和集成工作之后,能从这些重叠池中盈利捕获多少。[CM036, CM037, CM038, CM039, CM040, CM041]
| 驱动因素 / 约束因素 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 视频体量与档案搜索痛点 | 驱动因素 | 当前 | 媒体、企业和安防团队靠人工打标、抽帧越来越不划算 | 量化每个工作流节省的工时,并与 API 和计算成本对比 |
| AI 驱动的检测、向量嵌入与推理 | 驱动因素 | 当前 / 近期 | 支持搜索、自动审核、告警、结构化输出和语料库级推理;这些过去只靠元数据很难跑通 | 在客户自有视频上基准测试准确率、延迟和检索质量 |
| 智慧城市、监控与边缘计算扩张 | 驱动因素 | 当前 / 中期 | 扩大原始摄像头 / 视频数据量,也抬高实时或事后分析需求 | 把隐私安全的分析用例与受限生物识别监控拆开 |
| 多模态与生成式 AI 预算扩张 | 驱动因素 | 当前 / 中期 | AI 平台团队会转向具备视频能力的模型,而不是只用文本工具 | 确认视频预算是新增、重新分配,还是实验性试点 |
| 企业视频云端现代化 | 驱动因素 | 当前 | 支撑 API 采用,并接入内容管理和协作平台 | 梳理集成伙伴、数据驻留和企业安全要求 |
| EU AI Act 与生物识别限制 | 约束因素 | 2025-2026 年实施 | 限制 CCTV 抓取、生物识别、情绪识别和高风险用途,或抬高合规成本 | 评估面向生物识别、执法、工作场所和 EU 部署的产品控制 |
| AI 项目失败与 ROI 怀疑 | 约束因素 | 当前 | 许多 AI 项目因数据、工作流、基础设施和范围错配而失败,买方可能要求 demo 之外的证明 | 收集试点转生产转化率、回本周期和可引用 ROI 证据 |
| 集成、数据隐私、误报与计算 / 存储成本 | 约束因素 | 当前 | 即使用例有吸引力,部署进 VMS、MAM 和数据平台环境也会变慢 | 尽调实施负担、误报率、数据留存控制,以及高视频工作负载下的毛利率 |
驱动因素和约束因素刻意成对呈现,因为市场需求可以增长,但部署仍会卡在合规、集成和 ROI 关口。
[CM036, CM037, CM038, CM039, CM040, CM041]2.5 展示项
03竞争格局
3.1 格局:直接对手、巨头、相邻玩家与替代方案
Twelve Labs 所在的视频 AI 格局拥挤但碎片化。最接近的直接对手不是数字人或文生视频厂商,而是 Coactive、Hive、Reka、Memories.ai、Vidrovr,以及试图把非结构化视频变成可搜索、可查询、可治理数据的内部团队。大型科技巨头很重要,因为 Google Cloud Video Intelligence、Gemini、Azure AI Video Indexer、Amazon Rekognition、OpenAI 和 Meta 能把视频理解打包进既有采购和开发者入口。Runway、Synthesia、HeyGen 和 Pika 等相邻生成工具争夺 AI 视频预算和买方注意力,但它们的公开产品界面以创作为中心,而不是企业档案理解。AssemblyAI 和 Deepgram 等音频优先厂商,在转录、摘要或语音智能足够时就是替代方案。开源的 VideoLLaMA、InternVideo 和 Video-ChatGPT 让复杂买方仍可相信内部自建,而人工打标仍是受监管或低量工作流的后备选项。[CP015, CP017, CP020, CP022, CP023, CP024]
| 竞争对手 / 替代方案 | 类别 | 公开规模或融资信号 | 目标细分市场 | 产品范围 | 相对 Twelve Labs 的差异化 | 局限 / 尽调问题 |
|---|---|---|---|---|---|---|
| Twelve Labs | 视频原生基础模型 API | 2026 年 $100M Series B;公开 Developer 定价 | 开发者、媒体、体育、广告、政府、企业视频档案 | Marengo 搜索 / 向量嵌入,加 Pegasus 分析 | 一个 API 覆盖深度时序视频检索与视频转文本 | 需要私有胜率、实际企业定价和独立基准测试 |
| Google Gemini + Video Intelligence | 大科技多模态 + 传统 CV | Google Cloud 平台;token 和按分钟定价 | 云开发者与企业工作负载 | Gemini 视频理解 + Video Intelligence 标注 | 采购触达强,多模态生态广 | 传统 CV 和通用模型界面可能达不到视频原生检索效果 |
| OpenAI GPT-4o / Sora | 大科技多模态 + 视频生成 | OpenAI API 生态;token 定价;Sora API 文档 | 开发者、创作者、构建多模态应用的企业 | 通用多模态推理和生成视频 | 开发者心智强,多模态模型迭代很快 | 没有公开证据证明其专用长档案语义视频搜索达到同等水平 |
| Microsoft Azure AI Video Indexer 服务 | 云视频分析既有厂商 | Azure 按分析预设和输入分钟定价 | 企业媒体、合规、企业视频档案 | 转写、主题、OCR、人脸、场景、标签 | Azure 采购能力强,媒体工作流集成深 | 对基础模型检索和向量嵌入的专精程度较低 |
| Amazon Rekognition Video | 云 CV / 审核既有厂商 | AWS 按分钟定价;AWS 账户集成 | 安防、审核、媒体分析、AWS 用户 | 物体 / 人员 / 文本 / 活动检测和审核 | AWS 分发、账单和数据重力 | 传统识别服务;新客户使用流式视频能力受限 |
| Coactive | 直接视觉搜索同业 | $30M Series B;企业 demo 驱动 GTM | 拥有图像 / 视频库的媒体、零售、平台 | 无元数据多模态视觉搜索和激活 | 直击人工元数据和视觉数据激活痛点 | 定价和正面对比质量未公开 |
| Hive | 直接审核 / 内容 AI 同业 | 用量定价;企业 / 视频特殊费率 | 信任与安全、审核、内容平台 | 跨模态审核、搜索、生成、识别 | 审核分类体系强,内容安全工作流成熟 | 对企业档案长视频推理的定位不够清晰 |
| Runway | 相邻生成式视频 | 2026 年 $315M Series E;套餐从 $12/month 起 | 创作者、工作室、广告、企业创意团队 | 生成视频和世界模拟工具 | 争夺 AI 视频预算和创意端心智 | 不是直接的档案搜索或视频理解 API |
| Synthesia / HeyGen / Pika | 相邻 AI 视频创作 | Synthesia 从 $18/month 起;HeyGen 据报道 ARR 为 $200M;Pika 2.5 生成 | 营销、培训、销售赋能、创作者团队 | 数字人、本地化、生成式商务视频 | 相邻预算池大,企业采用度高 | 主要是创作,而不是理解既有视频 |
| Reka / Memories.ai 阵营 | 多模态基础模型同业 | Reka 据报道完成 $110M 轮融资、估值 >$1B;Memories.ai 开始冒头 | 开发者、视觉记忆、搜索、机器人 / 安防 / 媒体 | 多模态 API、视觉记忆、视频搜索和推理 | 凭更广模型组合,可能匹配视频推理 | 包装、定价和生产环境证据仍稀缺 |
| 开源 + 内部自建 | 替代方案 / 潜在进入路径 | VideoLLaMA3、InternVideo、Video-ChatGPT 的 GitHub 项目 | AI 基础设施团队和研究能力强的企业 | 自托管模型与检索栈 | 控制、定制和潜在成本优势 | 需要承担评估、服务、治理和集成负担 |
| 人工打标 / 传统 DAM-MAM | 现状替代方案 | 私有预算和劳动力 / 流程成本 | 受监管、低容量或对准确率敏感的档案 | 人工元数据、规则和既有资产系统 | 可信、可审计,且已经嵌入流程 | 成本高、速度慢,长尾语义检索弱 |
画像行结合官方产品 / 定价页、公司融资公告、独立新闻和作者分类;私有收入、胜率和企业折扣数据仍不可得。
[CP015, CP017, CP019, CP020, CP021, CP022]序数评分把 Twelve Labs 的专业化,与超大云厂商分发和相邻创作工具区分开。
x=视频理解深度,y=分发广度,采用作者基于公开产品、定价和融资证据得出的 1-10 序数评分;没有公开基准提供可共用的数字坐标轴。
[CP039]3.2 能力与定价:Twelve Labs 的差异点
Twelve Labs 最强的差异化,是把视频原生嵌入 / 搜索与视频转文本分析打包在一起。Marengo 和 Pegasus 处理视频的时间、语音、音频和视觉结构;云视频 API 对标签、OCR、人脸、场景、审核和转录仍很有用,但通常被组织成经典标注服务或通用多模态模型调用。定价强化了这种比较。Twelve Labs 公布了索引、搜索、Analyze 输入、输出 token 和嵌入基础设施的用量计量项;Google、AWS 和 Azure 的许多视频功能按分钟定价;Gemini 和 OpenAI 更偏向模型 token 定价;创意工具则发布创作者或企业订阅。这让同口径成本比较变难,但也意味着采购团队可以多家并用,为每个工作流选择足够便宜、足够好的工具。矩阵里无支持的单元格是有意保留的:公开来源很少披露实际折扣、延迟或正面对比的检索质量。[CP001, CP002, CP003, CP005, CP006, CP007]
| 购买标准 | Twelve Labs | Google Gemini / Video Intelligence | OpenAI GPT-4o / Sora | Azure AI Video Indexer | AWS Rekognition Video | Coactive / Hive | Runway / Synthesia / HeyGen / Pika | 开源 / 内部自建 |
|---|---|---|---|---|---|---|---|---|
| 原生语义视频检索 | 支持:Marengo 搜索 / 向量嵌入 | 部分支持:Video Intelligence 标签;Gemini 可对视频推理 | 部分支持 / 未知:多模态模型,但没有专用档案搜索证明 | 部分支持:视频洞察 / 搜索,传统索引器 | 部分支持:标签 / 审核 / 搜索原语 | Coactive 支持;Hive 对内容搜索 / 审核部分支持 | 不支持档案检索 | 可行,但需买方自建 |
| 视频转文本推理 / 分析 | 支持:Pegasus | 部分支持:Gemini 视频理解 | 部分支持:GPT-4o 视频输入和模型推理;Sora 生成另算 | 部分支持:主题 / 转写 / 情感 / 实体 | 有限:输出检测结果,不做深层推理 | 部分支持 / 未知:因供应商而异 | 不支持或间接支持 | 可行,但需买方自建 |
| 传统 CV 标签、OCR、人脸、审核 | 部分支持:非公开核心定位 | Video Intelligence 强支持 | 部分支持 / 未知 | 强支持 | 强支持 | Hive 强;Coactive 视觉搜索更宽 | 不支持审核用例 | 可用多个模型拼出 |
| 生成视频 / 数字人 | 不支持 | Video Intelligence 不支持;Gemini 模型生态更宽 | Sora 支持 | 不支持 | 不支持 | 不支持或不明确 | 强支持 | 可用单独模型拼出 |
| 企业云采购 | 起步中:直销和 AWS 渠道分发 | 强支持 | OpenAI 生态和合作伙伴渠道强支持 | 强支持 | 强支持 | 因厂商而异;多数是专业厂商销售 | 有企业档位,但偏创意工作流 | 内部采购 / 控制,但运维负担重 |
| 成本可预测性 | 混合:多种计量 + 企业定制 | 混合:按分钟 + token 模型 | 混合:token / 视频生成计量 | 混合:分钟 / 预设定价 | 混合:按分钟 / API 定价 | 企业视频未知 / 定制 | 订阅 / 点数套餐;企业定制 | 控制力高,但有隐性基础设施成本 |
| 开放 / 自托管控制 | 公开 SaaS/API 不支持 | 托管云,控制有限 | 托管 API,控制有限 | Azure 托管,控制有限 | AWS 托管,控制有限 | 多为托管平台 | 多为托管 SaaS | 天然支持 |
| 仅音频替代是否足够 | 转写足够时显得过重 | 可用语音组件 | 语音 / 多模态生态 | 支持音频洞察 | 需要单独 AWS 服务 | 非主要用途 | 非主要用途 | 仅音频场景下 AssemblyAI / Deepgram 更便宜 |
“不支持”表示本轮没有公开来源把该能力作为主要用例支撑;“部分支持”表示买方能拼出工作流,但公开证据不足以证明语义视频理解达到与 Twelve Labs 相当的水平。
[CP001, CP002, CP003, CP005, CP006, CP007]| 供应商 / 套餐 | 公开计费单位或套餐 | 包含能力 | 未知项 / 折扣提示 | 竞争影响 |
|---|---|---|---|---|
| Twelve Labs Developer | Marengo $0.042/min;Search 每 1k 次查询 $4;Pegasus 输入 $0.0292/min;输出每 1k tokens $0.0075 | 索引、向量嵌入、搜索、视频分析、文本输出 | 企业承诺用量条款和实际折扣未公开 | 开发者入口透明,但多计量用法增加比价摩擦 |
| Twelve Labs Free | 最高 10 小时 / 600 分钟;90 天索引访问权 | 可试用 Marengo/Pegasus 的索引和分析 | 免费索引不长期保留 | 开发者低门槛评估 |
| Google Cloud Video Intelligence | 按分钟计价的视频标注;部分功能有免费额度 | 标签、镜头、露骨内容、语音、OCR、物体、Logo、人脸 / 人物检测 | 批量折扣和 Gemini 组合工作流成本不可见 | 对比专用 API 的低成本经典 CV 基准 |
| Google Gemini API | 按模型分层的 token 计费 | 通用多模态推理,文档包含视频理解 | 真实视频 token 化和档案级经济性需要用工作负载测试 | 可能吸收围绕视频片段的推理任务 |
| OpenAI API / Sora | 模型按 token 定价;Sora 视频生成 API 文档 | 通用多模态 API 加生成式视频 | 视频理解的档案级经济性和企业折扣未知 | 很可能带来新进入者压力,尤其针对已使用 OpenAI 的团队 |
| Azure AI Video Indexer | 按预设 / 输入分钟计价 | 音频 / 视频洞察、转录、OCR、标签、人脸、主题 | 区域条款和 Azure 承诺用量价格各异 | 企业采购和 Microsoft 捆绑优势 |
| AWS Rekognition Video | 视频分析按分钟计价 | 标签、审核、文本、人脸、名人 / 人物 / 路径追踪,以及镜头 / 技术线索 | 更广泛的 AWS 架构和流式访问条件各异 | 经典识别场景里的强 AWS 默认选项 |
| Runway | 创作者套餐 $12/month 起;企业销售 | 生成图像 / 视频和创意工具 | 点数消耗和企业使用条款各异 | 更像相邻预算竞争者,而不是检索同类 |
| Synthesia | 套餐目前 $18/month 起;企业层级 | AI 虚拟形象、语音、本地化、商业视频 | 分钟 / 席位 / 企业条款各异 | 争夺商业视频制作预算 |
| HeyGen | 免费版加创作者 / Pro / 商业版定价 | AI 视频生成和虚拟形象工作流 | 点数使用和企业条款各异 | 已有规模的相邻视频厂商,据报道 ARR 有动能 |
| Coactive / Hive | 演示驱动或按用量计价;Hive 视频特价 | 视觉搜索、无元数据分析、审核 | 企业定价和帧采样经济性未公开 | 直接同类的经济性需要客户报价 |
| AssemblyAI / Deepgram | 语音 / 音频 API 定价页 | 语音转文本、音频智能、语音 API | 附加项、声道、实时层级各异 | 在以转录为核心的工作流里,纯音频方案会挤压完整视频分析 |
所有价格均为 2026-07-21 观察到的公开标价或官网定价;私有承诺用量折扣、最低消费和实际用量组合未披露。
[CP001, CP005, CP006, CP007, CP008, CP009]功能广度解释了为什么买方可能多平台并用,而不是标准化到单一供应商。
矩阵标签来自公开产品页的分类判断;「不支持」表示已保留的公开来源未证明该单元格是主要能力。
[CP040]3.3 分发、切换成本与护城河持久性
Twelve Labs 有可信的技术护城河,但没有可与超大规模云厂商相比的分发护城河。AWS、Microsoft、Google、OpenAI 和 Meta 能把平台控制转化为默认选项、采购便利、安全审查捷径、承诺用量折扣和模型打包。Twelve Labs 可用专精、视频优先 API、开发者友好的入门定价,以及其嵌入和分析在长视频、多模态视频检索上显著更好的证明来反击。尽管如此,护城河仍容易被多家并用削弱:客户可以把部分档案索引在 Twelve Labs,在 Rekognition 或 Azure 上跑经典 CV,用 Gemini/OpenAI 做推理,并把只需转录的工作流留给语音 API。只有当索引、评估集、合规审查和应用集成嵌入业务后,切换成本才会上升。因此,核心尽调问题不是有没有竞争对手,而是 Twelve Labs 是否能赢下经常性的生产工作负载,并且原生视频理解是否明显胜过更便宜、已打包或内部托管的替代方案。[CP030, CP032, CP033, CP035, CP036, CP037]
| 护城河主张或风险 | 威胁来源 | 严重程度 | 证据 | 缓释措施或尽调问题 |
|---|---|---|---|---|
| 视频原生模型专精 | Gemini、OpenAI、Reka、Meta 多模态模型 | 高 | 通用多模态模型已经处理或瞄准视频输入和推理 | 委托独立基准测试,覆盖长视频检索、时间推理、延迟和成本 |
| 开发者友好的 API 与定价 | 云厂商按分钟 API 和 Mixpeek 成本控制定位 | 中 | Twelve Labs 标价透明但计量项多;既有云厂商发布简单的按分钟计费单位 | 在承诺用量条款下测试真实工作负载 |
| 企业分发 | AWS、Azure、Google、OpenAI 平台默认选项 | 高 | 存量厂商嵌在既有采购、合规、计费和云数据引力里 | 衡量 AWS / 合作伙伴渠道带来的销售周期差异和附加率 |
| 专精型直接同类 | Coactive、Hive、Reka、Memories.ai 与 Vidrovr | 中 | 同类厂商切入视觉搜索、审核、多模态 API、视觉记忆或国防工作流 | 收集正面竞争输赢和客户用例分层 |
| 创意视频预算邻接 | Runway、Synthesia、HeyGen、Pika | 中 | 相邻厂商已拿到融资或已有规模,争夺 AI 视频心智 | 客户访谈中拆分档案理解预算和生成式视频预算 |
| 内部自建 / 开源 | VideoLLaMA3、InternVideo、Video-ChatGPT | 中 | 开源项目让自托管原型具备可信度 | 量化评估、服务、治理和维护的总体拥有成本 |
| 人工元数据替代 | 传统 DAM/MAM 和人工打标 | 低至中 | 现状仍可信、可审计,但语义规模化能力弱 | 梳理模型输出需要人工复核或审计轨迹的工作流 |
| 定价压力和多供应商并用 | 所有计量型 API 和专精 SaaS 替代方案 | 高 | 买家可以把搜索、CV、生成和音频拆给不同供应商 | 要求提供留存、量级承诺和应用层粘性的证据 |
严重程度由作者基于公开证据评估;拿到私有客户输赢、生产基准和承诺用量定价数据后,应重新校准。
[CP028, CP030, CP032, CP033, CP035, CP037]简表显示 Twelve Labs 的竞争耐久性:专业化亮点被分发和定价压力抵消。
KPI 数值是作者的定性评分,不是经审计的运营指标;用于概括证据账本和未解缺口。
[CP041]3.4 反向观点:商品化、定价压力与证据缺口
反向情景是,在 Twelve Labs 锁定持久应用层之前,视频理解先变成更广泛多模态平台的一个功能。Mixpeek 直接攻击成本可预测性以及存储 / 控制权取舍。Google、Microsoft 和 AWS 可以压低价格或打包经典视频分析;OpenAI 和 Meta 则可以持续改进通用多模态模型,吸收更多视频推理任务。开源项目让有基础设施能力的团队保留可信的内部自建威胁。由此形成的尽调议程很具体:获取相对云巨头和直接对手的赢单 / 输单数据,比较实际承诺用量价格而不是标价,并用独立基准测试 Marengo 3.0 和 Pegasus 1.5 相对 Gemini、OpenAI、Reka 和开源系统的表现。没有这些私人或基准输入,本章可以支持细致的竞争观点,但不能证明持久定价伞。还要补充一个时间维度:Twelve Labs 今天可能在生产视频检索上享有专业玩家领先,但最大平台可以等品类成熟后,把足够可用的能力打包进既有 AI 套件。这是执行窗口问题,不是赢家通吃的二元市场。合同证据应作裁判。[CP028, CP030, CP032, CP033, CP035, CP038]
3.5 展示项
04财务情况
4.1 收入模型与定价架构
Twelve Labs 变现的是视频理解 API,而不是按席位优先的 SaaS 应用。公开定价界面展示了一组用量计量项:Marengo 视频索引、月度嵌入基础设施、Search API 查询、按模态计费的 Embed API 调用,以及 Pegasus 分析和输出 token。更多索引分钟、搜索和分析片段会带来更多可计费事件,因此客户活动到收入之间有清晰桥梁。企业合同增加第二层:高量客户可以从标价转向承诺用量或定制计划,AWS Marketplace 也为 Bedrock 式消费提供合作伙伴分发界面。该模型在财务上有吸引力,因为用量定价部分匹配算力 COGS,但它也让实际定价、折扣和工作负载组合成为关键尽调变量。官方定价是标价,不是 ARR、毛利率或净留存率;因此,在 Twelve Labs 提供客户级收入明细表之前,本章把所有私人财务指标都视为未披露。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入来源 | 机制 | 计费单位 | 公开状态 | 收入质量 | 尽调要求 |
|---|---|---|---|---|---|
| Marengo 视频索引 | 一次性将上传视频索引成可搜索表示 | $/video minute | 开发者标价已披露 | 如果用量增长且毛利率覆盖计算成本,质量较好 | 按客户和队列提供索引分钟数 |
| 向量嵌入基础设施 | 为生成的向量嵌入及下游搜索 / 分析提供月度服务 | $/indexed minute/month 计费 | 标价已披露 | 有经常性,但绑定保留索引 | 披露保留索引分钟数和流失客户索引删除情况 |
| Search API | 跨已索引视频库做语义搜索 | $/1,000 queries | 标价已披露 | 用量匹配度高;利润取决于查询成本 | 提供查询量、延迟层级和单次查询成本 |
| Embed API | 为视频、音频、图像和文本输入生成向量嵌入 | 按分钟或请求计费 | 各模态标价已披露 | 开发者收入可能具备可扩展性 | 按模态和工作负载规模拆分收入 |
| Pegasus Analyze / Segment | 视频转文本分析、结构化抽取和片段定义 | $/video minute 加输出 tokens | 标价和片段乘数已披露 | 用量匹配度强,但计算负载重 | 按用例提供 Pegasus 毛利率 |
| 企业合同 | 自定义定价、更高限额、微调和承诺用量条款 | 合同 / 承诺 | 自定义;无公开标价 | 若有最低承诺,收入质量可能较高 | 披露 ACV、期限、折扣和续约条款 |
收入来源来自官方标价和伙伴分销页面;当前组合和实际收入均为私有未披露。
[CI001, CI003, CI004, CI005, CI006, CI007]| 套餐或计量项 | 公开价格 / 条款 | 标价 vs. 实际 | 折扣或未知项 | 来源 |
|---|---|---|---|---|
| 免费计划 | 600 分钟视频索引;90 天索引访问权 | 标价权益 | 付费转化率未知 | Twelve Labs 定价 |
| Developer Marengo 索引 | 每视频分钟 $0.042 | 标价 | 量级折扣未披露 | Twelve Labs 定价 / 计算器 |
| 向量嵌入基础设施 | 每索引分钟每月 $0.0015 | 标价 | 实际存储 / 向量嵌入成本未知 | Twelve Labs 定价 / 计算器 |
| Search API | 每 1,000 次查询 $4 | 标价 | 查询组合和缓存情况未知 | Twelve Labs 定价计算器 |
| Embed API | 视频 $0.042/min;音频 $0.0083/min;图像 $0.10/1k;文本 $0.07/1k | 标价 | 实际综合费率未知 | 定价计算器和 UsagePricing |
| Pegasus Analyze | 每输入分钟 $0.0292,1k 输出 tokens $0.0075;Segment 按片段定义倍增 | 标价 | 工作负载复杂度可能改变可计费分钟数 | 定价页 / 计算器 |
| AWS Marketplace | 按合同或用量计价的市场上架;可能另收 AWS 基础设施费用 | 伙伴渠道条款 | AWS 私有报价和云条款未知 | AWS Marketplace |
标价表;实际企业折扣、抵扣额度和量级承诺未公开。
[CI002, CI003, CI004, CI005, CI006, CI007]客户视频活动先映射到多种用量计费器,再由企业合同和计算成本决定毛利。
定性桥接;标价计费器公开,但收入结构和毛利率未披露。
[CI003, CI004, CI005, CI006, CI007, CI008]4.2 单位经济模型与利润率驱动
核心单位经济问题是,可计费分钟和查询能否覆盖视频摄取、嵌入、存储、检索和生成的成本。公开 AI 基础设施基准带有警示意味:上线后推理和可观测性成本仍会变化,AI 原生毛利率常被认为低于经典 SaaS 毛利率。Twelve Labs 有两个缓释因素。第一,定价页披露的用量单位足够细,可以让重度用户为重度工作负载付费。第二,如果 Trainium 优化和 SageMaker 韧性降低训练与服务成本,Amazon 关系可能改善基础设施经济性。两点都不能证明利润率质量。缺失的尽调材料是按产品线分列、带日期的成本台账:每索引分钟成本、每 Pegasus 分钟成本、缓存命中率、AWS 额度和承诺支出条款、客户支持负担,以及按企业队列划分的利润率。在这些数据到来之前,标价应被视为变现框架,而不是有吸引力毛利率的证明。[CI009, CI015, CI018, CI019, CI029, CI030]
| 指标 | 数值 | 置信度 | 重要性 | 尽调要求 |
|---|---|---|---|---|
| 毛利率 | 低 | 检验用量定价能否覆盖 AI 推理和存储 COGS | 按 Marengo、Pegasus 和企业客户队列提供毛利率 | |
| 每索引分钟成本 | 低 | 直接对照 $0.042/min 索引价格的 COGS | 按摄取、向量嵌入、存储和检索导出成本台账 | |
| 每 Pegasus 分析分钟成本 | 低 | Pegasus 计算负载重,片段乘数可能改变经济性 | 按工作负载提供每分钟成本和输出 token 利润率 | |
| 推理在 AI 基础设施支出中的占比 | 基准 55–80%,公司值为空 | 中 | 揭示生产流量会让 COGS 持续随量波动的风险 | 按训练、推理、存储、可观测性拆分 Twelve Labs 支出 |
| AI 原生毛利率基准 | 基准 50–65%,公司值为空 | 中 | 基准显示毛利率低于传统 SaaS | 将公司毛利率与基准对账,并解释 AWS 影响 |
| CAC 回本 / 销售效率 | 低 | 企业合同可能掩盖长周期和高支持成本 | 提供新增 ARR、销售与营销支出、回本周期和销售周期 | |
| 净收入留存(NRR) | 低 | 客户扩大视频库后,用量扩张应体现在 NRR 中 | 按客户 logo 队列和产品线提供 GRR/NRR | |
| AWS 抵扣或折扣贡献 | 低 | 抵扣额度可能暂时抬高毛利率和现金跑道 | 提供 AWS 订单、抵扣额度、承诺支出和到期日 |
公司未知指标有意保留为空值;每个空值都对应具体尽调路径。
[CI004, CI007, CI015, CI018, CI019, CI029]毛利路径取决于标价、模态组合、推理成本、AWS 优化和企业支持负载。
没有公司毛利数据;节点把公开定价和外部 2026 年 AI 单位经济性基准合在一起。
[CI018, CI019, CI024, CI029, CI030, CI031]4.3 资本充足性与融资依赖
财务章节只能通过本地取证的主张引用融资时间线。最新公开融资证据显示 July 2026 完成 $100 million Series B,市场数据报道显示累计融资约 $207 million。公告中的资金用途——R&D、模型开发、地理扩张,以及 New York 和 London 办公室——指向增长投入,而不是自给自足声明。这笔总资本注入可能改善公司的谈判现金跑道,但手头现金、净烧钱、现金跑道月数、承诺云支出和债务均未公开。Twelve Labs 的资本强度集中在前沿模型 R&D、视频推理、企业支持和云承诺,而不是库存或借贷资产。AWS 关系有战略价值,但也可能带来公开文章看不见的合同支出义务或集中度风险。投资人应在审阅银行流水、董事会批准预算、AWS 订单表和收入 / 烧钱轨迹后,再承保资本充足性。[CI016, CI017, CI018, CI020, CI021, CI022]
| 项目 | 公开数值 / 状态 | 置信度 | 含义 | 尽调路径 |
|---|---|---|---|---|
| 最新总融资 | 2026 年 7 月 1 日宣布 $100M Series B | 高 | 增强资本缓冲,但不等于在手现金 | 确认总 / 净到账、交割日和当前银行余额 |
| 累计融资 | Tracxn 报道六轮融资共 $207M | 中 | 显示此前已有较大稀释,也有投资人支持 | 用董事会材料核对股权结构表和优先股条款 |
| 在手现金 | 低 | 只看融资总额无法计算现金跑道 | 提供银行流水和受限现金明细 | |
| 月度净烧钱 | 低 | 现金跑道和下一轮时间点的核心输入 | 提供过去 12 个月月度经营计划和实际数 | |
| 现金跑道(月) | 低 | 公开来源未披露现金除以烧钱速度 | 用现金、承诺支出、收入回款和招聘计划计算 | |
| 资金用途 | 研发、SF/Seoul 扩张,以及 New York/London 新办公室 | 中 | 增长投入很可能抬高运营费用 | 审查招聘计划、办公室成本和模型训练预算 |
| 债务 / 项目融资 | 未发现公开债务或授信额度 | 低 | 没有公开证据不等于没有义务 | 索取债务明细、云承诺、租约和担保 |
| 下一轮触发因素 | ARR 规模、毛利率和烧钱倍数未公开 | 低 | 无法只靠公开信息承销融资依赖 | 用实际 ARR 和毛利率桥接建模基准 / 下行情景 |
融资事实在本章本地引用;除非另有说明,现金、烧钱速度、现金跑道和债务仍属私有信息。
[CI016, CI017, CI020, CI027, CI037, CI038]已知公开融资数据精确,但当前运营指标缺失;基准只能勾勒可能的毛利压力。
区间来自公开事实或外部基准,不是 Twelve Labs 管理层指引;不主张当前 ARR 区间。
[CI016, CI020, CI026, CI030, CI031]Series B 资金流向研发、AWS 优化计算、企业部署和全球扩张;ARR 与毛利证据跑出来后,才会触发下一轮融资决策。
现金和烧钱速度未公开;流程展示资金用途类别和尽调闸口。
[CI016, CI017, CI018, CI024, CI037, CI038]4.4 公开财务缺口与投资判断
公开记录支持存在一个连贯的收入引擎,但不能证明该引擎的规模或质量。Latka 的历史 $4.2 million 2023 收入数据点只适合作为过时信号,而且其“未融资自举”标签与广泛报道的风险融资轮冲突,削弱了内部一致性。官方页面和第三方资料没有披露当前 ARR、毛利率、净收入留存率、回本周期、客户集中度、云额度或烧钱倍数。这让投资结论带条件:Twelve Labs 值得因用量制包装、企业分发和新近 Series B 获得认可,但在私人指标证明视频推理利润率可以扩展之前,财务承保应维持“继续研究”。第一份尽调包应包括按产品划分的 ARR、毛利率桥、头部客户集中度、月度烧钱、AWS 承诺、实际折扣,以及把消费增长与算力成本相连的两年预测。该资料包还应把董事会预测与计量用量队列对齐,让增长和成本假设可审计。[CI026, CI027, CI028, CI036, CI037, CI038]
| 缺失指标 | 公开证据状态 | 影响 | 精确尽调路径 |
|---|---|---|---|
| 当前 ARR / 收入运行率 | 官方未披露;Latka 只有 2023 年收入数据点 | 卡住收入规模和倍数分析 | 按月、产品、客户和地区获取 ARR 桥接 |
| 收入组合 | 计量项标价公开;索引 / 搜索 / 分析 / 企业的组合未公开 | 卡住收入质量和毛利率分析 | 按 SKU、用量计量项和企业计划导出收入 |
| 毛利率 / COGS | 无公开毛利率;AI 基准只是外部可比 | 卡住单位经济模型和估值置信度 | 提供毛利率桥接和云成本台账 |
| 实际定价与折扣 | 标价公开;私有报价和 AWS 抵扣额度未公开 | 可能让重度用户变得不赚钱 | 审查前 20 大合同、折扣、抵扣额度和最低消费承诺 |
| CAC、回本周期、销售周期 | 没有公开销售效率指标 | ARR 扩大前,企业 GTM 可能消耗大量资本 | 提供队列 CAC、回本周期、管线转化率和销售周期 |
| 烧钱速度和现金跑道 | 已披露融资;现金和烧钱速度未披露 | 阻碍资本充足性评估 | 提供现金、烧钱速度、招聘计划和已承诺云支出 |
| 债务、租赁和承诺 | 未发现公开义务 | 隐性承诺可能消耗 Series B 资金 | 提供债务明细、租赁、AWS 合同和表外义务 |
| 客户集中度 / NRR | 没有公开客户经济性 | 若模式健康,用量计费应体现扩张 | 提供头部客户 ARR、GRR、NRR 和流失原因 |
该缺口表把私有未披露指标转成尽调请求;空值表示已审阅公开来源均未披露该指标。
[CI026, CI027, CI028, CI036, CI037, CI038]4.5 展示项
05产品与技术
5.1 产品模块与客户工作流
Twelve Labs 更适合被理解为视频理解 API 和平台,而不是通用视频生成工具。客户工作流从导入视频或把系统指向视频开始,随后选择模型模态,通过异步任务或分析端点处理资产,再在下游应用中使用搜索、结构化元数据、摘要、分类或合规审核。Marengo 3.0 是检索和嵌入层;Pegasus 1.5 是生成式视频转文本和基于时间的元数据层。这个分工在商业上重要,因为买方可以从搜索和索引起步,再扩展到摘要、审核、切分或自动化工作流决策。公开证据支持开发者优先界面,包括 REST/JSON API、Python 和 JavaScript SDK、GitHub 仓库、npm 包以及 AWS Bedrock 接入。主要产品缺口不是 API 是否存在,而是私人客户能否在自己的视频库上反复获得精度、召回率、延迟、成本和人工审核结果。承保时,实际测试应看这些模块能否缩短真实客户任务,而不是迫使团队围绕脆弱的概念验证重建存储、元数据和审核员工作流。[CE001, CE002, CE003, CE007, CE008, CE009]
| 模块或资产 | 主要用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Marengo 3.0 | 开发者、媒体搜索团队、数据团队 | 文档和 Bedrock 材料中的当前生产模型 | 面向语义视频、音频、图像和文本检索的多模态嵌入 | 在客户素材库上验证私有精确率 / 召回率、延迟和成本 |
| Pegasus 1.5 | 剪辑师、合规分析师、开发者 | 有 TBM 发布证据的当前生成式视频转文本模型 | 基于 schema 的带时间戳元数据和长视频分析 | 按工作流取得误报率和幻觉率 |
| Search API | 开发者和档案运营人员 | 公开产品 / API 界面 | 基于自然语言和多模态的已索引视频搜索 | 与买方原有搜索和超大规模云厂商替代方案做基准对比 |
| Embed API | ML / 数据团队 | 公开 API / 模型文档和 AWS 模型文档 | 用于相似搜索、聚类和检索的可移植嵌入 | 确认向量数据库成本和刷新策略 |
| SDK 和开发者中心 | 应用工程师 | Python 和 JavaScript SDK,以及文档和代码库 | 缩短首次集成时间 | 审计版本节奏、问题响应和 SDK 兼容性 |
| Bedrock 部署 | AWS 企业买家 | 合作伙伴文档确认可用 | 为 AWS 账户缩短采购和治理路径 | 确认区域可用性、配额、定价和 SLA 归属 |
各行综合公开产品、文档和合作伙伴证据;成熟度指公开界面成熟度,不等于私有收入或续约证据。
[CE002, CE003, CE007, CE010, CE015, CE043]| 用户任务 | 当前工作流痛点 | Twelve Labs 解决方案 | 可衡量收益信号 | 局限 |
|---|---|---|---|---|
| 语义档案搜索 | 手工标签和转录文本搜索会漏掉视觉 / 音频上下文 | Marengo Search API 覆盖视觉、音频和文本模态 | 官方基准和产品说法显示检索更丰富 | 独立、买方特定的召回率仍未验证 |
| 结构化媒体元数据 | 人工切分长视频并手动录入元数据 | Pegasus 1.5 基于时间的元数据提取,输出 JSON schema | PRWeb 和官方文章称,最长两小时视频可输出带时间戳结果 | 需要在每个领域 schema 上验证 |
| 合规审查 | 人工审查员观看大规模内容库 | Pegasus 加 Mux 工作流标记上下文化合规事件 | Mux 官方文章称,它能加速大规模视频审查 | 高风险决策仍需要人工审查 |
| 推荐用嵌入 | 团队需要从视频资产生成可搜索向量 | Marengo Embed API 生成跨模态嵌入 | 512 维说法可能降低存储成本 | 没有公开的客户侧向量成本基准 |
| 开发者原型搭建 | 团队需要快速 API 访问和示例 | REST API、Python SDK、JavaScript SDK、开发者中心 | 文档展示索引 / 任务示例和包安装 | 支持质量和 SDK 采用指标为私有 |
收益列记录公开证据信号,不代表有保证的客户 ROI。
[CE005, CE008, CE009, CE013, CE014, CE020]技术栈拆成视频输入、模态处理、Marengo 检索、Pegasus 分析和应用输出。
分层综合自公开文档和合作伙伴描述,并非披露的内部参考架构。
[CE003, CE007, CE009, CE011, CE031, CE035]典型实施从视频摄取推进到处理、查询 / 分析调用、审核和下游自动化。
流程抽象自文档和产品页示例;客户实际流程会因 API 和部署路径而异。
[CE001, CE008, CE009, CE020, CE032, CE044]5.2 架构、部署与技术依赖
可见架构有四层:客户视频和元数据输入;跨视觉、音频和转录信号的模态处理;通过 Marengo 嵌入 / 搜索和 Pegasus 分析执行模型;再通过 API、SDK、Bedrock、合作伙伴面板或客户应用交付应用输出。Twelve Labs 没有公开完整训练栈,但来源显示其对超大规模云厂商和加速器基础设施有显著依赖。AWS 描述了 Bedrock 分发、SageMaker HyperPod 训练、MediaConvert 视频处理和以 S3 为中心的集成。Twelve Labs 也把 NVIDIA 加速计算作为部署和性能伙伴。这些是优势,因为它们降低采纳摩擦,并给企业买方熟悉的采购路径。它们也是依赖:宕机、Bedrock 模型覆盖、云成本、GPU/Trainium 经济性和合作伙伴路线图选择,都可能塑造客户体验和毛利率。因此,依赖图谱应与采购条款一并测试,因为最佳技术路径可能在直接 API、Bedrock 和私有企业部署之间不同。[CE011, CE015, CE016, CE017, CE018, CE019]
| 层级或组件 | 作用 | 依赖 | 风险 |
|---|---|---|---|
| 客户视频输入 | 原始资产、URL、上传文件、音频、图像、文本或 schema 提示词 | 客户数据权利和视频质量 | 源质量差或权利不清,会拉低输出并带来法律风险 |
| 模态选择 | 选择视觉、音频和转录处理 | Twelve Labs 模型选项和搜索选项 | 模态选错会增加成本或漏掉信号 |
| Marengo 嵌入层 | 将多模态内容转换为可搜索向量 | Twelve Labs 模型运行时和向量 / 索引基础设施 | 官方基准需要买方特定复现 |
| Pegasus 分析层 | 生成摘要、回答、分段和结构化元数据 | 模型质量、prompt / schema 设计和长上下文处理 | 幻觉和 schema 漂移需要审查控制 |
| AWS / Bedrock 路径 | 企业采购和托管推理路径 | Amazon Bedrock、SageMaker HyperPod、MediaConvert 与 S3 | 云故障、区域、配额或经济性都可能限制采用 |
| NVIDIA 加速 | GPU 支持的训练 / 推理性能路径 | NVIDIA GPU 软件和基础设施 | 加速器可用性和成本仍是外部依赖 |
架构根据公开文档、合作伙伴页面和研究论文推断;Twelve Labs 未公开完整内部系统设计。
[CE011, CE015, CE016, CE017, CE018, CE019]技术依赖图突出客户数据、Twelve Labs 模型、云 / 加速器伙伴和控制平面依赖。
依赖箭头表示公开证据中的依赖,不是独家供应关系。
[CE015, CE016, CE017, CE018, CE019, CE036]5.3 差异化、成熟度与发布阶段
产品看起来明显比研究演示更成熟。Marengo 3.0 有官方模型文档、生产 API 声明、Bedrock 文档、发布报道,以及嵌入效率和多模态检索改善的引用。Pegasus 1.5 有更清晰的应用切口:从长视频中生成 schema 定义、带时间戳的基于时间的元数据,直接映射到媒体运营、体育、合规和档案工作流。发布说明显示 2026 迭代活跃,包括批量分析和旧模型退役;NAB Show 报道则把 Twelve Labs 定位为走向全栈视频智能平台,而不只是模型基础设施。证据仍不均衡。多数性能声明来自官方或合作伙伴报道,公开记录还没有提供买方特定的可复现包、独立基准审计或客户侧错误率分布。因此,尽调应把差异化视为可信且有技术支撑,但在私人评估复现前不能完全承保。强技术尽调流程应复现一小组客户关键搜索、元数据 schema 和审核提示,再假设公开基准可以外推。[CE005, CE006, CE026, CE027, CE028, CE029]
| 日期或阶段 | 功能或里程碑 | 状态 | 影响 | 来源 |
|---|---|---|---|---|
| 2024-04 | Pegasus-1 技术报告 | 已发表研究论文 | 技术证明早于当前产品发布,并解释架构沿革 | arXiv |
| 2026-03-30 | Marengo 2.7 停止支持新的索引 / 搜索 / 嵌入 | 发布说明披露 | 客户必须迁移,并维持模型版本纪律 | Twelve Labs 文档 |
| 2026-04 | Pegasus 1.5 全面可用和 TBM | 已宣布发布 | 将分析从片段 QA 推向基于 schema 的长视频元数据 | PRWeb / Twelve Labs |
| 2026 NAB Show | 全栈平台定位和 Rodeo | 发布报道 | 应用层从仅 API 姿态向外扩展 | PRWeb / Sports Video Group |
| 2026 当前 | Twelve Labs 和 Amazon Bedrock 上的 Marengo 3.0 | 合作伙伴和媒体报道 | 扩大企业分发,也加深对 AWS 的依赖 | AWS / MarTech360 |
| 2026 当前 | SDK 和批量分析更新 | 文档可见 | 显示开发者界面仍在活跃迭代 | 发布说明 |
路线图表只使用公开发布证据;应索取私有承诺路线图和弃用政策。
[CE026, CE027, CE028, CE029, CE030, CE038]搜索 / 嵌入 / 文档的成熟度最强;独立基准和信任材料的证据较弱。
矩阵是定性成熟度记分卡,仅基于已抓取公开证据。
[CE010, CE015, CE020, CE021, CE026, CE037]5.4 信任、质量、合规与风险控制
Twelve Labs 对一家企业 API 私营 AI 初创公司而言有可信的信任信号:公司报告完成 SOC 2 Type 2,发布安全 / 隐私信息,记录模型输入限制,并提供许多企业安全团队已熟悉的 AWS 和 NVIDIA 合作路径。产品也天然适合内容合规:Pegasus 可以解释视觉、音频、对象、动作和时间语境,Mux 则为审核工作流提供视频基础设施。反向面是,视频理解输出仍可能错误、过度自信或对语境不敏感,尤其在审核、法律、监控或品牌安全场景。独立 AI 风险指引指出,人在回路验证和数据质量控制必不可少。公开来源也没有披露当前 SOC 2 报告、子处理方名单、模型训练数据使用条款、客户特定 SLA 或量化幻觉 / 误报率。这些不是致命缺口,但在高风险或受监管部署前都是门槛尽调项。投资含义很直接:信任控制对视频 AI 不是装饰;它们决定哪些工作负载能从实验进入受治理的生产。[CE013, CE014, CE020, CE021, CE022, CE023]
| 控制、认证或质量问题 | 状态 | 公开可见范围 | 缺口 |
|---|---|---|---|
| SOC 2 Type 2 | 公司博客称已完成 | 审计时点的平台安全 / 隐私承诺 | 需要当前报告、审计期间、排除系统和例外事项 |
| 隐私控制 | CSA 来源映射了行业控制 | 访问控制、加密、最小化、留存都属于判断标准 | 需要 Twelve Labs 特定的留存和子处理方细节 |
| 内容审核支持 | 作为 Mux 合规工作流提供支持 | 跨视觉、音频、动作、物体信号的上下文化视频理解 | 需要误报 / 漏报率和升级策略 |
| 可靠性 / 状态 | 第三方状态页列出组件和近期事故 | AIWatch 报道称,2026 年 7 月 16 日一场已解决事故影响了部分 API 功能 | 需要官方 SLA、正常运行时间历史和客户抵扣 |
| 幻觉 / 准确性控制 | 独立 AI 指引承认该风险 | 人在环验证是控制模式 | 需要按工作流设定的 QA 阈值和审查员工具 |
信任控制来自公开说法或行业控制类比;客户尽调应索取私有安全和可靠性材料。
[CE020, CE021, CE022, CE023, CE024, CE025]5.5 展示项
06客户情况
6.1 客户分层与买方图谱
公开客户证据支持分层客户基础,而不是单一横向买方。最强证明集中在媒体与娱乐、体育与广播、广告和品牌安全、商业视频、非营利档案、创作者营销合规、AI 训练数据工作流,以及开发者 / API 采纳。经济买方随细分市场变化:媒体运营、产品和技术负责人、内容运营、广告合规、电商推荐团队、档案所有者和开发者平台团队。用户通常是编辑、制片人、活动经理、数据团队或开发者,付款方则是企业采购团队、AWS Marketplace 买方或自助开发者。关键细微点在于,Twelve Labs 对安全、汽车、公共部门和政府有广泛解决方案叙事,但具名客户证明更集中在媒体、体育、广告、商业和档案工作流。这个点对尽调重要,因为分层证据不等于收入组合。具名引用证明产品解决了具体工作流痛点,但公开记录没有指出哪个垂直领域贡献最多签约额、哪个买方控制续约,或用量是否集中在少数档案量很大的账户。[CU001, CU002, CU009, CU013, CU014, CU016]
| 分层 | 买方 / 用户 / 付款方 | 主要用例 | 规模或证明 | 收入 / 战略价值 | 缺口 |
|---|---|---|---|---|---|
| 媒体与娱乐 | 媒体运营买方;剪辑师和授权团队;企业预算 | 档案搜索、场景元数据、宣传和授权工作流 | Twelve Labs 解决方案页面和 MLSE / SBS 证明 | 战略价值高,因为档案会变成可搜索库存 | 未披露分层收入或续约率 |
| 体育与广播 | 体育媒体运营;制作人和剪辑师;俱乐部或联盟预算 | 高光搜索、粉丝个性化、表现和档案搜索 | MLSE 和 Dyn 案例;体育解决方案页面 | 高,因为速度和个性化是重复工作流 | Dyn 公开证明缺少有日期的生产指标 |
| 广告 / 品牌安全 | 广告产品、合规和活动团队;出版商或市场平台付款方 | 品牌安全决策、上下文广告插播、创作者帖文验证 | Mantis 和 AffiliateNetwork 案例研究;广告解决方案页面 | 潜在价值高,因为它保护广告支出和合规 | 宣传的提升未披露样本量 |
| 电商视频 | 搜索和推荐团队;电商产品预算 | 视频上下文推荐和短片搜索 | GS SHOP 案例研究,含转化和订单指标 | 具战略性,因为它把视频 AI 绑到电商 KPI | 公开证据中只有一个具名电商客户 |
| 非营利 / 档案运营 | 传播和活动人员;非营利运营预算 | 搜索分散的一线影像和活动媒体 | UNICEF Korea 案例研究,含 8TB 档案和 95% 检索时间缩短 | 战略参考;收入价值未知 | 未披露续约或合同期限 |
| 开发者 / API 构建者 | 开发者和产品团队;自助或企业平台预算 | 构建语义搜索、Analyze、Embed 和自定义应用 | Developer Hub、免费定价计划、TechCrunch 报道 30,000+ 开发者 | 宽漏斗和生态价值 | 未披露开发者到付费转化 |
| 安全 / 监控 | 安全运营和分析师;机构 / 设施预算 | 事件重建、行为搜索、异常检测 | 官方解决方案页面和垂直领域需求报道 | 受监管环境中 ACV 可能较高 | 未找到具名公开安全客户 |
分层地图基于公开客户故事和官方解决方案页面;因未披露分层收入,收入价值为定性判断。
[CU001, CU009, CU013, CU014, CU016, CU018]公开证据支持一条路径:从手工处理视频的痛点,走向 API、云市场或合作伙伴主导的工作流嵌入。
旅程阶段综合公开案例路径,不是实测转化漏斗。
[CU003, CU005, CU007, CU010, CU011, CU019]6.2 采纳轨迹与生产证明
最好的客户证明不只是客户名录。Mantis 从 Q4 2025 概念验证,通过 AWS Marketplace 进入 March 2026 生产部署;GS SHOP 描述了从概念验证转向全面推出;UNICEF Korea 报告了带自动摄取的索引档案;MLSE 报告了明显的生产时间减少;Qencode 将 Twelve Labs 嵌入编码管线,作为原生智能输出;Protege 描述了两周交付视频数据集,而原本需要数月。这些故事显示出共同采纳路径:痛苦的人工搜索或审核、代表性内容测试、API 或合作伙伴集成,然后运营化上线。需要注意的是,部分引用,尤其是 SBS 和 Dyn,使用的是合作和机会语言,没有公开续约、合同价值或带日期的扩张证据。尽调中重要的区分不是具名和未具名客户,而是已测试内容、生产使用、嵌入式工作流、可衡量结果、续约和扩张。公开来源支持几个账户的前四个阶段,但很少支持最后两个。[CU003, CU004, CU005, CU006, CU007, CU008]
| 指标或信号 | 数值 | 日期 / 新鲜度 | 来源依据 | 置信度 | 影响 | 缺失分母 |
|---|---|---|---|---|---|---|
| 开发者采用 | 30,000+ 名开发者 | 2024 年 12 月报道,作为公开基线仍有参考价值 | TechCrunch 对 CEO 的采访 | 中 | 存在大型自助漏斗 | 未披露活跃开发者、付费转化和留存 |
| Mantis 采用路径 | 2025 年 Q4 PoC;2026 年 3 月投产;70–80+ 条 PoC 视频 | 以 2026 年案例研究为准 | Mantis 案例研究 | 中 | 有证据显示,可通过 AWS Marketplace 从试点推进到生产 | 合同规模和扩张量未知 |
| GS SHOP 成果 | +57.5% 下单客户;+29.4% 转化;+21.7% 点击 | 当前案例研究 | GS SHOP 案例研究 | 高 | 电商工作流能把视频智能绑到收入 KPI | 未披露实验设计和时长 |
| UNICEF Korea 档案部署 | 8TB+ 档案;约 200 小时 / 2TB 已索引;检索时间缩短 95% | 当前案例研究 | UNICEF Korea 案例研究 | 高 | 档案搜索能创造持久的日常工作流价值 | 授权价值和续约状态未知 |
| MLSE 工作流影响 | 16 小时降至 9 分钟;内容发现时间减少 97% | 当前案例研究 | MLSE 案例研究 | 高 | 强媒体运营生产率证明 | 未披露合同期限、NRR 或用户数 |
| Protege 交付速度 | 两周,对比传统交付 6+ 个月 | 当前案例研究 | Protege 案例研究 | 中 | AI 数据工作流可能因时间敏感的数据集创建而采购 | 重复频率和客户终端需求未知 |
| 免费计划 | 600 分钟免费;无需信用卡 | 2026 年抓取的定价页 | 定价页 | 高 | 低摩擦开发者试用路径 | 升级率和付费用量结构未知 |
这些数值都是公开客户声称或第三方报道的信号;不应解读为公司整体 ARR、活跃账户或留存指标。
[CU003, CU008, CU010, CU011, CU015, CU024]| 客户 | 分层 | 部署 / 用例 | 生产 / 试点 | 成果 | 局限 |
|---|---|---|---|---|---|
| Mantis Solutions / Reach PLC | 广告和出版商品牌安全 | 在 Amazon Bedrock 上用 Pegasus 自动化视频品牌安全与合规 | 截至 2026 年 3 月已投产 | PoC 测试 70–80+ 条视频;通过 AWS Marketplace 部署 | 未披露合同价值或续约 |
| Qencode | 视频基础设施平台 | 编码管线中的原生视频智能输出 | 已描述近似生产环境的平台集成 | 客户无需另建管线即可加入智能能力 | 新功能客户使用量未量化 |
| GS SHOP | 电商 / 直播购物 | 为 Short Pick 推荐和直播搜索提供视频上下文信号 | 已描述全面生产上线 | 下单客户 +57.5%,转化率 +29.4% | 实验时长和队列定义未披露 |
| UNICEF Korea | 非营利机构档案 | 可搜索的现场记录和活动媒体档案 | 已描述可投入生产的档案系统 | 检索时间减少 95%;约 200 小时 / 2TB 已索引 | 预算、续约和用户数未知 |
| MLSE | 体育和娱乐 | 面向体育制作和集锦工作流的语义搜索 | 引用的制作团队暗示已运营使用 | 从 16 小时降至 9 分钟,发现时间减少 97% | 未披露合同期限或扩张 |
| SBS | 广播媒体 | VFX 参考搜索、统计分析、短视频摘要 | 合作表述含糊 / 分阶段 | 可能复用媒体资产,并做场景级搜索 | 公开证明不够指向生产使用 |
| Dyn Media | 体育流媒体 | 搜索超越技术统计的数据片段,并支持编辑制作 | 已描述合作;生产指标未公开 | 解决每季 3,000+ 场直播赛事和手工搜索瓶颈 | 未披露明确上线日期或量化结果 |
| AffiliateNetwork | 创作者营销 / 广告合规 | 核验创作者视频是否符合品牌规则的 AI 帖文校验器 | 案例研究暗示已运营使用 | 面向 60,000+ 创作者生态,可秒级核验 | 未见留存或付费合同证据 |
| Protege | AI 训练数据 / 内容授权 | 在大型清权档案中找到精确片段 | 已描述运营合作 | 两周交付,对比传统路径 6+ 个月 | 终端客户集中度未披露 |
枚举不完整:仅覆盖已抓取来源中出现的具名公开案例研究和具名客户,并非完整客户群。
[CU001, CU003, CU005, CU007, CU008, CU010]示例性采用漏斗强调证据强度,而非实测转化率。
数值是证据强度指数,不是 Twelve Labs 转化或留存指标。
[CU003, CU007, CU008, CU010, CU011, CU015]命名生产状态和量化结果都可见时,证明最强。
矩阵评分是对已抓取公开证据的定性解读,不是客户健康度评分。
[CU001, CU003, CU005, CU008, CU010, CU011]6.3 留存、重复使用与满意度证据
留存质量是公开客户记录中最弱的一环。本章发现了强推荐语和运营结果,但没有公开 NRR、GRR、流失率、续约、队列留存、支持满意度或合同期限披露。有一些正向留存代理:客户描述了替代人工工作、把 Twelve Labs 嵌入持续运行的管线、自动索引新内容,以及位于日常创意工具内部的合作伙伴面板。这些代理意味着,一旦素材被索引、搜索成为运营一部分,工作流会有粘性。但它们不等于续约证明。独立满意度也偏薄:公开具名案例由公司发布,且有一份竞争对手比较可用,但已保留证据没有得到足够详细、可用于评分支持质量的第三方评论页面。客户可能喜欢产品,却仍把支出限制在试点、单一资料库或短期项目中;因此,尽调必须把使用热情与续约证据、预算扩张分开。[CU010, CU011, CU022, CU024, CU025, CU028]
| 指标 | 值 | 分群 | 置信度 | 尽调事项 |
|---|---|---|---|---|
| 净收入留存 | 全部客户 | 高:未披露 | 索取按队列、垂直行业、AWS Marketplace 与直销渠道拆分的 NRR | |
| 毛收入留存 | 全部客户 | 高:未披露 | 索取过去八个季度的 GRR 和 logo 留存桥表 | |
| 流失 / 部署失败 | 全部客户 | 中:未披露 | 索取流失 logo 清单、试点失败原因和复盘 | |
| 合同期限 | 企业客户 | 中:未披露 | 索取初始合同期中位数、续约期和预付用量承诺 | |
| 客户满意度 / NPS | 全部客户 | 中:未披露 | 索取 NPS、支持 CSAT、可回访客户清单和第三方评价导出 | |
| 重复使用代理指标 | 日常工作流中的自动摄取和搜索 | UNICEF、MLSE、Qencode、Avid 座谈会用户 | 中 | 核验 WAU、每索引分钟搜索次数和索引库扩张 |
| 开发者留存代理指标 | 免费计划和 Developer Hub 使用入口 | 开发者 | 中 | 索取免费分钟数到付费开发者和企业计划的队列转化 |
空值表示抓取来源中未找到公开指标;代理指标行不能替代续约或队列数据。
[CU022, CU024, CU025, CU034, CU039, CU040]没有公开真实留存队列;数值显示证据可见度,不代表客户留存。
未找到公开留存队列数据;0 表示证据不可得,代理值只表示相对公开可见度。
[CU034, CU039, CU040]6.4 扩张、集中度与反向观点
扩张可以来自更多索引分钟、额外模型、新用例、AWS Marketplace 采购、Bedrock 可用性、与数据平台、编辑面板和系统集成商的集成。即使没有披露 NRR,账户扩张逻辑也说得通。反向观点同样具体。公开证据没有披露客户数、生产账户数、头部客户占比、按垂直领域划分的收入或渠道集中度。一家竞争对手称 Twelve Labs 有多个计费计量项、仅云端部署、按区域处理限制,以及视频托管在 Twelve Labs 云环境中。这些问题对大型档案、受监管客户和有数据驻留要求的买方很重要。因此,在把公开牵引力视为持久收入前,尽调应索取队列留存、按索引分钟计的重复使用、续约率、AWS Marketplace 管线组合、安全 / 公共部门具名部署和收入集中度。结论是一种双面尽调姿态:公开证明强到足以支持推荐电话和资料室跟进,但还不足以假设高质量客户和用量数据。[CU019, CU020, CU021, CU022, CU025, CU027]
| 扩张驱动因素 | 集中风险 | 影响 | 尽调路径 |
|---|---|---|---|
| 更多索引分钟和保留嵌入向量 | 如果定价绑定索引分钟数,大型档案客户可能主导用量 | 上行空间高,但毛利率和集中度可能敏感 | 索取前 10 大客户和索引分钟队列的用量分布 |
| AWS Marketplace 与 Bedrock 渠道 | AWS 可能成为采购和基础设施上的大依赖 | 渠道加快企业采购,但可能集中销售管线 | 索取直销与 AWS Marketplace 的签约额、续约和毛利拆分 |
| 与 Databricks、Snowflake、Monks、Avid、Qencode 的伙伴集成 | 伙伴导入的用量可能是间接的,归因更难 | 能把触达面拓进企业工作流 | 索取伙伴来源 ARR、活跃部署和附加率 |
| 更多模型和用例 | 客户可能试点多种模式,却不续约生产负载 | 扩张说得通,但公开 NRR 尚未证明 | 索取按产品模块和续约队列拆分的扩张 ARR |
| 安全、政府和汽车需求 | 这些敏感垂直领域的具名证明稀疏 | ACV 可能很高,但采购和信任门槛也高 | 索取具名参考客户、部署状态、数据驻留控制和合规资料包 |
| 开发者自助漏斗 | 许多开发者可能停留在免费或试验阶段 | 广泛采用未必转成持久收入 | 索取免费转付费、开发者流失,以及按公司域名拆分的用量 |
| 仅云端部署和多计量定价摩擦 | 受监管客户可能更偏好自有存储或单租户替代方案 | 可能拖慢大型档案扩张,或触发竞品替换 | 用安全、法务和采购参考访谈验证买方异议 |
风险表结合了公开证据和明确证据缺口;头部客户集中度和渠道集中度未披露。
[CU019, CU021, CU024, CU025, CU032, CU035]6.5 展示项
07风险
7.1 监管与法律风险栈
法律风险不是泛泛的 AI 合规阴影,而是与公司的视频理解模态紧密绑定。Twelve Labs 的 AUP 已经禁止若干直接落入 EU AI Act 和隐私法危险区的用途,包括被禁止的 AI Act 实践、未经授权的人脸或身体特征识别、从图像或视频推断敏感属性、虚假信息、监控,以及缺少监督的高后果决策。这种缓释有用,但本质上是合同安排:它不能证明每个客户都有同意、数据来源、保留计划或部署控制。EU AI Act 将远程生物识别、情绪识别和生物特征分类列为高风险;GDPR/EDPB 指引也让生物识别和个人数据模型复用处在严格的目的、合法基础和匿名化分析之下。在美国,BIPA 式损害赔偿和 Colorado AI Act 义务,为就业、准入、安全或其他重大场景下的视频部署提供了原告或监管抓手。版权诉讼又增加了单独的数据集来源风险:近期 AI 训练案件区分了转化性训练论点,与非法获取或保留盗版材料。因此,Twelve Labs 的尽调路径应从数据集物料清单、退出 / 删除流程、EU AI Act 角色分类、生物识别用途护栏和赔偿限制开始,而不是假设模型质量本身就能承载风险。[CR001, CR002, CR003, CR004, CR005, CR006]
| 规则 / 许可 / 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 |
|---|---|---|---|---|---|---|---|
| EU AI Act 远程生物识别 / 生物特征分类 | 欧盟 | 高风险或禁止使用类别已生效或分阶段落地 | 中 | 极高 | AUP 禁止 AI Act 明令禁止的做法和未授权身份验证 | 客户部署仍可能触发提供方 / 部署方义务 | 把每个欧盟用例映射到 AI Act 角色、风险等级和合规证据 |
| GDPR / EDPB 个人数据和生物识别模型指引 | 欧盟 / EEA | 个人数据 AI 模型指引和 GDPR 权利仍然适用 | 中 | 高 | 隐私政策和 DPA 将 Twelve Labs 定位为客户内容的处理者 | 控制者同意、目的限制、删除和匿名化证据未公开 | 审查 DPA、删除证明、处理者分包商和生物识别合法依据模板 |
| Illinois BIPA 生物识别标识符和私人诉讼赔偿 | 伊利诺伊州 / 美国 | 法规和修正案诉讼仍在推进 | 中 | 高 | 除非法律允许,AUP 限制身份验证和敏感推断 | 上传视频中的人脸或声纹使用,可能给客户和供应商带来风险敞口 | 确认 BIPA 通知、书面授权责任分配、留存时间表和赔偿上限 |
| Colorado AI Act 高风险 AI 开发者 / 部署者义务 | 科罗拉多州 / 美国 | 义务随 2026 年框架和总检察长规则制定生效 | 中 | 高 | AUP 要求对重大决策保持监督 | 企业客户可能把视频 AI 归类为重大决策中的实质性因素 | 检查开发者文档、影响评估支持和歧视监控材料 |
| AI 版权训练数据案件,包括 Anthropic、Meta、OpenAI、Ross | 美国 | 2025-2026 年判例法和和解仍在演变 | 中 | 高 | 未公开数据集物料清单;条款限制竞争性合成训练用途 | 未经授权或盗版来源视频可能带来法定赔偿和禁令风险 | 要求提供数据集来源、许可记录、下架流程和输出相似性防护 |
法律清单按严重性排序,依据官方、监管和律所来源;可能性是尽调判断,并非公司披露。
[CR001, CR002, CR003, CR004, CR005, CR006]私有尽调前,法律 / 隐私 / IP 与云 / 模型可靠性落在剩余风险最高的单元格。
序数热力图把定性证据转换为类别,不是统计损失模型。
[CR001, CR004, CR011, CR018, CR021, CR029]7.2 运营、安全与模型质量风险
对一家私营 AI 初创公司而言,Twelve Labs 公布了相对成熟的安全姿态:静态和传输加密、最小权限、审计日志、漏洞扫描、事件响应、供应商风险审查、AWS 基础设施和 SOC 2 Type 2 公告。这些控制降低企业采购摩擦,但不等于视频认知的模型保证。最高运营风险在于,客户把搜索、切分、摘要或结构化抽取当作事实真相,用在误报、漏报或幻觉事件会产生法律或财务后果的工作流中。关于视频幻觉的学术工作指出,大型多模态模型可能给出看似可信但错误的答案;关于视觉提示注入的安全研究显示,图像 / 视频输入可以携带对抗性指令。SOC 2 可以证明信息安全控制,但 AI 特定控制仍需要评估集、红队日志、提示 / 模型变更管理、滥用监控、客户特定阈值和人在回路审核。在检查这些材料之前,尽管官方安全材料方向积极,剩余风险仍为中高。[CR015, CR016, CR017, CR018, CR019, CR020]
| 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 |
|---|---|---|---|---|---|
| 运营工作流中,视频理解产生幻觉或漏检事件 | 中 | 高 | 除公开模型声明外未知 | 误报 / 漏报可能影响客户信任或责任 | 需要按客户用例和复核阈值拆分的基准测试结果 |
| 视觉提示注入或对抗性视频内容 | 中 | 高 | 有公开 AUP 和安全控制,但无公开红队报告 | 攻击可操纵多模态模型解读 | 需要红队日志、输入清洗和事件升级证据 |
| 安全漏洞或不当访问客户视频 | 低-中 | 高 | SOC 2 Type 2、加密、最小权限、日志、AWS 控制 | SOC 2 本身不能证明 AI 专项治理 | 审查 SOC 2 报告、例外项、渗透测试和 ML 资产控制 |
| 被滥用于监控、深度伪造、虚假信息或高风险决策 | 中 | 高 | AUP 禁止多类滥用,并允许暂停服务 / 报告 | 执行效果取决于检测、客户审计权和日志 | 检查滥用监控、客户准入和下架历史 |
这些行结合 Twelve Labs 控制措施与独立 AI 安全和安全防护证据;缓释成熟度基于公开证据。
[CR015, CR016, CR017, CR018, CR019, CR020]监管、质量、安全和合作伙伴风险,会通过采用、毛利和融资信心传导。
有向边反映尽调中的因果逻辑,不是测得的弹性。
[CR012, CR017, CR020, CR024, CR027, CR030]7.3 合作伙伴、财务与执行风险
Series B 同时创造了现金跑道和集中度。多篇报道称 Twelve Labs 在 2026 融资 $100 million,由 NAVER Ventures 和 NEA 共同领投,Amazon 参与;Edaily 报道累计融资超过 $207 million,而 AWS 被描述为首选云供应商,未来模型将率先通过 AWS 发布,并在 Trainium 上优化。这在商业上可能很强,因为 AWS 可以通过 Bedrock 分发模型并承托算力规模,但它也让产品路线图、单位经济模型、发布顺序和买方感知都围绕单一云合作伙伴的芯片战略集中。公司也有一段反向平衡的 NVIDIA 历史:早期 NVIDIA 投资验证了技术雄心,但也凸显其暴露在 GPU 与 Trainium 基础设施竞赛之中。财务风险是资本强度,而不是即时偿付能力:视频基础模型需要大量训练和推理预算,而公开收入、毛利率、用量集中度和承诺云支出均未披露。执行风险同样具体。Jae Lee 是可见的创始人 CEO,融资叙事大量引用其投资逻辑,公司还在 San Francisco、Seoul、New York、London、开发者 API、Bedrock 分发和 Rodeo 等应用产品之间扩张。这样的广度带来关键人、招聘、产品聚焦和云成本风险,应设置明确的终止标准。[CR025, CR026, CR027, CR028, CR029, CR030]
| 依赖 | 交易对手 | 角色 | 集中度 | 失败情境 | 严重性 | 缓释措施 | 剩余风险敞口 |
|---|---|---|---|---|---|---|---|
| 首选云和 Trainium 优化 | AWS / Amazon | 云托管、Bedrock 分发、战略投资方 | 高 | AWS 条款或 Trainium 性能经济性限制发布节奏或利润率 | 高 | AWS 战略承诺和 Bedrock 渠道 | 最低消费、切换成本和路线图依赖未披露 |
| AI 加速器生态 | NVIDIA | 此前的战略投资方和 GPU 生态背书方 | 中 | GPU / Trainium 分歧抬高工程和采购复杂度 | 中 | 多投资方历史和 AWS 定制芯片路径 | 硬件可移植性和按加速器拆分的模型性能未公开 |
| 战略资本和韩国网络 | NAVER Ventures | Series B 共同领投方和早期支持者 | 中 | 投资方影响力或区域策略压缩伙伴选择空间 | 中 | NEA 共同领投和广泛投资团让治理更多元 | 董事会权利和集中度未公开 |
| 传播 / 分发和企业集成 | Autodesk、Bedrock、媒体 / 安全买方 | 进入生产工作流的路径 | 中 | 集成延迟或伙伴重新排序优先级会限制采用 | 中 | 产品发布和集成已有公开报道 | 合同条款、使用量和续约韧性未公开 |
依赖集中度根据 2026 年融资和产品分发报道推断;财务条款仍未公开。
[CR025, CR026, CR027, CR028, CR029, CR030]| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 创始人兼 CEO Jae Lee | 公开技术叙事和投资人信心高度集中在 Lee 身上 | 中 | 高 | 联合创始人、COO 和机构投资人提供一定后备力量 | 审查继任计划、留任方案和董事会紧急授权 |
| 工程 / 研究招聘 | 视频基础模型工作需要稀缺的多模态人才 | 中 | 高 | Series B 资金支持研发和全球扩张 | 检查招聘漏斗、离职率、薪酬消耗和论文 / 基准发布节奏 |
| 商业化执行 | 公司正从 API / 模型走向 Rodeo 等全栈应用 | 中 | 中 | Bedrock 和 Autodesk 集成拓宽分发 | 按产品线拆分 ARR、用量、续约和服务占比较高的收入 |
| 法务 / 信任组织 | EU AI Act、隐私和版权问题会扩大监管负担 | 中 | 中 | 公开法律条款、DPA 引用、AUP 和安全计划 | 确认法务、隐私、安全和事件响应是否有专门负责人 |
人员风险基于公开领导层能见度和扩张计划;私下组织架构深度仍是尽调缺口。
[CR033, CR034, CR035, CR036, CR037, CR038]Twelve Labs 的控制面把客户视频数据、AWS 基础设施、模型栈和法律护栏连在一起。
该图是尽调依赖模型,依据公开条款、融资报道和安全材料构建。
[CR015, CR025, CR026, CR028, CR031, CR032]7.4 缓释措施、监测指标与投资逻辑失效触发条件
可投路径不是等每个监管问题都消失,而是要求 Twelve Labs 拿出证据:识别、隔离风险的速度要快过产品扩张。最有用的尽调包应包括 SOC 2 报告访问权限、云承诺的经济账、按用例拆分的模型评估和幻觉基准、客户滥用 / 升级处理日志、数据集来源文件、DPA / 安全附录、EU AI Act 角色映射、生物识别和内容审核政策,以及董事会层面的接班计划。否决标准必须可量化。只要 Twelve Labs 无法证明训练和评估视频数据集的权利来源,企业合同在缺少配套控制的情况下把无上限生物识别 / IP 赔偿压到公司身上,AWS 最低消费或 Trainium 优化条款让毛利率结构性难看,模型评估日志显示监管或安全用例的误报 / 漏报率不可接受,或领导层过度集中且没有可信连续性方案,就应触发否决。反过来,如果公司能证明客户专属部署避开被禁止或高风险用途,重大流程强制人工复核,模型和滥用监控都有日志留痕,合作伙伴集中还有经济下限而不是开放式算力敞口,风险会实质下降。[CR036, CR037, CR038, CR039, CR040, CR041]
| 风险 | 可监控触发项 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 训练数据权利 | 数据集来源和许可审计 | 无法证明重要视频语料的权利、同意或删除路径 | 在完成整改,或赔偿 / 托管覆盖风险敞口前不投资 |
| 受监管生物识别用途 | EU AI Act/GDPR/BIPA 部署图谱 | 未经审查的人脸、声纹、敏感属性或重大决策用途 | 阻止受监管垂直领域扩张,并要求控制计划 |
| 模型可靠性 | 客户专属评估日志 | 高风险工作流中,误报 / 漏报超过客户定义容忍度 | 将用例限制在搜索辅助;不得进入自主或重大决策工作流 |
| 云集中度 | AWS 承诺和毛利率模型 | 最低消费或 Trainium 性能差距导致目标利润率无法达到 | 重估估值,或要求多云可移植性里程碑 |
| 安全 / 滥用 | SOC 2 例外项、红队报告、事件和滥用日志 | 重大未解决例外、提示注入漏洞,或反复发生绕过政策事件 | 敏感细分市场暂停部署,待整改完成 |
| 关键人物执行 | 继任和高管梯队审查 | 没有董事会批准的创始人 CEO 或关键研究负责人接续方案 | 把留任安排和治理保护作为融资前置条件 |
淘汰标准来自风险登记表中的尽调阈值;进入 IC 批准前,必须拿到私下证据支持这些阈值。
[CR039, CR040, CR041, CR042, CR043, CR044]7.5 附录
08估值
8.1 建议与价格纪律
建议:继续研究 / 观察,在假设约 $1B 入场价位跟踪,不应仅凭公开记录买入。最强证据是 Twelve Labs 在 2026 年从高质量战略和财务投资人处完成大额 Series B 轮,有差异化视频理解逻辑,并拿到可能影响推理经济性的 AWS 基础设施承诺。最弱证据是估值支撑:官方公告和已审阅报道没有披露投后估值,CB Insights 和 Tracxn 的估值或收入字段仍被隐藏,已抓取公开来源也没有补齐 ARR、毛利率、净烧钱、客户集中或优先权条款。因此,最终立场对价格敏感,而不是否定公司。如果私下尽调证明企业 ARR 有规模、重复扩张跑得通、Trainium 成本优势可持续,公司可以配得上溢价。没有这些证据时,公开证据支撑的是品类信念,而不是立刻按 $1B 承销。放到投委会语境,下一步不是对公司给出通过 / 否决,而是把私下证据转成价格、持股和下行保护。一个优先权普通、使用量加速的干净轮次,比又一篇媒体报道更快改变建议。[CV001, CV002, CV003, CV005, CV007, CV008]
| 维度 | 评估 | 证据基础 | 决策含义 |
|---|---|---|---|
| 建议 | 继续研究 / 跟踪 | Series B 轮规模大且具战略属性,但未公开估值或收入 | 没有私下证据,不按推定 $1B 买入 |
| 置信度 | 中低 | 融资事实交叉印证充分;财务指标不透明 | IC 前要求数据室确认 |
| 风险评级 | 高 | 算力经济性、大厂竞争、流动性集中和条款不透明仍是重大问题 | 分阶段尽调,不直接给出定价承诺 |
| 估值立场 | 偏高 / 未验证 | 可比公司体现溢价,但 Twelve Labs 缺少公开收入证据 | 入场必须绑定指标 |
| 决策触发 | 只有 ARR、利润率、留存和条款过线,才转向买入 | 尽调要求明确且可衡量 | 维持关系,并谈判信息权 |
该建议对价格敏感;抓取到的公开来源没有独立披露推定 $1B 背景。
[CV001, CV003, CV007, CV008, CV031, CV038]| 论点 | 证据支持 | 什么会改变判断 |
|---|---|---|
| 论点:视频是差异化 AI 模态 | Twelve Labs 描述了原生视频感知、记忆和推理架构 | 多个垂直领域出现客户工作负载证明,并有重复支出 |
| 论点:战略投资验证品类 | Amazon、NEA、NAVER 和既有投资者参与最新一轮 | 需要证明背书能转化为收入,而不只是基础设施赞助 |
| 论点:可比公司支持 AI 溢价定价 | Runway、Synthesia、Mistral 和 Perplexity 显示 AI 私募估值处于高位 | Twelve Labs 必须拿出可比企业级赢家的收入质量 |
| 反论点:估值未公开 | 官方和独立来源都未披露投后估值 | 已签署的投资条款清单和带投资者保护的股权结构表 |
| 反论点:收入不透明 | CB Insights 和 Tracxn 没有给出可用收入或倍数 | 按客户分群拆分的 ARR、NRR、毛利率和待履约订单 |
| 反论点:算力和大厂压力 | 反向来源指出算力昂贵,以及 Google / OpenAI 式竞争 | Trainium 基准测试、可持续云价格和可防守的准确率数据 |
每个论点都配有可证伪的尽调项,而不是被当成永久结论。
[CV005, CV006, CV007, CV014, CV017, CV024]决策先看战略验证,再过私有数据关口,之后才可能给出买入判断。
定性投委会逻辑,基于已获取的公开证据和私有数据缺口。
[CV001, CV005, CV007, CV022, CV031, CV039]投委会评分卡显示,强品类与高质量背书抵消了公开财务证据薄弱的短板。
序数 KPI 分数由公开证据综合而成;投委会阶段应替换为数据室指标。
[CV001, CV005, CV006, CV007, CV022, CV023]8.2 可比公司支撑与边界
可比公司给出很宽估值区间,但不能简单套读。Runway、Synthesia、Pika、Perplexity 和 Mistral 显示,只要投资人认为 AI 公司有品类领导地位,仍会支付高额溢价;但证据分成两类:有收入支撑的企业软件,和前沿级资本叙事。Synthesia 是最干净的企业视频可比项,因为 Sacra 报告其 ARR 可观、投后估值 $4B;Runway 是规模更大的视频 AI 创作 / 世界模型可比项;Pika 是更小的消费创作可比项,其被引用的风险直接映射到 Twelve Labs 的下行;Perplexity 和 Mistral 则说明,一旦投资人相信平台迁移正在发生,基础模型溢价可以走到多极端。Twelve Labs 卡在这些类别之间:它有基础模型叙事和战略投资人,但公开来源没有给出收入证据,无法把可比组合转成有把握的倍数。[CV012, CV013, CV014, CV015, CV016, CV017]
| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 适用性 | 局限 |
|---|---|---|---|---|
| Twelve Labs | 融资和披露指标 | $100M Series B;累计融资 $207M-$210M;公开估值未披露 | 直接标的和入场价锚点 | 缺少 ARR、收入倍数和优先权条款 |
| Runway | 视频 AI 生成 / 世界模型 | CB Insights:2026 年 2 月估值 $5.0B-$5.315B;2025 年收入 $90M | 专业视频 AI 估值上界可比对象 | 生成工作流不同于视频理解 API |
| Synthesia | 企业 AI 视频 SaaS | Sacra:2026 年 1 月投后 $4B;2025 年收入估计 $145.91M | 收入支撑最强的企业视频可比对象 | 数字人 / 培训 SaaS 不同于搜索 / 推理基础设施 |
| Pika | 消费者 / 创作者 AI 视频 | Sacra:2024 年估值 $470M;融资 $135M | 视频 AI 商品化和算力成本的下行可比对象 | 消费者生成业务模式不同于企业 API |
| Perplexity | AI 应用 / 搜索 | CB Insights:2025 年 9 月估值 $20B;2025 年收入参照 $100M | 当使用量可见时,AI 应用可拿到极端溢价 | 搜索应用,不是视频基础设施 |
| Mistral AI | 基础模型实验室 | CB Insights:2025 年 9 月估值 $11.7B-$13.7B;2026 年收入参照 $400M | 界定基础模型溢价和战略稀缺性 | 更宽的 LLM 平台,规模和资本需求不同 |
本章抓取的估值可比私有 AI 公司样本清单;数值混合了公司表述和分析师 / 市场数据页面。
[CV001, CV003, CV007, CV012, CV013, CV014]投资建议最受收入和毛利率证据牵动,而不是更多融资头条。
分数是投委会敏感度的 1-5 序数估计,不是公司指标。
[CV017, CV026, CV034, CV035, CV040, CV041]在确切持股和稀释未知前,区间围绕假定 $1B 入场价展示投资测算结果。
数值区间是为尽调框架设定的情景估计;公开来源未披露 Twelve Labs 的 ARR 或估值。
[CV013, CV015, CV016, CV031, CV032, CV033]8.3 下行、稀释与市场风险
反向逻辑不是 Twelve Labs 没有真实市场,而是 2026 年 AI 融资行情可能在收入质量可见前就为叙事付太多钱。市场来源描述了创纪录 AI 融资、资本极度集中、交易数量收缩、流动性选择性释放和显著 AI 估值溢价。这些条件解释了 Twelve Labs 为什么能募到战略轮,也让入场纪律更重要。已报道累计融资超过 $200M,确实可能带来优先权和期权池悬挂,但公开记录没有披露清算条款或所有权。Companies House 文件确认了英国实体,却不能证明经营收入或资本结构深度。AWS 战略背书若能降低成本并改善分销,就是护城河;若经济账靠云信用额度驱动、非排他,或不如云厂商原生视频产品,也会变成依赖风险。同一组证据也要求辛迪加流程有纪律:只有当优先权瀑布、云义务和客户经济性被绑定到可衡量下行触发器后,才为定价配额预留资本。[CV010, CV011, CV022, CV023, CV024, CV025]
| 触发项 | 阈值或事件 | 对论点的影响 | 行动含义 |
|---|---|---|---|
| 收入证据缺失 | NDA 下仍不披露当前 ARR、客户分群扩张或待履约订单 | 无法承销推定 $1B 价格 | 继续跟踪 / 不建议发投资条款清单 |
| 单位经济性弱 | 按工作负载拆分的毛利率没有随规模或 Trainium 优化改善 | 视频推理成本侵蚀软件式倍数 | 要求价格重置,否则放弃 |
| 战略依赖 | AWS 条款偏促销、不可迁移,或造成高额承诺支出 | 伙伴验证变成毛利率 / 依赖风险 | 要求合同审查和下行情景模型 |
| 股权结构包袱 | 参与型优先股、高额清算优先栈,或大规模期权池刷新 | 即使企业价值增长,普通股回报也受损 | 重新谈入场条件,或回避 |
| 竞争冲击 | 超大云厂商原生产品以更低价格达到同等检索质量 | 独立视频智能价值被压缩 | 暂停,直到拿到赢单 / 输单和基准证据 |
| 流动性压缩 | 二级或退出市场将应用 AI 倍数大幅下调 | 持有期和估值标记风险上升 | 只有拿到更高持股或更低价格才推进 |
淘汰标准把未解决的估值风险转成可监控的投资动作。
[CV017, CV022, CV023, CV024, CV029, CV030]8.4 情景与最终尽调要求
因此,决策框架是有条件的。乐观情景需要证明 Twelve Labs 不只是有前景的模型供应商,而是生产级视频智能层,能撬动扩张中的企业预算、强留存和可持续单元经济。基准情景假设公司仍有战略重要性,但私下可验证数据仍薄,投资人应守住准入、尽调权和价格纪律,而不是追逐新闻标题里的高估值。悲观情景假设 ARR 仍处早期,毛利率被视频推理压缩,大型科技公司竞争或客户集中限制退出选择。最高价值尽调不是再做一份通用市场扫描,而是向数据室索取 ARR、队列扩张、工作负载毛利率、云承诺、股权结构条款、销售管线转化和按垂直行业拆分的客户证据。买入建议必须要求这些材料达到明确门槛;如果公司拿不出来,就只应留在观察名单。[CV032, CV033, CV034, CV037, CV040, CV042]
| 情景 | 假设 | 估值 / 回报逻辑 | 概率信号 | 下行触发 |
|---|---|---|---|---|
| 乐观 | 企业 ARR 快速扩张;AWS 降低单位成本;留存强 | 若摊薄后退出估值达到 $4B 至 $6B,推定 $1B 入场价可继续复利 | 大客户扩展到多个视频工作流 | ARR 或利润率证据未出现 |
| 基准 | 品类真实,但公开指标仍不足 | 持续跟踪,直到私下指标支撑价格,或入场价重置到更低风险 | 战略投资者和市场可比公司保留期权价值 | 无法进入数据室,或条款差于标准 1x 非参与优先股 |
| 悲观 | 收入仍早期、毛利率弱,超大云厂商把视频理解商品化 | 推定 $1B 入场价可能遇到降价轮或二级退出持平 | 不利的市场集中和算力风险证据占主导 | AWS 经济性由抵扣额度驱动,或客户集中度高 |
情景数值是承销区间,不是公司披露的预测;实际回报需要股权结构和摊薄数据。
[CV017, CV022, CV023, CV029, CV030, CV032]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 估值和条款 | 投后估值、投前估值、期权池、清算优先权、按比例认购权、补充协议 | 决定推定 $1B 入场价能否产生 VC 回报 | 公司 CFO / 法务;审查融资文件 |
| 收入质量 | ARR、确认收入、待履约订单、NRR、总留存、客户集中度 | 区分战略热度和可重复的企业需求 | 财务数据室,加上头部客户分群分析 |
| 单位经济性 | 按索引、搜索、生成输出、存储和支持拆分的毛利率 | 视频 AI 可能很吃算力;倍数取决于毛利率路径 | 云账单和工作负载成本审查 |
| AWS 经济性 | Trainium 基准、抵扣额度、承诺支出、排他性和数据驻留条款 | 验证 AWS 伙伴关系改善还是束缚经济性 | 审查 MSA、订单表和基准测试 |
| 客户证明 | 具名生产客户、ACV、扩张、流失和部署状态 | 支撑乐观情景和退出准备度 | 客户访谈和合同抽样 |
| 竞争基准 | 准确率、延迟和价格对比 Google、OpenAI、Runway 及其他视频模型 | 检验面对超大云厂商和专业厂商时的防守能力 | 技术尽调和盲测基准 |
| 退出路径 | 战略收购方地图、二级市场兴趣、IPO 准备度和必要规模里程碑 | 决定持有期和目标回报 | 投行 / 二级市场核查和董事会计划审查 |
每个尽调问题都服务于一个目标:把建议从跟踪推进到买入、重定价或回避。
[CV034, CV035, CV036, CV037, CV042, CV043]8.5 附录
免责声明
本报告是基于截至 2026-07-21 的公开来源整理的研究与尽调辅助材料,不构成投资建议。私有财务指标、估值和交易条款不可得;凡标注为据报道或假定的数字,在任何投资决策前都需要一手确认。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | TwelveLabs is a video intelligence platform and API company focused on making video searchable, analyzable, and usable by AI systems. | 高 | SO001, SO012 |
| CO002 | TwelveLabs is headquartered in San Francisco and describes operations in Seoul, New York, Los Angeles, and London. | 高 | SO003, SO012 |
| CO003 | The careers page lists offices in San Francisco, Seoul, New York, London, and Pangyo with street-level addresses for each location. | 高 | SO003, SO012 |
| CO004 | Public company databases support a 2021 founding history, with Tracxn showing a March 31, 2021 legal-entity incorporation and Crunchbase listing Founded Mar 2021. | 中 | SO024, SO025 |
| CO005 | Jae Lee is consistently identified as TwelveLabs co-founder and CEO. | 高 | SO012, SO023, SO027, SO028 |
| CO006 | Jae Lee traces the company origin to 2021 work with four close friends from Korean Cyber Command, reinforcing a founder-origin story around video understanding. | 中 | SO027 |
| CO007 | CB Insights names Aiden Lee, Dave Chung, Jae Lee, SJ Kim, and Soyoung Lee as Twelve Labs founders. | 中 | SO026 |
| CO008 | TwelveLabs publicly highlights advisors including Fei-Fei Li, Silvio Savarese, Jeffrey Katzenberg, Alex Wang, Lukas Biewald, Nicolas Dessaigne, and Jay Simons. | 中 | SO002 |
| CO009 | The product surface centers on Search, Analyze, and Embed workflows built around the Marengo and Pegasus model families. | 中 | SO005, SO006 |
| CO010 | Marengo maps visual, audio, speech, and on-screen text signals into searchable video representations and supports cross-modal retrieval. | 高 | SO005, SO011, SO012 |
| CO011 | Pegasus is described as a video-language model that turns video representations into grounded descriptions, answers, summaries, and structured metadata. | 高 | SO005, SO008, SO012 |
| CO012 | Marengo 3.0 powers TwelveLabs Embed API and Search API and is claimed to use a 512-dimension embedding for lower storage and faster search. | 中 | SO007 |
| CO013 | Pegasus 1.5 shifts from clip-based question answering toward schema-first time-based metadata extraction across full videos. | 中 | SO008 |
| CO014 | TwelveLabs models are distributed through Amazon Bedrock and through TwelveLabs own API. | 高 | SO009, SO012 |
| CO015 | AWS is described as TwelveLabs preferred cloud provider under a multiyear commitment that includes optimizing video inference workloads on AWS Trainium chips. | 高 | SO012, SO014, SO016, SO032 |
| CO016 | TwelveLabs presents NVIDIA as an acceleration partner, and 2024 Series A materials say the platform integrates NVIDIA H100, L40S, Triton Inference Server, and TensorRT. | 中 | SO010, SO018 |
| CO017 | TwelveLabs announced a $100 million Series B on July 1, 2026 co-led by NEA and NAVER Ventures. | 高 | SO011, SO012, SO013, SO014, SO032 |
| CO018 | The Series B participant list includes Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures. | 高 | SO011, SO012, SO013 |
| CO019 | Series B proceeds are earmarked for R&D, Marengo and Pegasus advancement, Video Cognition System scaling, team building, and geographic expansion. | 高 | SO011, SO012, SO013 |
| CO020 | Current market-data sources support total equity funding around $207 million to $207.1 million after the Series B. | 高 | SO016, SO024, SO026 |
| CO021 | Crunchbase showed a stale or conflicting profile with $107.1 million raised and a last funding round in June 2025, materially below post-Series-B sources. | 中 | SO025 |
| CO022 | TwelveLabs Series A was a $50 million round co-led by NEA and NVIDIA NVentures. | 高 | SO017, SO018, SO019, SO020, SO030 |
| CO023 | Series A participation included previous investors such as Index Ventures, Radical Ventures, WndrCo, and Korea Investment Partners. | 高 | SO017, SO018, SO019 |
| CO024 | The 2024 Series A announcement said TwelveLabs planned to add more than 50 employees by year-end and nearly double headcount. | 中 | SO018, SO019 |
| CO025 | Series A materials said more than 30,000 users were using TwelveLabs APIs across sports, media and entertainment, advertising, automotive, and security. | 中 | SO018 |
| CO026 | In December 2022 Twelve Labs raised a $12 million seed extension led by Radical Ventures with Index Ventures, WndrCo, Spring Ventures, and angels participating. | 高 | SO022, SO033, SO019 |
| CO027 | The earlier $5 million seed round was led by Index Ventures with Radical Ventures, Expa, Techstars Seattle, and named angel investors participating. | 中 | SO021 |
| CO028 | Bloomberg-syndicated reporting in the Los Angeles Times says TwelveLabs had a team of around 200, evenly split between Seoul and San Francisco. | 高 | SO032, SO024 |
| CO029 | Reported customers include Hollywood studios, advertising companies, social media influencers, sports franchise owners, Maple Leaf Sports & Entertainment, AMC Global Media, and UNICEF. | 中 | SO032 |
| CO030 | No retained public source disclosed ARR, revenue run-rate, gross margin, burn, or runway; those cover metrics remain private-evidence diligence gaps. | 中 | SO011, SO012, SO024, SO026 |
| CO031 | The fetched official and wire Series B announcements did not disclose a post-money valuation, so any unicorn valuation should be treated as unverified until a citable source is obtained. | 中 | SO011, SO012, SO013 |
| CO032 | TwelveLabs planned New York and London expansion alongside continuing investment in San Francisco and Seoul. | 高 | SO012, SO013, SO003 |
| CO033 | TwelveLabs can be purchased or accessed through AWS Marketplace and Amazon Bedrock according to the AWS partner page. | 中 | SO009 |
| CO034 | The company cites demand and use cases across media, entertainment, advertising, government, security, sports, automotive, creators, and archives. | 高 | SO012, SO014, SO032 |
| CO035 | Both GlobeNewswire and the Los Angeles Times/Bloomberg report frame video as roughly 90% of world data yet largely opaque to machines. | 高 | SO012, SO032 |
| CO036 | No fetched source identified a lawsuit, sanction, enforcement action, or regulatory proceeding against TwelveLabs; adverse diligence is therefore centered on data conflicts and missing private metrics. | 中 | SO024, SO025, SO026, SO033 |
| CO037 | Key-person dependence is material because major financing, product, and investor narratives repeatedly quote or center Jae Lee as CEO and originator of the thesis. | 高 | SO011, SO012, SO023, SO027, SO028 |
| CO038 | The official about page says the team began with twelve members spanning language, video, machine learning, and perception expertise. | 中 | SO002 |
| CO039 | The World Economic Forum profile says Jae Lee serves on the board of the Republic of Korea Foundation Model Association and holds a UC Berkeley EECS degree. | 中 | SO028 |
| CO040 | CB Insights describes Twelve Labs as San Francisco-based, founded in 2021, backed by Amazon, Index Ventures, Korea Investment Partners, Naver Ventures, NEA, Quadrille, Radical, and Red Bull. | 中 | SO026 |
| CO041 | The current stage is best treated as private Series B after the July 2026 financing. | 高 | SO012, SO024, SO026 |
| CO042 | The Series B and careers evidence together support a 2026 operating footprint that has moved beyond a two-office San Francisco-Seoul setup into broader global coverage. | 高 | SO003, SO012, SO013 |
| CO043 | TwelveLabs received repeated external recognition before 2026, including CB Insights AI 100 appearances and Fast Company Most Innovative Companies according to the WEF profile. | 中 | SO028 |
| CO044 | TechCrunch reported that Google, Microsoft, and Amazon offer video-recognition services, creating a competitive context even as Lee argued TwelveLabs is differentiated by fine-tuning and context understanding. | 中 | SO033 |
| CO045 | The company took a first application-layer step with Rodeo shortly before the Series B, moving beyond model APIs toward applications. | 中 | SO012 |
| CO046 | TechCrunch reported in December 2024 that TwelveLabs added Yoon Kim, former SK Telecom CTO and a Siri architect, as president and chief strategy officer. | 中 | SO023 |
| CM001 | Twelve Labs offers a video intelligence platform and API for search, analysis, embeddings, and reasoning across video content. | 高 | SM001, SM002 |
| CM002 | Twelve Labs positions its Marengo and Pegasus models as video-native systems for multimodal embedding and video-language reasoning. | 中 | SM001 |
| CM003 | Current market reports produce materially different video-analytics estimates, making a range more defensible than one point TAM. | 高 | SM003, SM007, SM011, SM019, SM020, SM022, SM023 |
| CM004 | Grand View Research estimated the global video analytics market at USD 12.71 billion in 2024 and projected USD 37.84 billion by 2030 at a 19.5% CAGR. | 中 | SM003 |
| CM005 | Precedence Research estimated the global video analytics market at USD 18.53 billion in 2026 and USD 109.85 billion by 2035 at a 21.94% CAGR. | 中 | SM007 |
| CM006 | Mordor Intelligence projected the video analytics market at USD 15.04 billion in 2026 and USD 33.74 billion by 2030 at a 22.18% CAGR. | 中 | SM011 |
| CM007 | The Business Research Company reported the video analytics market at USD 11.59 billion in 2026 and USD 24.73 billion in 2030. | 中 | SM019 |
| CM008 | IMARC estimated the global video analytics market at USD 9.8 billion in 2025 and USD 35.3 billion by 2034, implying a 14.81% CAGR during 2026-2034. | 中 | SM020 |
| CM009 | Polaris reported a 2026 video analytics market size of USD 17.62 billion, a 2034 forecast of USD 72.43 billion, and a 19.3% CAGR. | 中 | SM023 |
| CM010 | Precedence Research estimated the multimodal AI market at USD 3.43 billion in 2026 and USD 51.76 billion by 2035 at a 35.34% CAGR. | 中 | SM008 |
| CM011 | MarketsandMarkets projected the multimodal AI market to reach USD 4.5 billion by 2028 at a 35.0% CAGR and listed Twelve Labs among providers. | 中 | SM006 |
| CM012 | Market.us projected the multimodal AI market to reach USD 26.5 billion by 2033 from USD 1.4 billion in 2023 at a 34.2% CAGR. | 中 | SM024 |
| CM013 | Precedence Research estimated the generative AI market at USD 55.51 billion in 2026 and USD 1,206.24 billion by 2035. | 中 | SM009 |
| CM014 | Fortune Business Insights estimated the generative AI market at USD 161 billion in 2026 and USD 1,260.15 billion by 2034, materially above Precedence's 2026 value. | 中 | SM010 |
| CM015 | Grand View Research estimated the computer vision market at USD 19.82 billion in 2024 and USD 58.29 billion by 2030 at a 19.8% CAGR. | 中 | SM004 |
| CM016 | MarketsandMarkets projected the video surveillance market to reach USD 88.06 billion by 2031 from USD 56.11 billion in 2025 at a 7.8% CAGR. | 中 | SM005 |
| CM017 | Grand View Research reported the sports analytics market at USD 7.0 billion in 2026 and USD 23.1 billion by 2033 at an 18.5% CAGR. | 中 | SM012 |
| CM018 | MarketsandMarkets projected sports analytics to grow from USD 2.29 billion in 2025 to USD 4.75 billion by 2030 and AI in sports to reach USD 2.61 billion by 2030. | 中 | SM013 |
| CM019 | Mordor Intelligence projected the enterprise video market at USD 28.98 billion in 2026 and USD 46.93 billion by 2031, with video analytics growing faster than the overall market. | 中 | SM025 |
| CM020 | Grand View Research reported enterprise video at USD 16.39 billion in 2021, with video content management and marketing/client engagement segments expected to grow at double-digit CAGRs. | 中 | SM014 |
| CM021 | Precedence Research estimated automotive AI at USD 5.80 billion in 2026 and USD 58.99 billion by 2035, with autonomous driving applications growing rapidly. | 中 | SM015 |
| CM022 | The most defensible included market boundary is video search, retrieval, analysis, embeddings, video-to-text, and corpus reasoning rather than base camera or storage hardware. | 高 | SM001, SM002, SM005, SM011 |
| CM023 | Relevant adjacencies for Twelve Labs include media operations, security analytics, sports analytics, advertising video workflows, automotive perception, and enterprise developer platforms. | 高 | SM012, SM014, SM015, SM019, SM025 |
| CM024 | Excluded spend includes cameras, monitors, storage appliances, generic video conferencing, and broad AI spend that does not involve video understanding. | 中 | SM005, SM014, SM025 |
| CM025 | Because published market definitions overlap, TAM, SAM, and SOM should be treated as separate lenses rather than additive layers. | 高 | SM003, SM007, SM008, SM009, SM011, SM019 |
| CM026 | A broad TAM proxy for Twelve Labs is the 2026 generative-AI market estimate of USD 55.51 billion, but it is intentionally broader than serviceable video understanding. | 中 | SM009 |
| CM027 | A core SAM proxy is Mordor's 2026 global video analytics estimate of USD 15.04 billion. | 中 | SM011 |
| CM028 | A narrow beachhead proxy is Precedence's 2026 multimodal AI estimate of USD 3.43 billion. | 中 | SM008 |
| CM029 | A reasonable 2026 video-analytics range is low USD 11.59 billion, base USD 15.04 billion, and high USD 18.53 billion across three market publishers. | 中 | SM019, SM011, SM007 |
| CM030 | Media and entertainment buyers for video intelligence are likely media operations, archive, product, post-production, and compliance teams that need searchable and summarized video libraries. | 中 | SM001, SM014, SM025 |
| CM031 | Security and surveillance buyers are public safety, security operations, facilities, and smart-city teams using video analytics for detection, alerts, incident review, and monitoring. | 中 | SM005, SM011, SM023 |
| CM032 | Sports teams, leagues, and broadcasters are plausible buyers because sports analytics includes player performance, tactical analysis, broadcast management, and video analytics use cases. | 中 | SM012, SM013 |
| CM033 | Advertising and marketing buyers can use video AI for content analysis, personalized video workflows, campaign asset classification, and creative review. | 中 | SM009, SM014 |
| CM034 | Automotive ADAS and autonomous-driving teams are plausible users of video search and reasoning because automotive AI and computer vision markets depend on perception and visual data workflows. | 中 | SM015, SM004 |
| CM035 | The adoption path for video-understanding AI runs from video-corpus pain to API pilot, benchmark, integration and governance review, workflow rollout, and ROI-based expansion. | 中 | SM001, SM002, SM016, SM018 |
| CM036 | Growth in camera networks, smart cities, and video data creates demand for automated analysis and reduces reliance on manual monitoring. | 中 | SM003, SM005, SM023 |
| CM037 | AI improves video analytics by enabling object detection, behavior analysis, predictive insights, automated monitoring, embeddings, and searchable video moments. | 高 | SM001, SM002, SM011, SM023 |
| CM038 | Cloud, edge, and SaaS deployment trends support video analytics and enterprise video adoption while changing integration requirements. | 中 | SM011, SM025, SM014 |
| CM039 | Multimodal and generative AI growth expands the set of video search, reasoning, summarization, and generation use cases available to enterprise buyers. | 中 | SM008, SM009, SM010, SM024 |
| CM040 | The EU AI Act bans or restricts several sensitive video-AI practices, including untargeted scraping of CCTV for facial recognition databases, emotion recognition in workplaces and education, and real-time remote biometric identification for law enforcement in public spaces. | 中 | SM017 |
| CM041 | NIST's AI Risk Management Framework reinforces that trustworthy AI deployments require structured risk identification, measurement, management, and governance. | 中 | SM018 |
| CM042 | RAND reported that more than 80% of AI projects fail and cited misdefined problems, inadequate data, technology chasing, infrastructure gaps, and excessive task difficulty as root causes. | 中 | SM016 |
| CM043 | Polaris identified data privacy, VMS integration complexity, higher compute and storage requirements, and false positives or false alarms as constraints on video analytics adoption. | 中 | SM023 |
| CM044 | BCG warned that few executives view GenAI cost as their top solution-selection concern even though expanding usage can raise cost discipline issues. | 中 | SM026 |
| CM045 | Stanford HAI reported rising AI incidents, rare standardized responsible-AI evaluations, and 59 U.S. federal AI-related regulations in 2024. | 中 | SM027 |
| CM046 | Buyer, user, and payer roles for video intelligence often split across workflow owners, technical integrators, and budget owners, making integration evidence central to adoption. | 中 | SM002, SM014, SM023, SM025 |
| CM047 | Public sources reviewed do not isolate Twelve Labs' serviceable revenue share, penetration, customer-budget conversion, or production deployment rate. | 中 | |
| CP001 | Twelve Labs prices Marengo video indexing at $0.042 per minute, Search API usage at $4 per 1,000 queries, Pegasus input video at $0.0292 per minute, and Pegasus output text at $0.0075 per 1,000 tokens on its Developer plan. | 高 | SP001, SP033 |
| CP002 | Twelve Labs positions Marengo as a video foundation model for analyzing frames, temporal relationships, speech, and sound for retrieval tasks. | 高 | SP002, SP004 |
| CP003 | Twelve Labs positions Pegasus as a video-first language model that uses visual, audio, and speech information for video-to-text analysis. | 高 | SP001, SP005 |
| CP004 | Twelve Labs documentation and release-note surfaces identify Marengo and Pegasus as active model families, supporting the chapter baseline of a video-native search-plus-analysis API. | 中 | SP003, SP004, SP005 |
| CP005 | Google Cloud Video Intelligence remains a per-minute video-annotation incumbent with labels, shots, explicit content, speech, text, object, face, person, and logo capabilities rather than a Twelve-style video-native foundation-model API. | 高 | SP006, SP007 |
| CP006 | Google Gemini creates a broader multimodal threat because its developer docs support video understanding while its paid API pricing is token-based by model tier. | 高 | SP008, SP009 |
| CP007 | OpenAI is an adjacent and likely direct entrant because its API pricing is model-token based while its Sora docs expose video generation workflows. | 高 | SP010, SP011 |
| CP008 | Azure AI Video Indexer combines minute-based pricing with audio/video insight extraction for transcription, labels, OCR, faces, topics, and scenes, giving Microsoft a classic enterprise-video analytics alternative. | 高 | SP012, SP013 |
| CP009 | Amazon Rekognition Video combines AWS distribution with stored and streaming video analysis and minute-based pricing for recognition, moderation, and related video-analysis tasks. | 高 | SP014, SP015 |
| CP010 | Runway competes adjacently for creative video budgets because it publishes image-and-video plans starting from $12 per month rather than archive-search pricing. | 中 | SP016 |
| CP011 | Runway announced a $315 million Series E in February 2026 to scale world simulation, making it one of the best-capitalized adjacent video-AI companies. | 高 | SP016, SP036 |
| CP012 | Synthesia is an adjacent AI-avatar competitor whose pricing page says plans now start from $18 per month and whose feature surface centers avatars, voices, localization, and enterprise video creation. | 高 | SP017, SP018 |
| CP013 | HeyGen is an adjacent identity-first video generator with a free tier plus creator, pro, and business pricing plans for generated videos and avatar workflows. | 高 | SP019, SP020 |
| CP014 | HeyGen reported $200 million ARR in 2026, signaling that adjacent AI-video creation can command enterprise budget even when it does not solve Twelve Labs search use cases directly. | 高 | SP019, SP039 |
| CP015 | Coactive is a direct visual-search peer because it markets a multimodal AI layer for media and announced a $30 million Series B to analyze images and videos without manual metadata. | 高 | SP021, SP022 |
| CP016 | Coactive frames the buyer pain as a no-metadata future, which directly overlaps Twelve Labs value propositions around making video searchable without manual tagging. | 高 | SP021, SP022 |
| CP017 | Hive offers a content-understanding and moderation alternative with usage-based pricing and video moderation handled through special rates or sales contact. | 高 | SP023, SP024 |
| CP018 | Hive is stronger as a trust-and-safety or moderation substitute than as a deep semantic video-reasoning API because its public positioning emphasizes moderation, search, and generation across content types. | 中 | SP023, SP024 |
| CP019 | Vidrovr remains relevant as a defense and real-time video-AI specialist after CesiumAstro announced its acquisition to enhance space communications systems and build a planetary intelligence layer. | 中 | SP034 |
| CP020 | Reka is a multimodal foundation-model competitor because it markets API-accessible infrastructure to tag, reason over, search, and clip large volumes of video. | 高 | SP025, SP026 |
| CP021 | Reka reportedly raised $110 million and topped a $1 billion valuation, making it a well-capitalized multimodal entrant even if its video packaging is less specialized than Twelve Labs. | 高 | SP025, SP037 |
| CP022 | Memories.ai is a newer visual-memory competitor that markets AI video analysis and long-lived visual memory infrastructure. | 中 | SP027 |
| CP023 | AssemblyAI is an audio-first substitute that can capture transcript, summarization, and audio intelligence workstreams inside video pipelines without handling full visual semantics. | 中 | SP028 |
| CP024 | Deepgram is an audio-first substitute with scalable speech-to-text, text-to-speech, and voice-agent pricing that can undercut full-video analysis when audio transcripts are sufficient. | 中 | SP029 |
| CP025 | VideoLLaMA3 is an open-source developer substitute positioned as frontier multimodal foundation models for image and video understanding. | 中 | SP030 |
| CP026 | InternVideo is an open-source developer substitute around video foundation models and data for multimodal understanding. | 中 | SP031 |
| CP027 | Video-ChatGPT is an open-source video conversation model, showing that a build-your-own path exists for teams willing to absorb integration and evaluation burden. | 中 | SP032 |
| CP028 | Mixpeek presents an adverse alternative to Twelve Labs by contrasting preflight estimates and object-storage/vector-store control against Twelve Labs multi-meter usage. | 中 | SP033 |
| CP029 | The most direct product differentiation is that Twelve Labs combines native video embeddings/search and video-to-text analysis, while many incumbents split classic CV labeling, general multimodal reasoning, or generative-video creation into separate products. | 中 | SP001, SP004, SP005, SP006, SP007, SP011, SP017 |
| CP030 | Google, Microsoft, AWS, and OpenAI have distribution advantages through established cloud or developer platforms that Twelve Labs must offset with video-specialized accuracy, latency, and workflow fit. | 中 | SP006, SP008, SP010, SP012, SP014 |
| CP031 | Runway, Synthesia, HeyGen, and Pika are adjacent budget competitors because they sell generated-video creation rather than retrieval over existing enterprise video archives. | 中 | SP016, SP017, SP019, SP038 |
| CP032 | Multi-homing is structurally easy for many buyers because Twelve Labs, cloud APIs, and adjacent video tools are metered or subscription APIs rather than deeply exclusive data platforms. | 中 | SP001, SP006, SP009, SP010, SP012, SP015, SP016, SP019 |
| CP033 | Internal build is a real substitute for technical buyers because open-source video-language projects exist, but it trades vendor margin for model hosting, retrieval infrastructure, data governance, and evaluation burden. | 中 | SP030, SP031, SP032 |
| CP034 | Manual tagging and rules-based metadata remain a status-quo substitute, but Coactive and Twelve Labs both attack the same manual-metadata bottleneck, implying pressure on older DAM/MAM workflows rather than only peer APIs. | 中 | SP001, SP021, SP022 |
| CP035 | Pricing pressure is plausible because cloud video APIs publish per-minute rates near comparable order-of-magnitude units while Twelve Labs charges $0.042 per indexing minute plus separate search, infrastructure, and analysis meters. | 中 | SP001, SP006, SP012, SP015, SP033 |
| CP036 | Twelve Labs introduces retention and infrastructure considerations because its Free plan keeps indexes for 90 days and its Developer plan includes monthly embedding infrastructure fees. | 中 | SP001 |
| CP037 | Enterprise trust is an incumbent advantage for Microsoft, AWS, and Google because their video offerings sit inside existing cloud procurement, compliance, and billing relationships. | 中 | SP006, SP012, SP014 |
| CP038 | OpenAI and Meta increase commoditization risk because both are pushing natively multimodal models that can absorb more video-understanding tasks over time. | 中 | SP040, SP041 |
| CP039 | The competitive positioning map places Twelve Labs high on video-understanding depth but below hyperscalers on distribution breadth, while cloud incumbents score higher on distribution and adjacent generators score higher on content creation. | 中 | SP001, SP002, SP006, SP010, SP012, SP014, SP016, SP017, SP019, SP021 |
| CP040 | The feature breadth matrix shows no competitor is uniformly strongest across semantic retrieval, classic CV, generated-video creation, open-source control, and audio-only substitution. | 中 | SP001, SP006, SP008, SP011, SP012, SP014, SP016, SP017, SP019, SP023, SP030, SP031, SP032 |
| CP041 | The moat KPI view supports a mixed durability score: video-native model focus and developer pricing help Twelve Labs, while hyperscaler distribution, open-source substitution, and multi-homing cap the moat. | 中 | SP001, SP002, SP006, SP010, SP014, SP030, SP033 |
| CP042 | The relevant landscape spans direct video-understanding peers, cloud incumbents, generative-video adjacencies, audio-first substitutes, open-source/internal build, manual tagging, and likely multimodal entrants. | 中 | SP001, SP006, SP010, SP016, SP021, SP023, SP028, SP030, SP040, SP041 |
| CP043 | The strongest adverse displacement evidence is not a single benchmark loss but the ability of Mixpeek, cloud providers, and open-source stacks to attack cost predictability, procurement convenience, or customization. | 中 | SP006, SP012, SP014, SP030, SP033 |
| CP044 | Twelve Labs is well funded enough to compete with larger platforms after a July 2026 $100 million Series B, but its absolute balance sheet remains smaller than hyperscaler or OpenAI/Meta ecosystems. | 高 | SP001, SP035 |
| CP045 | OpenAI GPT-4o is a direct future threat to video understanding because OpenAI describes it as accepting text, audio, image, and video inputs. | 高 | SP040, SP010 |
| CP046 | Meta Llama 4 is a likely entrant threat because Meta describes the herd as natively multimodal and priced compellingly, supporting an open-model route into video reasoning. | 中 | SP041 |
| CP047 | Pika is an adjacent video-generation competitor whose homepage markets AI videos, automated workflows, agents, and Pika 2.5 generation rather than archive understanding. | 中 | SP038 |
| CI001 | Twelve Labs presents its commercial model as flexible plans that let customers start free, pay as they go, or scale into enterprise contracts. | 高 | SI001, SI002 |
| CI002 | The Free plan gives users 600 minutes of video indexing and keeps free-plan index access for 90 days. | 高 | SI001, SI002 |
| CI003 | The disclosed Developer price for Marengo video indexing is $0.042 per minute. | 高 | SI001, SI002, SI020 |
| CI004 | The disclosed embedding infrastructure service fee is $0.0015 per indexed minute per month. | 高 | SI001, SI002, SI020 |
| CI005 | The disclosed Search API usage price is $4 per 1,000 queries. | 高 | SI002, SI020 |
| CI006 | The Embed API is priced by input type, including video minutes, audio minutes, image requests, and text requests. | 中 | SI002, SI020, SI022 |
| CI007 | Pegasus Analyze is billed on input video duration and output text tokens, with Segment calls multiplying billed duration by segment definitions. | 高 | SI001, SI002, SI020 |
| CI008 | Enterprise monetization is custom-priced and can include higher usage, custom terms, and fine-tuning discussions rather than public list rates. | 高 | SI001, SI002, SI020 |
| CI009 | The pricing calculator illustrates how 600 Marengo minutes convert to $25.20 of video indexing plus $0.90 of infrastructure before query charges. | 中 | SI002 |
| CI010 | AWS Marketplace lists Twelve Labs multimodal models with contract-based pricing and warns that additional AWS infrastructure costs may apply. | 高 | SI007, SI008 |
| CI011 | Twelve Labs documentation says customers use APIs to search, analyze, generate embeddings, or reason across a video knowledge store. | 高 | SI004, SI005, SI006 |
| CI012 | Marengo is positioned as the embedding/search model that analyzes multiple modalities in video content. | 高 | SI003, SI005 |
| CI013 | Pegasus is positioned as a video-to-text generative model for summaries, answers, and structured analysis. | 高 | SI003, SI006 |
| CI014 | The public pricing surface supports at least six monetization lines: indexing, infrastructure, search queries, embeddings, Pegasus analysis/tokens, and enterprise contracts. | 中 | SI001, SI002, SI020 |
| CI015 | Twelve Labs' revenue model is usage-aligned rather than seat-first, so revenue quality depends on whether per-minute and per-query prices cover compute-heavy workloads after discounts. | 中 | SI001, SI020, SI025, SI028 |
| CI016 | On July 1, 2026, Twelve Labs announced a $100 million Series B co-led by NEA and NAVER Ventures with Amazon and other investors participating. | 高 | SI012, SI013, SI015 |
| CI017 | Reported Series B proceeds are intended for research and development, continued San Francisco and Seoul investment, and new offices in New York and London. | 高 | SI012, SI015 |
| CI018 | Funding coverage says Twelve Labs inference workloads will be optimized on AWS Trainium and that new models will launch first on AWS. | 高 | SI012, SI014, SI015 |
| CI019 | AWS's startup story says SageMaker HyperPod can reduce training time by up to 40% through cluster health monitoring and job resiliency. | 中 | SI009 |
| CI020 | Tracxn reports Twelve Labs has raised $207 million over six rounds and is at Series B stage. | 中 | SI018, SI016 |
| CI021 | Companies House lists Twelve Labs Limited as an active UK private limited company incorporated on 23 October 2025. | 高 | SI010, SI011 |
| CI022 | Companies House lists Twelve Labs Limited's first accounts as made up to 31 December 2025 and due by 30 September 2026. | 高 | SI010, SI011 |
| CI023 | Companies House lists the first confirmation statement date as 22 October 2026, due by 5 November 2026. | 高 | SI010, SI011 |
| CI024 | VAST Data announced a February 2026 partnership giving Twelve Labs a customer-managed deployment path for sovereign or sensitive video environments. | 中 | SI023 |
| CI025 | PRWeb's NAB 2026 release describes Twelve Labs as moving from model/API infrastructure toward a full-stack platform for production video workflows. | 中 | SI024 |
| CI026 | Latka reports Twelve Labs reached $4.2 million of revenue in 2023, but its page also calls the company bootstrapped, which conflicts with widely reported venture financing. | 中 | SI019, SI012, SI018 |
| CI027 | The reviewed official pricing, company, and funding sources do not disclose current ARR, revenue run-rate, gross margin, NRR, CAC, burn, or runway. | 中 | SI001, SI003, SI012, SI017, SI018 |
| CI028 | CB Insights and Tracxn maintain financial-profile pages for Twelve Labs, but the fetched public extracts do not provide audited financial statements or current margin details. | 中 | SI017, SI018 |
| CI029 | DA Digital Applied argues that AI services typically carry 50–60% gross margins rather than classic SaaS 80–90% because each query incurs inference COGS. | 中 | SI025 |
| CI030 | KnowledgeLib's 2026 AI-native SaaS benchmark cites 50–65% gross margins and materially higher infrastructure cost shares than traditional SaaS. | 中 | SI028 |
| CI031 | Spheron states that inference is now the ongoing production cost center and estimates 55–80% of enterprise AI GPU spend goes to inference. | 中 | SI026 |
| CI032 | JustSoftLab warns that GenAI cost overruns often come from data preparation, evaluation and observability, and re-embedding cycles rather than only headline model prices. | 中 | SI027 |
| CI033 | UsagePricing independently summarizes Twelve Labs as a pay-as-you-go API metered by minute of video processed. | 中 | SI020 |
| CI034 | ToolRadar's July 2026 pricing review corroborates the free, Developer, and Enterprise plan structure while framing hidden costs as a buyer consideration. | 中 | SI022 |
| CI035 | F6S describes Twelve Labs as a multimodal AI video-understanding platform and API for developers and enterprises with free and pay-as-you-go positioning. | 中 | SI021 |
| CI036 | No reviewed public source discloses Twelve Labs customer concentration, gross retention, net revenue retention, enterprise discounting, or realized price per minute. | 中 | SI001, SI002, SI017, SI018, SI019 |
| CI037 | No reviewed public source discloses debt, credit facilities, cloud-credit balances, or project-finance obligations for Twelve Labs. | 中 | SI010, SI011, SI012, SI017, SI018 |
| CI038 | A reasonable financial diligence view is that the $100 million Series B improves near-term capital adequacy, but actual runway remains unknowable without cash balance and monthly net burn. | 中 | SI012, SI015, SI025, SI026 |
| CI039 | Twelve Labs' capital intensity is primarily compute and R&D intensity rather than inventory or credit-book intensity, with AWS Trainium potentially reducing but not eliminating variable COGS. | 中 | SI009, SI018, SI025, SI026 |
| CI040 | The financial underwriting blocker is not evidence of weak monetization mechanics; it is the absence of private metrics needed to test realized revenue quality, margins, and burn efficiency. | 中 | SI001, SI002, SI027, SI028 |
| CE001 | The Twelve Labs API is positioned to extract information from video and make it available to applications through a REST/JSON interface. | 中 | SE012 |
| CE002 | Twelve Labs’ public product surface centers on three customer-facing verbs: search, analyze, and embed video. | 高 | SE001, SE002 |
| CE003 | Marengo 3.0 powers Twelve Labs’ Embed API and Search API as the embedding model for semantic video retrieval. | 高 | SE003, SE009 |
| CE004 | The Marengo documentation describes Marengo as an embedding model for comprehensive video understanding across visuals, audio, and text. | 中 | SE009 |
| CE005 | Twelve Labs claims Marengo 3.0 uses a 512-dimension embedding and argues this lowers vector storage and query costs relative to higher-dimensional alternatives. | 中 | SE003 |
| CE006 | Amazon Bedrock documentation says Marengo Embed 3.0 can generate embeddings from video, text, audio, image, or multi-input text-with-images inputs. | 中 | SE023 |
| CE007 | Pegasus is Twelve Labs’ generative video-to-text model, and the current docs list Pegasus 1.5 for prompt-based general analysis and segmentation/time-based metadata workflows. | 高 | SE010, SE004 |
| CE008 | Pegasus 1.5 introduces time-based metadata extraction where users define a JSON schema and receive timestamped structured metadata from video up to two hours long. | 高 | SE004, SE024 |
| CE009 | The developer hub and API examples show a workflow that creates an index, creates an asynchronous task for a video, waits for processing, then queries or analyzes results. | 中 | SE005, SE014 |
| CE010 | Twelve Labs supports first-party Python and JavaScript/Node developer paths through SDK documentation and public SDK repositories. | 高 | SE005, SE013, SE017, SE018 |
| CE011 | The modality docs state that model options and search options can be configured around visual, audio, and transcription inputs. | 中 | SE015 |
| CE012 | The Twelve Labs GitHub organization exposes public repositories for Python, JavaScript, evaluation, embeddings, and integration tooling. | 中 | SE016 |
| CE013 | The Python SDK documentation and repository describe a pip-installable SDK and note model file requirements including files up to 4 GB. | 中 | SE014, SE017 |
| CE014 | The JavaScript SDK repository and npm package document installation with npm install twelvelabs-js and link to the same model capability constraints. | 中 | SE018, SE019 |
| CE015 | Amazon materials say Twelve Labs models are available through Amazon Bedrock as well as Twelve Labs’ own API distribution. | 高 | SE020, SE022, SE029 |
| CE016 | AWS says Twelve Labs uses Amazon SageMaker HyperPod for model training and AWS services for cloud video processing infrastructure. | 中 | SE020, SE021 |
| CE017 | AWS describes Twelve Labs using AWS Elemental MediaConvert and Amazon S3 integration to avoid maintaining some video processing infrastructure itself. | 中 | SE021 |
| CE018 | Amazon Bedrock model documentation maps Pegasus to InvokeModel/streaming operations and Marengo Embed models to StartAsyncInvoke operations. | 中 | SE022 |
| CE019 | Twelve Labs’ NVIDIA partner page says running models on NVIDIA GPUs is intended to improve latency, throughput, and production video-insight performance. | 中 | SE008 |
| CE020 | The Mux partnership post positions Pegasus for large-scale compliance and moderation because it interprets visuals, actions, objects, audio, and temporal context. | 中 | SE007 |
| CE021 | Twelve Labs states it completed a SOC 2 Type 2 audit as part of its commitment to data security and privacy for video foundation-model customers. | 中 | SE006 |
| CE022 | Twelve Labs cautions that the SOC 2 certification reflects the audit timing and that security remains an ongoing process for customers to review. | 中 | SE006, SE033 |
| CE023 | Cloud Security Alliance frames privacy controls around access controls, encryption, data minimization, and data-retention schedules, which are relevant diligence asks for customer video handling. | 中 | SE033 |
| CE024 | Independent AI-risk guidance emphasizes human-in-the-loop verification and data quality because generative AI systems can hallucinate or produce inaccurate outputs. | 中 | SE032 |
| CE025 | AIWatch lists Twelve Labs service components for Marengo 3.0 and Pegasus 1.5 and reports a resolved July 16, 2026 incident affecting some API features for 20 minutes. | 中 | SE030 |
| CE026 | Twelve Labs release notes state that Marengo 2.7 was sunset on March 30, 2026 and that batch analysis requires Pegasus 1.5. | 中 | SE011 |
| CE027 | The NAB Show 2026 announcement describes Twelve Labs moving from model/infrastructure provider toward a full-stack video intelligence platform. | 中 | SE025 |
| CE028 | Sports Video Group reports that Twelve Labs’ core 2026 model lineup includes Marengo 3.0, Pegasus 1.5, and Rodeo as an application-layer product. | 中 | SE029 |
| CE029 | MarTech360 reports Marengo 3.0 became generally available through Twelve Labs and Amazon Bedrock. | 中 | SE031 |
| CE030 | The Pegasus-v1 technical report identifies Pegasus-1 as a multimodal language model specialized in video content understanding and natural-language interaction. | 中 | SE026, SE027 |
| CE031 | The Pegasus-1 paper describes an encode-align-decode framework that uses a Marengo video encoder and ASR data to produce video embeddings for language interaction. | 中 | SE027 |
| CE032 | Public product and AWS materials map the product to semantic search, classification, summarization, metadata generation, creative optimization, and compliance workflows. | 中 | SE001, SE007, SE020 |
| CE033 | Marengo and SDK docs disclose practical processing constraints, including four-hour audio/video limits in documentation and up to 4 GB files in SDK guidance. | 中 | SE009, SE017, SE018 |
| CE034 | Pegasus docs list English as fully supported and multiple other languages as partially supported for visual/audio processing, prompt understanding, and output generation. | 中 | SE010 |
| CE035 | The API introduction says Twelve Labs is REST-oriented, JSON-based, programming-language compatible, and usable through SDKs, Postman, and other clients. | 中 | SE012, SE013 |
| CE036 | The architecture is materially dependent on cloud and accelerator partners: Amazon Bedrock/SageMaker/MediaConvert for AWS pathways and NVIDIA GPUs for accelerated deployments. | 中 | SE008, SE020, SE021 |
| CE037 | The public record does not provide an independently audited head-to-head benchmark suite for Marengo 3.0 and Pegasus 1.5 across buyer-specific workflows. | 中 | SE003, SE004, SE026 |
| CE038 | Recent release notes indicate active platform iteration in 2026, including batch analysis, updated SDK references, and retirement of older Marengo 2.7 indexing paths. | 中 | SE011 |
| CE039 | No fetched source disclosed a public end-to-end SLA, error budget, or customer-specific uptime commitment for Twelve Labs APIs. | 低 | |
| CE040 | Visible GitHub organization text shows public SDK and evaluation repositories, but the public repository footprint appears modest relative to hyperscaler developer ecosystems. | 中 | SE016, SE017, SE018 |
| CE041 | Public SOC 2 messaging does not expose the full audit report, data-retention schedules, subprocessor list, or model-training data-use terms needed for enterprise risk review. | 中 | SE006, SE033 |
| CE042 | The Mux moderation post and Thomson Reuters accuracy guidance together support treating video-AI moderation as a decision-support system requiring human review for high-stakes outputs. | 中 | SE007, SE032 |
| CE043 | Twelve Labs differentiates Marengo as an any-to-any retrieval layer spanning text, audio, image, and video rather than a transcript-only or frame-only tool. | 中 | SE002, SE003, SE023 |
| CE044 | Compliance, content review, search, classification, and video-to-text analysis all use the same model family rather than a standalone point solution for each workflow. | 中 | SE001, SE007, SE020 |
| CU001 | Twelve Labs publicly lists at least nine named customer stories across advertising, video infrastructure, commerce, nonprofit archives, sports, broadcast, creator marketing, and AI-data workflows. | 高 | SU001, SU030 |
| CU002 | Mantis Solutions is Reach PLC’s technology division and Reach PLC is described as the UK’s largest commercial news publisher with more than 100 regional brands. | 中 | SU002 |
| CU003 | Mantis moved from a Q4 2025 proof-of-concept over 70–80+ videos to a production deployment in March 2026 through AWS Marketplace procurement. | 中 | SU002, SU015 |
| CU004 | Mantis uses Pegasus on Amazon Bedrock for automated video brand-safety and compliance decisions with pass/fail outputs, scores, reasoning, and flagged thumbnails. | 中 | SU002, SU015 |
| CU005 | Qencode embedded TwelveLabs Marengo and Pegasus directly into its encoding pipeline so customers can add video intelligence without a second pipeline. | 中 | SU003, SU021, SU025 |
| CU006 | Qencode’s CEO said the company evaluated the AI video understanding market and chose TwelveLabs because the models indexed the full multimodal signal rather than only frames or transcripts. | 中 | SU003, SU021 |
| CU007 | GS SHOP initiated a Marengo proof-of-concept on Amazon Bedrock and then shifted from experiment to full-scale production rollout. | 中 | SU004, SU015 |
| CU008 | GS SHOP reported a 57.5% lift in total ordering customers, a 29.4% conversion-rate lift, a 21.7% unique-click lift, watch time rising from 6.3 seconds to 8.0 seconds, and search time falling from 1–2 hours to seconds. | 高 | SU004, SU001 |
| CU009 | UNICEF Korea had more than 8TB of media across decades and implemented a TwelveLabs-powered archive system through Letsur using TwelveLabs APIs. | 中 | SU005, SU030 |
| CU010 | UNICEF Korea reported a 95% reduction in content retrieval time and approximately 200 hours / 2TB of video indexed and made instantly accessible. | 高 | SU005, SU001 |
| CU011 | MLSE reported that TwelveLabs reduced video search and retrieval effort from 16 hours to 9 minutes and produced a 97% reduction in content discovery time. | 高 | SU006, SU030 |
| CU012 | SBS describes a partnership to use Marengo 2.7 for VFX reference search and Pegasus 1.2 for statistical analysis and short-form summarization, but public language is more roadmap-like than production-metric specific. | 中 | SU007 |
| CU013 | Dyn Media covers more than 3,000 live events per season and partnered with TwelveLabs to reduce manual search and make sports moments easier to find. | 中 | SU008, SU012 |
| CU014 | AffiliateNetwork connects brands with more than 60,000 creators and uses TwelveLabs-style unified video understanding to verify creator posts in seconds rather than slower general-purpose pipelines. | 中 | SU009, SU013 |
| CU015 | Protege and TwelveLabs positioned their joint workflow as delivering targeted video datasets in two weeks versus a traditional 6+ month process. | 中 | SU010 |
| CU016 | TwelveLabs officially targets media and entertainment, sports and broadcasting, advertising, and security workflows with separate solution pages. | 高 | SU011, SU012, SU013, SU014 |
| CU017 | The advertising solution page claims contextual ad breaks and brand-safety analysis can lift completion rates 10–20% in side-by-side tests, but the public page does not disclose sample size or customers for that metric. | 中 | SU013 |
| CU018 | The security solution page frames buyer value around behavioral surveillance, evidence search and reconstruction, and anomaly or pattern detection across archived footage. | 中 | SU014 |
| CU019 | TwelveLabs’ AWS partnership gives customers two enterprise procurement and deployment paths: AWS Marketplace and native access through Amazon Bedrock. | 中 | SU015, SU022, SU027 |
| CU020 | The NVIDIA partner page positions TwelveLabs as a managed SaaS backed by NVIDIA GPU acceleration on AWS rather than customer-hosted infrastructure. | 中 | SU016 |
| CU021 | TwelveLabs partner pages with Databricks and Snowflake extend distribution into enterprise data workflows and vector/search infrastructure. | 中 | SU017, SU018, SU029 |
| CU022 | Monks and the Avid Media Composer panel indicate a systems-integrator and editor-workflow route into broadcast, sports, streaming, and post-production operations. | 中 | SU019, SU020 |
| CU023 | The Avid panel page says indexing can run at about 50x real time, making an hour of footage searchable in under a minute. | 中 | SU020 |
| CU024 | TwelveLabs supports a developer-led motion through a Developer Hub, SDK setup, Analyze, Embed, and Search examples, and a free plan with 600 minutes and no credit card. | 高 | SU023, SU024 |
| CU025 | TwelveLabs’ pricing page discloses usage-based meters for Pegasus input video, output text, indexing, and infrastructure, which means production economics depend on indexed minutes, API usage, and retained indexes. | 中 | SU023 |
| CU026 | The July 2026 funding release says Marengo 3.0 and Pegasus 1.5 are distributed through Amazon Bedrock and the TwelveLabs API. | 高 | SU026, SU027, SU033, SU034 |
| CU027 | The July 2026 release says TwelveLabs has deep traction in media and entertainment and demand from public sector, advertising, security, sports, and automotive, but it does not disclose customer counts or retention metrics. | 中 | SU027, SU033, SU034 |
| CU028 | TechCrunch reported in December 2024 that TwelveLabs had 30,000-plus developers using the platform, ranging from individuals experimenting to major enterprises integrating the technology. | 中 | SU029 |
| CU029 | TechCrunch reported that TwelveLabs had enterprise, media, and entertainment clients and cited Databricks and Snowflake integrations as examples of strategic enterprise distribution. | 中 | SU029, SU017, SU018 |
| CU030 | The Los Angeles Times/Bloomberg article reported that customers include Hollywood studios, advertising companies, social media influencers, sports franchise owners, MLSE, AMC Global Media, and UNICEF. | 中 | SU030 |
| CU031 | CB Insights identifies Twelve Labs as a multimodal AI-models-and-APIs company for developers to understand, search, analyze, and generate insights from video content. | 中 | SU031 |
| CU032 | Mixpeek’s adverse comparison says TwelveLabs uses multiple cost meters, a cloud-only deployment model, processed-in-TwelveLabs regions, and video hosted in TwelveLabs cloud, which could complicate cost forecasting or data-residency reviews. | 中 | SU032 |
| CU033 | Mixpeek also concedes that TwelveLabs is strong for pure video-understanding APIs, deep video specialization, clean SDKs, and video-specific documentation. | 中 | SU032, SU024 |
| CU034 | No fetched public source discloses TwelveLabs net revenue retention, gross retention, churn, renewal rate, contract length, or cohort retention by customer segment. | 中 | SU001, SU023, SU027, SU029, SU030 |
| CU035 | No fetched public source discloses revenue concentration, top-customer percentage, top-ten customer percentage, or segment revenue mix for TwelveLabs. | 中 | SU001, SU027, SU029, SU030, SU031 |
| CU036 | The named customer proof is stronger for media, sports, commerce, advertising, nonprofit archives, creator marketing, and AI-data workflows than for security, automotive, or public-sector named deployments. | 中 | SU001, SU014, SU027, SU030 |
| CU037 | Several customer stories provide production or operational signals, but SBS and Dyn are more ambiguous because their public stories describe partnership phases and opportunities without dated renewal or revenue proof. | 中 | SU007, SU008, SU002, SU004, SU005, SU006 |
| CU038 | A self-serve and marketplace motion can expand the top of funnel, but enterprise durability still depends on customers moving from free/developer experiments to governed production deployments. | 中 | SU015, SU023, SU024, SU032 |
| CU039 | Public evidence supports land-and-expand vectors through more content volume, additional models, new modes, partner workflow embeds, and adjacent use cases rather than through disclosed seat expansion or NRR. | 中 | SU002, SU003, SU004, SU017, SU018, SU020, SU023 |
| CU040 | Independent customer satisfaction remains thin: the retained public sources include strong named testimonials, but not independently verifiable G2, Capterra, renewal, or support-quality review data. | 中 | SU001, SU029, SU032 |
| CR001 | The EU AI Act identifies remote biometric identification, emotion recognition, and biometric categorisation as high-risk AI use cases. | 高 | SR011, SR012 |
| CR002 | The EU AI Act prohibits certain AI practices including untargeted scraping to create facial-recognition databases and real-time remote biometric identification in public law-enforcement contexts. | 高 | SR011, SR012 |
| CR003 | The EU AI Act requires high-risk AI systems to have risk mitigation, data governance, logging, documentation, user information, human oversight, robustness, cybersecurity, and accuracy controls before market use. | 高 | SR011, SR012 |
| CR004 | The European Commission states that GPAI enforcement powers, including fines, enter application from 2 August 2026. | 高 | SR013, SR014 |
| CR005 | GDPR Article 9 covers special categories of personal data, and biometric data used for identification creates heightened privacy risk for video AI workflows. | 高 | SR015, SR030 |
| CR006 | GDPR Article 22 rights around automated individual decision-making are relevant when video AI becomes a substantial factor in consequential outcomes. | 高 | SR016, SR030 |
| CR007 | Colorado SB24-205 requires developers and deployers of high-risk AI systems to use reasonable care against algorithmic discrimination and support impact assessments and consumer protections. | 高 | SR018, SR019 |
| CR008 | Illinois BIPA remains a material biometric privacy risk because biometric identifiers and information can support private litigation and statutory damages. | 高 | SR020, SR023, SR024 |
| CR009 | AI copyright litigation in 2026 includes cases testing whether model training, retention of copied works, and AI outputs create infringement liability. | 高 | SR021, SR022 |
| CR010 | Norton Rose Fulbright reports that the Anthropic litigation distinguished fair-use training arguments from the risk of storing pirated copies. | 高 | SR021, SR022 |
| CR011 | Twelve Labs' public sources do not disclose a dataset bill of materials or training-video license register. | 中 | SR002, SR003, SR004 |
| CR012 | Twelve Labs' AUP prohibits customer uses involving unauthorized identity verification from faces or other physical characteristics. | 中 | SR004 |
| CR013 | Twelve Labs' AUP restricts sensitive-attribute inference from images or videos unless permitted and legally authorized. | 中 | SR004 |
| CR014 | Twelve Labs' AUP prohibits uses that facilitate spyware, communications surveillance, unauthorized monitoring, disinformation, and EU AI Act prohibited practices. | 高 | SR004, SR011 |
| CR015 | Twelve Labs' security page discloses encryption in transit and at rest, least privilege, access reviews, audit logging, vulnerability scanning, and incident response planning. | 高 | SR001, SR005 |
| CR016 | Twelve Labs announced a SOC 2 Type 2 certification, which supports baseline enterprise security assurance. | 高 | SR005, SR001 |
| CR017 | Baker Tilly warns that SOC 2 reports are evolving for AI controls, so classic security attestation does not by itself prove AI model governance. | 高 | SR028, SR005 |
| CR018 | The HAVEN paper states that hallucination in large multimodal models for video understanding limits reliability and applicability. | 高 | SR025, SR026 |
| CR019 | Video-hallucination research evaluates causes, aspects, and formats across thousands of questions and multiple large multimodal models. | 中 | SR025 |
| CR020 | Cloud Security Alliance describes visual prompt injection as a multimodal attack vector that can embed adversarial instructions in images or video-like inputs. | 高 | SR027, SR026 |
| CR021 | Twelve Labs' AUP gives the company contractual tools to investigate, remove content, report suspected illegal activity, suspend access, or terminate service for policy violations. | 中 | SR004 |
| CR022 | Public sources reviewed for this chapter did not identify a specific Twelve Labs security breach or outage. | 中 | SR001, SR005, SR028 |
| CR023 | Twelve Labs' security materials state that it uses AWS infrastructure and evaluates third-party vendor risks. | 中 | SR001 |
| CR024 | Model quality, prompt-injection, and abuse-control risks transmit into customer trust because video search outputs may be used in enterprise workflows. | 中 | SR025, SR027, SR004 |
| CR025 | Multiple 2026 reports state that Twelve Labs raised a $100 million Series B round. | 中 | SR007, SR008, SR009, SR010 |
| CR026 | Edaily reports Twelve Labs' cumulative funding exceeded $207 million after the 2026 Series B. | 中 | SR010 |
| CR027 | AWS is reported as Twelve Labs' preferred cloud provider under a multiyear commitment that includes Trainium optimization and first AWS availability for future models. | 中 | SR007, SR009, SR010 |
| CR028 | Twelve Labs' models are reported as distributed through Amazon Bedrock and Twelve Labs' own API. | 中 | SR007, SR010 |
| CR029 | Startup Fortune frames the AWS-Trainium relationship as strategically significant because it moves a visible video AI workload onto Amazon's custom silicon. | 中 | SR009 |
| CR030 | Edaily reports that Twelve Labs was the first Korean AI startup to receive direct investment from NVIDIA, while NAVER Ventures co-led the 2026 Series B. | 中 | SR010, SR007 |
| CR031 | Twelve Labs plans to use the 2026 funding for R&D, San Francisco and Seoul expansion, and new offices in New York and London. | 中 | SR007, SR010 |
| CR032 | Twelve Labs announced Pegasus 1.5, Rodeo, and Autodesk Flow Capture integration at NAB Show 2026. | 中 | SR031, SR007 |
| CR033 | Twelve Labs' public funding narrative quotes CEO Jae Lee's thesis that video, not language, is the substrate of machine intelligence. | 中 | SR007, SR010 |
| CR034 | Startup Fortune reports Twelve Labs was founded in 2021 by Jae Lee and four co-founders. | 中 | SR009 |
| CR035 | Public sources reviewed do not disclose Twelve Labs' revenue, ARR, gross margin, customer concentration, or AWS minimum-spend terms. | 中 | SR007, SR008, SR009, SR010 |
| CR036 | Twelve Labs' enterprise terms incorporate a Data Processing Addendum and AI standard clauses, indicating contractual mitigation for privacy and AI use. | 高 | SR003, SR002 |
| CR037 | Twelve Labs' AUP requires appropriate human oversight for decisions with consequential impact on legal, financial, employment, human-rights, or injury-related outcomes. | 中 | SR004 |
| CR038 | The public record supports policy-level mitigation but not customer-specific proof that high-risk uses are actually blocked, reviewed, or monitored. | 中 | SR004, SR001, SR003 |
| CR039 | A training-data provenance failure would be a thesis-break risk because AI copyright cases show statutory-damages and injunction exposure can be material. | 高 | SR021, SR022 |
| CR040 | A regulated biometric deployment without consent, lawful basis, impact assessment, and human review should be treated as a no-go condition. | 高 | SR011, SR012, SR015, SR018, SR020 |
| CR041 | An AWS commitment that prevents attractive gross margins would materially weaken the investment case because video foundation models are compute-intensive. | 中 | SR009, SR010, SR027 |
| CR042 | A material SOC 2 exception, prompt-injection exploit, or repeated policy-bypass incident should trigger deployment pause in sensitive customer segments. | 中 | SR005, SR027, SR028, SR004 |
| CR043 | Customer-specific model evaluation logs are required before allowing autonomous or consequential use of video-understanding outputs. | 中 | SR025, SR026, SR004 |
| CR044 | A credible succession and retention plan is a required mitigation because the public narrative concentrates technical vision and fundraising credibility around CEO Jae Lee. | 中 | SR007, SR010, SR009 |
| CR045 | Residual risk falls if Twelve Labs can show dataset provenance, customer-specific high-risk use gates, logged abuse monitoring, model benchmarks, and economically bounded cloud commitments. | 中 | SR003, SR004, SR011, SR025, SR027, SR009 |
| CV001 | Twelve Labs announced a $100M Series B on July 1, 2026 co-led by NEA and NAVER Ventures. | 高 | SV001, SV002 |
| CV002 | The Series B participant list included Amazon, Radical Ventures, Korea Investment Partners, Index Ventures, Quadrille Capital, and Red Bull Ventures. | 高 | SV001, SV002, SV010 |
| CV003 | CB Insights reported TwelveLabs had raised $210.12M across 11 rounds, while Tracxn reported $207M across six rounds. | 高 | SV010, SV011 |
| CV004 | Twelve Labs said Series B proceeds would fund R&D, San Francisco and Seoul investment, and new offices in New York and London. | 高 | SV001, SV002 |
| CV005 | Twelve Labs and AWS described AWS as preferred cloud provider with a multiyear Trainium optimization commitment. | 高 | SV001, SV002, SV004 |
| CV006 | Twelve Labs positions Marengo and Pegasus as perception and reasoning models for searchable, structured video intelligence. | 中 | SV001, SV002 |
| CV007 | The fetched official announcement and independent 2026 coverage did not disclose a post-money valuation for the Series B. | 高 | SV001, SV002, SV007 |
| CV008 | CB Insights listed TwelveLabs Series B valuation as masked and revenue as 0 FY undefined in the public profile. | 中 | SV010 |
| CV009 | Tracxn displayed Twelve Labs valuation and revenue multiple fields as access-gated numeric placeholders rather than public figures. | 中 | SV011 |
| CV010 | Companies House lists TWELVE LABS LIMITED as an active UK private limited company incorporated on 23 October 2025. | 中 | SV012 |
| CV011 | Companies House filing history shows a 23 October 2025 incorporation with model articles and £1 statement of capital plus a 14 November 2025 accounting-period filing. | 高 | SV012, SV013 |
| CV012 | Runway announced $315M of Series E funding in February 2026 to scale world-model development. | 中 | SV014 |
| CV013 | CB Insights reported Runway valuation in February 2026 at $5.0B to $5.315B and 2025 revenue at $90M. | 中 | SV015 |
| CV014 | Sacra estimated Synthesia revenue at $145.91M for 2025 and described a 2026 revenue track near $200M. | 中 | SV016 |
| CV015 | Sacra reported Synthesia closed a January 2026 Series E at a $4B post-money valuation after a $2.1B Series D in 2025. | 中 | SV016 |
| CV016 | Sacra reported Pika valuation at $470M in 2024 with $135M of funding and possible reports up to $700M. | 中 | SV017 |
| CV017 | Sacra identified Pika risks from commoditization by OpenAI and Google and from compute economics that can outrun subscription pricing. | 中 | SV017 |
| CV018 | CB Insights reported Perplexity at a $20B September 2025 valuation and $100M 2025 revenue, with a displayed 180x revenue indicator in a secondary-market row. | 中 | SV019 |
| CV019 | CB Insights reported Mistral AI valuation at $11.723B to $13.723B in September 2025 and 2026 revenue at $400M. | 中 | SV020 |
| CV020 | CB Insights reported AssemblyAI had raised $108.12M but did not expose a public valuation in the fetched profile. | 中 | SV021 |
| CV021 | Runway and Twelve Labs both sit in video-centric AI, but Runway is more creation/world-model oriented while Twelve Labs is video understanding and retrieval oriented. | 中 | SV001, SV014, SV015 |
| CV022 | CB Insights reported private AI companies raised $226B in Q1 2026, with mega-rounds accounting for 94% of AI funding and capital increasingly top-heavy. | 中 | SV022 |
| CV023 | CB Insights reported Q1 2026 venture funding was record-high but deal count declined, indicating concentration rather than broad recovery. | 中 | SV023 |
| CV024 | CB Insights warned that for startups outside frontier model developers, differentiation windows were narrowing as the largest model developers pulled ahead. | 中 | SV022 |
| CV025 | Crunchbase reported AI captured $242B, or 80% of global venture funding in Q1 2026, and that the unicorn board added $900B of value in the quarter. | 中 | SV030 |
| CV026 | Carta reported AI startups had higher valuations than non-AI peers in 2025, including a 38% Series A premium and a 193% Series E-plus premium. | 中 | SV026 |
| CV027 | SVB reported $4.4T of value locked in US private unicorns and that five AI companies outvalued all dot-com era IPOs. | 中 | SV025 |
| CV028 | Forbes and TrueBridge described 2026 venture as more selective, with capital concentrated around companies and investors with the strongest networks and track records. | 中 | SV029 |
| CV029 | Startup Fortune framed the Twelve Labs round as a test of whether video search can become fast and accurate at enterprise scale rather than merely another expensive AI compute story. | 中 | SV004 |
| CV030 | Eastern Herald highlighted direct competition with Google video understanding capabilities and stated no valuation was disclosed. | 中 | SV007 |
| CV031 | A base-case investment view is to track or research more at a presumed $1B entry because public evidence supports category quality but not valuation proof. | 中 | SV001, SV007, SV010, SV011, SV022, SV023 |
| CV032 | A bull case could support a $1B entry only if private diligence confirms rapid enterprise revenue, durable gross margin improvement, and AWS-enabled cost advantages. | 中 | SV001, SV005, SV016, SV026 |
| CV033 | A bear case makes a $1B entry unsupported if ARR is immaterial, gross margins are compute-constrained, or strategic partnerships fail to convert into repeatable enterprise demand. | 中 | SV004, SV007, SV017, SV022 |
| CV034 | Entry discipline should require actual ARR, net retention, gross margin by workload, burn, cloud commitments, and customer concentration before moving from track to buy. | 中 | SV010, SV011, SV016, SV026 |
| CV035 | Preferred terms, liquidation stack, option pool refresh, and investor pro rata rights cannot be evaluated from fetched public sources. | 低 | |
| CV036 | After more than $200M of reported funding, dilution and preference overhang are material diligence items even if the latest round is strategically validating. | 中 | SV001, SV010, SV011 |
| CV037 | Exit readiness is more likely to depend on strategic M&A, secondaries, or selective IPO reopening than on a near-term broad public-market window. | 中 | SV023, SV029, SV030 |
| CV038 | Public sources support Twelve Labs as a credible strategic AI company more strongly than they support an immediately investable valuation. | 中 | SV001, SV002, SV004, SV007, SV010, SV022 |
| CV039 | The final recommendation is research-more or track rather than buy until private metrics corroborate the mandate valuation and downside terms. | 中 | SV007, SV010, SV011, SV023, SV026 |
| CV040 | A thesis-break trigger should fire if private ARR is below the level needed to underwrite at least a plausible forward revenue multiple for the presumed entry price. | 中 | SV016, SV019, SV020, SV026 |
| CV041 | A thesis-break trigger should fire if AWS economics are only promotional credits or capacity dependence rather than durable gross-margin improvement. | 中 | SV001, SV004, SV017 |
| CV042 | The company should provide the latest post-money valuation, liquidation preferences, option pool, pro rata rights, and full security stack. | 低 | |
| CV043 | The company should provide current ARR, revenue growth, customer concentration, gross margin, net burn, and runway. | 低 | |
| CV044 | The company should provide AWS commercial terms, committed spend, credit expiry, Trainium benchmark data, and data-residency obligations. | 低 | SV001, SV002 |
| CV045 | The valuation case should assume a private-market hold period because 2026 market sources describe strong AI funding but selective liquidity and concentrated outcomes. | 中 | SV023, SV025, SV029, SV030 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | TwelveLabs | TwelveLabs: Video Intelligence Platform & API | Designed for organizations working with video at scale – turning raw, passive footage into a strategic asset teams can actually use. |
| SO002 | TwelveLabs | About TwelveLabs: Video-Native AI Company | Our team began with twelve members, each bringing a diverse blend of research expertise spanning language, video, machine learning, and perception. |
| SO003 | TwelveLabs | TwelveLabs Careers: AI and Machine Learning Jobs | Our Offices |
| SO004 | TwelveLabs | TwelveLabs Press: News and Media Resources | Catch our team on stage, talks, panels, and conversations on the future of video understanding. |
| SO005 | TwelveLabs | Video AI Platform: Search, Analyze & Embed - TwelveLabs | TwelveLabs models can see and reason about video like no other AI – and they set the standard for a new era of video data interaction. |
| SO006 | TwelveLabs | Video Foundation Models: Marengo & Pegasus - TwelveLabs | Modeling the world. Remodeling video. |
| SO007 | TwelveLabs | Marengo 3.0: Real-World Multimodal Embedding AI | Marengo 3.0 is the foundation model powering Twelve Labs' Embed API and Search API. |
| SO008 | TwelveLabs | Building Pegasus 1.5: From Clip-Based QA to Time-Based Metadata | Pegasus 1.5 represents a fundamental shift. |
| SO009 | TwelveLabs | AWS + TwelveLabs: Multimodal Video AI on Amazon Bedrock | You can start building with TwelveLabs directly through the AWS Marketplace or seamlessly within Amazon Bedrock. |
| SO010 | TwelveLabs | NVIDIA + TwelveLabs: GPU-Accelerated Video AI Infrastructure | TwelveLabs pairs state-of-the-art video understanding with NVIDIA’s accelerated computing platform. |
| SO011 | TwelveLabs | TwelveLabs Raises $100M to Build Video Superintelligence | We raised $100 million to accelerate this work. |
| SO012 | GlobeNewswire | TwelveLabs Raises $100 Million in Series B Funding to Build Video Superintelligence | AWS is TwelveLabs' preferred cloud provider, and the two companies have deepened their strategic partnership with a multiyear commitment. |
| SO013 | The SaaS News | TwelveLabs Raises $100M Series B | TwelveLabs, a San Francisco-based video intelligence company, has raised $100 million in a Series B funding round. |
| SO014 | Sports Video Group | TwelveLabs Raises $100 Million in Series B Funding | AWS is TwelveLabs’ preferred cloud provider under a multiyear commitment that includes optimizing video inference workloads on AWS Trainium chips. |
| SO015 | PYMNTS | Twelve Labs Raises $100 Million to Fund Bet on Video AI | Twelve Labs Raises $100 Million to Fund Bet on Video AI. |
| SO016 | Digital Today | Twelve Labs raises $100 million in additional funding, expands AWS alliance | The funding brings Twelve Labs' total amount raised to more than $207 million. |
| SO017 | TwelveLabs | Our Series A to Build the Future of Multimodal AI | Series A funding co-led by New Enterprise Associates (NEA) and NVIDIA's NVentures. |
| SO018 | PRWeb | Twelve Labs Earns $50 Million Series A Co-led by NEA and NVIDIA's NVentures | Twelve Labs plans to add more than 50 employees by the end of the year. |
| SO019 | SiliconANGLE | Twelve Labs raises $50M for multimodal AI foundation models | The round follows $12 million raised as an extension to its seed round in late 2022 and brings the total raised to more than $77 million. |
| SO020 | The SaaS News | Twelve Labs Secures $50 Million in Series A | Twelve Labs Secures $50 Million in Series A. |
| SO021 | TwelveLabs | TwelveLabs Raises $5M to Simplify Video Search | The company has raised a $5 million seed round. |
| SO022 | TwelveLabs | TwelveLabs Raises $12M for Context-Aware Video AI | Twelve Labs went on to raise $17 million in venture capital — $12 million of which came from a seed extension round. |
| SO023 | TechCrunch | Twelve Labs is building AI that can analyze and search through videos | TwelveLabs on Thursday announced that it’s adding a president to its C-suite: Yoon Kim. |
| SO024 | Tracxn | Twelve Labs | Twelve Labs has raised a total funding of$207M over 6 rounds. |
| SO025 | Crunchbase | Twelve Labs - Crunchbase Company Profile & Funding | Twelve Labs has raised $107.1M. |
| SO026 | CB Insights | TwelveLabs - Products, Competitors, Financials, Employees, Headquarters Locations | Twelve Labs has now raised a total of $207.1M in total equity funding. |
| SO027 | NEA | NEA Invests in Twelve Labs' Multimodal AI | CEO Interview | I started Twelve Labs back in 2021 with four of my best friends. |
| SO028 | World Economic Forum | Jae Lee | Jae Lee is the Co-Founder and CEO of Twelve Labs. |
| SO029 | The Org | Twelve Labs | The Org | Twelve Labs is helping developers build programs that can see, hear, and understand the world as we do. |
| SO030 | VCA Online | Twelve Labs Earns $50 Million Series A Co-led by NEA and NVIDIA's NVentures | Twelve Labs Earns $50 Million Series A Co-led by NEA and NVIDIA's NVentures. |
| SO031 | Radical Ventures | Twelve Labs | Twelve Labs. |
| SO032 | Los Angeles Times | San Francisco video search startup raises $100 million from Amazon and VCs | Lee, whose team of around 200 is evenly split between Seoul and San Francisco. |
| SO033 | TechCrunch | Twelve Labs lands $12M for AI that understands the context of videos | Google, as well as Microsoft and Amazon, offer services that recognize objects, places and actions in videos. |
| SM001 | Twelve Labs | TwelveLabs: Video Intelligence Platform & API | |
| SM002 | Twelve Labs Docs | Introduction | TwelveLabs | |
| SM003 | Grand View Research | Video Analytics Market Size & Share | Industry Report, 2030 | |
| SM004 | Grand View Research | Computer Vision Market Size, Share & Trends Report, 2030 | |
| SM005 | MarketsandMarkets | Video Surveillance Market Size Report 2025 - 2031 | |
| SM006 | MarketsandMarkets | Multimodal AI Market by Offering and Data Modality - Global Forecast to 2028 | |
| SM007 | Precedence Research | Video Analytics Market Size to Surpass USD 109.85 Bn By 2035 | |
| SM008 | Precedence Research | Multimodal AI Market Size to Hit USD 51.76 Billion by 2035 | |
| SM009 | Precedence Research | Generative AI Market Size to Hit USD 1,206.24 Bn By 2035 | |
| SM010 | Fortune Business Insights | Generative AI Market Size, Share, Value Report [2026-2034] | |
| SM011 | Mordor Intelligence | Video Analytics Market Size, Growth Trends 2030 - Industry Forecast | |
| SM012 | Grand View Research | Sports Analytics Market Size And Share Report, 2026-2033 | |
| SM013 | MarketsandMarkets | Sports Analytics Market by Offering and Application - Global Forecast to 2030 | |
| SM014 | Grand View Research | Enterprise Video Market Size & Share | Industry Report, 2030 | |
| SM015 | Precedence Research | Automotive Artificial Intelligence (AI) Market Size to Hit USD 58.99 Billion by 2035 | |
| SM016 | RAND Corporation | Why AI Projects Fail and How They Can Succeed | |
| SM017 | European Commission | AI Act | |
| SM018 | National Institute of Standards and Technology | AI Risk Management Framework | |
| SM019 | The Business Research Company | Video Analytics Market Size, Share, Growth Report 2026 | |
| SM020 | IMARC Group | Video Analytics Market Report 2026-2034 | |
| SM021 | Allied Market Research | Video Analytics Market Size, Share & Forecast - 2027 | |
| SM022 | Data Bridge Market Research | Video Analytics Market Size, Share, Growth & Forecast 2033 | |
| SM023 | Polaris Market Research | Video Analytics Market Size, Share & Industry Forecast 2034 | |
| SM024 | Market.us | Multimodal AI Market | |
| SM025 | Mordor Intelligence | Enterprise Video Market Size, Share Analysis & Research Report, 2031 | |
| SM026 | Boston Consulting Group | BCG AI Radar: From Potential to Profit with GenAI | |
| SM027 | Stanford HAI | The 2025 AI Index Report | |
| SP001 | Twelve Labs | TwelveLabs Pricing: API Plans and Costs | Developer pricing lists $0.042/minute video indexing, $4/1000 search queries, and Pegasus analyze pricing. |
| SP002 | Twelve Labs | Marengo 3.0: Real-World Multimodal Embedding AI | Marengo is described as a video foundation model for frames, temporal relationships, speech, and sound. |
| SP003 | Twelve Labs Docs | Release notes | Release notes document current Twelve Labs model and API changes. |
| SP004 | Twelve Labs Docs | Marengo | Marengo is the model family for embeddings and search. |
| SP005 | Twelve Labs Docs | Pegasus | Pegasus is the video-first language model for text generation from video. |
| SP006 | Google Cloud | Video Intelligence API pricing | Prices are per minute and partial minutes are rounded up to the next full minute. |
| SP007 | Google Cloud Documentation | Video Intelligence API documentation | Video Intelligence API documentation covers labels, shots, explicit content, speech, text, objects, and other annotations. |
| SP008 | Google AI for Developers | Gemini API video understanding | Gemini API documentation demonstrates uploading video files and querying them. |
| SP009 | Google AI for Developers | Gemini Developer API pricing | Gemini pricing is token-based by model tier, with paid input and output token rates. |
| SP010 | OpenAI | Pricing | OpenAI API | OpenAI API pricing is model- and token-based. |
| SP011 | OpenAI Developers | Video generation with Sora | OpenAI provides an API guide for video generation with Sora. |
| SP012 | Microsoft Azure | Pricing – Azure AI Video Indexer | Azure AI Video Indexer pricing is estimated by analysis preset and input minutes. |
| SP013 | Microsoft Learn | What is Azure AI Video Indexer? | Azure AI Video Indexer extracts insights from audio and video files. |
| SP014 | Amazon Web Services | Amazon Rekognition – Video | Amazon Rekognition Video analyzes stored and streaming video for objects, people, text, activities, and moderation. |
| SP015 | Amazon Web Services | Amazon Rekognition pricing | Rekognition pricing includes video analysis by minute and face metadata storage. |
| SP016 | Runway | AI Image and Video Pricing from $12/month | Runway lists AI image and video plans starting from $12/month. |
| SP017 | Synthesia | Synthesia Pricing - Compare Free and Paid Plans | Synthesia pricing page says plans now start from $18/month. |
| SP018 | Synthesia | Discover Synthesia unique features | Synthesia markets AI avatars, voices, localization, and enterprise video features. |
| SP019 | HeyGen | Pricing Plans for Creators and Marketers | HeyGen pricing includes a free plan and paid creator/pro/business tiers for video generation. |
| SP020 | HeyGen | Free AI Avatar Generator | HeyGen markets a large AI avatar catalog and video creation features. |
| SP021 | Coactive | The Contextual Intelligence Layer for Modern Media | Coactive describes a multimodal AI platform for modern media workflows. |
| SP022 | Coactive | Coactive AI Series B Funding Round | Coactive announced $30 million in Series B funding to analyze images and videos. |
| SP023 | Hive | Pricing | Hive | Hive pricing is usage-based and video moderation requires special rates or sales contact. |
| SP024 | Hive | AI to Understand, Search, and Generate Content | Hive positions itself as AI to understand, search, and generate content. |
| SP025 | Reka | Reka homepage | Reka describes scalable infrastructure to tag, reason over, search, and clip large volumes of video. |
| SP026 | Reka Docs | Reka API Documentation | Reka documentation positions the service as an API for multimodal AI. |
| SP027 | Memories.ai | Memories.ai — AI Video Analysis & Visual Memory Platform | Memories.ai markets AI video analysis and a visual memory platform. |
| SP028 | AssemblyAI | AssemblyAI pricing | AssemblyAI publishes production-ready AI model pricing for speech and audio use cases. |
| SP029 | Deepgram | Deepgram Pricing | Deepgram lists scalable speech-to-text, text-to-speech, and voice-agent API pricing. |
| SP030 | GitHub | VideoLLaMA3 repository | VideoLLaMA3 describes frontier multimodal foundation models for image and video understanding. |
| SP031 | GitHub | InternVideo repository | InternVideo describes video foundation models and data for multimodal understanding. |
| SP032 | GitHub | Video-ChatGPT repository | Video-ChatGPT is described as a video conversation model for meaningful conversations about videos. |
| SP033 | Mixpeek | Mixpeek vs. Twelve Labs: 2026 Video AI Comparison & Alternative Guide | Mixpeek contrasts its preflight estimate and storage path with Twelve Labs multi-meter pricing. |
| SP034 | CesiumAstro | CesiumAstro Announces Acquisition of Vidrovr | CesiumAstro announced acquisition of Vidrovr to enhance space communications systems and build a planetary intelligence layer. |
| SP035 | The SaaS News | TwelveLabs Raises $100M Series B | TwelveLabs raises $100M in Series B funding led by NEA and NAVER Ventures. |
| SP036 | Runway | New Funding to Scale World Simulation | Runway announced $315 million in Series E funding led by General Atlantic. |
| SP037 | Investing.com via Wayback | Reka AI raises $110 million, valuation tops $1 billion | Reka AI raises $110 million and valuation tops $1 billion. |
| SP038 | Pika | Pika homepage | Pika markets AI videos, automated workflows, agents, and Pika 2.5 generation. |
| SP039 | Yahoo Finance | HeyGen Doubles to $200M ARR in Eight Months | HeyGen doubles to $200M ARR in eight months on the rise of identity-first AI video. |
| SP040 | OpenAI | Hello GPT-4o | GPT-4o accepts as input any combination of text, audio, image, and video and generates text, audio, and image outputs. |
| SP041 | Meta AI | The Llama 4 herd | Meta describes Llama 4 models as natively multimodal AI with compelling price. |
| SI001 | TwelveLabs | TwelveLabs Pricing: API Plans and Costs | Start free, speed up, or scale. Build, launch, and grow with flexible plans that match your momentum. |
| SI002 | TwelveLabs | TwelveLabs Pricing Calculator: Estimate API Costs | Marengo Video indexing $0.042/min; Infrastructure $0.0015; Search API usage $4/1K queries. |
| SI003 | TwelveLabs | TwelveLabs: Video Intelligence Platform & API | |
| SI004 | TwelveLabs Docs | Introduction | TwelveLabs | Upload your videos and use the API to search, analyze, generate embeddings, or reason across an entire knowledge store. |
| SI005 | TwelveLabs Docs | Marengo | TwelveLabs | |
| SI006 | TwelveLabs Docs | Pegasus | TwelveLabs | |
| SI007 | AWS Marketplace | AWS Marketplace: TwelveLabs Multimodal Foundation Models | Pricing is based on the duration and terms of your contract with the vendor. |
| SI008 | AWS Marketplace | TwelveLabs Pegasus 1.2 (Amazon Bedrock Edition) | Pricing is based on actual usage, with charges varying according to how much you consume. |
| SI009 | Amazon Web Services | Twelve Labs pioneers AI video intelligence on AWS | Businesses use SageMaker HyperPod to train FMs for weeks or even months while actively monitoring cluster health. |
| SI010 | Companies House | TWELVE LABS LIMITED overview - Find and update company information | Company status Active; Company type Private limited Company; Incorporated on 23 October 2025. |
| SI011 | Companies House | TWELVE LABS LIMITED filing history | |
| SI012 | Business Insider Markets | TwelveLabs Raises $100 Million in Series B Funding to Build Video Superintelligence | TwelveLabs ... announced it has raised $100 million in Series B funding. |
| SI013 | SiliconANGLE | TwelveLabs raises $100M to bring superintelligence to AI video models | |
| SI014 | Startup Fortune | Twelve Labs raises $100 million as Amazon bets its Trainium chips on video AI | |
| SI015 | Sports Video Group | TwelveLabs Raises $100 Million in Series B Funding | The company plans to use the funding for research and development ... and opening new offices in New York and London. |
| SI016 | Parsers VC | Twelve Labs – Funding, Valuation, Investors, News | |
| SI017 | CB Insights | TwelveLabs Stock Price, Funding, Valuation, Revenue & Financial Statements | |
| SI018 | Tracxn | Twelve Labs - Company Profile & Team | Total funding of $207M over 6 rounds. |
| SI019 | Latka | Twelve Labs Revenue 2023: $4.2M ARR | In 2023, Twelve Labs's revenue reached $4.2M. |
| SI020 | UsagePricing | Twelve Labs Pricing | Twelve Labs sells video-understanding foundation models ... as a pay-as-you-go API metered by the minute of video processed. |
| SI021 | F6S | Twelve Labs Reviews and Pricing 2026 | |
| SI022 | ToolRadar | TwelveLabs Pricing 2026: Plans, Hidden Costs & Cheaper Alternatives | Plans, hidden costs, and cheaper alternatives compared; Pricing verified Jul 2026. |
| SI023 | VAST Data | VAST Data and TwelveLabs Partner to Advance Secure Video Intelligence | |
| SI024 | PRWeb | TwelveLabs Unveils the Next Era of Video Intelligence at NAB Show 2026 | |
| SI025 | DA Digital Applied | AI Unit Economics: Pricing & Margins for AI Services | Inference is real COGS; margins 50-60% not 80-90%. |
| SI026 | Spheron | AI Inference Cost Economics in 2026: GPU FinOps Playbook | Inference is the cost center now; industry analysts estimate 55-80% of enterprise AI GPU spend goes to inference. |
| SI027 | JustSoftLab | Calculating the cost of generative AI — and how to keep it under control | |
| SI028 | KnowledgeLib | AI-Native SaaS Benchmarks 2026: GPU Costs, Inference Margins & Pricing | AI-native SaaS companies ... compressing gross margins to 50–65%. |
| SI029 | TwelveLabs Docs | Release notes | TwelveLabs | |
| SE001 | Twelve Labs | Video AI Platform: Search, Analyze & Embed - TwelveLabs | Use the TwelveLabs Video Understanding Platform to search, analyze, and embed video. |
| SE002 | Twelve Labs | Video Foundation Models: Marengo & Pegasus - TwelveLabs | Marengo transforms text, audio, image, and video into numerical representations; Pegasus integrates visual, audio, and speech information for text generation. |
| SE003 | Twelve Labs | Marengo 3.0: Real-World Multimodal Embedding AI | Marengo 3.0 is the foundation model powering Twelve Labs' Embed API and Search API. |
| SE004 | Twelve Labs | Building Pegasus 1.5: From Clip-Based QA to Time-Based Metadata | With Pegasus 1.5, you define a schema for what matters in the video, run time-based metadata extraction, and evaluate the output. |
| SE005 | Twelve Labs | TwelveLabs Developer Hub: APIs, SDKs and Docs | The developer hub shows index creation, task creation, and SDK examples for the TwelveLabs API. |
| SE006 | Twelve Labs | Our SOC 2 Type 2 Certification | Twelve Labs has successfully completed its SOC 2 Type 2 audit, marking a significant milestone in our commitment to data security and privacy. |
| SE007 | Twelve Labs | Solving Compliance Video Intelligence with Mux & TwelveLabs | The TwelveLabs Pegasus model directly addresses this challenge by understanding video through visuals, actions, objects, audio, and contextual meaning over time. |
| SE008 | Twelve Labs | NVIDIA + TwelveLabs: GPU-Accelerated Video AI Infrastructure | TwelveLabs pairs state-of-the-art video understanding with NVIDIA’s accelerated computing platform to deliver fast, high-accuracy insights from video. |
| SE009 | Twelve Labs Docs | Marengo | TwelveLabs | Marengo is an embedding model for comprehensive video understanding. |
| SE010 | Twelve Labs Docs | Pegasus | TwelveLabs | Pegasus is a generative model for video-to-text generation. |
| SE011 | Twelve Labs Docs | Release notes | TwelveLabs | Batch analysis requires Pegasus 1.5; Marengo 2.7 has been sunset. |
| SE012 | Twelve Labs Docs | Introduction | TwelveLabs | The API is organized around REST and returns responses in JSON format. |
| SE013 | Twelve Labs Docs | TwelveLabs SDKs | TwelveLabs | TwelveLabs provides client SDKs that enable you to integrate and utilize the platform within your application. |
| SE014 | Twelve Labs Docs | Python SDK | TwelveLabs | The TwelveLabs Python SDK provides a robust interface for interacting with the TwelveLabs Video Understanding Platform. |
| SE015 | Twelve Labs Docs | Modalities | TwelveLabs | Model options include visual, audio, and transcription, with search options specifying which modalities to use. |
| SE016 | GitHub | Twelve Labs Inc. GitHub organization | Popular repositories include official TwelveLabs SDKs for Python and JavaScript and evaluation tooling. |
| SE017 | GitHub | GitHub - twelvelabs-io/twelvelabs-python: Official TwelveLabs SDK for Python | Install the latest version of the twelvelabs package: pip install twelvelabs. |
| SE018 | GitHub | GitHub - twelvelabs-io/twelvelabs-js: Official TwelveLabs SDK for Javascript | Install the latest version of the twelvelabs-js package: npm install twelvelabs-js. |
| SE019 | npm | twelvelabs-js | The npm package documents the twelvelabs-js installation path and model capability links. |
| SE020 | Amazon Web Services | TwelveLabs video understanding models are now available in Amazon Bedrock | TwelveLabs has introduced Marengo, a video embedding model, and Pegasus, a video language model. |
| SE021 | Amazon Web Services | Twelve Labs pioneers AI video intelligence on AWS | Twelve Labs leverages AWS Elemental MediaConvert for cloud-based video transcoding. |
| SE022 | Amazon Web Services Docs | TwelveLabs models - Amazon Bedrock | Pegasus supports InvokeModel and streaming operations; Marengo Embed models support StartAsyncInvoke operations. |
| SE023 | Amazon Web Services Docs | TwelveLabs Marengo Embed 3.0 - Amazon Bedrock | Marengo Embed 3.0 generates enhanced embeddings from video, text, audio, image, or multi-input inputs. |
| SE024 | PRWeb | TwelveLabs Launches Pegasus 1.5, Turning Raw Video Into Structured, Queryable Data at Scale | Pegasus 1.5 introduces Time Based Metadata Extraction with timestamped, structured metadata from video content up to two hours long. |
| SE025 | PRWeb | TwelveLabs Unveils the Next Era of Video Intelligence at NAB Show 2026 | The releases showcase TwelveLabs’ evolution from a model and infrastructure provider to a full-stack platform. |
| SE026 | arXiv | Pegasus-v1 Technical Report | Pegasus-1 is a multimodal language model specialized in video content understanding and interaction through natural language. |
| SE027 | arXiv | Pegasus-1 Technical Report HTML | The framework is designed to encode, align, and decode video using Marengo embeddings and ASR data. |
| SE028 | SiliconANGLE | TwelveLabs raises $100M to bring superintelligence to AI video models | The company’s flagship products include the Marengo model family and Pegasus 1.5. |
| SE029 | Sports Video Group | TwelveLabs Raises $100 Million in Series B Funding | Both models are distributed through Amazon Bedrock and TwelveLabs’ own API. |
| SE030 | AIWatch | Is Twelve Labs Down? Operational | Recent incidents include some API features experiencing issues on Jul 16 that resolved after 20 minutes. |
| SE031 | MarTech360 | TwelveLabs Launches Marengo 3.0 on TwelveLabs & Amazon Bedrock | TwelveLabs announced general availability of Marengo 3.0 on TwelveLabs and Amazon Bedrock. |
| SE032 | Thomson Reuters | Accuracy in AI: Reducing hallucinations at work | Human-in-the-loop verification, data quality, and strategic use of AI tools reduce hallucination risk. |
| SE033 | Cloud Security Alliance | SOC 2 Privacy: Key Criteria | Cloud architectures require strong access controls, encryption, data minimization, and automated data retention schedules. |
| SU001 | Twelve Labs | TwelveLabs Case Studies: Enterprise Video AI Results | The page lists customer stories including Mantis, Qencode, GS SHOP, UNICEF Korea, MLSE, SBS, Dyn Sport, AffiliateNetwork, and Protege. |
| SU002 | Twelve Labs | TwelveLabs and Mantis Solutions case study | The deal closed and moved to production in March 2026 — approximately six months from first engagement to live deployment. |
| SU003 | Twelve Labs | TwelveLabs and Qencode case study | Customers who are already encoding with Qencode can access video intelligence without changing anything else about how they work today. |
| SU004 | Twelve Labs | GS SHOP case study | Total ordering customers +57.5%, conversion rate +29.4%, unique clicks +21.7%, average video watch time 6.3s to 8.0s. |
| SU005 | Twelve Labs | UNICEF Korea case study | The impact included a 95% reduction in content retrieval time and approximately 200 hours / 2TB of video indexed. |
| SU006 | Twelve Labs | MLSE case study | MLSE was able to turn 16 hours of video search and retrieval efforts into 9 minutes. |
| SU007 | Twelve Labs | SBS case study | The case study describes a scene search service using Marengo 2.7 and later Pegasus-powered statistical analysis and summarization phases. |
| SU008 | Twelve Labs | Dyn Sport case study | Dyn needed to cover over 3,000 live events per season and reduce dependency on manual search. |
| SU009 | Twelve Labs | AffiliateNetwork case study | AffiliateNetwork connects brands with 60,000+ creators and processes thousands of creator videos and millions of views daily. |
| SU010 | Twelve Labs | Protege case study | Two weeks. That is how long it took to deliver what would typically need 6+ months. |
| SU011 | Twelve Labs | Video AI for Media and Entertainment | TwelveLabs makes every frame of your library searchable in natural language, with scene-level structure for producers, editors, and licensing teams. |
| SU012 | Twelve Labs | Video AI for Sports and Broadcasting | The page promises editorial teams can publish 50 to 100 clips per editor per day. |
| SU013 | Twelve Labs | Video AI for Advertising | The page says multimodal evaluation can lift completion rates 10-20% in side-by-side tests. |
| SU014 | Twelve Labs | Video AI for Security | The page positions TwelveLabs for natural-language footage search, incident reconstruction, and anomaly or pattern detection. |
| SU015 | Twelve Labs | TwelveLabs and AWS partnership | You can start building with TwelveLabs directly through the AWS Marketplace or seamlessly within Amazon Bedrock. |
| SU016 | Twelve Labs | TwelveLabs and NVIDIA partnership | The page says TwelveLabs SaaS is backed by NVIDIA GPU acceleration on AWS. |
| SU017 | Twelve Labs | TwelveLabs and Databricks partnership | Databricks is working toward native model access through Mosaic AI Model Serving. |
| SU018 | Twelve Labs | TwelveLabs and Snowflake partnership | The partnership focuses on video intelligence inside Snowflake workflows while maintaining governance, compliance, and scalability. |
| SU019 | Twelve Labs | TwelveLabs and Monks partnership | Monks continues embedding TwelveLabs models as core components of its Monks.Flow platform. |
| SU020 | Twelve Labs | TwelveLabs Panel for Avid Media Composer | Index at about 50x real-time; an hour of footage is searchable in under a minute. |
| SU021 | Twelve Labs | TwelveLabs and Qencode blog | The integration turns a standard media processing job into an automated content operations workflow. |
| SU022 | Twelve Labs | Marengo and Pegasus on Amazon Bedrock tutorial | The tutorial shows developers how to build searchable video libraries and generate rich descriptive metadata through Amazon Bedrock. |
| SU023 | Twelve Labs | TwelveLabs Pricing | The Free plan provides 600 minutes and does not require a credit card. |
| SU024 | Twelve Labs | TwelveLabs Developer Hub: APIs, SDKs and Docs | The Developer Hub exposes Analyze, Embed, and Search examples plus API-key and SDK setup. |
| SU025 | Qencode | Qencode company blog / platform profile | Qencode says its clients range from startups to large enterprises and its platform spans transcoding, live streaming, storage, delivery, player, and analytics. |
| SU026 | Sports Video Group | TwelveLabs Raises $100 Million in Series B Funding | TwelveLabs core models include Marengo 3.0 and Pegasus 1.5, and both are distributed through Amazon Bedrock and TwelveLabs own API. |
| SU027 | GlobeNewswire | TwelveLabs Raises $100 Million in Series B Funding to Build Video Superintelligence | The release says TwelveLabs has deep traction in media and entertainment and demand from advertising, security, sports, and automotive. |
| SU028 | PRWeb | Twelve Labs earns $50 million Series A | The 2024 release describes enterprise momentum around multimodal AI foundation models for video. |
| SU029 | TechCrunch | Twelve Labs is building AI that can analyze and search through videos | Lee said TwelveLabs had 30,000-plus developers and clients in enterprise, media, and entertainment spaces. |
| SU030 | Los Angeles Times / Bloomberg | Video search startup raises $100 million from Amazon, VCs | Customers include Hollywood studios, advertising companies, social media influencers, sports franchise owners, MLSE, AMC Global Media, and UNICEF. |
| SU031 | CB Insights | TwelveLabs company profile | CB Insights describes Twelve Labs as developing multimodal AI models and APIs that help developers understand, search, analyze, and generate insights from video content. |
| SU032 | Mixpeek | Mixpeek vs Twelve Labs: 2026 Video AI Comparison & Alternative Guide | Mixpeek argues Twelve Labs has multiple pricing meters, cloud-only deployment, and video hosted in Twelve Labs cloud. |
| SU033 | Yahoo Finance | TwelveLabs raises $100 million Series B | Enterprises are rapidly moving from experimentation to production-scale deployment of video understanding technology. |
| SU034 | Markets Insider | TwelveLabs Raises $100 Million in Series B Funding | The reprinted release says additional verticals including advertising, security, sports, and automotive continue to drive demand. |
| SU035 | SiliconANGLE | Twelve Labs raises $50M for multimodal AI foundation models | SiliconANGLE covered Twelve Labs raising $50 million for multimodal AI foundation models co-led by NEA and NVIDIA. |
| SR001 | Twelve Labs | TwelveLabs Security and Compliance | TwelveLabs leverages Amazon Web Services (AWS) and we utilize hardening practices from the Center for Internet Security (CIS) Benchmarks. |
| SR002 | Twelve Labs | TwelveLabs Privacy Policy | The privacy policy governs personal information and customer content processed through Twelve Labs services. |
| SR003 | Twelve Labs | TwelveLabs Enterprise Terms of Service | Attachments incorporated into the agreement include Bonterms AI Standard Clauses, the Acceptable Use and Conduct Policy, Supplemental Terms, and a Data Processing Addendum. |
| SR004 | Twelve Labs | Acceptable Use and Conduct Policy - TwelveLabs | The policy restricts identifying or verifying people from faces or other characteristics unless permitted and legal, and bars prohibited practices under Article 5 of the EU AI Act. |
| SR005 | Twelve Labs | Our SOC 2 Type 2 Certification | Twelve Labs announced completion of a SOC 2 Type 2 audit for its AI video understanding platform. |
| SR006 | Twelve Labs | Video-to-Text Arena: Compare Video Language Models | Twelve Labs compares video language models across tasks in a Video-to-Text Arena. |
| SR007 | Sports Video Group | TwelveLabs Raises $100 Million in Series B Funding | TwelveLabs' core models include Marengo 3.0 and Pegasus 1.5, and AWS is described as the preferred cloud provider. |
| SR008 | The SaaS News | TwelveLabs Raises $100M Series B | TwelveLabs raised $100M in Series B funding. |
| SR009 | Startup Fortune | Twelve Labs raises $100 million as Amazon bets its Trainium chips on video AI | The tension is hard to miss: Nvidia backed Twelve Labs when it was smaller, while AWS signed a multiyear contract around Trainium. |
| SR010 | Edaily | Twelve Labs Secures 150 billion won in Series B Funding… Strengthens Partnership with AWS | Twelve Labs has designated AWS as its preferred cloud provider and plans to optimize its video inference models for AWS's Trainium. |
| SR011 | EUR-Lex | Regulation (EU) 2024/1689 Artificial Intelligence Act | The regulation lays down harmonised rules on artificial intelligence and high-risk AI systems. |
| SR012 | European Commission | AI Act | High-risk use cases include AI systems used for remote biometric identification, emotion recognition and biometric categorisation. |
| SR013 | European Commission | Guidelines for providers of general-purpose AI models | From 2 August 2026, the Commission's enforcement powers enter into application. |
| SR014 | European Commission | General-purpose AI obligations under the AI Act | GPAI model providers have obligations around documentation, copyright policy, and safety for systemic-risk models. |
| SR015 | GDPR.eu | Art. 9 GDPR - Processing of special categories of personal data | Article 9 concerns processing of special categories of personal data. |
| SR016 | GDPR.eu | Art. 22 GDPR - Automated individual decision-making, including profiling | Article 22 concerns automated individual decision-making, including profiling. |
| SR017 | Federal Trade Commission | Artificial Intelligence | The FTC maintains artificial-intelligence resources for technology industry guidance and enforcement. |
| SR018 | Colorado General Assembly | SB24-205 Consumer Protections for Artificial Intelligence | The act requires developers and deployers of high-risk AI systems to use reasonable care to protect consumers from algorithmic discrimination. |
| SR019 | Colorado Attorney General | Colorado Automated Decision-Making Technology & Chatbot Safety Rulemaking | Colorado AG rulemaking covers automated decision-making technology and chatbot safety requirements. |
| SR020 | Justia | 2025 Illinois Compiled Statutes 740 ILCS 14 Biometric Information Privacy Act | BIPA covers biometric identifiers and biometric information under Illinois civil-liability law. |
| SR021 | Norton Rose Fulbright | AI in litigation series: An update on AI copyright cases in 2026 | Numerous copyright infringement cases ask whether training AI on copyrighted works is fair use and who bears liability for infringing outputs. |
| SR022 | Holland & Knight | Major Publishers Challenge AI Training Practices in Landmark Copyright Suit Against Meta | Major publishers challenged AI training practices in a landmark copyright suit against Meta. |
| SR023 | WilmerHale | Seventh Circuit Weighs in on Critical BIPA Retroactivity Question | The Seventh Circuit weighed in on a critical BIPA retroactivity question. |
| SR024 | Davis Wright Tremaine | UPDATE: Seventh Circuit Holds That BIPA Amendment Limiting Damages Applies Retroactively | The BIPA amendment limiting damages applies retroactively, according to the Seventh Circuit update. |
| SR025 | arXiv | Exploring Hallucination of Large Multimodal Models in Video Understanding: Benchmark, Analysis and Mitigation | The paper states that hallucination in video modality limits reliability and applicability. |
| SR026 | International AI Safety Report | International AI Safety Report 2026 | The 2026 report reviews general-purpose AI risks and safety evidence. |
| SR027 | Cloud Security Alliance | Image-Based Prompt Injection: Hijacking Multimodal LLMs Through Visually Embedded Adversarial Instructions | The research note addresses hijacking multimodal LLMs through visually embedded adversarial instructions. |
| SR028 | Baker Tilly | Evolving SOC 2 reports for AI controls | SOC 2 reporting is evolving to address AI controls. |
| SR029 | State of Surveillance | The EU AI Act Takes Full Effect in August: What It Bans | The explainer focuses on what the EU AI Act bans for biometric surveillance. |
| SR030 | European Data Protection Board | Opinion 28/2024 on certain data protection aspects related to AI models | The EDPB opinion addresses data protection aspects related to processing personal data in the context of AI models. |
| SR031 | PRWeb | TwelveLabs Unveils the Next Era of Video Intelligence at NAB Show 2026 | Pegasus 1.5 introduces Time-Based Metadata Extraction and Rodeo brings AI agents into video production workflows. |
| SV001 | TwelveLabs | TwelveLabs Raises $100M to Build Video Superintelligence | We raised $100 million to accelerate this work. |
| SV002 | GlobeNewswire | TwelveLabs Raises $100 Million in Series B Funding to Build Video Superintelligence | AWS is TwelveLabs preferred cloud provider. |
| SV003 | The SaaS News | TwelveLabs Raises $100M Series B | |
| SV004 | Startup Fortune | Twelve Labs raises $100 million as Amazon bets its Trainium chips on video AI | If it does not, AWS is just hosting another well-funded AI startup with expensive compute. |
| SV005 | SiliconANGLE | TwelveLabs raises $100M to bring superintelligence to AI video models | |
| SV006 | Advanced Television | TwelveLabs raises $100m in Series B funding | |
| SV007 | The Eastern Herald | TwelveLabs Raises $100 Million and Names AWS as Preferred Cloud Partner for Video AI | No valuation was disclosed. |
| SV008 | Crypto Briefing | Twelve Labs lands $100M from Amazon, NEA and Naver to build AI for video archives | |
| SV009 | Los Angeles Times | San Francisco video search startup raises $100 million from Amazon and VCs | |
| SV010 | CB Insights | TwelveLabs Stock Price, Funding, Valuation, Revenue & Financial Statements | TwelveLabs has raised $210.12M over 11 rounds. |
| SV011 | Tracxn | Twelve Labs - 2026 Company Profile & Team | Twelve Labs has raised a total funding of $207M over 6 rounds. |
| SV012 | Companies House | TWELVE LABS LIMITED overview | Incorporated on 23 October 2025. |
| SV013 | Companies House | TWELVE LABS LIMITED filing history | Statement of capital on 2025-10-23 GBP 1. |
| SV014 | Runway | New Funding to Scale World Simulation | Today we are announcing $315 million in Series E funding. |
| SV015 | CB Insights | Runway Stock Price, Funding, Valuation, Revenue & Financial Statements | Runway valuation in February 2026 was $5,000 - $5,315M. |
| SV016 | Sacra | Synthesia revenue, valuation & funding | Synthesia closed a $200M Series E round that valued the company at $4B post-money. |
| SV017 | Sacra | Pika valuation, funding & news | Commoditization of AI video generation and unsustainable compute economics are highlighted risks. |
| SV018 | CB Insights | Pika Labs - Products, Competitors, Financials, Employees, Headquarters Locations | |
| SV019 | CB Insights | Perplexity Stock Price, Funding, Valuation, Revenue & Financial Statements | Perplexity valuation in September 2025 was $20,000M. |
| SV020 | CB Insights | Mistral AI Stock Price, Funding, Valuation, Revenue & Financial Statements | Mistral AI valuation in September 2025 was $11,723 - $13,723M. |
| SV021 | CB Insights | AssemblyAI Stock Price, Funding, Valuation, Revenue & Financial Statements | |
| SV022 | CB Insights Research | State of AI Q1 2026 Report | The AI market is becoming increasingly top-heavy. |
| SV023 | CB Insights Research | State of Venture Q1 2026 | This is not a broad market recovery. It is concentration at the top getting more extreme. |
| SV024 | CB Insights Research | State of Venture Q2 2026 | |
| SV025 | Silicon Valley Bank | State of the Markets Report H1 2026 | $4.4T of value is locked in US private unicorns. |
| SV026 | Carta | State of Private Markets: 2025 in Review | AI startups raised larger rounds and garnered higher valuations than non-AI counterparts. |
| SV027 | KPMG | Venture Pulse Q1 2026 | |
| SV028 | National Venture Capital Association | 2026 NVCA Yearbook | |
| SV029 | Forbes / TrueBridge | The State Of Venture Capital In 2026: Welcome To The Value Creation Era | The market also became more selective. |
| SV030 | Crunchbase News | Q1 2026 Shatters Venture Funding Records As AI Boom Pushes Startup Investment To $300B | AI shattered records last quarter, with $242 billion — 80% of total global venture funding in Q1 — going to companies in the sector. |