Reka AI
Reka AI 尽调报告
Reka 在高效多模态企业 AI 上有可信切入点,也拿到 Snowflake 和 NVIDIA 的真实战略验证;但公开财务披露偏薄,价格标签又大约在 $1B,公开证据下更适合继续研究,而不是现在投资。
封面要素
公司概况
Reka AI 是一家位于 Sunnyvale 的多模态模型公司,由前 DeepMind、Google、Meta 和 Baidu 研究人员于 2022 年创立。其商业产品面覆盖 Reka API、Reka Flash、Reka Vision 和 Reka Research,定位不靠纯粹前沿模型规模取胜,而是强调高效多模态推理、私有化部署与企业工作流适配。Snowflake 合作材料、Shutterstock 客户证明以及 Turing 伙伴证据支撑其真实市场相关性;2025 年 7 月由 NVIDIA 和 Snowflake 支持的融资则确认了战略兴趣。主要尽调约束在披露深度:公开证据无法证明经审计的收入质量、客户集中度、利润率结构,或相对于快速改进的前沿模型和开放权重替代品的持久护城河强度。
- 成立时间
- 2022-01-01
- 创始人
- Dani Yogatama, Cyprien de Masson d'Autume, Qi Liu, Mikel Artetxe, Yi Tay
- 创立地点
- Sunnyvale, California, USA
- 总部
- Sunnyvale, California, USA
- 产品
- Reka Flash 面向高效多模态 API 推理,Reka Vision 面向图像 / 视频索引与推理,Reka Research 面向基于网页和文档的企业研究工作流
- 客户
- 需要多模态搜索、视频理解、研究自动化以及私有或受治理部署的企业;也包括通过 Snowflake 和垂直伙伴触达的渠道型买家
- 商业模式
- 基于用量的模型 API 收入,加上围绕视觉、研究和受治理多模态工作流的企业软件与渠道部署
- 阶段
- late-stage private
- 融资情况
- 公开披露融资总额约 $168M-$170M,来自较早轮次与 2025 年 7 月 $110M 融资,最新估值超过 $1B
执行摘要
主要优势
- Reka 把高效多模态模型与视觉、研究场景的工作流产品结合起来,并不只是销售通用聊天端点。
- Snowflake 分销和 NVIDIA 背书融资,大幅提升了企业相关性,超过小型独立模型实验室通常能拿到的认可。
- 与更大的前沿模型同行相比,公司看起来技术密度高、资本效率相对更好,因此在成本敏感的企业部署里有可能守住细分位置。
主要风险
- 公开收入、利润率、烧钱速度、留存和客户集中度数据大多未披露,估值纪律只能依赖私下尽调。
- 核心竞争横跨 OpenAI、Anthropic、Google、Mistral、Cohere 和开放权重替代方案;基础模型层的切换成本仍低。
- 算力获取、安全 / 合规预期和伙伴渠道依赖,可能压缩利润率,或在市场趋同后削弱差异化。
未决问题
- 2025 和 2026 年审计收入、ARR、毛利率、烧钱速度、现金跑道,以及区分经常性软件经济性与算力密集服务的直接证据
- 客户数、集中度、续约行为,以及直接 Reka 收入与伙伴或渠道中介使用量之间的拆分
- 完整股权结构表、清算优先权、Snowflake 商业经济性,以及相对当前前沿模型和开放权重替代方案的最新独立基准
目录
01公司概览
1.1 身份、产品范围与商业模式
Reka AI 把自己定义为一家研究与产品公司,构建“面向物理 AI 时代的模型和基础设施”。这个定位比泛泛的 LLM 实验室更具体。官网和技术论文都把 Reka 描述为从底层开始做多模态,覆盖文本、图像、视频和音频,而不是事后薄薄补一层视觉能力。这一点具有商业含义:公司并不试图打赢广义消费者聊天机器人之战,而是把基础模型能力打包进开发者 API 和更高层企业工作流。在这些场景里,多模态上下文比单纯的原始规模更重要。 产品栈分三层。第一层,Reka 通过 API 暴露聊天模型,公开基线访问包括 reka-flash 和 reka-edge,定价形态类似传统按用量计费的模型平台。第二层,Reka 销售更高阶应用:Reka Vision 面向大型视频或图像语料库做搜索、推理、剪辑和告警;Reka Research 面向多步骤网页与私有文档研究。第三层,Reka 强调部署灵活性——API、私有云、本地部署、VPC 和气隙环境。对一家年轻初创公司来说,这一点异常突出,也贴合媒体、安全和受监管环境中的企业买家。 因此,公司的经济逻辑更像基础设施与应用软件的混合体。按用量定价的 API 变现直接开发者需求,企业层、微调和部署项目支撑更大的合同价值。Snowflake 集成把分销延伸进既有企业数据工作流,Guardian AI 等伙伴主导产品则让 Reka 把多模态推理嵌入面向客户的系统中间接变现。结果是一个比 OpenAI 式大众消费者分发更窄、但更贴近企业场景的商业模式。[CO003, CO004, CO005, CO006, CO007, CO008]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 备注 |
|---|---|---|---|---|
| 成立 | 2022 | 2022-2023 | 高 | 追踪机构和融资报道支持 |
| 总部 | Sunnyvale, California, USA | 2025 | 高 | 官方融资发布中明确披露 |
| 最新轮次 | $110M Series B | 2025-07-22 | 高 | NVIDIA 和 Snowflake 参投 |
| 最新估值 | >$1B | 2025-07-22 | 高 | Reuters 转发报道与追踪机构相互印证 |
| 前轮估值 | ~$300M | 2023 | 中 | 追踪机构 / Reuters 转发报道估计 |
| 总融资 | $168M-$170M | 2025-2026 | 中 | Tracxn 与 GetLatka 略有差异 |
| 员工规模 | 过去一年 20→50;到 2025 年末 / 2026 年 5 月为 60-64 | 2025-2026 | 中 | 私营公司追踪机构区间 |
| 核心商业产品 | Reka Flash、Reka Vision、Reka Research(商业产品) | 2025 | 高 | 官方产品和文档页面 |
| Go-to-market 模型 | API + 企业部署 + 嵌入式伙伴解决方案 | 2025-2026 | 高 | 定价 / 文档加伙伴发布 |
| 客户数量 | 未披露 | 2026 | 低 | 已有已知标识,但未披露总数 |
Headcount、prior valuation 和 total funding 是根据 Reuters 转发报道与追踪机构来源汇总的私营市场估计;customer count 仍未披露。
[CO001, CO002, CO008, CO017, CO019, CO020]公司把紧凑型多模态模型连接到更高层企业应用和灵活部署模式。
[CO009, CO010, CO015, CO024, CO026, CO035]公开可见的 KPI 重点显示:融资规模快速放大,但组织仍小,产品组合偏企业端。
融资和员工数按区间呈现,因为公开私营公司数据源之间略有差异。
[CO017, CO019, CO020, CO021, CO022, CO023]1.2 创始人、领导层与组织设计
Reka 由 Dani Yogatama、Cyprien de Masson d’Autume、Qi Liu、Mikel Artetxe 和 Yi Tay 于 2022 年创立;私营公司画像和官方引用都一致把 Yogatama 标为 CEO。公开材料没有给出完整董事会名单,但显示这是一家创始人主导的组织,即使达到独角兽状态后,团队规模仍异常小。小团队姿态并非偶然:外部报道和产品帖子反复强调资深技术密度、高效训练与推理,以及重视直接技术贡献、轻管理层级的文化。 创始人—市场匹配度很强。Shutterstock 的合作新闻稿称 Reka 由来自 DeepMind、Google Brain 和 FAIR 的科学家与工程师创立;2024 年技术报告则把五位具名创始人列入 Core、Flash 和 Edge 背后的核心作者组。研究履历与直接模型构建参与叠加,构成投资叙事的核心:买家和投资人本质上在押注一个精英小团队,相信它能用远少于最大实验室的资本,做出接近前沿的性能。 同一结构也放大关键人物风险。Yogatama 既是外部发言人,也是战略决策者;Yi Tay 反复作为科学可信度出现;更广泛的创始团队似乎与模型路线图高度绑定。由于公开治理披露很薄,外部投资人仍需尽调正式董事会控制权、留才方案和继任计划。不过目前,公司的小员工数和高创始人集中度更应被理解为有意的运营设计,而不只是组织不成熟。[CO011, CO012, CO013, CO021, CO022]
| 人物 | 角色 | 公开背景信号 | 匹配度证据 | 关键人物依赖 |
|---|---|---|---|---|
| Dani Yogatama | CEO 与联合创始人 | 官方发布中多次引用的创始人;独立报道显示曾与 DeepMind 相关 | 设定战略方向,是主要公开发言人 | 高 |
| Cyprien de Masson d'Autume(联合创始人) | 联合创始人 | 具名创始人和技术论文作者 | 直接参与核心模型开发 | 高 |
| Qi Liu | 联合创始人 | 追踪机构资料和论文作者列表中的具名创始人 | 支撑多模态模型 R&D 深度 | 中 |
| Mikel Artetxe | 联合创始人 | 具名创始人和技术论文作者 | NLP / 多模态研究可信度 | 中 |
| Yi Tay | 联合创始人 / 首席科学家信号 | 具名创始人,并因技术领导力被反复引用 | 科学可信度和模型路线图集中度 | 高 |
公开来源清楚识别了创始人和 CEO,但董事会构成和大多数高管履历披露仍较少。
[CO011, CO012, CO013]1.3 融资历史、战略投资人与资本形成
Reka 的资本故事不长,但很关键。Tracker 数据指向 2023 年 Series A,规模约 $58M-$60M,估值约 $300M;随后是 2025 年 7 月 $110M 的 Series B,把估值推到 $1B 以上。Snowflake 出现在两轮融资里:2023 年先是投资人与合作伙伴,2025 年又与 NVIDIA 一同参与。重复参与说明 Reka 通过了战略投资人的一个关键尽调门槛:其模型不只是值得发合作新闻,也足以获得资产负债表支持。 Snowflake 2023 年公告把合作框定为让客户在 Snowflake 内运行和微调 Reka;后续 Snowflake 材料又把 Cortex 内对 Reka Flash 和 Core 的支持扩展开来。战略解释是,Reka 给 Snowflake 提供多模态模型库存和企业定制可选项,Snowflake 不必把每项能力都内部自建。对 Reka 来说,Snowflake 带来企业分销、治理可信度,以及进入数据云账户的下游路径。 不利皱褶在于,Snowflake 也曾在 2024 年探索以超过 $1B 收购 Reka,谈判后来停止。这笔未成交易看起来并不致命——双方仍保持合作——但它确实暴露了战略张力:Reka 价值足以吸引收购接触,却选择或接受继续独立。独立保留了股东上行空间,但也让公司必须在没有大型平台所有者庇护的情况下,为人才、算力和 GTM 扩张继续融资。[CO014, CO015, CO016, CO017, CO018, CO019]
| 利益相关方 | 角色 | 经济或战略重要性 | 尽调问题 |
|---|---|---|---|
| NVIDIA | Series B 投资人 | 验证算力 / 基础设施相关性和多模态 thesis | 确认除资本之外的商业协作 |
| Snowflake | Series A 投资人、Series B 投资人、伙伴、前潜在收购方 | 分发、产品嵌入和战略可选性 | 审查商业最低承诺、排他性和刷新权利 |
| DST Global | Series A 领投方 / 投资人 | 2023 轮早期财务支持方 | 评估治理权利和 follow-on 意愿 |
| Radical Ventures | Series A 领投方 / 投资人 | AI 专业支持方,支持早期技术 thesis | 澄清董事会或观察员权利 |
| Nat Friedman | Angel / 战略投资人 | 体现创始人-市场网络深度 | 了解非正式招聘或 GTM 支持 |
| Shutterstock | 客户和数据授权伙伴 | 锚定媒体 / 档案用例和数据访问 | 衡量收入集中度和续约风险 |
| Turing | 嵌入式市场进入伙伴 | 展示物理安全部署规模 | 验证合同经济性和留存 |
私营 cap table 百分比未披露;本表聚焦战略上可见的利益相关方,以及它们带来的具体 diligence 问题。
[CO014, CO015, CO017, CO019, CO020, CO033]融资、合作、产品和战略控制事件显示,公司从实验室成立到企业商业化的切换很快。
公开报道披露月份但未披露具体日期时,使用月度日期。
[CO014, CO017, CO023, CO024, CO029, CO033]1.4 商业牵引、部署与里程碑
到 2025 年中,Reka 已经从“模型实验室”状态进入可识别的商业部署阶段。最清晰的公开客户证明是 Shutterstock:2024 年 6 月合作使 Shutterstock 同时成为数据授权方和付费客户,使用 Reka 丰富其图像和视频库的元数据。这一点重要,因为它把 Reka 的多模态卖点连到一个真实生产档案库,背后有法律授权约束、元数据经济性,以及重视质量而非新奇感的企业买家。 第二个主要证明点是物理安全。Reka Vision 被定位为一层智能能力,贴在既有视频管理系统旁边,而不是要求推倒重来。官方产品材料称,在部署中案件解决速度快 65%,误报最多减少 95%;Turing 合作则称 Guardian AI 在 13,000+ 站点、10M+ 日事件的覆盖范围内运行于 Reka Vision 之上,并已被美国执法客户使用。这些结果能否泛化仍未证明,但它们确实显示 Reka 在瞄准带有可衡量 ROI 的运营工作流,而不是泛用聊天机器人。 里程碑记录强化了一个分阶段商业化的公司图景:2022 年成立,2023 年外部融资与 Snowflake 合作,2024 年 Shutterstock 合作与技术报告发布,2025 年 Vision 和 Research 发布 / GA,加上独角兽融资和 Turing 合作。国防安全页面又补上一条战略方向线索:Reka 正有意争取主权和气隙部署,在这些场景里,模型效率、隐私和部署可移植性与排行榜性能同样重要。[CO023, CO024, CO025, CO026, CO027, CO028]
| 日期 | 事件 | 类型 | 金额 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2022 | Reka 在 Sunnyvale 成立 | 成立 | 公司创建 | 创始团队 | 独立多模态模型实验室启动 |
| 2023-06 | Series A 融资 | 融资 | $58M-$60M,估值约 $300M | DST Global、Radical Ventures、Snowflake 与 Nat Friedman | 初始资本化,并建立 Snowflake 战略联系 |
| 2023-06 | Snowflake 宣布投资和伙伴关系 | 伙伴关系 | 未披露的战略投资 | Snowflake 和 Reka | 使 Reka 能在 Snowflake 内运行和微调 |
| 2024-04-18 | Core、Flash 和 Edge 技术论文发布 | 产品 | arXiv 报告发布 | Reka 研究团队 | 为多模态栈建立技术可信度 |
| 2024-06-04 | Shutterstock 伙伴关系宣布 | 伙伴关系 | 多年数据授权 + 客户关系 | Shutterstock 和 Reka | 增加企业媒体证明和训练数据访问 |
| 2024-05/06 | Snowflake 收购谈判出现后停止 | 反向 | 报道称讨论价 >$1B,未达成交易 | Snowflake 和 Reka | 显示战略价值,但保持独立 |
| 2025-07-22 | Series B / growth 轮宣布 | 融资 | $110M,估值 >$1B | NVIDIA、Snowflake、现有投资人 | Reka 成为 unicorn,并为商业化融资 |
| 2025-07 | Reka Vision 和 Reka Research 被突出为 GA 产品 | 产品 | 商业平台已进入市场 | Reka | 从实验室转向应用型企业产品 |
| 2025-07 | Turing 在 Reka Vision 上推出 Guardian AI | 规模 | 伙伴覆盖 13,000+ 站点和每日 10M+ 事件 | Turing 和 Reka | 验证物理安全用例的规模化 |
部分日期只到月份,因为私营公司披露不完整;时间线优先列出会改变商业姿态或估值的事件。
[CO001, CO014, CO016, CO017, CO023, CO024]1.5 图表
02市场分析
2.1 市场边界与现状替代方案
分析 Reka 时,不应把它看成服务整个 AI 市场。相关市场是企业多模态模型层:一套软件和服务,让组织能在生产治理约束下,对文本、图像、视频、音频和私有文档进行推理。Gartner 的多模态预测以及更广义基础模型报告中的分类法,都支撑这个框架。落到实践中,Reka 销售的是纯文本 LLM 不够用的工作流——安全录像搜索、媒体档案打标、多模态研究,以及嵌在专有数据旁边、受治理的企业 AI。 这个定义包括四个支出桶。第一是模型消费本身:API 或托管模型用量。第二是在模型上叠加的企业应用价值,例如 Reka Vision 和 Reka Research。第三是在私有云、VPC、本地环境或气隙环境中运行这些系统所需的部署与治理工作。第四是相邻的算力和数据基础设施支出,用来把多模态用例的性能推到可生产化。 边界也排除了几类相似但不同的类别。消费者聊天机器人不是核心市场。通用办公 copilots 作为替代品有间接意义,但不定义 Reka 买家的待办任务。纯机器人硬件、泛化云 IaaS 和一次性标注服务也只是相邻,而非核心。现状替代方案往往是人工审核、关键词元数据系统、狭窄计算机视觉点工具,或使用 hyperscaler 与开源组件内部自建。[CM001, CM002, CM003, CM004, CM005, CM006]
| 细分 / 品类 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 Reka 的相关性 |
|---|---|---|---|---|
| 企业多模态模型使用 | 用于文本-图像-视频-音频推理的 API / 模型消费 | 消费者聊天机器人订阅 | CTO / 平台负责人 | 核心 |
| 多模态应用 | 视频搜索、研究 agents、元数据增强、事件工作流 | 通用办公生产力 copilots | 业务单元负责人 / CIO | 核心 |
| 受治理部署 | 私有云、VPC、on-prem、air-gapped deployment 工作 | 没有模型层的商品化云计算 | 安全 / IT / 采购 | 高 |
| 相邻数据基础设施 | 多模态工作流的 embedding、数据移动、索引、连接器工作 | 独立存储和摄像头硬件 | 数据 / 基础设施负责人 | 中 |
| 现状替代方案 | 人工审核、关键词元数据、传统计算机视觉、内部自建 | N/A | 运营负责人 / 分析师经理 | 竞争基线 |
本表定义 Reka 可触达的品类,而不是整个生成式 AI 经济。
[CM001, CM002, CM004, CM005, CM006, CM007]即便底层多模态能力相近,不同工作流也会经过不同的经济赞助人。
[CM019, CM020, CM021, CM024, CM025, CM026]类别扩张要转化为供应商增长,必须依次处理算力、治理和替代风险。
漏斗是基于产品和部署材料综合出的采用路径,不是实测转化数据集。
[CM026, CM027, CM029, CM031, CM033, CM034]2.2 规模测算视角与可服务支出
公开市场和分析师来源很少单独拆出 Reka 瞄准的精确切片,因此单一 headline TAM 会夸大精度。最佳自上而下锚点是 Gartner 的 2026 年 AI 支出预测:AI 模型支出 $32.6B,AI 软件支出 $453.2B,总 AI 市场为 $2.596T。Gartner 还称,未来几年多模态能力将渗透企业软件,这意味着模型层价值会越来越从纯文本用量迁往混合模态工作流。 第二个视角是结构性而非数值型。ResearchAndMarkets 按数据模态、垂直行业、技术和地区把多模态 AI 映射到 2035 年;Business Research Company 则把基础 AI 模型分为语言、视觉、多模态、语音和代码应用。这些结构支撑一种判断:Reka 的商业产品面横跨多个终端市场——媒体与娱乐、政府 / 公共部门、制造、零售和企业知识工作流——而不是单一狭窄垂直。 第三个视角受证据约束,并且更贴近 Reka。如果 Gartner 2026 年 AI 模型支出池最终只有 5-15% 分配给与 Reka 相关的企业多模态推理、视频理解和文档扎根研究类别,那么在加入部署服务或周边软件之前,模型层支出约为 $1.6B 到 $4.9B。这不是精确预测;它是一个保守的 SAM 式启发法,用来界定 Reka 未来几年可以合理进攻的类别,而不假设它会成为通用型 hyperscaler。[CM010, CM011, CM012, CM013, CM014, CM015]
| 视角 | 发布方 / 依据 | 年份 | 地理范围 | 数值 | 方法 | 置信度 | 限制 |
|---|---|---|---|---|---|---|---|
| AI 模型支出 | Gartner | 2026 | 全球 | $32.6B | AI 模型市场的直接预测 | 高 | 不限于多模态 |
| AI 软件支出 | Gartner | 2026 | 全球 | $453.2B | AI 软件市场的直接预测 | 高 | 远宽于 Reka 所在细分 |
| 多模态 AI 市场结构 | ResearchAndMarkets | 2025-2035 | 全球 | N/A | 按模态、垂直行业和地区细分至 2035 年 | 中 | 抓取页面披露结构,不披露 headline 数值 |
| Foundation-model 应用广度 | Business Research Company | 2026 | 全球 | N/A | 按模型类型、部署、应用和终端使用行业细分 | 中 | 品类页面没有隔离 Reka 类供应商 |
| Reka-relevant SAM 低情境 | Agent 估计 | 2026 | 全球 | $1.6B | 假设 Gartner AI 模型支出的 5% 映射到与 Reka 相关的企业多模态工作流 | 低 | 启发式估计,不是第三方预测 |
| Reka-relevant SAM 基准情境 | Agent 估计 | 2026 | 全球 | $3.3B | 假设 Gartner AI 模型支出的 10% 映射到与 Reka 相关的工作流 | 低 | 启发式估计,不是第三方预测 |
| Reka-relevant SAM 高情境 | Agent 估计 | 2026 | 全球 | $4.9B | 假设 Gartner AI 模型支出的 15% 映射到与 Reka 相关的工作流 | 低 | 启发式估计,不是第三方预测 |
前两行是来源直接支持的自上而下锚点;最后三行是基于 Gartner 支出类别和 Reka 产品边界得出的证据受限估计。
[CM010, CM011, CM012, CM015, CM016, CM017]规模测算把 Gartner 支出锚点与更窄的多模态采用视角结合起来,而不是只复述表格。
底层数值是受证据约束的估算,不是第三方披露数字。
[CM010, CM011, CM013, CM014, CM016, CM017]Reka 的保守 SAM 式区间采用全球 AI 模型支出的一小部分,而不是整个 AI 堆栈。
每个点估算都是 Gartner AI 模型支出的启发式份额,用来框住一个合理的多模态企业类别,而不是声称精确。
[CM016, CM017, CM018]2.3 买家、用户与付款方分层
买家地图比泛泛的“企业 AI”更专业。在媒体和档案用例里,买家很可能是内容平台、媒体运营或数据产品负责人,希望把大型图像 / 视频库变现或变得可搜索;Shutterstock 是最清晰的公开证明点。在物理安全用例里,买家通常是安全或运营负责人,负责告警质量、调查速度和摄像头网络 ROI;用户是调查员、调度员或操作员,而经济赞助人可能坐在公共安全预算负责人、企业安全负责人或 CIO 位置。在企业研究用例里,买家通常是 CTO、首席数据官或 AI 平台负责人,希望让团队同时查询外部网页来源和内部文件。 采用路径同样按工作流分化。买家通常在高价值人工流程变得太慢或太嘈杂时启动:清洗安全录像、丰富元数据,或跨许多来源综合复杂研究。随后,他们会在受限数据集或站点覆盖范围上测试窄试点,评估准确率和运营节省,之后才扩大到生产规模部署。Snowflake 和 Reka 的部署材料说明,治理和数据本地性从第一次会议开始就是采购关键项,而不是事后补丁。 结论是,Reka 更可能在用户痛点尖锐且多模态证据不可或缺的场景胜出,而不是在买家只想要低成本通用助手时胜出。这会收窄漏斗,但也能提高付费意愿,因为解决方案替代人工、压缩周期,或解锁此前难以变现的资产。[CM019, CM020, CM021, CM022, CM023, CM024]
| 细分 | 买方 | 用户 | 付款方 / 预算负责人 | 工作流 | 采用触发因素 |
|---|---|---|---|---|---|
| 媒体档案 / 内容平台 | 内容平台负责人 | 元数据 / 档案团队 | CIO 或内容运营预算 | 打标签、搜索、剪辑生成 | 大型多媒体库尚未充分变现 |
| 实体安防 / 公共安全 | 安防运营负责人 | 调查员、调度员、分析师 | 安防负责人 / 公共安全负责人 | 视频搜索、告警、事件摘要 | 人工回看录像太慢,噪音也太多 |
| 企业研究 | AI 平台负责人 | 研究员、分析师、知识工作者 | CTO / CDO | Web + 私有文档研究 | 高价值问题仍要人工综合判断 |
| 受监管 / 主权部署 | 项目负责人 | 安全环境内的领域专家 | CIO / 任务负责人 | 本地或隔离环境中的多模态推理 | 数据不能离开受控边界 |
| 数据云嵌入式 AI | 数据平台负责人 | 开发者和分析师 | 数据 / 平台预算 | 让模型贴近受治理数据运行 | 现有企业数据卡在信任边界后面 |
预算负责人根据每个产品替代或加速的工作流推断。
[CM019, CM020, CM021, CM022, CM023, CM024]2.4 增长驱动、约束与时机
几股力量支撑采用。Gartner 的 2030 多模态预测显示,企业软件会越来越在一个工作流里摄取图像、视频、音频和文本。IDC 的 FutureScape 表述指向 AI 从试点走向更广泛编排和由信任驱动的转型。ARK 的基础设施研究认为,训练和推理成本在使用量扩张的同时快速下降;NVIDIA 则称,基于 Blackwell 的推理提供商在某些开放模型部署中,可相较 Hopper 将每 token 成本最多降低 10x。如果这些效率曲线延续,Reka 这类较小提供商就能服务更丰富的多模态工作负载,而不必拥有 hyperscaler 级资产负债表。 约束同样实质。Control Risks 认为,算力获取越来越受出口管制、电力、水和地缘政治许可塑造,而不只是钱。Snowflake 自己的企业 AI 材料强调,数据跨信任边界移动会抬高安全和运营开销。对 Reka 来说,这些约束会转化为真实 GTM 摩擦:企业在把敏感录像或内部知识交给多模态模型前,需要隐私、治理、延迟和支持承诺。 第三个约束是市场结构。开源和开放权重模型改进很快,hyperscaler 平台也越来越在客户数据旁边提供多模态工具。这意味着 Reka 的市场在增长,替代压力也在增长。当前采用窗口有利,因为买家仍需要专业编排、部署和工作流打包。但随着时间推移,除非 Reka 持续在效率、部署灵活性和领域化产品 UX 上拉开差异,否则模型层商品化可能压缩利润率。[CM027, CM028, CM029, CM030, CM031, CM032]
| 驱动因素 / 约束 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 企业软件走向多模态 | 驱动因素 | 2025-2030 | 把品类需求从纯文本助手扩展出去 | 哪些客户工作流今天就需要视频 / 音频? |
| AI 从试点转向编排 | 驱动因素 | 2026+ | 信任门槛解决后,可支撑更广的生产环境采用 | 哪些证据能证明试点转生产? |
| 单 token 推理成本下降 | 驱动因素 | 2026+ | 更丰富的多模态工作负载更容易跑通 | Blackwell 式成本节省,Reka 能吃到多少? |
| 数据本地化和治理需求 | 驱动因素 + 约束 | 当前 | 利好可部署供应商,但也抬高销售复杂度 | 哪些部署模式已经产生收入? |
| 计算许可和电力约束 | 约束 | 2026+ | 在非优先地区,扩张可能放慢或成本升高 | Reka 对稀缺 GPU 供给有多依赖? |
| 开源模型进步 | 约束 | 当前 | 迫使买方拿专业供应商与自建方案比较 | 哪些功能仍难以在内部复制? |
| 超大云厂商捆绑 | 约束 | 当前-2028 | 可能压缩定价,挤掉独立模型预算 | Reka 在 Snowflake 内的包装有多强防御力? |
驱动因素和约束按时间线串起来,章节才能判断采用节奏,而不只是判断方向。
[CM027, CM028, CM029, CM030, CM031, CM032]2.5 图表
03竞争对手
3.1 竞争格局与对手类别
Reka 的竞争集合比简单的“其他模型实验室”更广。第一类是前沿平台既有巨头:OpenAI、Anthropic 和 Google,它们都提供广泛多模态能力、大型分销足迹,以及快速迭代的模型家族。第二类是面向企业的挑战者,例如 Cohere、Mistral 和 Aleph Alpha,更直接地围绕隐私、定制和受控部署竞争。第三类是由 Llama 和 Gemma 带动的开放权重替代,让成熟买家可以内部自建或通过托管推理提供商完成,而不是付费给专业供应商。 Reka 夹在这些类别之间。它太小,无法在支出上压过前沿既有巨头;但围绕多模态视频和研究工作流的产品化程度,又高于许多通用 API 供应商。Snowflake 的投资和 Cortex 分销有助于抵消公司的规模差距,因为 Reka 可以在既有企业数据平台内部被评估,而不必走单独采购流程。这个渠道优势有意义,但不能改变一个事实:买家常常可以同时拿 Reka 与多个 API 提供商做并行测试。 Tracxn 的画像列出数百个活跃竞争者,进一步强化核心结论:Reka 不在赢者通吃市场。它所在的是拥挤、快速趋同的市场,真正的问题不是“谁有模型?”,而是“谁能用可接受的经济性、信任和部署适配,解决买家的精确工作流?”[CP001, CP002, CP003, CP004, CP005, CP006]
| 竞争对手 | 类别 | 规模 / 画像 | 目标细分市场 | 差异化 | 与 Reka 对比时的局限 |
|---|---|---|---|---|---|
| OpenAI | 前沿在位者 | 覆盖广泛的多模态 API 平台 | 开发者 + 企业 | 实时多模态宽度、工具生态 | 对隔离部署或视频特定工作流投入不如 Reka 聚焦 |
| Anthropic | 前沿在位者 | 偏重推理的企业平台 | 企业知识 / 编码 / 搜索 | 长上下文、强推理、企业连接器 | 在主权部署上的公开表述不如 Reka 明确 |
| Google Gemini | 前沿在位者 | 深生态 + 企业阶梯 | 开发者 + Workspace / 云买方 | 强多模态 + Agentic 工作流集成 | 捆绑生态可能盖过工作流特定专业化 |
| Mistral | 企业挑战者 | 多模型与 studio 平台 | 想要灵活性 / 自托管选项的构建者 | 多种模型变体、Agentic 平台、混合姿态 | 公开材料对打包视频工作流的强调较少 |
| Cohere | 企业挑战者 | 安全企业 AI 栈 | 大型企业 | 私有部署、企业搜索导向 | 视频密集型多模态专业化的公开证据较少 |
| Aleph Alpha | 主权专业厂商 | 欧洲 SLLM / 主权聚焦 | 受监管公共部门 / 行业 | 数据主权和领域专业化 | 前沿多模态宽度证据较少 |
| Llama / Gemma | 开放权重替代品 | 生态支持广泛的开放模型 | 成熟的内部自建团队 | 低成本试验和部署灵活性 | 要求买方投入更多集成工作 |
画像强调每个竞争对手在企业多模态采购流程中的核心战略姿态。
[CP001, CP002, CP003, CP004, CP005, CP006]这张图按企业部署控制力和通用多模态平台能力广度为供应商打分。
序数评分综合了部署控制力和能力广度的公开证据;它不是基准测试输出。
[CP002, CP003, CP004, CP005, CP013, CP026]3.2 能力宽度与定价对比
单看原始模型宽度,既有巨头仍在定节奏。OpenAI 的 GPT-4o 家族强调全模态输入输出、实时语音和工具访问。Anthropic 的 Claude 家族强调推理、长上下文、强视觉和企业连接器。Google 的 Gemini 栈强调先进多模态理解、长周期智能体工作流,以及从免费到企业级的价格阶梯。Mistral 定位为拥有多种模型变体的智能体生产平台,Cohere 聚焦安全企业 AI 和私有部署。Aleph Alpha 又不同:它不太像前沿规模 API 竞赛参与者,更像面向高度受控环境的欧洲主权提供商。 Reka 无法在每个维度击败所有对手,因此比较必须按标准拆开。按原始前沿声望看,它落后于最大实验室。按 token 价格看,Reka Flash 明显便宜于高端前沿模型,并且配有明确的视频和研究产品打包。按部署灵活性看,Reka 比许多源自消费者业务的供应商更能竞争,因为它公开销售私有云、本地和气隙模式。按视频中心工作流看,它比那些近期才深化多模态工具的文本优先 API 供应商更专业。 定价方向上对 Reka 有利,但并非在所有场景都一锤定音。只有当每美元能力仍足够高时,token 价格才重要;大型供应商又常通过捆绑、企业点数或相邻平台锁定给折扣。因此,当工作流适配和部署约束占主导时,买家更可能选择 Reka;当他们想要一个覆盖广泛通用 AI 工作负载组合的供应商时,更可能选择既有巨头。[CP009, CP010, CP011, CP012, CP013, CP014]
| 采购标准 | Reka | OpenAI | Anthropic | Google Gemini | Mistral | Cohere | Aleph Alpha |
|---|---|---|---|---|---|---|---|
| 以视频为中心的工作流产品 | 是 | 部分 | 部分 | 部分 | Unknown | Unknown | 否 |
| 私有 / 本地 / 主权姿态 | 是 | 部分 | 部分 | 企业特定 | 是 | 是 | 是 |
| 企业研究工作流包装 | 是 | 基于工具 | 研究 / 搜索 | Agentic / 工具 | Agent 平台 | 企业知识工具 | 领域工作流 |
| 同一系列内的开放权重自托管替代品 | 否 | 否 | 否 | Gemma 相邻 | 部分模型 / 平台灵活性 | 否 | 否 |
| 通过数据云伙伴分发 | 通过 Snowflake | 否 | 通过云伙伴 | 通过 Google Cloud | 通过云 / 自托管 | 企业直销 | 直销 / 主权 |
矩阵标记战略上重要的能力,不代表绝对基准领先。“是”表示公开证据支持,不代表各供应商质量相同。
[CP009, CP010, CP011, CP012, CP013, CP014]| 供应商 / 产品 | 公开标价 / 模型 | 包装姿态 | 包含能力 | 未知项 | 影响 |
|---|---|---|---|---|---|
| Reka Flash / Research | Flash 每 1M token 输入 $0.80 / 输出 $2.00;Research 每 1k 请求 $25 | 按用量计费 + 企业层 | 聊天、视觉、研究、视频定价 | 大批量企业折扣未披露 | 面向专业工作流的成本位置看起来有利 |
| OpenAI | GPT-5.4 mini 每 1M 输入 $0.75 / 输出 $4.50;工具另收费 | 广泛模型菜单 + 工具 | 实时语音、图像、Web 搜索、容器 | 当前页面摘录未列出 GPT-4o 直接价格 | 通用平台很强,但针对工作流适配未必最便宜 |
| Anthropic | Claude 3.5 Sonnet 每 1M $3 / $15,200K 上下文 | 免费到企业的阶梯 | 强推理、视觉、连接器 | 最新旗舰定价组合变化很快 | 以企业姿态提供高价推理 |
| Google Gemini | 免费、付费和企业阶梯 | API + 企业 Agent 平台 | 长上下文、工具、批处理、企业功能 | 各模型 token 价格比较随层级变化 | 低摩擦入口会挤压专业供应商 |
| Cohere | 定制企业定价 / Model Vault 实例 | 企业合同驱动 | 私有部署、搜索、托管模型 | 这里未披露可比公开 token 价格 | 竞争点在企业包装,而不是透明 token 价格 |
这些价格跨能力层级并非完全可比;它们只是公开标价和包装姿态的方向性指标。
[CP014, CP015, CP016, CP017, CP018]供应商差异更多体现在部署和工作流包装,而不是是否“有一个模型”。
分数为序数:3=公开强调强,2=有实质支持,1=支持有限 / 间接。
[CP009, CP012, CP013, CP014, CP015, CP021]3.3 分销力量、切换成本与多供应商并用
模型 API 比传统企业软件更容易多供应商并用,因此纯模型层的切换成本有限。买家通常只需适度工程投入,就能把同一套 prompt flow 放到 OpenAI、Anthropic、Gemini、Mistral、Cohere 和 Reka 上测试。这个现实削弱了“独立模型 API 本身就是持久护城河”的说法。 粘性开始变重要的地方在模型之上。Reka Vision 和 Reka Research 嵌入工作流逻辑、索引、告警和部署模式,比基础 chat-completions 端点更难替换。一旦客户把视频索引接入既有 VMS,或在安全评审下把研究工作流连到内部文件,重新认证成本会显著上升。Snowflake 分销又增加一层粘性:模型越靠近受治理的企业数据,买家越会优化安全和运营便利,而不是只看原始基准差异。 反向力量来自大型供应商的捆绑能力。OpenAI、Google 和 Anthropic 可以把研发摊到更宽产品线。Snowflake 本身也能通过在 Cortex 中提供多个第三方模型来调节需求。开放模型进一步降低切换门槛,因为成熟团队拥有可信的内部自建退路。换句话说,Reka 可以围绕工作流和部署创造局部锁定,但不能假设模型品类层面存在全局锁定。[CP019, CP020, CP021, CP022, CP023, CP024]
Reka 的竞争就绪度不只看原始模型质量;渠道、部署和工作流嵌入同样重要。
KPI 标签概括了章节证据支持的竞争姿态,不是经审计业务指标。
[CP019, CP020, CP021, CP023, CP024, CP032]3.4 护城河持久性与竞争风险
Reka 今天的护城河是复合体,而不是单一堡垒。一层是效率:公司反复强调紧凑、成本敏感的多模态模型,而不是暴力堆规模。第二层是产品聚焦视频密集和文档扎根的企业工作流。第三层是在私有、本地和主权设置中的部署可信度。第四层是通过 Snowflake 获得渠道杠杆。这些层叠加起来,为一部分买家形成了差异化价值主张。 主要风险是趋同。开放权重模型继续改进,NVIDIA 生态信息也清楚表明,低成本推理正在变得更容易,受益者不只是 Reka。与此同时,前沿供应商正在加入更好的视觉、工具、搜索和企业集成,侵蚀“专业多模态”切入点。如果所有主要平台最终都能提供合格的视频推理和受治理部署,Reka 的溢价就必须来自更优工作流 UX、更好的运营指标或更深领域调优。 最稳健的解释是,Reka 拥有可防守的近期利基,而不是不可攻破的长期垄断。它能在买家当下需要多模态能力加部署灵活性的地方取胜。但护城河持久性取决于它能否在大型对手或内部自建替代把品类商品化之前,更快把这个利基转化为客户数据、工作流嵌入和可重复的垂直打法。[CP026, CP027, CP028, CP029, CP030, CP031]
| 护城河主张 | 威胁 | 严重性 | 缓解措施 / 尽调问题 |
|---|---|---|---|
| 高效多模态经济性 | 开源和 Blackwell 驱动的成本下降被所有人拿到 | 高 | 量化 Reka 特有的利润率优势,而不是泛泛行业趋势 |
| 视频和多模态工作流专业化 | 前沿在位者加入可比的视频 / 搜索工作流 | 高 | 跟踪在位者仍无法匹配的客户成果 |
| 部署灵活性 | 大厂加深私有云和主权选项 | 中 | 确认受监管场景中的生产部署 |
| Snowflake 渠道准入 | Snowflake 继续提供许多模型,削弱排他性 | 中 | 厘清与 Snowflake 的收入依赖和商业权利 |
| 小团队技术密度 | 人才挖角或创始人集中拖慢执行 | 中 | 审查留任方案和招聘管线 |
严重性反映的是:如果 Reka 不能持续增加工作流特定价值,各项风险会以多快速度侵蚀差异化。
[CP026, CP027, CP028, CP029, CP030, CP031]3.5 图表
04财务
4.1 收入模型与变现栈
Reka 的公开价格卡显示,公司不是靠一个无差别聊天端点变现。基础层是按用量付费的 API,并按能力做明确价格区分:Edge 是低成本或端侧选项,Flash 是主流工作马,Core 是高端层,Research 则按每千次请求为多步骤网页与文档工作单独定价。在这个基础模型层之上,Vision 围绕已索引视频分钟数、搜索、图像存储、打标和剪辑生成引入另一套收入逻辑。这一点重要,因为公司变现的不只是推理,还有工作流上下文,而不只是 prompt 量。 第二个重要层是打包。Vision 有标准计量价格的开发者层,但企业层转向月度开票、批量折扣、灵活存储和专属支持。Research 同样打包了更高价值工作流,其单位不只是 tokens,而是完成的研究请求。这些产品面合在一起,指向一种混合模型:用自助式用量降低摩擦,再用企业合同把支持、留存和存储纳入经济方程。公开证据支撑这种宽度,但没有披露收入结构、实际折扣率,或销售中有多少来自产品化软件、多少来自专业化赋能。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入流 | 机制 | 公开单位 / 状态 | 收入质量推读 | 关键尽调问题 |
|---|---|---|---|---|
| 基础模型 API(Edge / Flash / Core) | 按用量计价的 API | 公开 token 和媒体价格表 | 已确认真实变现面,但实际折扣和模型组合未知 | 索取 Edge、Flash、Core、图像、视频和音频用量的月度收入拆分 |
| Reka Research | 按请求计费的 Agentic 工作流 | 公开标价为每 1k 请求 $25-$60 | 更高价值的工作流包装可见,但请求量和企业转化未知 | 索取月度请求量、企业账户挂载率和平均实际价格 |
| Reka Vision 开发者层 | 按视频、图像、搜索、标签和剪辑用量计量 | 公开自助价格表 | 按标价看,用量经济性可见;但毛利取决于存储、索引和支持负载 | 索取按工作负载类型拆分的索引分钟数、图像数量、查询量和毛利 |
| Reka Vision 企业层 | 合同软件 / 用量混合 | 月度开票、批量折扣、无速率限制、可选经常性存储 | 企业变现明确存在,但合同结构和最低承诺未披露 | 索取前 10 大企业合同、折扣政策、存储承诺和支持义务 |
| Snowflake 渠道 | 通过 Cortex / 伙伴生态间接分发 | 产品可用性和支持已确认;经济条款未披露 | 渠道可降低直接 GTM 成本,但计费机制可能在 Snowflake 而不是 Reka | 索取收入分成、转介绍或市场计费条款,以及渠道来源 ARR |
本表区分公开标价、合同销售动作和渠道变现;实际 ASP 与产品组合未公开披露。
[CI001, CI002, CI003, CI004, CI005, CI006]| 产品 | 公开价格 / 单位 | 计费动作 | 已知信息 | 未知信息 | 来源锚点 |
|---|---|---|---|---|---|
| Reka Edge | 1M 输入 token $0.10;输出 $0.005;每张图像 $0.03 | 自助用量 | 低成本入门层和端侧定位写得很明确 | 实际用量、客户集中度和利润率 | API 定价 |
| Reka Flash | 1M 输入 token $0.80;输出 $2.00;图像 $0.01;视频分钟 $0.06;音频分钟 $0.015 | 自助用量 | 主流主力 SKU 和多模态计量写得很明确 | 折扣或捆绑后的净有效价格 | API 定价 |
| Reka Core | $6.00 输入锚点,加上高价媒体费用 | 自助用量 / 高端 SKU | 公司公开保留 Flash 之上的高端层 | 输出 token 的实际收入和企业包装 | API 定价 |
| Reka Research | 按推理层级,每 1k 请求 $25 / $35 / $60 | 自助工作流定价 | Research 按完成请求量变现,不只是按 token | 企业挂载率和最低承诺 | API 定价 |
| Reka Vision 开发者 | 视频分钟索引 $0.05;搜索 $0.005;上传 1M 张图像 $10;图像存储 $50 / 1M 张 / 月;输出 token $2 / 1M | 基于点数的自助服务 | Vision 的摄取、查询、存储和输出经济性都看得见 | 试用流量有多少转为生产使用 | Vision 定价 |
| Reka Vision 企业版 | 定制安排;按月开票;批量折扣;专属支持;无速率限制 | 合同 / 企业 | 企业销售动作显然不止自助服务 | ACV 区间、折扣深度、存储承诺和支持负担 | Vision 定价 |
所有价格都是公开标价或已披露的合同销售动作描述;并非客户实际价格。
[CI001, CI002, CI003, CI004, CI005, CI006]Reka 通过多层变现:基础模型用量、工作流软件、企业 Vision 合同和合作伙伴渠道分销。
这座桥映射公开变现结构,不代表已披露的产品级收入组合或利润率贡献。
[CI001, CI002, CI003, CI005, CI006, CI008]4.2 企业分销与 GTM 代理指标
Snowflake 在财务上重要,因为它不只是一个 logo 投资人。Snowflake 自己的公告称,客户可以把 Reka 带入 Snowflake 账户,在 Cortex 中运行或支持 Reka 模型,并把多模态分析留在受治理的企业边界内。这种安排可能会把发现和采购嵌入既有数据平台关系,从而降低 Reka 在部分账户上的直接获客负担。它也意味着,即便买家首先通过 Snowflake 而不是 Reka 原生 API 体验 Reka,渠道仍可产生间接变现。 最清晰的具名终端客户证明仍是 Shutterstock:它一边向 Reka 授权训练数据,一边聘用 Reka 来丰富其图像和视频库的元数据。这种双重角色在财务上重要:它证明了真实企业用例,但也暗示有些商业关系可能把数据权利、工作流软件和模型用量捆在一起,公开价格卡看不到这些安排。GTM 上行空间很清楚——可信的企业参考客户、战略平台伙伴,以及媒体和安全工作流中的产品—市场匹配。GTM 警示也同样重要:没有公开来源披露销售周期长度、ACV、胜率、渠道组合,或 Snowflake 是通过收入分成、转介还是简单模型可用性来计费。[CI011, CI012, CI013, CI014, CI015, CI016]
| 指标 | 公开值 / 估计 | 置信度 | 重要性 | 具体尽调问题 |
|---|---|---|---|---|
| 2025 收入 | $10.9M 估计 | 低 | 收入规模有公开线索,但来自第三方估计,不是备案数字 | 获取管理层认证的 2025 收入,以及按产品拆分的月度运行率 |
| 员工数 | 2025 年末至 2026 年 5 月追踪快照显示 60-64 人 | 中 | 用来框定收入和融资额对应的经营规模 | 要求提供当前组织架构、职能拆分,以及各团队全口径薪酬 |
| 隐含人均收入 | 每名员工 $0.170M-$0.182M(10.9 / 64 至 10.9 / 60) | 中 | 早期企业 AI 公司的方向性运营效率代理指标 | 要求按产品线提供月度人均收入和完全摊薄员工数 |
| 隐含估值 / 收入 | 按 $1B 估值和 $10.9M 收入估计,约 92x | 中 | 投资人押注的是未来杠杆,而不是当前现金产出 | 要求提供董事会融资材料,说明估值方法和前瞻计划 |
| 隐含融资 / 收入 | 按 $168M-$170M 累计融资和 $10.9M 收入估计,约 15.4x-15.6x | 中 | 相对于已披露商业化进展,资本强度偏高 | 要求按年份拆分资本投放,并说明主要投资项的预期回收 |
| 客户数 | 低 | 没有客户数,ACV 分布和集中度风险都看不见 | 要求披露活跃付费客户数、前 10 大客户集中度,以及按队列拆分的 ARR | |
| 毛利率 | 低 | 毛利质量决定定价体现的是软件杠杆,还是计算成本转嫁 | 要求按 API、Vision、Research 和渠道交付工作负载拆分毛利率 | |
| CAC / 回本周期 / NRR | 低 | 销售效率和留存决定增长能否持续 | 要求按渠道提供 CAC、回本周期、NRR、总留存率和续约率 |
标为估计或不可得的行是跨来源代理值,不是经审计的公司披露。
[CI015, CI017, CI018, CI019, CI020, CI021]4.3 公开收入信号与单位经济估算
公开财务可见度很薄,但不是零。最强的收入顶线信号是 GetLatka 估计 Reka 在 2025 年达到 $10.9M 收入。这不是公司申报收入,不应被视为经审计事实,但方向上有用,因为 tracker 来源大体同意 Reka 现在是独角兽,约有 60-64 名员工,并披露了两轮机构融资。如果这些数字方向正确,公司已经从收入前叙事跨入可衡量商业化,但相对于已融资本规模仍非常早期。 同一组 tracker 数字只能支撑粗略代理指标。用 $10.9M 估算收入除以 60-64 名员工,隐含每员工收入约 $170k-$182k;对早期企业 AI 供应商来说,这个水平可观,但仍明显低于成熟软件基准。把同一收入估算与约 $168M-$170M 总融资配对,隐含已融资本超过年收入 15x;与 $1B 估值配对,则隐含估值 / 收入倍数约 92x。这些衍生指标都不应被误认为真实单位经济——它们是跨来源估算——但它们清楚框定了承销问题:投资人买的是未来平台杠杆,而不是当前披露的现金生成。[CI017, CI018, CI019, CI020, CI021, CI036]
| 输入项 | 公开信号 | 状态 / 置信度 | 承销解读 | 具体尽调问题 |
|---|---|---|---|---|
| 最新融资 | 2025 年 7 月由 NVIDIA 和 Snowflake 支持的 $110M 融资 | 高 | 公司拿到了有分量的新资本,也获得战略背书 | 要求提供交割后的股权结构表、轮次条款,以及任何影响未来融资的投资人权利 |
| 投后估值 | 2025 年超过 $1B,2023 年约 $300M | 高 | 估值扩张快于已披露财务指标 | 要求提供估值桥、可比公司,以及本轮使用的内部 KPI 门槛 |
| 累计融资 | 两轮已披露融资合计 $168M-$170M | 中 | 资产负债表支持力度不小,但总额仍来自追踪数据源 | 要求提供完整融资历史,包括任何 SAFE、风险债或老股交易部分 |
| 在手现金 | 低 | 没有实际现金余额,就无法负责任地估计现金跑道 | 要求提供最新资产负债表,以及截至 2026-06-21 或最近结账日的月末现金 | |
| 月度烧钱 / 现金跑道 | 低 | 公开承销无法判断当前资本能撑 12 个月还是 36 个月 | 要求提供月度烧钱、总烧钱与净烧钱,以及 24 个月经营计划 | |
| 资金用途 | 技术开发、更广的企业采用、招聘和基础设施扩容 | 中 | 资本看起来用于增长,而不是修补资产负债表 | 要求按计算、员工、GTM、合作伙伴项目和数据成本拆分详细预算 |
| 债务 / 项目融资义务 | 未发现公开披露 | 低 | 没披露不等于不存在 | 要求提供债务明细、云承诺义务,以及任何保底基础设施合同 |
这里有意压缩历史融资轮次,承销重点放在当前资本是否足够,以及仍未披露的内容。
[CI022, CI023, CI024, CI025, CI026, CI027]公开可观察的财务锚点很少,但仍能框出 Reka 当前规模相对于已融资本和估值的位置。
区间结合了追踪器和新闻来源。收入和员工数是估算,不是公司申报指标。 每个条目的 detail 会说明混合单位。
[CI017, CI018, CI019, CI020, CI022, CI023]4.4 成本结构、资本强度与资本充足性
Reka 官方叙事反复强调高效训练和服务基础设施、更低算力需求以及紧凑多模态模型。这个叙事可能成立,但公开证据仍只支持方向,不支持收益幅度。第三方行业证据在这里有用:NVIDIA 和 Snowflake 都描述了一个由 Blackwell 级硬件和系统优化显著降低每 token 成本、提升吞吐,并让数据贴近受治理企业环境的世界。ARK 和 Control Risks 提供反向权重。它们认为推理成本确实在快速下降,但这种优势会消耗,因为许多提供商都能接入同一条效率曲线,而电力、水、监管和算力获取仍约束现实部署。 因此,资本充足性是关键未解问题。公司背后显然有一次有意义的资产负债表事件——在 2023 年 Series A 之上又拿到 $110M 融资——但没有公开来源披露账上现金、月度烧钱、 runway、债务或已承诺 capex。最可防守的解读因此很窄:Reka 有足够外部融资和伙伴支持,继续扩张产品、人才和基础设施,但没有负责任的公开依据来声称这笔资本能撑多久,或下一次融资触发点何时出现。管理层称募资用于技术开发和更广泛企业采用;投资人仍需要真实烧钱地图。[CI022, CI023, CI024, CI025, CI026, CI027]
| 缺失的私有指标 | 重要性 | 当前公开信号 | 精确尽调路径 |
|---|---|---|---|
| 按 API、Vision、Research 和渠道拆分的收入结构 | 决定耐久性、合同质量,以及对纯用量波动的暴露 | 只有目录价和产品可用性是公开的 | 要求提供过去 12 个月收入拆分、按产品拆分的毛利率,以及渠道来源 ARR |
| ARR、总留存率和 NRR | 看早期企业账户是扩张还是流失 | 未发现公开留存指标 | 要求按年份批次提供队列 ARR 表、总留存率、净留存率和客户 logo 流失 |
| 按工作负载拆分的毛利率 | 区分软件杠杆与计算、存储和支持成本转嫁 | 未发现公开毛利率披露 | 要求按 API、Vision、Research、支持和 Snowflake 交付工作负载拆分毛利率 |
| 客户数和集中度 | 决定 ACV、集中度风险和收入质量 | Shutterstock 和 Turing 被点名,但客户数未披露 | 要求提供活跃付费客户数、前 10 大客户,以及收入集中度明细 |
| CAC、销售周期、回本周期和渠道经济性 | 决定 Snowflake 和企业 GTM 能否带来高效增长 | 未发现公开销售效率指标 | 要求按渠道提供销售漏斗、平均销售周期、CAC、回本周期和伙伴分成条款 |
| 现金、烧钱、现金跑道和云承诺 | 决定资本充足性和下一轮融资时间 | 新融资是公开的,但现金和烧钱不是 | 要求提供最新现金余额、月度烧钱、预测现金跑道和基础设施承诺 |
这些是用公开资料完成财务承销的主要阻碍;每一行都需要管理层披露,而不是继续做网页研究。
[CI021, CI035, CI036, CI037, CI038, CI041]即便 Reka 比大型实验室更高效,这套经济堆栈仍会先把资本导向算力、存储、数据、人才和企业支持,然后才沉淀成耐久利润率。
这是基于产品架构和行业经济性绘制的潜在现金流压力结构图,不是公司特定成本桶披露。
[CI022, CI026, CI028, CI029, CI030, CI031]公开的单位经济路径从低摩擦产品访问走向企业合同,但关键转化和利润率检查点仍未披露。
这张图映射公开产品和渠道结构所隐含的经济检查点。它不是披露的漏斗,也不是量化队列模型。
[CI005, CI011, CI021, CI035, CI040, CI042]4.5 财务结论、风险与尽调阻断点
故事中可投资的部分很直接。Reka 的变现面比纯前沿模型实验室更宽:公开 API 定价、打包的 Vision 和 Research 产品、具名企业用户,以及与 Snowflake 的可信分销关系。公司资本化水平看起来也足以在近期继续产品开发和 GTM 扩张。对一个紧凑团队来说,这些信号并不轻。 阻断点在于,收入质量披露仍严重不足。最好的公开收入数字来自第三方估算,客户群没有列明,也没有公开来源披露 ARR、毛利率、留存、定价兑现或销售效率。由于整个行业的算力经济性都在压缩,证明责任会从原始模型效率转向持久合同质量、客户扩张和运营杠杆。管理层提供这些数据室材料之前,本章财务结论可以对变现潜力保持建设性,但不足以支持完整承销。正确尽调姿态不是怀疑 Reka 能否为产品收费,而是怀疑这种收费模式在规模化后究竟有多可重复、多高毛利、多资本高效。[CI035, CI036, CI037, CI038, CI039, CI040]
4.6 图表
05产品与技术
5.1 以客户工作流看产品面
Reka 现在更像一家多模态基础设施供应商,而不是泛用聊天机器人实验室,并且有清晰买家路径。在核心 API 侧,公开自助产品面围绕 reka-flash 和 reka-edge,通过 OpenAI 兼容聊天接口暴露。这一点重要,因为开发者采用摩擦被降低:团队可以复用 OpenAI SDK 模式,切换 base URL,并从文本、图像、短视频、音频和 PDF 输入开始,而不必学习全新请求形态。公开模型菜单刻意比完整研究叙事更窄,说明公司在策划一个实用自助产品,而不是一次性暴露每个内部模型产物。 在基础聊天路径之上,Reka 已把多模态转成产品化工作流。Vision 通过明确的上传、索引、搜索、Q&A、剪辑生成和元数据打标端点,处理长视频和图像档案;Research 增加跨网页和私有文档的扎根浏览;Speech 则面向高量离线任务,提供带时间戳的多语言转写和翻译。按工作流看,Reka 销售的是把大型非结构化媒体语料库转成可搜索、可自动化系统的方法,而不只是一个更好的 prompt 框。 因此,产品组合很好映射到企业媒体、安全、机器人和受监管知识工作用例,但不太适合消费者 AI 分发。公开产品面强调可部署 API、领域工作流和集成资产,而不是精致的大众市场助手。这个产品定位在文档、Labs 页面和独立评论中保持一致:当买家更关心视频、边缘推理或受治理部署,而不是广泛消费者采用或排行榜品牌力时,Reka 才会赢。[CE001, CE002, CE003, CE004, CE005, CE006]
| 模块 / SKU | 主要用户或买方 | 当前公开成熟度 | 差异化 | 主要尽调缺口 |
|---|---|---|---|---|
| Reka Chat / reka-flash(聊天产品) | 需要通用多模态聊天的 API 开发者 | 公开自助基线 | OpenAI 兼容接口,支持文本、图像、短视频、音频和 PDF 输入 | 公开文档没有揭示最新版本的完整企业支持范围,也没有足够的独立基准深度 |
| Reka Edge / reka-edge-2603(边缘模型) | 需要本地、低延迟视觉推理的团队 | 公开自助 + 本地部署路径 | 面向 token 高效边缘推理和 vLLM/HF 部署打造的 7B 级视觉模型 | 商业授权、支持条款,以及自托管生产参考客户仍只披露了一部分 |
| Reka Vision | 拥有大型图像 / 视频语料库的媒体、安全和运营团队 | 有详细 API 文档的公开产品 | 索引搜索、问答、打标、片段生成和图像搜索,而不是只做通用聊天式多模态 | 公开文档描述了能力,但没有披露经审计的正常运行时间、准确性 SLA 或广泛客户基准数据 |
| Reka Research | 知识工作和研究团队 | 有定价和功能文档的公开产品 | 基于网页和私有文档的智能体,带并行思考模式 | 当前输出质量的独立验证少于公司自有证据 |
| Reka Speech | 有大批量转录或翻译任务的企业 | 已公开发布的能力;在自助文档中的露出少于 Chat/Vision | 针对离线吞吐优化、带时间戳的 850M 多语言模型 | 走向广泛公开 API 可用性和客户参考的路径,不如 Chat/Vision 成熟 |
| MCP / n8n / GitHub 部署资产 | 应用型开发者和自动化团队 | 公开生态资产,不同触点成熟度不一 | 让 Reka 可用于智能体 IDE、低代码流程和本地运行时 | 一些适配器仍落后于核心产品;n8n 文档把 Speech、Research 和 Text 支持标为即将推出 |
状态反映截至 2026-06-21 的公开文档;缺口标出买方仍需向管理层或通过试用获取的内容。
[CE001, CE002, CE005, CE008, CE017, CE020]| 用户任务 | 当前工作流痛点 | Reka 解决方案 | 可衡量收益或解读 | 局限 |
|---|---|---|---|---|
| 围绕短多模态输入提问 | 团队往往需要分开的图像、音频和文本流水线 | 单一 OpenAI 兼容界面上的 Chat API | 多模态内容适配同一种请求模式,集成摩擦更低 | 短视频指南仍会把更长素材推向单独的 Vision 工作流 |
| 搜索长视频归档 | 人工审看或脆弱的 CV 技术栈无法扩展到数小时素材 | Vision 上传 + 索引 + 语义搜索 | 带时间戳的分块检索和可选生成报告,构成可用的检索层 | 需要摄取 / 索引步骤,公开自助速率限制也较低 |
| 回答关于长片段的问题 | 直接把完整视频塞进原始 LLM 提示并不现实 | 基于已索引资产的 Vision 问答 | 长视频检索与对话层分离,上下文才可控 | 买方仍需在自己的素材上测试真实准确性 |
| 为媒体工作流生成片段和元数据 | 编辑或审核员要手工识别高光和描述符 | 片段生成 + 元数据打标 | 结构化标签和自动片段任务,让技术栈更接近工作流软件,而不是基础推理 | 输出质量和政策适配仍由公司描述,而非广泛独立基准验证 |
| 跨网页和私有文档做有依据的研究 | 分析师手工浏览、对比和总结来源 | Reka Research,带浏览、文档搜索和并行思考 | 相比普通聊天,事实锚定更好,准确性 / 成本权衡也可调整 | 基准测试和信任仍高度依赖公司自有证据 |
| 在设备附近运行多模态推理 | 云端往返带来延迟、隐私风险和部署复杂度 | Reka Edge 本地 / 本地机房 / 离线路径 | 边缘封装很适合机器人、监控和其他物理 AI 闭环 | 开放权重商业条款和生产支持范围仍是尽调项 |
收益被表述为从公开文档和合作伙伴文章推导出的工作流层面解读;除非来源明确注明,否则不是经审计的客户 ROI 陈述。
[CE003, CE004, CE006, CE007, CE008, CE015]Reka 的理想工作流从非结构化媒体或文档开始,按需加入索引和工具使用,最终落到企业行动,而不只是聊天答案。
这条流程把多个公开 API 压缩成一张面向买方工作流的运行图。
[CE003, CE004, CE005, CE006, CE007, CE008]5.2 架构与运营模型
Chat、Vision、Research 和 Edge 之间的架构拆分,是理解 Reka 如何看待多模态工作负载的最清晰线索。Chat 针对直接对话式推理优化,包括短视频片段;Vision 则为更长视觉语料引入单独摄取和索引层。长视频路径在文档里很明确:按文件或 URL 上传,可选择把资产分组,完成索引,再对结果分块执行语义搜索、问答、剪辑生成和打标。这与单次端到端 chat completion 调用的运营模型有实质差异,也反映了一种判断:企业视频检索需要预处理和检索结构,而不是把原始上下文硬塞进去。 在模型层,Edge 是最能说明问题的产物。Reka 称它是约 7B 级视觉语言模型,把 ConvNeXt V2 视觉编码器与 6B+ 语言骨干结合,并设计成每个图像 tile 只发出 64 个 tokens。这种 token 纪律是公司效率论点的核心:如果高分辨率视觉输入能被紧凑表示,延迟、上下文压力和服务成本会同时改善。Edge 发布材料把它定位为面向机器人、摄像头、车辆和可穿戴设备的物理 AI 基础设施,而不是只跑在云端的助手。 Research 和 Flash 也显示出同样偏好:重结构化推理,而不是暴力堆规模。Research 使用工具增强浏览,覆盖网页和私有文档;Flash 3.1 被描述为一个 21B 推理模型,先通过强化学习改进,再通过 Llama 兼容发布和量化工作降低自托管难度。结果是一套产品架构,试图把价值转移到编排、检索和高效模型打包上,而不是正面参加最大参数竞赛。[CE004, CE005, CE006, CE007, CE008, CE017]
| 层或组件 | 在技术栈中的作用 | 关键依赖 | 运营优势 | 主要风险 |
|---|---|---|---|---|
| OpenAI 兼容 Chat API | 处理对话式多模态请求和基础开发者入门 | OpenAI SDK 模式和 Reka API 密钥 | 很多团队已经熟悉客户端界面,因此采用更快 | 各模型功能并非完全一致;函数调用目前仅 Flash 支持 |
| Vision 摄取和索引 | 把长视频或图像集转成可搜索资产 | 上传流水线、索引任务、存储、分组 | 将检索和生成拆开,长媒体可被反复查询 | 相比一次性聊天,流水线更复杂;自助服务也暴露每日请求上限 |
| Vision 检索和问答 | 运行语义搜索、带时间戳检索和长视频问答 | 已索引分块、阈值、分组过滤和报告生成 | 比纯上下文窗口提示更贴合企业归档工作流 | 客户素材下的边缘案例和更大规模准确性仍需行业验证 |
| Edge 本地运行时 | 在设备附近或私有基础设施中执行视觉推理 | HF / vLLM 工具、受支持硬件、量化 | 降低往返延迟,并支持隐私敏感部署 | 商业自托管条款和支持边界没有完全公开 |
| Research 推理层 | 结合浏览、文档工具和并行候选答案解析 | 工具、resolver 模型、定价模式 | 让研究任务的准确性 / 成本权衡显性化 | 仍高度依赖公司运行的基准和持续演化的智能体行为 |
| 智能体和自动化适配器 | 在 MCP 客户端和 n8n 中暴露 Vision 及相关工作流 | reka-mcp、n8n node、SDK 仓库 | 比原始 REST 端点更能改善开发者体验 | 适配器覆盖不均,一些新界面落后于主 API |
这张表映射公开运营模型,而不是未公开的内部基础设施;各行聚焦外部可见层和依赖。
[CE001, CE004, CE005, CE006, CE007, CE010]Reka 把工作流产品和开发者适配器叠在紧凑型多模态模型之上,并为长媒体配置独立的检索 / 索引基础设施。
这套堆栈反映公开产品表面和文档,不代表未公开内部基础设施。
[CE001, CE004, CE010, CE017, CE020, CE022]5.3 部署模式、集成与开发者体验
对一家年轻模型公司来说,Reka 的部署叙事异常居中。quickstart 文档同时覆盖托管 API 使用和本地 Edge 部署,包括 Apple Silicon 本地运行指引,以及面向更高吞吐服务的 Linux CUDA 与 vLLM 指引。Hugging Face 模型卡进一步说明,本地部署不是营销后的补充:Flash 3.1 以 Llama 兼容格式发布,GitHub 组织也展示了 vllm-reka 插件、SDK、reka-mcp、clip 示例和官方 n8n node 等支持资产。换句话说,Reka 试图在开发者已有工作环境中与他们相遇,而不是强迫他们进入封闭专有运行时。 同样模式也出现在智能体工具中。Vision 的 MCP server 明确面向 Claude Code、Codex、Cursor 及类似智能体客户端,通过智能体原生控制平面暴露视频上传、索引、搜索、Q&A、转写检查和目标检测。n8n node 把同一逻辑推进低代码自动化;剪辑、图像提示、长视频 Q&A 和目标检测已经有文档,而 Research、Speech 和 Text 支持仍标为后续推出。这些产品面让应用型构建者更容易理解平台,比原始端点目录更清楚。 Reka 也依赖分销和互操作伙伴,而不是只坚持直接 API 采用。Snowflake 把 Reka 多模态能力嵌入 Cortex;NVIDIA 的 VSS blueprint 为 Vision 提供大规模视频部署的标准流水线;Oracle 国防生态纳入则释放出对安全和类主权环境的 GTM 兴趣。取舍在于,伙伴触达可以加速企业采用,但产品成功也会部分依赖第三方平台及其采购周期。[CE001, CE010, CE011, CE012, CE013, CE014]
| 日期或阶段信号 | 功能或里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2024-04 技术报告 | Core / Flash / Edge 作为从零训练的多模态模型家族推出 | 历史基础 | 公司一开始就押注完整多模态模型,而不是后来补上视觉能力 | arXiv 技术报告 |
| 2025 推理发布周期 | Flash 3.1 RL 升级和 Llama 兼容本地部署路径 | 已公开发布 | 强化小模型和智能体规划器叙事 | Flash 3.1 文章 + 模型卡 |
| 2025 研究发布周期 | Research-Eval 基准和 Parallel Thinking 模式 | 已公开发布 | 释放出持续投入有依据研究质量的信号,而不只是基础聊天 | Research-Eval + Parallel Thinking 文章 |
| 2025 语音发布 | 用于转录和翻译、带时间戳的 Speech 模型 | 已公开宣布 | 将技术栈扩展到音频密集型企业工作负载 | Reka Speech 文章 |
| 2025 合作伙伴扩张 | Snowflake Cortex、NVIDIA VSS 和 Oracle 防务分销信号 | 公开可见 | 表明路线图正向企业与类主权渠道倾斜 | Snowflake、NVIDIA 和 Oracle 页面 |
| 2026 生态成熟 | MCP 服务器加 n8n 节点,部分界面仍标为“soon” | 公开但不均衡 | 开发者体验在改善,但适配器覆盖仍落后于核心 API | MCP 文档 + n8n 代码库 |
各行把公开可见发布和集成资产作为路线图信号;它们不是管理层提供的未来承诺。
[CE010, CE013, CE014, CE015, CE016, CE018]Reka 的产品交付依赖第一方模型资产、本地运行时工具和企业分销伙伴的组合。
这些依赖是商业和运营依赖,不是源代码依赖。
[CE012, CE014, CE015, CE016, CE029, CE039]5.4 效率策略与前沿规模的取舍
Reka 的技术策略最好理解为拒绝单纯用模型尺寸与既有巨头对打。Edge 发布、Labs 页面、量化发布和 Flash 3.1 文章都强调效率、延迟、紧凑视觉表示和可部署性。Edge 被营销为比可比开放权重同类更快,图像 tokens 更少、延迟更低,并且能实际兼容本地硬件;Flash 3.1 被描述为通过强化学习改进的推理模型,再借助 Llama 兼容和量化发布路径更容易运行。这是一套连贯产品哲学:更小、更可控的多模态模型,应当先于巨型通用系统赢下一些企业任务。 这种方法在视频密集和边缘密集工作流里的优势很明显。如果买家需要对摄像头录像做自然语言搜索、目标定位、长视频 Q&A,或靠近设备做离线多模态推理,Reka 的打包看起来比许多文本优先既有巨头更锋利。arXiv 报告和 Edge 材料也给了公司一个可信叙事:其较小模型可以在多模态任务中越级发挥。 但前沿既有巨头仍在泛化宽度、企业支持成熟度和独立基准可见度上占优。ChatGPT Enterprise 公开强调 rollout 支持和 SLAs,Gemini 则公开强调长周期智能体与多模态宽度。Artificial Analysis 跟踪 Reka Flash,但目前第三方对最新 Edge、Vision 和 Research 产品面的覆盖,仍比领先既有巨头更薄。因此,Reka 的效率论点是真实的,但它必须持续比大型供应商向下扩张、开放模型追赶更快地转化为工作流级结果。[CE018, CE021, CE024, CE025, CE026, CE027]
Reka 在可部署多模态工作流比通用前沿广度更重要的场景里最强。
序数评分:3 = 公开成熟度强,2 = 有实质但不完整,1 = 早期或公开来源验证较薄。
[CE020, CE021, CE024, CE027, CE033, CE034]5.5 信任、可靠性与待尽调缺口
公开材料里,信任与可靠性这层已经有实质内容,但还不完整。积极的一面是,Reka 记录了结构化 JSON 错误、用于日志关联的 request ID、针对 400/401/404/429/500 状态的明确客户端动作,以及可见的 Vision 速率限制响应头。隐私政策也比泛泛营销话术更具体:付费 API 内容不会用于模型训练,除非客户主动选择加入;免费或促销使用可能会被使用;上传文档可能临时存放在安全 S3 中,配有会过期的链接和限时删除;政策还给出了漏洞报告联系方式。这些都是企业开发者可用的运营信号。 Vision 的元数据标注能力又增加了一层运营控制:它暴露 violence、profanity、adult content、drugs、alcohol、gambling、political markers 等面向内容的字段,也包括描述性标签和偏传播性的字段。这说明 Reka 理解真实部署需要分类和策略钩子,而不只是开放式生成。 不过,抓取到的材料没有呈现谨慎买家在大规模迁移敏感工作负载前常要看的经审计信任中心证据:公开 SOC 2 或 ISO 引用、公开 uptime 历史,或可与大型在位厂商相比的明确支持承诺。独立评测也强调,Reka 仍以企业和开发者为中心,集成工作和重新验证开销都是实际成本。技术叙事很强;运营证明层在改善,但公开成熟度还没有同等跟上。[CE009, CE029, CE030, CE031, CE032, CE035]
| 控制或信号 | 公开状态 | 范围 | 帮助解决的问题 | 缺口或注意事项 |
|---|---|---|---|---|
| 付费请求训练默认设置 | 已记录 | API 内容 | 付费请求不会用于训练,除非客户选择加入 | 免费或促销使用可能用于改进,因此环境选择很重要 |
| 临时文件暂存和删除 | 已记录 | 上传文档 / 连接文件 | 政策描述了安全 S3 暂存、过期链接和限时删除示例 | 运营落地停留在政策层面,公开材料中没有独立审计 |
| 结构化 API 错误和请求 ID | 已记录 | 开发者运营 | 支持调试、重试逻辑,以及用关联请求 ID 升级处理 | API 可运营性文档没有配套公开正常运行时间历史 |
| Vision 速率限制标头 | 已记录 | Vision 自助端点 | 让请求预算和退避逻辑清晰可见 | 公开配额较低,会把严肃工作负载推向企业计划 |
| 政策敏感媒体的内容标签字段 | 已记录 | Vision 元数据输出 | 提供暴力、粗口、成人内容、毒品、酒精、赌博和政治标记 | 分类质量和误报率没有独立公开 |
| 经审计的信任 / 状态 / 认证证据 | 已抓取材料中未露出 | 企业尽调 | 受监管和对正常运行时间敏感的买方会重视这些证据 | 本次审阅来源中没有露出公开 SOC 2、ISO 或状态中心证据 |
各行区分已记录的控制与缺失的公开证据;“未露出”指在抓取的公开材料中未发现,不代表该控制在内部不存在。
[CE009, CE029, CE030, CE031, CE032, CE039]5.6 证据图表
06客户
6.1 客户细分与理想客户画像
Reka 的公开界面显示,它主要卖给技术买家,而不是大众消费者。首页和 Vision 材料反复把平台定位给安全、媒体、国防和其他物理世界工作流里的企业、创作者和开发者;快速上手和定价页则给 API 原生构建者提供自助路径。实际看,至少形成四条客户线:想要 OpenAI 兼容多模态 API 的开发者;需要对视频做搜索、告警和调查的企业安全与运营团队;希望大规模做元数据增强或内容理解的媒体与数据平台买家;以及把 Reka 能力打包进自身产品套件的平台 / 渠道伙伴。因此,最有吸引力的 ICP 是这样的组织:拥有大量非结构化视觉或多模态数据集,工程团队熟悉 API 集成,并且有明确的受治理部署业务场景。国防、公共安全或受监管企业场景里,需要 on-premise、VPC 或 air-gapped 部署的买家尤其契合。小型非技术团队仍可通过 playground 和预付 API credits 试用,但公开材料显示,他们会更早撞上实际限制,因为速率限制、实施工作和仅企业可用功能,会把严肃生产使用推向更高接触度的销售动作。[CU001, CU002, CU003, CU004, CU005, CU006]
| 分层 | 买方 / 用户 / 付款方 | 主要用例 | 公开证据 | 主要缺口 |
|---|---|---|---|---|
| 开发者 / API 原生团队 | 工程负责人 / 开发者 / 产品或工程预算 | OpenAI 兼容的多模态聊天、研究和自动化工作流 | 快速入门、API 定价、MCP、GitHub 和 n8n 资产 | 没有公开披露从沙盒使用转为付费生产的转化数据 |
| 企业安全 / 公共安全运营方 | 安全运营负责人 / 调查员 / 安全或 IT 预算 | 面向视频的自然语言搜索、告警、片段检索和事件报告 | Vision 页面、Ohio 警方案例、Turing Guardian AI、防务安全材料 | 公开证据集中在少数案例,缺少广泛续约数据 |
| 媒体与内容平台 | 数据 / 产品团队 / 运营用户 / 产品或数据预算 | 元数据丰富、搜索相关性和大型多媒体库理解 | Shutterstock 客户与案例研究材料 | 公开详细披露的主要媒体客户只有一家 |
| 通过伙伴进入的受治理数据云企业 | 数据平台团队 / 分析师和构建者 / 平台预算 | 在既有数据资产内做多模态分析 | Snowflake 投资、Cortex 集成和文档 | 公开证据显示渠道可用,不显示活跃终端账户数量 |
| 防务 / 受监管项目 | 项目负责人 / 安全运营人员 / 任务或基础设施预算 | 隔离网络或本地部署的视觉智能 | 防务安全页面和 Oracle 生态清单 | 生产客户和采购速度未公开披露 |
分层把直接 API 买方与渠道驱动或垂直化客户区分开;公开证据在安全、媒体和受治理企业部署上最强。
[CU001, CU003, CU004, CU005, CU006, CU016]| 指标 | 数值 | 日期 / 状态 | 来源 | 置信度 | 含义 | 缺失分母 |
|---|---|---|---|---|---|---|
| 自助入门 | 免费账户 + 预付费额度 + API 密钥 | 当前公开文档 | Reka 快速入门 / 常见问题 / 定价 | 高 | 开发者评估门槛低 | 未披露免费到付费转化 |
| Vision 公开请求上限 | 限额:100 image uploads/day; 50 video uploads/day; 10 clip jobs/day | 当前公开文档 | Reka Vision 速率限制页面 | 高 | 自助使用是真实存在的,但有边界;严肃工作负载很可能需要企业计划 | 未披露企业上限或平均生产工作负载 |
| Shutterstock 元数据覆盖面 | 案例研究范围内有 550M 图像 / 视频资产;每年新增 60M+ 资产 | 2024 案例研究叙述 | Shutterstock 博客 | 中 | 如果推出足够广,候选扩张基数很大 | 未披露 Reka 处理的资产占比 |
| Turing 已安装基础 | 13,000+ 站点部署;每天处理 10M+ 事件 | 伙伴公告 | Reka / Turing 合作帖子 | 中 | 渠道伙伴让 Reka 触达大规模安防足迹 | 未披露该基础中有多少使用 Guardian AI |
| Ohio 警方部署代理指标 | 案件解决速度提升 65%;运营成本节省 42%;警员满意度 89% | 公司声称的部署结果 | Reka VMS 集成文章 | 低中 | 表明可量化 ROI 能支撑试点后的扩张 | 单一案例研究;未披露独立验证或样本量 |
| Snowflake 渠道扩张 | Reka 模型已在 Cortex 中可用,用于多模态分析 | 当前公开产品可用性 | Snowflake 博客和文档 | 中高 | 渠道能扩大触达,不必每次都直接双边销售 | 未披露活跃账户数或消耗量 |
轨迹证据依赖公开工作流和渠道信号,而不是已披露客户标识数或 ARR。缺失分母使留存或渗透分析无法做扎实。
[CU004, CU005, CU008, CU012, CU015, CU020]典型 Reka 客户旅程从技术评估开始,随后分叉为自助实验、顾问式试点或渠道牵头采用。
[CU003, CU004, CU005, CU024, CU025, CU027]6.2 具名客户与伙伴证明:证据真正说明了什么
最干净的直接客户证明是 Shutterstock。Reka 和 Shutterstock 都称,双方达成多年安排:Shutterstock 向 Reka 授权训练数据,同时聘用 Reka 增强 Shutterstock 图像和视频库附带的元数据。这个信号有意义,因为它证明了真实媒体数据运营中的付费工作流价值,而不只是标识展示。不过,它没有披露合同金额、跨业务单元 rollout 广度,或关系是否已从元数据增强扩展出去。Snowflake 是另一类证明点。Snowflake 先投资 Reka,随后公开把 Reka 模型集成进 Cortex,让 Snowflake 客户能在数据原本所在的位置运行多模态分析。这验证了渠道可信度和企业治理定位,但不等于证明 Snowflake 本身是 Reka 的大型直接应用客户。Turing 又不同:Reka Vision 在 Turing 的监控平台之上驱动 Guardian AI,为 Turing 的装机基础提供自然语言搜索和 agentic 事件工作流。合在一起,这三个例子说明,Reka 在接入既有平台、目录或摄像头网络时更容易赢;它们还不能证明公司拥有庞大的独立客户名单,或网站上每个标识都有广泛生产使用。[CU010, CU011, CU012, CU013, CU014, CU015]
| 名称 | 分层 | 部署 / 用例 | 生产与试点 | 证据证明了什么 | 局限 |
|---|---|---|---|---|---|
| Shutterstock | 媒体 / 内容平台 | 向 Reka 授权数据,并用 Reka 丰富其图像和视频库的元数据 | 生产商业关系 | 确认一个具名付费客户用例,绑定元数据运营和搜索相关性 | 未披露合同金额、推出广度或续约历史 |
| Snowflake | 数据云 / 渠道伙伴 | 让在受治理数据资产内工作的客户可在 Snowflake Cortex 中使用 Reka 多模态模型 | 生产渠道集成 | 确认通过大型平台做企业分销,叙事围绕治理和安全 | 不显示有多少 Snowflake 账户活跃使用 Reka,也不显示多少转化为持久支出 |
| Turing | 物理安防平台伙伴 | Guardian AI 基于 Turing 平台和 Reka Vision 构建,用于自然语言监控搜索和告警 | 生产伙伴发布 | 确认嵌入一个有装机基础的平台,且有真实终端用户工作流 | 公开证据不显示 Turing 站点中有多少采用了 Guardian AI |
| Orange Village / Ohio 警方案例 | 公共安全运营方 | 在既有摄像头网络上做调查工作流,将 Reka Vision 叠加到监控栈 | Reka 声称的生产案例研究 | 表明 Reka 能在调查中支撑可量化 ROI | 客户确认是间接的,且集中在公司撰写材料 |
最好的证据点横跨直接客户使用、渠道嵌入和垂直解决方案包装。它们都真实存在,但性质并不相同,公开披露也无法汇总成一个经验证的客户总数。
[CU010, CU011, CU013, CU014, CU016, CU018]公开证明质量在具名案例之间差异很大;Shutterstock 是最强的直接客户信号,Snowflake 和 Turing 则更能证明分发和嵌入式工作流。
单元格是作者对公开证据深度的定性评估,不是私下尽调结论。
[CU010, CU011, CU014, CU016, CU019, CU022]6.3 购买动作、部署模式与渠道角色
Reka 似乎采用分层 go-to-market 模型。低端入口处,开发者可以创建免费账号、预付 credits、生成 API key,并从 OpenAI 兼容调用或 n8n node、工作流模板等社区自动化资产开始。这降低了构建者和小团队的评估摩擦。但公开文档也显示了自助路径的边界:Research 的内部数据访问仅限企业,Vision 在公开 keys 上的速率限制较低,更高 quota 或专门部署模式需要直接联系。面对安全和国防买家,销售动作会更偏咨询式。Reka 的国防页面描述了资格筛选、环境评估、在客户边界内试点部署和运营交接。VMS 集成文章明确建议把 Vision 叠加到既有摄像头栈上,先从一组高价值摄像头开始,再横向扩展。渠道很重要,因为它们缩短了这段旅程。Snowflake 让客户在受治理的数据云环境中消费多模态模型;Turing 把 Vision 打包进已部署在数千个站点的监控产品;GitHub 加 n8n 则为技术熟练的采用者打开生态路径。因此,扩张逻辑不像病毒式席位增长,更像是在窄试点证明 ROI 后,扩大摄像头覆盖、索引更多媒体、增加更多工作流并加深集成。[CU009, CU024, CU025, CU026, CU027, CU028]
| 扩张驱动 / 风险 | 当前证据 | 影响 | 尽调路径 |
|---|---|---|---|
| 用有限试点落地,再扩到更多摄像头、影像或工作流 | VMS 集成文章建议先做高价值子集,再横向推出 | ROI 可衡量则为正面;试点卡住则放慢 | 要求试点到推出的转化和平均扩张周期 |
| 借 Snowflake 和 Turing 撬动渠道 | 公开集成把 Reka 延伸到数据云和监控环境 | 能在没有直接销售的情况下加速触达,但也可能模糊直接客户归属 | 拆分渠道来源收入、用量和集中度 |
| 安全 / 防务部署复杂度 | 本地、VPC 和隔离网络支持适合高价值账户,但拉长采购 | ACV 潜力更高,但周期更长、买方池更窄 | 按部署模型要求销售周期长度和赢率 |
| 客户标识与生产的模糊性 | 官网和伙伴生态讲出了强叙事,但广泛客户名册细节有限 | 如果把客户标识当成活跃续约客户,可能夸大成熟度 | 要求生产推荐证明清单,附上线日期和扩张证据 |
| 收入集中度风险 | 未公开披露头部账户、渠道组合或客户数量分布 | 少数战略账户或伙伴可能贡献过大的收入份额 | 要求前十大客户集中度和伙伴收入依赖 |
扩张逻辑看起来由工作流和渠道驱动;但集中度仍是重大未知,因为没有公开申报文件或客户指标。
[CU016, CU018, CU025, CU026, CU027, CU028]Reka 的采用路径先从广泛技术兴趣收窄到更高价值的试点和生产部署,再靠渠道和工作流增长重新放大。
[CU009, CU024, CU025, CU026, CU028, CU029]6.4 留存信号、采用障碍与公开证据缺口
公开留存证据很薄。Reka 没有披露客户数、NRR、GRR、流失率、平均合同期限或收入集中度,具名证明也不包含续约历史。可用的代理信号都是间接的:Shutterstock 的用例嵌在元数据增强运营里,Snowflake 把 Reka 放进更大的企业平台,Turing 加俄亥俄州执法案例则暗示监控调查工作流有深度。这些信号支持一种判断:集成完成后,切换成本可能变得有意义;但它们不能替代 cohort 或收入留存数据。主要采用障碍也能在公开记录中看到。小型买家要面对预付 credits、技术集成工作、有限的公开速率上限,以及更奖励工程成熟度的 API-first 界面。即便是一篇来自竞争对手的正向第三方定价评测,也把 Reka 描述为强大的原始模型层,而不是开箱即用的支持产品,并指出可变 token 经济性和定制集成会让低技术团队更难做预算和部署。因此,最大的 diligence 缺口不是 Reka 能否做有意思的多模态工作;而是公司是否已把这种能力转化为多元化、可续约、生产级的客户基础,而不只是少数被充分宣传的伙伴和案例。[CU031, CU032, CU034, CU035, CU036, CU039]
| 指标 | 数值 / 状态 | 分层 | 置信度 | 尽调要求 |
|---|---|---|---|---|
| NRR / GRR | 所有客户分层 | 低 | 要求按直接、渠道和企业部署路径拆分群组留存 | |
| 客户标识流失 / 续约率 | 直接企业和渠道账户 | 低 | 要求续约排期、总流失和可引用续约 | |
| 工作流黏性代理指标 | Reka 嵌入元数据、数据云或摄像头工作流时为中高 | Shutterstock / Snowflake / Turing 类账户 | 中 | 验证生产集成在首次部署后是否扩张 |
| 公开用户满意度证据 | 稀疏;最强引用证据来自 Ohio 警方和 Shutterstock 高管评论 | 安全和媒体案例 | 低中 | 要求独立客户推荐证明和上线后 KPI 复盘 |
| 采用摩擦 | API 优先界面、预付费额度、公开速率限制和集成工作会增加小团队摩擦 | 开发者和 SMB 潜在客户 | 中 | 询问各分层的入门漏斗指标和平均投产时间 |
公开来源未披露群组、收入留存或续约数据的地方,null 是有意保留。黏性评分依赖工作流嵌入代理指标,而不是已报告的收入留存。
[CU008, CU024, CU031, CU032, CU034, CU036]对公开记录最充分的客户和渠道案例给出的定性留存信号评分。
分数是 0-100 的定性代理指标,衡量工作流嵌入、切换成本和扩张潜力;Reka 未公开披露真实队列留存或 NRR。
[CU013, CU018, CU021, CU023, CU031, CU039]6.5 证据图表
07风险
7.1 监管、版权与数据治理暴露
多模态模型提供商面临的法律和治理负担已经从理论问题变成具体问题。Reka 自身条款在一个窄点上有帮助:公开条款称,付费 API 请求不会用于模型训练,除非客户明确选择加入;免费使用可用于改进模型,聚合匿名使用数据仍可能为运营目的保留。这对企业采购有方向性帮助,但还不能证明公司已经具备监管机构和大客户现在期待的稳健训练数据 lineage、权利管理或删除工作流。European Commission 的 GPAI 指引和 AI Act 文本说明了问题为何重要。通用 AI 模型提供商必须维护技术文档、执行版权政策并发布训练内容摘要;被认定有系统性风险的模型还要承担额外的事件报告、风险缓释和网络安全义务。与此同时,EDPB、ICO 和 California 隐私制度都指向第二层暴露:个人数据的合法基础、生物识别或敏感数据处理、数据主体权利,以及可解释治理。实际风险不是 Reka 今天显然无法履行这些义务;而是本章审阅的公开记录还没有证明公司如何在视频、图像、音频和文档工作负载中把这些义务运营化。这个缺口很重大,因为一次隐私或版权争议就可能拖慢企业采用,其速度远快于模型质量改进带来的加速。[CR001, CR002, CR003, CR004, CR005, CR006]
| 风险 / 规则 | 司法辖区 | 当前公开信号 | 可能性 | 严重性 | 缓解成熟度 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| AI Act GPAI 透明度和版权义务 | 欧盟 | 条款和委员会指引要求技术文档、版权政策和训练内容摘要 | 高 | 高 | 早期 | 高 | 对照 Chapter V 义务审查 Reka 训练数据摘要、版权政策和事件工作流 |
| 模型足迹跨过阈值后的系统性风险升级 | 欧盟 | 系统性风险提供方必须通知委员会、缓解风险、报告事件并落实网络安全保护 | 中 | 高 | Unknown | 高 | 要求内部判断:当前或规划中的任何模型家族是否可能触发 GPAI 系统性风险待遇 |
| 训练数据版权挑战 | 美国 / 全球 | 版权局已提升训练和输出问题的优先级;公开记录无法证明 Reka 的授权或退出运营 | 高 | 高 | Unknown | 高 | 检查来源日志、供应商授权、退出受理和赔偿条款 |
| 个人数据合法依据和删除权 | 欧盟 / 英国 / 加州 | EDPB、ICO 和 CCPA 指引都强调合法依据、透明度和用户权利 | 高 | 高 | 早期 | 高 | 审查 DPA/DPIA 包、删除工作流、数据主体请求指标和留存设置 |
| 生物识别或监控专项审查 | 英国 / 欧盟 / 客户特定 | ICO 标出生物识别指南;Reka 营销安全和视频用例 | 中 | 高 | Unknown | 中高 | 测试部署手册是否区分搜索 / 摘要与身份敏感型生物识别用例 |
各行按最影响决策的公开法律与监管暴露排序;公开记录无法证明运营控制时,剩余暴露仍高。
[CR001, CR002, CR003, CR005, CR006, CR007]残余风险集中在治理、算力获取和竞争基准压力上,而不是某个孤立问题。
数值是作者根据引用的公开记录综合出的风险排序;它们是序数,不是统计概率。
[CR041, CR018, CR023, CR037, CR039, CR040]7.2 基础设施、算力与依赖风险
Reka 的战略依赖一个命题:即便整个行业竞相部署更大、更多工作负载,高效多模态模型仍能保持商业吸引力。公开宏观证据更像警告,而不是安慰毯。IEA 称,2025 年 data-centre 用电需求激增 17%,到 2030 年 data-centre 用电将翻倍,AI-focused 站点则增长三倍;同时,chips、transformers、turbines 和电网连接都在收紧。RAND 得出类似尖锐结论,估计到 2027 年全球 AI data-centre 电力需求为 68 gigawatts,并警告电力短缺可能把基础设施推向海外,带来安全和出口管制后果。DOE 和 BloombergNEF 进一步强化了同一问题的运营版本:gigawatt-scale 项目已经撞上电网 lead time 和 reserve-margin 压力。对 Reka 来说,这有两层含义。第一,即便模型效率真实存在,公司仍处在与大型竞争对手相同的上游算力、电力和托管瓶颈中。第二,本章公开记录仍未披露 Reka 是否有预留 GPU capacity、优先云经济性,或有电力背书的托管承诺来保护供应紧张时的服务质量。结果是典型的不对称风险:公司可以赢下产品评估,但如果上游 capacity 成为真正 choke point,仍可能输掉 gross margin 或交付可靠性。[CR013, CR014, CR015, CR016, CR017, CR018]
| 故障模式 | 公开证据 | 可能性 | 严重性 | 缓解成熟度 | 剩余暴露 | 未解决缺口 |
|---|---|---|---|---|---|---|
| 电力 / 电网约束延迟容量 | IEA、RAND、DOE 和 BNEF 都指向数据中心电力需求上升和基础设施瓶颈 | 高 | 高 | 低中 | 高 | 没有公开证据显示有电力保障的托管或容量承诺 |
| GPU / 芯片或组件短缺 | IEA 和 RAND 将 AI 扩展与先进芯片及组件供应限制相连 | 中高 | 高 | Unknown | 高 | 未公开披露供应商集中度、预留权或备用供应策略 |
| 多模态越狱或有害输出事件 | CSA / Enkrypt 报告显示,在对抗性输入下多模态漏洞更高 | 中 | 高 | Unknown | 中高 | 此处审阅范围内,没有 Reka 模型公开安全红队或事件报告指标 |
| 企业安全审查卡住 | Nudge Security 强调采购团队会问同一组问题:认证、供应链、泄露历史和 GDPR 姿态 | 中 | 中高 | 早期 | 中 | 已审阅公开来源未清楚显示具名认证或近期独立测试 |
| 发布治理回归 | 公开变更日志显示产品和模型更新活跃,抬高回归管理负荷 | 中 | 中 | Unknown | 中 | 需要证明各版本发布中的回滚、基准测试和支持响应纪律 |
运营行把上游算力约束与模型质量、安全故障模式放在一起;严重性反映对交付、利润率和企业信任的影响。
[CR012, CR013, CR014, CR015, CR016, CR017]| 依赖 | 交易对手 / 集群 | 角色 | 集中度信号 | 失败场景 | 严重性 | 缓解 | 剩余暴露 |
|---|---|---|---|---|---|---|---|
| 上游算力 | GPU 供应商 / 托管栈 | 训练和服务容量 | 行业集中度高 | 容量配给或成本飙升挤压利润率并拖慢交付 | 高 | 效率叙事加潜在多供应商对冲 | 高 |
| 电力和电网接入 | 区域公用事业 / 站点运营方 | 支撑大规模推理托管 | 电网项目和储备裕度已经吃紧 | 新工作负载无法按时部署到需要的地方 | 高 | 使用低功耗模型和地理上灵活的托管 | 高 |
| 企业隐私基准 | OpenAI / Anthropic / Mistral / Cohere | 竞争性买方预期 | 竞争对手公开营销强控制 | 安全或法务审查偏向更大或更适合私有部署的供应商 | 高 | Reka 付费方案选择加入机制及企业条款 | 中高 |
| 开放权重替代品 | Meta Llama / Google Gemma / Mistral Small(开放权重模型) | 低价或可自托管的替代方案 | 开放模型质量和可部署性快速提升 | 多模态工作负载面临价格压缩或切换成本下降 | 高 | 靠工作流封装、视频工具和支持服务拉开差距 | 中高 |
| 战略渠道和投资者 | Snowflake / Nvidia 生态 | 分发、可信度和基础设施入口 | 杠杆明显,但可能集中 | 合作伙伴优先级转移,或经济性变差 | 中高 | 分散直销企业关系和托管选项 | 中高 |
本表聚焦外部依赖;即使产品质量不变,这些依赖也会改变定价权、分销触达或交付可靠性。
[CR018, CR019, CR020, CR021, CR022, CR023]下行情形从治理或产能约束开始,传导到销售摩擦、利润率压力、融资需求和估值压缩。
连线表达业务上的方向性传导,不代表确定因果。
[CR018, CR021, CR023, CR038, CR041, CR046]Reka 所处的依赖栈包括监管机构、电力和算力提供商、合作伙伴渠道,以及可信度越来越高的开源模型替代品。
这里的依赖是商业、法律和基础设施依赖,不是源代码依赖。
[CR023, CR024, CR027, CR028, CR030, CR032]7.3 竞争、客户集中与融资风险
Reka 并不是在一个高效多模态模型自然就能拿到溢价的真空市场里竞争。竞争对手现在恰好在宣传企业买家评估小供应商时最在意的控制项:OpenAI 称客户保留对 inputs 和 outputs 的所有权与控制权,Anthropic 主打企业搜索、SSO、基于角色的控制、合规 API,以及默认不使用客户内容训练模型;Mistral 强调 self-hosted 和 hybrid deployment,并让客户完全拥有数据;Cohere 则营销 VPC 或 on-prem deployment 以及 training opt-out。同时,open-weight 压力也变得更严肃。Meta 把 Llama 定位为 open-source,Gemma 突出 cloud-to-device deployment 和 safety classifiers,Mistral Small 3.1 称可在单张 RTX 4090 上运行,且采用 Apache 2.0 license。这种组合压缩了小型闭源提供商仅靠模型访问取胜的空间。融资有帮助,但不能消除风险。公开报道确认 Reka 完成了由 Nvidia 和 Snowflake 支持的 $110 million 融资,估值 $1 billion;追踪机构则把公司规模放在约 60-64 名员工、2025 年收入约 $10.9 million。这些数字本身还不是推翻 thesis 的证据,但它们意味着估值仍假设未来 operating leverage、多元化企业合同,以及资本市场继续愿意为 AI infrastructure 暴露买单。公开来源仍没有显示 top-customer concentration、retention 或预留 channel economics,因此下行情境仍对少数账户和伙伴关系高度敏感。[CR025, CR026, CR027, CR028, CR029, CR030]
| 角色 / 职能 | 依赖或缺口 | 发生可能性 | 严重性 | 缓释措施 | 剩余暴露 | 尽调路径 |
|---|---|---|---|---|---|---|
| 合规 / 隐私负责人 | 必须覆盖多模态产品在 AI Act、GDPR、英国和加州规则下的义务 | 中高 | 高 | 借助外部律师和企业条款 | 高 | 要求披露隐私和数据权利流程的具名负责人、外部顾问及运营指标 |
| 安全 / 安保工程 | 多模态模型需要红队测试、发布治理和事件响应 | 中 | 高 | 借力框架指引和采购压力 | 中高 | 要求披露红队节奏、事件日志、回滚流程和认证路线图 |
| 销售 / 客户成功带宽 | 客户集中或漫长安全审查会拉伸小团队 | 中高 | 中高 | 合作伙伴渠道有助于分发 | 中 | 要求提供前 10 大账户地图、支持配比和续约流程 |
| 财务 / 基础设施规划 | 高估值和算力依赖要求严格管理资本开支和 runway | 中 | 高 | 近期融资提供了时间缓冲 | 中高 | 要求提供关于 burn、按工作负载拆分的毛利率和下一轮融资触发条件的董事会材料 |
| 创始人 / 关键人集中 | 公开追踪员工数约 60-64 人,说明领导层集中度不低 | 中 | 中高 | 扩充高管梯队并下放 ownership | 中 | 要求提供组织架构、招聘计划,以及产品、基础设施和企业职能的继任覆盖 |
各行把公开追踪数据和产品节奏转成可能的管理带宽约束,而不是假装公开记录能证明内部组织质量。
[CR033, CR034, CR035, CR036, CR037, CR038]7.4 人员、安全与执行扩张风险
Reka 同时要做几件难事,执行风险被放大:交付接近前沿的多模态产品,满足对隐私敏感的企业买家,支持伙伴分发,并跟上快速变化的监管边界。公开 tracker 数据显示,团队只有约 60-64 人;以当前产品面来看很 impressive,但也提示管理带宽、安全运营、支持覆盖和合规专门化可能被摊得很薄。外部安全证据也并不舒服。Cloud Security Alliance 转载的多模态安全报告显示,在对抗条件下,基于图像的 prompt injection 和多模态 jailbreaks 会显著增加有害输出和危险信息泄露。MIT 的 AI risk repository 则把问题进一步扩大:有害内容、隐私泄露、模型利用和 misinformation 不是彼此分离的边缘案例,而是一组反复出现、会相互叠加的风险。Reka 自身公开 change log 显示,功能和模型仍在持续变化;这在战略上是正面的,但也提高了对严格 release governance 和 regression testing 的需求。这些并不能证明 Reka 不安全。它证明的是,投资人应该把公司承销为一个攻击面持续扩大的实时运营系统,而不只是一个模型 benchmark 故事。在这种规模的公司里,一个高级合规、安全、平台或企业支持缺口,都可能级联成更慢的交易、更弱的事件响应和更高的 churn 风险。[CR011, CR012, CR013, CR014, CR015, CR031]
7.5 下行情境、终止标准与尽调问题
正确理解风险栈的方式,是把它看成传导链,而不是一串互不相关的警告。训练数据 provenance 或隐私权治理一次失手,就可能拖延企业采购;采购延迟会让估值更依赖未来融资轮;如果算力和电力稀缺压缩 gross margin,或 open-weight 对手缩小产品 moat,未来融资会更难;小团队也就更难吸收事件响应、企业安全要求或渠道摩擦。这条逻辑给出了清晰的 diligence 门槛。投资人应要求能补上公开缺口的具体证据:训练数据 provenance 与 opt-out 记录、具名安全认证和近期 pen-test 输出、已承诺算力或托管 capacity、top-customer 与伙伴集中度,以及与监管地图匹配的合规招聘计划。如果管理层无法展示这些项目,审慎姿态就不应是轻微谨慎,而是降低 conviction score 或明确等待。反过来,如果 Reka 能在保持低成本推理和多元需求的同时证明企业级治理,同样的公开风险就会变得可管理,而不是致命。投资 case 因此取决于运营证明,而不是对 AI 市场的泛泛乐观。[CR018, CR023, CR038, CR039, CR041, CR042]
| 风险 | 可监控触发条件 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 训练数据 / 版权治理 | 来源证据和退出处理 | 尽调资料室没有记录数据集沿袭、权利分析或下架流程 | 暂停承保,或要求投资前完成法律整改 |
| 隐私 / 生物识别暴露 | 数据权利运营指标 | 没有 DPIA 式流程、删除 SLA,或针对监控部署的敏感数据边界 | 降低信心,并收窄目标客户论证 |
| 算力 / 电力入口 | 预留容量和托管经济性 | 未为预计未来 12-18 个月增长锁定 GPU 或托管容量 | 假设毛利率受压、融资需求上升 |
| 客户集中 | 头部账户收入占比 | 前 3 大客户或渠道贡献过大,且缺少多年留存证据 | 下调收入耐久性评级 |
| 安全与信任状态 | 认证 / 渗透测试 / 事件准备度 | 企业扩张期间没有可信安全路线图或独立测试证据 | 预期销售周期拉长、成交速度放慢 |
| 团队和执行带宽 | 组织深度和招聘计划 | 虽然产品线很宽,但合规、基础设施和企业支持没有清晰梯队 | 在补齐人员前,把管理带宽视为可能打破论点的风险 |
触发条件被写成可监控的尽调阈值,而不是叙事化担忧,方便投资团队决定继续、暂缓还是停止。
[CR018, CR023, CR037, CR038, CR041, CR042]7.6 证据图表
08估值
8.1 建议摘要与入场纪律
Reka 的公开 case 起点是一笔真实融资,而不是传闻:公司宣布 2025 年 7 月完成 $110 million 融资,NVIDIA 和 Snowflake 参投;Reuters-syndicated 报道和私营公司 trackers 则把公司估值放在 $1 billion 以上。这足以把 unicorn mark 当真,但不足以把它视为明显有吸引力。最好的公开商业信号仍来自 GetLatka 的估计:Reka 2025 年收入约 $10.9 million,团队约 60 人;Tracxn 则把公司规模放在约 64 名员工、累计融资约 $168 million。合起来看,这轮融资意味着约 92x trailing revenue,以及超过 15x cumulative funding to trailing revenue。这是 venture-style 预期,不是当下 fundamentals。 与其把这轮理解成干净的 trailing-multiple comp,不如理解成战略期权价值押注。Snowflake 不只是财务 logo;其自身 partner 材料称,客户可以把 Reka 的多模态助手带入 Snowflake 内的企业数据,这给 step-function revenue growth 提供了一个可信渠道。关于 2024 年 Snowflake 与 Reka 在 $1 billion 左右进行收购谈判的公开报道,也增加了背景:战略买家显然认为这项资产重要,但谈判仍然停止。这说明存在战略相关性,而不是已被证明的估值底部。因此,我们的建议是 **CONDITIONAL MONITOR / PASS AT CURRENT PRICE**:投资人不应把当前 mark 当作便宜,但如果 diligence 能证明更高得多的 forward run-rate、强 gross margins 和干净的 late-stage preference stack,就应继续保持接触。[CV001, CV002, CV003, CV004, CV005, CV006]
| 维度 | 当前判断 | 证据锚点 | 决策含义 |
|---|---|---|---|
| 投资建议 | 有条件跟踪 / 以当前价格放弃 | 公开基准情景仍低于本轮估值 | 只有尽调补齐核心缺口,才继续保持接触 |
| 信心 | 中 | 融资轮、追踪器和可比公司数据真实存在,但关键经济性仍是私有信息 | 不要用稀疏公开数据过度拟合精确结论 |
| 风险评级 | 高 | 隐含约 92x trailing 倍数,披露很薄 | 需要明确下行保护或更强证明 |
| 估值立场 | 偏高但并不荒谬 | 战略伙伴价值解释了部分溢价 | 当前估值已经计入重大前瞻执行 |
| 可由公开数据支撑的区间 | ~$0.2B-$1.65B;基准 ~$0.5B-$0.9B | 牛市情景需要 $60M-$75M 收入;基准情景支撑不了本轮估值 | 只有偏牛判断才适合承保当前价格 |
| 上调门槛 | 需要 >$30M-$40M run-rate、>65% 毛利率、干净条款 | 这些指标会显著压低隐含前瞻倍数 | 若拿不到,就等待或重新谈判入场纪律 |
本表把公开证据转成决策姿态;各项判断是综合研判,不是管理层指引。
[CV001, CV002, CV005, CV006, CV034, CV043]决策流把最新融资、偏薄的公开基本面、战略渠道上行空间和价格纪律连到「观望」建议。
流程突出最可能推动建议变化的变量,而非公司所有属性。
[CV001, CV005, CV007, CV008, CV028, CV034]IC 风格记分卡,衡量价格支撑、渠道杠杆、证据质量和退出可选性。
分数是基于保留证据综合出的 0-10 启发式估算,而非外部评级。
[CV005, CV008, CV034, CV037, CV041, CV042]8.2 投资 thesis 与反 thesis
可投资 thesis 很直接。Reka 拥有真实的多模态产品 wedge、可信的战略支持者,以及多数小型模型公司没有的渠道故事。Snowflake 的产品和投资材料显示,Reka 可以出现在受治理的企业数据栈内部,而不是只作为独立 API 销售。公司也把新资本定位用于技术开发和企业采用,这与投资人在这一阶段希望看到的方向一致。在这种解读下,这轮融资买的是三个相互叠加的期权:Snowflake 带动的分发、视频 / 文档密集型工作流 specialization,以及高效多模态模型产生比市场假设更好 gross margins 的可能性。 反 thesis 同样强。Bessemer 警告称,private AI cloud valuations 在 public cloud multiples 更接近历史常态时,可能又一次起泡;Equidam 的说法更严厉:当 compute-heavy 成本和 cloud-credit 结构被隐藏时,AI revenue multiples 会主动误导。可比公司集合也强化了这种谨慎。Cohere、Glean、Harvey 和 Databricks 都是在披露远高于 Reka 公开水平的 ARR、run-rate revenue、用户规模或 workflow embed 后,才拿到更高的 private marks。Anthropic 的 2024 融资过程进一步说明,AI marks 可能受 thresholds、cloud contracts 和 SPV structures 塑造,而不是由干净 price discovery 决定。正确 framing 因此不是 Reka 质量低;而是公开证明集更支持一个 strategic-upside story,而不是 fundamental value story。[CV008, CV009, CV010, CV011, CV012, CV013]
| 论点 | 当前证据 | 什么会改变判断 |
|---|---|---|
| Snowflake 牵头的企业分发可以加速规模化 | Snowflake 产品和投资材料显示,GTM 集成真实存在 | 需要渠道贡献的 pipeline、转化率和收入分成证据 |
| 高效多模态产品可能带来有吸引力的毛利率 | 管理层和合作伙伴叙事指向效率,而不是蛮力扩张 | 需要按产品和工作负载拆分、经审计的毛利率桥接 |
| 私有 AI 可比公司可以支撑溢价估值 | Cohere、Glean、Harvey 和 Databricks 都表明,私有 AI 可以获得溢价定价 | 需要证明 Reka 能迅速从约 $10.9M 收入提升到 $40M+ |
| 当前私有 AI 市场可能过热 | BVP 认为,在公开市场倍数正常化时,私有 AI cloud 可能已经起泡 | 更冷静的私有市场,或 Reka 增长放慢,都会压缩估值 |
| AI 轮次条款可能扭曲 headline valuation | Semafor 和 Equidam 都认为,绑定云服务的条款和成本结构会误导简单倍数 | 需要实际的 2025 年优先权栈和算力承诺 |
| 战略选择权真实存在,但不是估值底 | 收购谈判报道显示买方有兴趣,但谈判仍然停止 | 需要多个战略买家或可重复入站兴趣的证据 |
论点对价格敏感:同一家公司的质量,在一个入场估值下可能有吸引力,换一个估值就可能偏贵。
[CV008, CV009, CV010, CV011, CV012, CV014]8.3 Bull / Base / Bear 情境分析
情境分析应被视为可支撑价值的护栏,而不是伪精确 DCF。当前公开锚点是 >$1 billion 融资,对应约 $10.9 million 的 2025 年估计收入;核心问题是,怎样的 forward revenue 路径能让这个价格显得正常。在 $1 billion 估值下,公司在 $40 million 收入上是 25x,在 $50 million 上是 20x,在约 $66.7 million 上是 15x。这些门槛才是真正的 underwriting test。 Bull case 假设 Snowflake 和直接企业渠道快速转化,把 Reka 在未来 12-24 个月内推向约 $60 million-$75 million 收入,同时 gross margins 保持在 65% 以上,产品组合转向更高价值的工作流软件。按 18x-22x revenue 计算,可支撑约 $1.1 billion-$1.65 billion,使当前 mark 变得 plausible,虽然仍谈不上明显便宜。Base case 假设转化更温和:约 $35 million-$50 million 收入和 14x-18x revenue,只能支撑约 $0.5 billion-$0.9 billion。Bear case 假设收入仅达到 $20 million-$30 million,margins 仍受算力拖累,或渠道依赖弱于预期;按 10x-14x revenue,可支撑约 $0.2 billion-$0.4 billion。换言之,公开数据下的回报不对称:如果执行极佳就有上行,但 base case 仍低于本轮估值。[CV029, CV030, CV031, CV032, CV033, CV034]
| 情景 | 收入假设 | 倍数假设 | 可支撑估值区间 | 概率信号 | 关键条件 |
|---|---|---|---|---|---|
| 牛市 | 近期前瞻收入 $60M-$75M | 18x-22x | ~$1.1B-$1.65B | 需要超计划执行 | Snowflake / 直销企业渠道完成转化,毛利率保持软件式水平 |
| 基准 | 近期前瞻收入 $35M-$50M | 14x-18x | ~$0.5B-$0.9B | 公开数据最能支撑 | 商业化增长真实,但不足以完全支撑当前价格 |
| 熊市 | 近期前瞻收入 $20M-$30M | 10x-14x | ~$0.2B-$0.4B | 如果转化或经济性不达预期,概率不低 | 算力 / 存储拖累、伙伴依赖,或客户扩张乏力 |
情景数值是基于公开收入锚点和可比倍数区间推导出的可支撑价值区间,不是目标价。
[CV029, CV030, CV031, CV032, CV033, CV034]区间图展示可支撑价值带与当前轮次标记的对比。
这些是公开数据支撑区间,不代表内在价值或预期交易价格。
[CV031, CV032, CV033, CV034, CV046]8.4 可比估值框架
私营可比集合分成两类。第一类是已经披露实质规模的 enterprise-AI application 或 platform 公司。Cohere 2025 年 ARR 约 $240 million,估值背景约 $7 billion,意味着 high-20s multiple;Glean 的 ARR 轨迹到 $300 million、估值 $7.2 billion,也说明市场会在公司跨过远大于 Reka 公开收入估计的门槛后给予 premium marks。Harvey 和 Databricks 从不同角度说明同一点:类别领导者绝对可以达到 $11 billion 或 $134 billion 估值,但前提是 workflow embed、客户渗透或收入规模比 Reka 公开呈现的深得多。 第二类是 framing comps,而不是干净 multiple comps。Anthropic 的融资史显示,cloud-linked contracts 和 threshold-based terms 会扭曲 headline valuation 讨论。Mistral 的 2026 valuation rumor 和 Aleph Alpha 更早的 sovereign-AI 融资表明,frontier 或 sovereign narratives 能吸引巨额资本,但这些公司更接近 foundation-model 或 state-backed strategic stories,而不是 Reka 当前的 enterprise-multimodal wedge。因此,公开锚点更适合用来保持纪律。按这里保留来源,Snowflake 约 17x revenue,NVIDIA 约 24x,Palantir 约 69x,C3.ai 约 4x。Reka 隐含约 92x trailing mark,高于所有这些公司,包括 Palantir 这个离群值。这不能证明本轮定价错误——private AI 可以按 future optionality 定价——但它确实证明当前 mark 已经嵌入大量未来成功。[CV012, CV013, CV014, CV015, CV016, CV017]
| 可比公司 | 指标 / 估值锚点 | 隐含倍数或规模备注 | 为什么相关 | 局限 |
|---|---|---|---|---|
| Reka(当前) | 基于约 $10.9M 估算 2025 年收入,估值 >$1B | ~92x trailing 收入 | 本章的直接锚点 | 收入来自第三方估算,未经审计 |
| Cohere | $240M ARR 与约 $7B 估值背景 | ~29x ARR 经验值 | 企业模型 / API 可比公司,带私有部署姿态 | ARR 与融资日期并不完全同步 |
| Glean | 2026 年 5 月 ARR 达 $300M;2025 年 6 月估值 $7.2B | 获得溢价估值前,ARR 规模大得多 | 企业 AI 应用可比公司 | 当前 ARR 和最新估值来自不同观察日期 |
| Harvey | $11B 估值;1,300 家组织和 100,000+ 名律师 | 更深工作流嵌入支撑溢价估值 | 产品变得像操作系统时,展示垂直 AI 天花板 | 保留的官方来源没有披露干净 ARR 数字 |
| Anthropic | 2024 年目标估值 $15B-$20B,并有阈值 / 信贷复杂性 | 不是干净倍数可比 | 说明 frontier AI 估值可能被结构性扭曲 | 基础模型经济性与 Reka 差异很大 |
| Mistral | 传闻 2026 年融资 €20B,此前 Series C 为 €11.7B | frontier / 主权溢价框架 | 可作为欧洲基础模型参考 | 基于传闻,且保留来源没有收入锚点 |
| Aleph Alpha | 2023 年 $500M Series B | 融资规模可比,不是干净收入倍数可比 | 主权企业 AI 参考点 | 公开收入 / 估值同步性较弱 |
| Databricks | 2025 年 12 月 run-rate 收入 $4.8B、估值 $134B | ~28x run-rate 收入 | 规模化 data+AI 平台天花板 | 规模和成熟度远高于 Reka |
| Snowflake | FY2026 收入 $4.684B,对应市值 $80.51B | ~17.2x 收入 | 与 Reka 渠道背景相连的公开 governed-data-platform 锚点 | 流动性公开市场倍数,且规模成熟 |
| Palantir | FY2025 收入 $4.475B,对应市值 $307.98B | ~68.8x 收入 | 公开 AI / 平台溢价异常值 | 产品套件更宽,政府业务占比也高于 Reka |
| C3.ai | FY2025 收入 $389.1M,对应市值 $1.49B | ~3.8x 收入 | 公开企业 AI 应用估值底 | 公开公司重估,增长质量也不同 |
各行混合了干净收入倍数锚点和框架性可比公司;多个私有 AI 轮次缺少完全同步的估值和收入日期,因此本表应方向性使用,不宜机械套用。
[CV012, CV013, CV014, CV016, CV017, CV019]条形图对比 Reka 的隐含过去十二个月倍数与选定私营、公开 AI / 平台锚点。
数值是基于保留公开和私营来源搭出的四舍五入收入倍数启发式估算;它们体现方向,不是精确交易可比公司。
[CV005, CV013, CV021, CV024, CV025, CV026]8.5 关键驱动、打破 thesis 的触发项与最终尽调问题
最能证明本轮估值合理的驱动项都可衡量。第一,管理层需要证明 Snowflake 和直接企业渠道很快能把公司推到至少 $30 million-$40 million 的 forward run-rate,并且之后有可信路径超过 $50 million。第二,公司需要证明产品组合真的能创造软件式经济性,而不只是价格更高的 compute resale:gross margins 高于 65%、support cost 稳定,Vision 式工作负载中的 storage / indexing drag 有限。第三,投资人需要确信客户基础没有过度集中在少数 design partners、channel relationships 或一次性 lighthouse deployments。 下行因素同样清楚。如果当前轮次带有结构性保护,实质上压低新钱或 common 的位置;如果 gross margin 更接近 infrastructure 而不是 software;如果 Snowflake-sourced demand 很浅或经济成本很高;或如果 top-customer concentration 很高,那么当前价格的容错空间就太小。公开来源仍未披露经审计 ARR、net retention、burn、customer concentration、channel economics 或 2025 preference stack。这个缺失证据不是脚注;它决定了这是战略叙事,还是可承销投资。因此,正确下一步是一张带硬性 pass/fail 门槛的短 diligence 清单,而不是更多讲故事。今天最合理的上行退出仍是另一轮私募融资或战略出售,而不是近期 IPO,因为 Reka 公开可见的收入基础仍大幅落后于这里审阅的最小 public software anchors。[CV035, CV036, CV037, CV038, CV039, CV040]
| 触发条件 | 阈值 / 事件 | 如何传导到论点 | 行动含义 |
|---|---|---|---|
| 前瞻收入不达标 | 下一次融资检查点时 run-rate 仍低于 $25M | 即使用乐观前瞻倍数支撑,当前估值仍然偏高 | 放弃,或要求实质更好的入场条款 |
| 毛利率不达标 | 产品毛利率低于 50%-55% | 高效模型故事更像算力转售,而不是软件杠杆 | 下调估值区间,并建模 down-round |
| 优先权负担 | 2025 年融资轮有参与型优先股、ratchet,或重 seniority | 即使经营进展真实,普通股 / 新资金上行也会被封顶 | 任何正面判断前,都要求法律条款审查 |
| 渠道集中 | Snowflake 或单一伙伴贡献了 pipeline 或已签 ARR 的过大份额 | 战略价值从杠杆变成依赖 | 折价渠道溢价,并收紧情景权重 |
| 客户集中 / 留存薄弱 | 前五大账户主导 ARR,或扩张低于健康 SaaS 常态 | 估值被少数续约绑架 | 把增长视为脆弱,而不是可复利 |
这些阈值旨在把模糊的后期 AI 故事转成清晰的尽调通过 / 失败门槛。
[CV033, CV034, CV035, CV036, CV039, CV040]| 主题 | 缺失证据 | 为什么重要 | 负责人 / 尽调路径 |
|---|---|---|---|
| 当前收入和留存 | 月度 ARR / 收入桥接、队列留存,以及按产品拆分的扩张 | 没有这些,当前轮次只能建立在陈旧或估算的 top-line 数据上 | CFO 和收入运营材料包 |
| 毛利率和 burn | 按产品拆分的毛利率,加上算力、存储、支持和 burn 桥接 | 决定 AI 收入是否是有价值收入 | CFO 加基础设施负责人 |
| 优先权栈 | Series B 条款清单、清算优先权、反稀释和 side letter 摘要 | 干净的 headline valuation 仍可能带来薄弱投资人经济性 | 外部律师和董事会材料 |
| 渠道经济性 | Snowflake 来源 pipeline、已签 ARR、收入分成和折扣规则 | 检验战略渠道是增厚毛利率还是稀释毛利率 | CRO / 合作伙伴负责人 |
| 客户集中 | 前 10 大客户组合、最大部署、续约日历和用例集中度 | 集中度会把一条有希望的增长曲线变成悬崖风险 | CRO 加客户成功复核 |
这些是从公开框架走向可承保价格决策所需的最低私有数据要求。
[CV037, CV038, CV039, CV040, CV042]免责声明
本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;作出任何投资决定前,应直接向管理层核验,并查阅一手文件。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Reka was founded in 2022. | 高 | SO015, SO021 |
| CO002 | Reka is headquartered in Sunnyvale, California, USA. | 高 | SO004, SO020 |
| CO003 | Reka operates as an AI research and product company rather than as a pure research lab. | 高 | SO004, SO021 |
| CO004 | Reka’s core model stack is natively multimodal across text, image, video, and audio. | 高 | SO004, SO013 |
| CO005 | The company frames its mission as building models and infrastructure for the physical AI era. | 中 | SO001 |
| CO006 | Reka’s publicly available baseline chat models include reka-flash and reka-edge. | 中 | SO002 |
| CO007 | Reka Flash is the workhorse multimodal model behind Reka’s product offerings. | 中 | SO004 |
| CO008 | By July 2025 Reka described Reka Vision and Reka Research as generally available multimodal platforms. | 中 | SO004 |
| CO009 | Reka Vision is positioned as a platform for visual understanding, search, and reasoning across large video and image corpora. | 高 | SO003, SO024 |
| CO010 | Reka Research browses the web and private documents and offers enterprise deployment options including private cloud and on-premise. | 高 | SO006, SO025 |
| CO011 | Public profiles identify the founders as Dani Yogatama, Cyprien de Masson d’Autume, Qi Liu, Mikel Artetxe, and Yi Tay. | 高 | SO013, SO021 |
| CO012 | Dani Yogatama is the publicly identified CEO and co-founder of Reka. | 高 | SO019, SO021 |
| CO013 | Official and partner materials tie Reka’s founding team to DeepMind, Google Brain, and FAIR research pedigrees. | 高 | SO013, SO019 |
| CO014 | Snowflake announced an investment in Reka and a partnership in 2023. | 高 | SO011, SO021 |
| CO015 | Snowflake said its customers would be able to run and fine-tune Reka inside Snowflake accounts. | 中 | SO011 |
| CO016 | Snowflake later expanded Cortex support to Reka Flash and developed support for Reka Core. | 中 | SO012 |
| CO017 | Reka announced a $110M funding round on 2025-07-22 backed by NVIDIA and Snowflake. | 高 | SO004, SO014 |
| CO018 | Reka said the 2025 funding would accelerate technical development and scale its multimodal platforms for wider enterprise adoption. | 中 | SO004 |
| CO019 | Reuters-syndicated and tracker sources place Reka’s July 2025 valuation above $1B and about triple its 2023 level. | 高 | SO015, SO021 |
| CO020 | Private-market trackers place Reka’s total disclosed funding at roughly $168M-$170M across two rounds. | 中 | SO020, SO021 |
| CO021 | Reuters-syndicated reporting said Reka expanded from 20 to 50 employees over the year before the July 2025 round. | 中 | SO015 |
| CO022 | Private-company trackers subsequently placed Reka at roughly 60-64 employees by late 2025 / May 2026. | 中 | SO020, SO021 |
| CO023 | Shutterstock became both a data-licensing partner and a paying customer in June 2024. | 高 | SO004, SO019 |
| CO024 | Turing launched Guardian AI on top of Reka Vision across a footprint of 13,000+ sites and 10M+ daily events. | 高 | SO004, SO007 |
| CO025 | Turing and Reka said Guardian AI was already being used by law-enforcement officers in the United States. | 中 | SO007 |
| CO026 | Reka Vision is designed as an intelligence layer that can integrate with existing VMS deployments and run in cloud, VPC, on-premise, or air-gapped environments. | 高 | SO009, SO010 |
| CO027 | Reka markets up to 95% fewer false alarms and 65% faster case resolution from Reka Vision deployments. | 中 | SO009 |
| CO028 | Reka has a specific defense and security offering for qualified sovereign, on-premise, and air-gapped programs. | 中 | SO010 |
| CO029 | Reka’s April 2024 technical report says Core, Flash, and Edge were trained from scratch and that Flash and Edge deliver state-of-the-art results for their compute class. | 中 | SO013 |
| CO030 | The same report says Reka Core performed competitively to GPT-4V on image QA and outperformed Gemini Ultra on the Perception-Test video benchmark. | 中 | SO013 |
| CO031 | Reka Flash 3.1 is a 21B-parameter model that improved 10 points on LiveCodeBench v5 from the prior Flash version. | 中 | SO005 |
| CO032 | Reka Quant is described as near-lossless 3.5-bit compression of Flash 3.1 with only 1.6 average performance degradation. | 中 | SO005 |
| CO033 | Snowflake reportedly held acquisition talks to buy Reka for more than $1B before the process ended without a transaction. | 中 | SO016, SO018 |
| CO034 | Dani Yogatama said Snowflake and Reka decided it made sense to move independently while continuing collaboration. | 中 | SO015 |
| CO035 | Reka’s business model combines API usage, enterprise deployments, and partner-embedded multimodal applications rather than a mass-market consumer chatbot. | 中 | SO011, SO022, SO023, SO025 |
| CO036 | Reka publishes usage-based API pricing for chat and research plus enterprise and developer tiers for Vision. | 中 | SO022, SO023 |
| CO037 | Vision pricing includes a free evaluation tier with 180 minutes of indexed video and an enterprise option with no rate limits and monthly invoicing. | 中 | SO023 |
| CO038 | Reka Research is priced from $25 per 1,000 requests and Reka Flash chat pricing lists $0.80 input and $2.00 output per 1M tokens. | 中 | SO022 |
| CM001 | Reka’s core market is enterprise multimodal AI workflows rather than the entire generative-AI market. | 中 | SM001, SM005, SM011 |
| CM002 | That market includes model consumption, multimodal applications, and governed deployment work. | 中 | SM005, SM011, SM012 |
| CM003 | Consumer chatbot subscriptions are adjacent to Reka’s market but are not the core job-to-be-done evidenced by Reka’s public products. | 中 | SM011, SM012, SM016 |
| CM004 | Status-quo substitutes include manual review, keyword metadata systems, narrow computer-vision tools, and internal buildouts on hyperscaler platforms. | 中 | SM010, SM011, SM014, SM019 |
| CM005 | The Business Research Company categorizes foundation AI models across language, vision, multimodal, speech, and code segments. | 中 | SM005 |
| CM006 | ResearchAndMarkets maps the multimodal AI market by type, offering, data modality, technology, and vertical through 2035. | 中 | SM004 |
| CM007 | Richer deployment work matters because Reka markets private cloud, on-premise, VPC, and air-gapped options alongside its models. | 中 | SM011, SM012, SM016 |
| CM008 | Snowflake’s multimodal positioning reinforces that data-adjacent deployment is part of the commercial market, not just a technical feature. | 高 | SM010, SM015 |
| CM009 | A broad AI TAM would overstate precision for Reka because public sources mostly size the full stack rather than specialist multimodal enterprise niches. | 中 | SM002, SM004, SM005 |
| CM010 | Gartner forecasts total worldwide AI spending of $2.595667T in 2026. | 中 | SM002 |
| CM011 | Gartner forecasts AI-model spending of $32.604B in 2026. | 中 | SM002 |
| CM012 | Gartner forecasts AI-software spending of $453.209B in 2026. | 中 | SM002 |
| CM013 | Gartner says 80% of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024. | 中 | SM001 |
| CM014 | IDC describes 2026 as a period when AI scales from pilots to enterprise transformation. | 中 | SM003 |
| CM015 | Public sources support several top-down lenses for Reka’s market, but not a precise third-party SAM. | 中 | SM002, SM004, SM005 |
| CM016 | A conservative low-case SAM for Reka-like multimodal workflows is about $1.6B if they capture only 5% of Gartner’s 2026 AI-model-spend pool. | 低 | SM002, SM011, SM012 |
| CM017 | A base-case SAM heuristic for Reka-like workflows is about $3.3B if they map to 10% of 2026 AI-model spend. | 低 | SM002, SM011, SM012 |
| CM018 | A high-case SAM heuristic for Reka-like workflows is about $4.9B if they map to 15% of 2026 AI-model spend. | 低 | SM002, SM011, SM012 |
| CM019 | In media use cases, the buyer is likely a content-platform or archive owner seeking better metadata, search, or clip generation. | 中 | SM011, SM013 |
| CM020 | In physical-security use cases, the buyer is likely a security-operations owner and the user is an investigator or operator. | 中 | SM014, SM016 |
| CM021 | In enterprise-research use cases, the buyer is likely a CTO, CDO, or AI-platform owner enabling analysts and knowledge workers. | 中 | SM012, SM017 |
| CM022 | Shutterstock is the clearest public proof point that media-archive buyers can become both customers and data partners. | 中 | SM013 |
| CM023 | Turing is the clearest public proof point that security buyers will pay for multimodal event search and alerting workflows. | 中 | SM014 |
| CM024 | Adoption typically starts when manual review or weak metadata creates an acute operational bottleneck. | 中 | SM011, SM014, SM016, SM017 |
| CM025 | Enterprises are likely to pilot a narrow dataset or footprint before committing to broad rollout. | 中 | SM011, SM014 |
| CM026 | Governance review is an early-stage adoption gate because deployment mode and data movement are core purchasing variables. | 中 | SM010, SM015, SM016 |
| CM027 | Structural growth drivers include rising multimodal penetration in enterprise software and AI shifting from pilots to enterprise transformation. | 高 | SM001, SM003 |
| CM028 | ARK argues enterprise token demand has risen 28x since December 2024, indicating sharp growth in model consumption workloads. | 中 | SM007 |
| CM029 | NVIDIA says Blackwell-based inference providers can reduce cost per token by up to 10x versus Hopper in some open-model deployments. | 中 | SM009 |
| CM030 | Snowflake argues Blackwell-class compute adjacent to enterprise data reduces security risk and operational overhead from data movement. | 中 | SM010 |
| CM031 | Control Risks warns that compute access in 2026 is constrained by export controls, infrastructure bottlenecks, and geopolitical permission. | 中 | SM006 |
| CM032 | For Reka-like vendors, governance and deployment requirements are commercial constraints as much as technical ones. | 中 | SM010, SM011, SM015, SM016 |
| CM033 | Open-source and open-weight models improve the market opportunity for buyers while simultaneously increasing substitution pressure on specialist vendors. | 中 | SM008, SM009, SM024, SM025 |
| CM034 | Hyperscaler and platform bundling can compress stand-alone model budgets by giving buyers multimodal tooling inside existing data or cloud contracts. | 中 | SM010, SM015, SM019, SM020, SM021 |
| CM035 | Reka’s market outlook is favorable near term, but long-term margin durability depends on whether product packaging and deployment flexibility outrun model commoditization. | 中 | SM008, SM009, SM011, SM015 |
| CP001 | Reka competes against frontier incumbents, enterprise-focused challengers, and open-weight substitutes rather than a single peer group. | 中 | SP005, SP022 |
| CP002 | OpenAI, Anthropic, and Google are the most relevant frontier-platform incumbents in Reka’s competitive set. | 中 | SP005, SP006, SP009, SP011 |
| CP003 | Cohere, Mistral, and Aleph Alpha compete more directly on enterprise privacy, deployment control, or sovereign positioning. | 中 | SP013, SP015, SP017 |
| CP004 | Llama and Gemma are credible open-weight substitutes for buyers willing to build or customize internally. | 中 | SP018, SP019 |
| CP005 | Tracxn lists hundreds of active competitors around Reka, underscoring a crowded market rather than a winner-take-all field. | 中 | SP022 |
| CP006 | Reka is more specialized around multimodal video and research workflows than many generalist API vendors. | 中 | SP002, SP003, SP004, SP006 |
| CP007 | Snowflake distribution gives Reka a route into enterprise data-platform accounts that offsets some scale disadvantage. | 高 | SP020, SP021, SP025 |
| CP008 | Despite that channel help, buyers can still compare multiple model vendors in parallel because the category is structurally crowded. | 中 | SP005, SP022 |
| CP009 | OpenAI markets GPT-4o as a model that accepts text, audio, image, and video input. | 中 | SP007 |
| CP010 | OpenAI’s API platform also offers multimodal, realtime, and web-search tooling around its model menu. | 中 | SP006 |
| CP011 | Anthropic positions Claude 3.5 Sonnet as a fast reasoning model with strong vision and a 200K context window. | 高 | SP009, SP010 |
| CP012 | Google positions Gemini around advanced multimodal understanding, long-horizon tasks, and multi-step problem-solving. | 中 | SP011 |
| CP013 | Mistral positions Studio around building, deploying, and governing agentic AI with hybrid and self-hosted options. | 中 | SP013 |
| CP014 | Cohere positions Command around secure enterprise AI and private deployment rather than broad consumer reach. | 高 | SP015, SP016 |
| CP015 | Aleph Alpha positions itself around sovereign European domain-specific models for regulated environments. | 中 | SP017 |
| CP016 | Reka’s public product stack emphasizes video reasoning, multimodal research, and enterprise deployment flexibility. | 中 | SP002, SP003, SP004, SP024 |
| CP017 | Reka Flash public list pricing is $0.80 per million input tokens and $2.00 per million output tokens. | 中 | SP001 |
| CP018 | Reka Research public list pricing starts at $25 per 1,000 requests. | 中 | SP001 |
| CP019 | Claude 3.5 Sonnet public pricing is $3 per million input tokens and $15 per million output tokens. | 中 | SP009 |
| CP020 | OpenAI’s current API menu publishes multimodal and reasoning model pricing but does not make workflow-specific switching costs inherently high. | 中 | SP006 |
| CP021 | Gemini uses a free, paid, and enterprise ladder, signaling that Google competes through ecosystem entry points as much as raw token price. | 中 | SP012 |
| CP022 | Cohere’s pricing page emphasizes contract-led enterprise packaging and managed instances rather than transparent consumer-style token menus. | 中 | SP016 |
| CP023 | Reka’s public pricing appears cheaper than premium frontier vendors on raw list-token terms, but not necessarily on outcome-adjusted workflow cost. | 中 | SP001, SP006, SP009, SP012 |
| CP024 | Pure model APIs are relatively easy to multi-home compared with traditional enterprise software. | 中 | SP005, SP006, SP012, SP014 |
| CP025 | Switching costs rise when a vendor owns indexing, alerting, deployment, or workflow logic above the base model. | 中 | SP003, SP004, SP020, SP025 |
| CP026 | Snowflake distribution can increase stickiness because governed data access and operational convenience matter alongside model quality. | 高 | SP020, SP021, SP025 |
| CP027 | Open-weight models lower switching barriers by giving sophisticated buyers a credible internal-build fallback. | 中 | SP018, SP019 |
| CP028 | Hyperscaler and platform bundles can compress stand-alone model budgets because buyers may already be paying for adjacent cloud or data-platform services. | 中 | SP006, SP011, SP020, SP021, SP026 |
| CP029 | Reka’s moat is composite: efficiency, workflow packaging, deployment flexibility, and channel access. | 中 | SP002, SP003, SP004, SP021 |
| CP030 | Efficiency alone is not a durable moat because industry-wide inference costs are falling for many providers. | 中 | SP002, SP005, SP018, SP019 |
| CP031 | Video-centric and governed-deployment workflows are the strongest areas where Reka can still differentiate from generic APIs. | 中 | SP003, SP004, SP020, SP021 |
| CP032 | Convergence risk is high because frontier incumbents keep adding stronger vision, search, and enterprise features. | 中 | SP006, SP009, SP011, SP013 |
| CP033 | Open-source improvement cuts both ways for Reka: it lowers infrastructure cost but also strengthens substitute options. | 中 | SP018, SP019 |
| CP034 | Aleph Alpha shows that sovereign and regulated buyers can choose purpose-built regional specialists instead of broader global APIs. | 中 | SP017 |
| CP035 | Reka’s near-term niche looks defendable, but long-term moat durability depends on embedding workflows faster than the base-model layer commoditizes. | 中 | SP003, SP004, SP020, SP021 |
| CP036 | OpenAI also packages enterprise deployment guidance, support, and change-management services around its models, strengthening incumbent distribution power. | 中 | SP026 |
| CI001 | Reka sells API access on a pay-as-you-go basis with no upfront commitment disclosed on the public price page. | 中 | SI001 |
| CI002 | Reka Research is publicly priced at $25 standard, $35 low-parallel-thinking, and $60 high-parallel-thinking per 1,000 requests. | 中 | SI001 |
| CI003 | Reka Flash is publicly listed at $0.80 per 1M input tokens and $2.00 per 1M output tokens, with separate image, video, and audio metering. | 中 | SI001 |
| CI004 | Reka Edge is the low-cost tier at $0.10 per 1M input tokens and $0.005 output tokens, indicating deliberate price segmentation below Flash. | 中 | SI001 |
| CI005 | Reka Vision separates a developer self-service tier from an enterprise tier that uses monthly invoicing, bulk discounts, flexible storage, and dedicated support. | 中 | SI002 |
| CI006 | Vision developer pricing monetizes multiple activities separately, including video indexing, search, image upload and storage, output tokens, and clip generation. | 中 | SI002 |
| CI007 | Vision enterprise storage can carry recurring costs while self-service storage auto-deletes after 30 days at no storage fee. | 中 | SI002 |
| CI008 | Reka markets higher-order products above the base model API, including Reka Research and the Vision API. | 中 | SI003, SI004, SI025 |
| CI009 | Reka Research browses the web and private documents, while Vision provides video and image management, semantic search, QA, clip generation, and tagging. | 中 | SI003, SI004 |
| CI010 | Reka's 2025 funding announcement says Flash is the workhorse of the offering and that Vision and Research had recently gone into general availability. | 中 | SI005, SI006 |
| CI011 | Snowflake said in 2023 that its customers would be able to bring Reka to their data and run and fine-tune it within their Snowflake accounts. | 中 | SI007 |
| CI012 | Snowflake later expanded the partnership so Flash was supported in Cortex and support for Core was being developed, extending Reka's enterprise distribution inside Snowflake. | 高 | SI008, SI009 |
| CI013 | Snowflake's product and documentation pages position multimodal models inside a secure enterprise perimeter, which is consistent with Reka targeting governed workloads rather than consumer traffic. | 中 | SI009, SI011 |
| CI014 | Snowflake's quickstart and lifecycle materials show that customers can build multimodal analysis, GPU training, and real-time or batch inference workflows inside Snowflake, making it a plausible indirect GTM channel for Reka. | 中 | SI012, SI013 |
| CI015 | Shutterstock is both a training-data partner and a paying customer that retains Reka to enhance image and video metadata. | 中 | SI019 |
| CI016 | Official 2025 materials name Shutterstock and Turing Video as Vision users but do not disclose contract value, customer count, or deployment volume for Reka itself. | 中 | SI005, SI006 |
| CI017 | GetLatka estimates that Reka generated $10.9M of revenue in 2025. | 低 | SI020 |
| CI018 | Investing reported that Reka expanded from 20 to 50 employees over the prior year, while GetLatka and Tracxn place employee count around 60 to 64 in late 2025 to May 2026. | 中 | SI014, SI020, SI021 |
| CI019 | Using the $10.9M revenue estimate and 60-64 employee range implies roughly $170k-$182k of annual revenue per employee. | 中 | SI020, SI021 |
| CI020 | Using the same $10.9M revenue estimate against $168M-$170M of cumulative funding implies about 15.4x-15.6x capital raised to annual revenue. | 中 | SI020, SI021 |
| CI021 | Public sources do not disclose ARR, gross margin, NRR, CAC, payback, or customer concentration for Reka. | 中 | SI020, SI021 |
| CI022 | Multiple sources corroborate that Reka raised $110M in July 2025 from investors including NVIDIA and Snowflake. | 高 | SI005, SI006, SI014 |
| CI023 | Multiple sources corroborate that the 2025 round valued Reka at more than $1B, versus about $300M in the 2023 round. | 高 | SI014, SI020, SI021 |
| CI024 | GetLatka and Tracxn disagree slightly on cumulative funding, with $170M versus $168M respectively. | 中 | SI020, SI021 |
| CI025 | The public financing history still appears to be only two disclosed institutional rounds: a 2023 Series A and a 2025 Series B. | 中 | SI020, SI021 |
| CI026 | CRN and MarketScreener both reported that Snowflake explored buying Reka for more than $1B in 2024. | 高 | SI017, SI018 |
| CI027 | Those acquisition talks did not close, and public reporting describes the companies as continuing to collaborate independently afterward. | 中 | SI014, SI017, SI018 |
| CI028 | Reka says the 2025 proceeds will accelerate technical development and scale its multimodal platforms for broader enterprise adoption. | 中 | SI005, SI006 |
| CI029 | NVIDIA says Blackwell-based inference providers are reducing cost per token by up to 10x versus Hopper and showed 2.5x better throughput per dollar in one case. | 中 | SI022 |
| CI030 | Snowflake says Blackwell integration can deliver up to 5x higher inference throughput and lower total cost of ownership through faster enterprise AI workflows. | 中 | SI013 |
| CI031 | ARK says AI training costs are falling about 75% per year and inference costs about 95% per year for frontier-capable models. | 中 | SI024 |
| CI032 | Control Risks argues that access to AI compute is constrained not only by money but also by power, water, regulation, and geopolitics. | 中 | SI023 |
| CI033 | Because Reka sells video, image, and research workflows, likely cost buckets extend beyond token inference to storage, indexing, data handling, and enterprise support. | 中 | SI002, SI009, SI019 |
| CI034 | The Vision enterprise tier's monthly invoicing, bulk discounts, recurring storage, and dedicated support imply that realized gross margin will depend heavily on workload mix and support intensity. | 中 | SI002 |
| CI035 | Public evidence supports multiple monetization surfaces—base API, Research requests, Vision software, and Snowflake-mediated distribution—but not the revenue mix across them. | 中 | SI001, SI002, SI003, SI007, SI008 |
| CI036 | Revenue quality is harder to underwrite than revenue existence because the best public top-line number is a third-party estimate rather than a company disclosure. | 低 | SI020 |
| CI037 | There is no responsible public basis for estimating cash, burn, or runway because the company discloses financing events but not balance-sheet detail. | 中 | SI005, SI020, SI021 |
| CI038 | The financial bull case is efficient-model monetization with strong strategic partners; the bear case is that the capital base is growing faster than public unit-economics disclosure. | 中 | SI014, SI020, SI021, SI023 |
| CI039 | If the $1B valuation and $10.9M revenue estimate are both directionally right, the implied valuation-to-revenue multiple is roughly 92x. | 中 | SI014, SI020 |
| CI040 | Snowflake's documentation frames model selection around performance per credit and in-perimeter deployment, implying that some indirect Reka economics may be mediated by Snowflake's platform model rather than Reka's native API list price. | 中 | SI010, SI011 |
| CI041 | Snowflake files annual reports with the SEC, which highlights the disclosure asymmetry between Reka's private financial reporting and its largest public-channel partner. | 低 | SI026 |
| CI042 | The World Economic Forum says AI data-centre investment is outpacing grid build-out, making power-grid connectivity a binding constraint for AI infrastructure scaling. | 中 | SI027 |
| CE001 | Reka’s public chat API is OpenAI-compatible and documented around the same client pattern, with requests sent to https://api.reka.ai/v1. | 中 | SE001, SE002 |
| CE002 | The public baseline models always available for self-serve access are reka-flash and reka-edge (including reka-edge-2603). | 中 | SE001, SE005 |
| CE003 | The Chat API supports image_url, video_url, audio_url, and pdf_url content types inside a single conversation surface. | 中 | SE002, SE003 |
| CE004 | Reka documents the Chat API as best for short videos, while longer videos should move into the Vision API upload-and-index workflow. | 中 | SE003, SE008 |
| CE005 | Vision’s video-management surface supports upload by file or URL, optional grouping, indexing, thumbnail generation, and absolute start timestamps. | 中 | SE006 |
| CE006 | Vision video search is built around natural-language queries over indexed videos, with thresholds, group filters, timestamps, explanations, and optional generated reports. | 中 | SE007 |
| CE007 | Vision video Q&A uses an indexed-video chat endpoint for longer footage, whereas short videos can stay in the base Chat API. | 中 | SE003, SE008 |
| CE008 | Reka’s public visual workflow surface includes clip generation, metadata tagging, and image search in addition to long-video search and Q&A. | 中 | SE009, SE010, SE011 |
| CE009 | Vision metadata tagging exposes policy-relevant fields such as violence, profanity, adult content, drugs, alcohol, gambling, political, plus descriptive and performance-oriented fields. | 中 | SE010 |
| CE010 | The Vision MCP server exposes upload, listing, indexing, search, Q&A, transcript/caption inspection, and object-detection capabilities inside agent clients such as Claude Code, Codex, and Cursor. | 高 | SE012, SE027 |
| CE011 | Reka’s quickstart documents both local Edge execution on Apple Silicon Macs and higher-throughput Linux CUDA serving with vLLM, including a cited 40-50 tokens-per-second test on 3090 GPUs. | 中 | SE001 |
| CE012 | Reka’s public developer ecosystem includes an active GitHub organization with vllm-reka, reka-mcp, SDKs, clip examples, and other integration assets updated through mid-2026. | 中 | SE027 |
| CE013 | The official n8n node already supports clipping from video URLs, image and short-video prompting, long-video Q&A, and object detection, while Research, Speech, and Text are still marked as “soon.” | 中 | SE028 |
| CE014 | Snowflake positions Reka models inside Cortex for governed multimodal analysis of images and video close to enterprise data. | 中 | SE033 |
| CE015 | Reka says its Vision Agent integrates with NVIDIA’s Video Search and Summarization blueprint so customers can add natural-language search, Q&A, and event detection without replacing existing video systems. | 高 | SE031, SE032 |
| CE016 | Oracle’s defense-ecosystem cohort includes Reka, signaling go-to-market relevance in secure and mission-readiness environments rather than consumer-first distribution. | 中 | SE030 |
| CE017 | Reka Research combines web browsing, private-document search, and document analysis tools, and Parallel Thinking runs multiple candidate generations before a resolver model selects the final answer. | 中 | SE019, SE021 |
| CE018 | Parallel Thinking is priced as low and high modes at $35 and $60 per 1,000 requests and is reported to improve Research-Eval high-mode accuracy from 59.1 to 63.3 and SimpleQA to 94.8. | 中 | SE019 |
| CE019 | Research-Eval is a 374-question benchmark designed specifically for search-augmented LLMs, with reported frontier-model scores between 26.7% and 59.1%. | 中 | SE020 |
| CE020 | Reka Speech is an 850M multilingual transcription and translation model, built for high-volume offline jobs with timestamps and reported as 8x-35x faster than Whisper-based alternatives on the cited H100 test workload. | 中 | SE018 |
| CE021 | Reka’s 2024 technical report says Core, Flash, and Edge were trained from scratch for text, image, video, and audio, with Core competitive with GPT-4V-class systems and Edge/Flash outperforming larger models in their compute classes. | 中 | SE025 |
| CE022 | Reka Edge is described as a roughly 7B-class model built from a ConvNeXt V2 vision encoder of about 657M-660M parameters plus a 6B-plus language backbone. | 高 | SE015, SE017 |
| CE023 | Reka Edge is designed to emit only 64 tokens per image tile so that high-definition visual inputs consume less context and memory. | 高 | SE015, SE016, SE017 |
| CE024 | Reka markets Edge as the fastest vision-language model in the 7B-8B class and says it is about 2.4x faster or lower-latency than peers on average across requests. | 中 | SE015, SE017 |
| CE025 | The Edge launch post claims about 3x fewer input tokens than comparable models, 5.46 images per second, 0.522 seconds TTFT, and up to 2.3x higher throughput after quantization with over 98% retained performance. | 中 | SE017 |
| CE026 | Reka Quant is released as an open-source quantization stack, with the post claiming near-lossless 3.5-bit Flash quantization and far lower average benchmark degradation than native llama.cpp baseline methods. | 中 | SE023 |
| CE027 | Flash 3.1 is presented as a 21B reasoning model improved through reinforcement learning, released in a Llama-compatible format, and reported as about 10 points better than Flash 3 on LiveCodeBench v5. | 高 | SE022, SE029 |
| CE028 | Function calling is currently documented only for Reka Flash, so advanced tool-use exposure is not yet uniform across the public model family. | 中 | SE004 |
| CE029 | Reka publicly documents structured JSON API errors, X-Request-ID correlation, explicit client actions for 400/401/404/429/500 cases, and retry/backoff guidance. | 中 | SE013 |
| CE030 | Vision self-serve rate limits are explicit but modest, including 50 uploads or searches per day and 10 clip jobs per day per API key, with enterprise plans positioned for higher quotas. | 中 | SE014 |
| CE031 | Reka’s privacy policy says paid API content is not used for model training unless customers opt in, while free or promotional usage may be used to improve models. | 中 | SE024 |
| CE032 | The privacy policy says uploaded files may be staged in secure Amazon S3 with expiring links and automatic deletion within a set period such as 24 hours. | 中 | SE024 |
| CE033 | OpenAI’s enterprise surface publicly emphasizes rollout guidance, analytics, 24/7 support with SLAs, and AI advisors, indicating a more mature public enterprise-support posture than Reka currently shows. | 中 | SE034 |
| CE034 | Google’s Gemini page publicly emphasizes advanced multimodal understanding, long-horizon tasks, and strong benchmark visibility, underscoring how large incumbents still lead on generalized breadth and public proof points. | 中 | SE035 |
| CE035 | Artificial Analysis tracks Reka Flash across cost, speed, latency, and context dimensions, but the fetched independent provider page still shows parts of evaluation coverage as forthcoming or unavailable. | 中 | SE026 |
| CE036 | Independent review coverage consistently frames Reka as enterprise- and developer-centric, strong for physical AI and multimodal media work, but demanding more integration and revalidation than turnkey consumer tools. | 低 | SE036 |
| CE037 | Public evidence shows Reka’s current buyer-facing stack is built around Chat, Edge/local deployment, Vision, Research, and Speech rather than a broad consumer-assistant suite. | 中 | SE001, SE015, SE018, SE019 |
| CE038 | Reka’s strongest technical wedge is deployable multimodal efficiency—local Edge, video-specific Vision workflows, and governed enterprise integrations—rather than generalized frontier-scale breadth. | 中 | SE015, SE017, SE033 |
| CE039 | The fetched public material documents privacy defaults, tagging controls, errors, and rate limits, but it does not surface public SOC 2, ISO, or status-center evidence comparable to what cautious regulated buyers often request. | 低 | SE013, SE014, SE024 |
| CU001 | Reka's public positioning emphasizes enterprise use in security, media, and defense rather than consumer distribution. | 中 | SU001, SU002 |
| CU002 | Reka Vision is presented as a product for enterprises, creators, and developers who need multimodal search, Q&A, and clip generation over visual content. | 中 | SU002 |
| CU003 | The public customer lanes visible in fetched material are direct API developers, enterprise security operators, media-data buyers, and channel partners that embed Reka in their own products. | 中 | SU001, SU002, SU004, SU015, SU020, SU023 |
| CU004 | A developer can create a free account, obtain an API key, and start using Reka through an OpenAI-compatible API with pay-as-you-go pricing. | 中 | SU004, SU005 |
| CU005 | Enterprise-scale usage requires a higher-touch motion because public docs route customers to contact Reka for higher limits, volume discounts, and some deployment options. | 中 | SU005, SU006, SU008 |
| CU006 | Reka publicly supports cloud, on-premise, VPC, and air-gapped deployment patterns for security-sensitive workloads. | 高 | SU003, SU029 |
| CU007 | Reka Research support for internal data sources is currently described as enterprise-only rather than generally available to all developers. | 中 | SU006 |
| CU008 | Public Vision rate limits are modest at 100 image uploads per day, 50 video uploads per day, and 10 clip requests per day per API key. | 中 | SU008 |
| CU009 | Developer-distribution channels extend beyond the core docs through MCP and n8n assets that let builders use Reka inside agents and automations. | 中 | SU009, SU023, SU024, SU025 |
| CU010 | Shutterstock is explicitly described by both Reka and Shutterstock as a customer that uses Reka to enhance metadata for its image and video library. | 高 | SU010, SU011, SU012 |
| CU011 | The Shutterstock proof is about metadata enrichment and search/discoverability over a content library, not about a generic chatbot deployment. | 中 | SU011, SU012, SU014 |
| CU012 | Shutterstock's case-study narrative says Reka would enhance metadata across 550 million image and video assets and that more than 60 million new assets are added annually. | 中 | SU012 |
| CU013 | The Shutterstock relationship proves paid workflow utility for a named media platform but does not disclose contract value, rollout breadth, or renewal history. | 中 | SU011, SU012, SU013 |
| CU014 | Snowflake publicly says its customers can bring Reka's multimodal assistant to their data within their own Snowflake account. | 高 | SU015, SU016 |
| CU015 | Snowflake later announced that Cortex supports Reka Flash and is developing support for Reka Core for multimodal analysis. | 高 | SU016, SU017 |
| CU016 | The Snowflake proof is strongest as a governed distribution channel because customers consume Reka capabilities inside the Snowflake Data Cloud rather than necessarily through a direct Reka application contract. | 中 | SU015, SU016, SU019 |
| CU017 | Snowflake's multimodal docs and quickstart show buyers can build image and audio analysis workflows inside the Snowflake environment instead of adopting a separate front-end product from Reka. | 中 | SU017, SU018 |
| CU018 | Public Snowflake materials do not disclose how many end accounts actively use Reka-powered features or what their spend looks like. | 中 | SU016, SU017, SU019 |
| CU019 | Reka and Turing say Guardian AI was built on top of Turing's platform and Reka Vision for the physical-security domain. | 中 | SU020, SU022 |
| CU020 | Reka states that Turing has over 13,000 site deployments and processes more than 10 million daily events, implying a potentially meaningful installed-base channel. | 中 | SU020 |
| CU021 | Guardian AI is described as enabling natural-language search, smarter alerts, and incident-report generation for Turing customers. | 中 | SU020, SU022 |
| CU022 | Fetched Reka materials say Guardian AI is already being used by law-enforcement officers in the United States and cite an Ohio police deployment example. | 中 | SU020, SU022 |
| CU023 | Reka's VMS-integration article claims the Orange Village / Ohio police deployment achieved 65 percent faster case resolution, 42 percent operational cost savings, and 89 percent officer satisfaction. | 低 | SU029 |
| CU024 | Reka's buying motion starts with self-serve evaluation but pushes larger customers toward sales-assisted limits, deployment scoping, and negotiated support. | 中 | SU004, SU005, SU006, SU008 |
| CU025 | Reka recommends layering Vision onto an existing video-management stack rather than replacing the VMS, which lowers rip-and-replace friction for security buyers. | 中 | SU029 |
| CU026 | Reka explicitly recommends starting security deployments with a high-value subset of cameras before broader rollout. | 中 | SU029 |
| CU027 | The defense-security motion is consultative, with qualification, environment assessment, pilot deployment inside the perimeter, and operational handover. | 中 | SU003 |
| CU028 | Public channels beyond direct sales include Snowflake for data-cloud buyers, Turing for surveillance buyers, Oracle ecosystem access for defense, and builder channels such as GitHub and n8n. | 中 | SU015, SU020, SU023, SU024, SU025, SU026 |
| CU029 | n8n and GitHub assets show that Reka is courting builders who want to automate video clipping, image/video Q&A, and agent workflows without waiting for bespoke enterprise integrations. | 中 | SU023, SU024, SU025 |
| CU030 | The MCP server lets customers or developers connect their own Reka API keys and search, index, and analyze videos from agent clients. | 中 | SU009, SU024 |
| CU031 | Outside a few named examples, public customer proof remains shallow because official surfaces mostly emphasize sectors, workflows, or partner narratives rather than a broad verified customer roster. | 中 | SU001, SU002, SU015, SU020, SU022 |
| CU032 | The fetched public record does not disclose total customer count, NRR, GRR, contract duration, or revenue concentration for Reka's customer base. | 中 | SU001, SU005, SU006, SU015, SU016, SU020, SU022, SU028 |
| CU033 | Reka's public materials and community assets are primarily legible to technical teams rather than nontechnical end users. | 中 | SU004, SU023, SU024, SU028 |
| CU034 | A competitor review argues that using Reka as a raw model API can create unpredictable budgeting and integration overhead for support-oriented teams. | 低 | SU027 |
| CU035 | An independent review frames Reka as strongest for organizations that need custom multimodal deployments at scale rather than casual plug-and-play use. | 低 | SU028 |
| CU036 | Prepaid credits, rate ceilings, and integration work are visible adoption barriers for smaller or less technical customers. | 中 | SU005, SU006, SU008, SU027 |
| CU037 | The clearest expansion logic is to land with one workflow and then expand into more media volume, more governed environments, or more departmental use cases once ROI is proven. | 中 | SU022, SU025, SU026, SU029 |
| CU038 | The strongest named proof points represent three different motions: Shutterstock as a direct customer, Snowflake as a platform channel, and Turing as a vertical solution partner. | 中 | SU010, SU015, SU020 |
| CU039 | Public evidence is insufficient to judge whether revenue is concentrated in a handful of strategic accounts or channel partners. | 中 | SU015, SU016, SU020, SU026, SU028 |
| CU040 | Because Reka sells into security, defense, and governed enterprise workflows, procurement is likely longer and higher-touch than for commodity developer APIs. | 中 | SU003, SU006, SU027 |
| CU041 | Oracle ecosystem inclusion is a go-to-market signal for defense-sector access, but it does not prove end-customer production usage of Reka. | 中 | SU003, SU026 |
| CU042 | Channel evidence can accelerate reach, but it also makes it harder to tell how much customer ownership and recurring spend sits directly with Reka rather than with partners. | 中 | SU015, SU016, SU020, SU026 |
| CR001 | Reka’s terms say free-tier use may be used to train, develop, and improve its machine learning models and related technologies. | 中 | SR001 |
| CR002 | Reka’s terms say paid API requests are not used for model training unless the customer has explicitly opted in. | 中 | SR001 |
| CR003 | Reka’s business terms define Aggregated Data as customer-usage information used in an aggregate and anonymized manner for operating and improving the service. | 中 | SR002 |
| CR004 | The European Commission says GPAI obligations under the AI Act entered into application on 2 August 2025. | 中 | SR003 |
| CR005 | Providers of general-purpose AI models must draw up technical documentation, implement a copyright policy, and publish a summary of training content under the EU AI Act regime described in the reviewed sources. | 高 | SR003, SR004 |
| CR006 | Providers of GPAI models with systemic risk face extra duties that include notifying the Commission, assessing and mitigating risk, reporting incidents, and implementing cybersecurity protections. | 高 | SR003, SR004 |
| CR007 | The U.S. Copyright Office has published Parts 1 and 2 of its AI report and released Part 3 on generative AI training in pre-publication form with no substantive analytical changes expected in the final version. | 中 | SR005 |
| CR008 | The Copyright Office’s Part 2 analysis says generative AI outputs are copyrightable only where a human author determines sufficient expressive elements and not through the mere provision of prompts. | 中 | SR006 |
| CR009 | The EDPB says AI-model governance must address whether a model is anonymous, whether legitimate interest is a lawful basis, and what happens if training data were processed unlawfully. | 中 | SR009 |
| CR010 | The ICO’s AI guidance directs organizations to AI and data protection guidance, a risk toolkit, and biometric recognition guidance for higher-risk uses. | 中 | SR007 |
| CR011 | California’s CCPA/CPRA framework gives consumers rights to know, delete, opt out, correct, and limit use of sensitive personal information. | 中 | SR008 |
| CR012 | NIST’s AI Risk Management Framework is a voluntary framework meant to incorporate trustworthiness into the design, development, use, and evaluation of AI systems. | 中 | SR010 |
| CR013 | The reviewed IEA source says electricity demand from data centres surged 17% in 2025 while global electricity demand grew 3%. | 中 | SR013 |
| CR014 | The reviewed IEA source says electricity consumption from data centres is set to double by 2030 and AI-focused data-centre power use is poised to triple. | 中 | SR013 |
| CR015 | The reviewed IEA source says AI deployment is increasingly hitting physical bottlenecks that include gas turbines, transformers, advanced chips, IT components, and grid connection capacity. | 中 | SR013 |
| CR016 | RAND estimates global AI data centres could require 68 gigawatts of power by 2027, close to California’s 2022 total power capacity. | 中 | SR014 |
| CR017 | RAND says inability to secure enough power could push AI data-centre buildout abroad, increasing security risk and undermining semiconductor export controls. | 中 | SR014 |
| CR018 | DOE and IEA together show that AI infrastructure expansion is already colliding with local-grid and component bottlenecks rather than scaling frictionlessly. | 高 | SR013, SR015 |
| CR019 | DOE says connection requests for hyperscale 300-1000MW facilities with one- to three-year lead times are stretching local grid capacity. | 中 | SR015 |
| CR020 | BloombergNEF projects data-centre power demand could hit 106 gigawatts by 2035. | 中 | SR016 |
| CR021 | BloombergNEF says PJM data-centre capacity could reach 31GW by 2030, nearly matching the 28.7GW of new generation expected over the same period, and ERCOT reserve margins could fall into risky territory after 2028. | 中 | SR016 |
| CR022 | Shaping Tomorrow highlights AI infrastructure concentration risk as systemic dependency on four vendors and flags regulatory fragmentation as a board-level AI risk. | 中 | SR017 |
| CR023 | The reviewed Senior Executive source recommends open-source models like Mistral or Llama and cloud-agnostic modular architectures as hedges against hyperscaler lock-in. | 中 | SR018 |
| CR024 | The Cloud Security Alliance post says the cited multimodal red-teaming report found tested models up to 60 times more prone to CSEM-related textual responses than comparable models under the report conditions. | 中 | SR011 |
| CR025 | The same multimodal safety source says tested models were 18 to 40 times more likely to produce dangerous CBRN information when prompted adversarially. | 中 | SR011 |
| CR026 | MIT’s AI Risk Repository groups AI risk into recurring categories that include harmful content, unfair treatment, privacy leakage, exploitable vulnerabilities, and misinformation. | 中 | SR012 |
| CR027 | OpenAI says enterprise customers receive ownership and control over their business inputs and outputs and support for compliance needs. | 中 | SR019 |
| CR028 | Anthropic’s enterprise materials advertise SSO, role-based access, a compliance API, HIPAA-ready offering, and no model training on customer content by default. | 中 | SR020 |
| CR029 | Mistral Studio markets hybrid, dedicated, and self-hosted deployment modes with full ownership of customer data. | 中 | SR021 |
| CR030 | Cohere’s security page says customers can opt out of model training and deploy through a VPC, on-premises setup, or dedicated Model Vault, while the API platform is SOC 2 Type II compliant. | 中 | SR024 |
| CR031 | Meta markets Llama as open-source AI. | 中 | SR025 |
| CR032 | Gemma markets open models that can run from cloud servers to laptops and phones and includes ShieldGemma 2 for policy-violating-content detection. | 中 | SR026 |
| CR033 | Mistral Small 3.1 is marketed as Apache 2.0-licensed, up to 128k context, roughly 150 tokens per second, and lightweight enough for a single RTX 4090. | 中 | SR022 |
| CR034 | Cohere Command markets private deployment and enterprise workflow integration rather than purely public self-serve inference. | 中 | SR023 |
| CR035 | SiliconANGLE reports that Reka raised $110 million backed by Nvidia and Snowflake and that the round valued the company at $1 billion. | 中 | SR027 |
| CR036 | GetLatka estimates that Reka had roughly 60 employees in 2026 and reached about $10.9 million of revenue in 2025. | 低 | SR028 |
| CR037 | Tracxn lists Reka at 64 employees as of May 2026, about $168 million of cumulative funding, and a current valuation of $1 billion. | 中 | SR029 |
| CR038 | Public tracker and press estimates imply investors are underwriting future scale-up rather than mature cash generation, because roughly $1 billion of valuation sits against publicly estimated 2025 revenue of about $10.9 million. | 中 | SR027, SR028, SR029 |
| CR039 | Publicly marketed privacy and deployment controls from larger rivals raise the enterprise benchmark that Reka must clear in security and procurement reviews. | 高 | SR019, SR020, SR021, SR023, SR024 |
| CR040 | Open-weight and self-hostable alternatives have become materially more credible because Llama is marketed as open-source, Gemma emphasizes open deployment, and Mistral Small 3.1 markets Apache 2.0 licensing with lightweight hardware needs. | 高 | SR022, SR025, SR026 |
| CR041 | Taken together, the AI Act, EDPB, ICO, and CCPA sources make training-data provenance, privacy rights handling, and documentation quality diligence-critical for a multimodal model provider. | 高 | SR003, SR004, SR007, SR008, SR009 |
| CR042 | The public sources reviewed for this chapter do not disclose Reka’s reserved GPU capacity, power-backed hosting commitments, or priority allocation rights. | 低 | |
| CR043 | The public sources reviewed for this chapter do not disclose top-customer concentration, top-channel concentration, or revenue retention metrics for Reka. | 低 | |
| CR044 | Nudge Security frames vendor-risk diligence around security certifications, supply chain detail, GDPR compliance, breach history, and application access, showing the scope of questions enterprise buyers are likely to ask. | 中 | SR030 |
| CR045 | Reka’s public change log shows ongoing product changes, including free Flash 3 chat access in April 2025 and adding Flash 3 to the API in March 2025. | 中 | SR031 |
| CR046 | If governance proof or compute capacity slips, the first business effect is likely slower enterprise conversion and higher infrastructure cost before it is a pure model-quality problem. | 中 | SR013, SR015, SR017, SR027 |
| CR047 | A smaller vendor facing stronger public privacy controls from rivals is likely to experience longer security and procurement cycles unless it can show equivalent enterprise safeguards. | 中 | SR019, SR020, SR021, SR024, SR030 |
| CR048 | A rational diligence stop-light should turn red if Reka cannot evidence training-data governance, named security controls, diversified customers, and committed compute capacity. | 中 | SR003, SR009, SR013, SR030 |
| CR049 | The most plausible downside scenario is a combination of compute or compliance friction, slower enterprise conversion, and renewed financing pressure rather than a single catastrophic product failure. | 中 | SR013, SR014, SR016, SR017, SR027, SR029 |
| CR050 | Visible public mitigations already include paid-plan training opt-in, a small-company lock-in hedge via open or modular architectures, and competitor-shaped demand for private deployment and governance controls. | 中 | SR001, SR018, SR021, SR024 |
| CV001 | Reka announced a $110 million financing in July 2025 backed by NVIDIA and Snowflake. | 中 | SV001, SV030 |
| CV002 | Reuters-syndicated coverage and private-company trackers place Reka's latest round at more than a $1 billion valuation. | 中 | SV027, SV002, SV003 |
| CV003 | GetLatka estimates that Reka generated $10.9 million of revenue in 2025 with about 60 employees. | 中 | SV002 |
| CV004 | Tracxn lists Reka as a Series B company with roughly 64 employees, about $168 million of funding, and a current valuation of $1 billion. | 中 | SV003 |
| CV005 | Using a $1 billion valuation and $10.9 million of estimated 2025 revenue implies an approximately 91.7x trailing revenue multiple for Reka. | 中 | SV002, SV027 |
| CV006 | Using roughly $168 million-$170 million of cumulative funding against $10.9 million of estimated 2025 revenue implies about 15.4x-15.6x funding-to-revenue. | 中 | SV002, SV003 |
| CV007 | Independent reporting in 2024 said Snowflake explored buying Reka for over $1 billion and later that the talks stopped. | 中 | SV028, SV029 |
| CV008 | Snowflake says customers will be able to bring Reka's multimodal assistant to their data, supporting the case that the partnership is a real distribution lever rather than only a capital-markets signal. | 中 | SV026 |
| CV009 | Bessemer wrote that the EMCLOUD index remained around historical norms while the private sector had arguably bubbled up again largely on the back of AI cloud. | 中 | SV004 |
| CV010 | Equidam argues that revenue multiples are especially dangerous for AI companies because compute-heavy cost structures make simple top-line shortcuts misleading. | 中 | SV005 |
| CV011 | Taken together, the BVP and Equidam lenses imply that Reka should be underwritten on forward revenue conversion and margin quality rather than on its trailing revenue estimate alone. | 中 | SV004, SV005 |
| CV012 | Sacra estimates Cohere reached $240 million of ARR in 2025, and BetaKit reported a February investor memo saying Cohere exceeded its internal $200 million target with quarter-over-quarter growth above 50% and gross margins around 70%. | 中 | SV006, SV007 |
| CV013 | Using a roughly $7 billion valuation context against $240 million of ARR implies a heuristic Cohere multiple of about 29x ARR. | 中 | SV006, SV007 |
| CV014 | Sacra says Glean reached $300 million of ARR by May 2026 after crossing $100 million in ARR in fiscal 2025, and that its valuation reached $7.2 billion in June 2025 after a $4.6 billion mark in September 2024. | 中 | SV009 |
| CV015 | Glean therefore shows that investors awarded multi-billion-dollar enterprise-AI application valuations only after ARR had scaled far beyond Reka's public revenue estimate. | 中 | SV009, SV002 |
| CV016 | Semafor reported that Anthropic aimed for a valuation between $15 billion and $20 billion in early 2024, with prior threshold terms and cloud-linked arrangements influencing its willingness to raise at a higher headline mark. | 中 | SV010 |
| CV017 | TechCrunch reported in June 2026 that Mistral was discussing a roughly €3 billion raise at about a €20 billion valuation after a €11.7 billion Series C mark in September 2025. | 中 | SV008 |
| CV018 | The same Mistral reporting ties part of that premium to sovereign-European positioning and major state or enterprise partnerships, which makes it a framing comp rather than a clean revenue-multiple comp for Reka. | 中 | SV008, SV011 |
| CV019 | Aleph Alpha raised a $500 million Series B in 2023 under a sovereignty-oriented positioning, illustrating that European enterprise AI narratives can attract large capital without mapping cleanly to Reka's current product and scale profile. | 中 | SV011 |
| CV020 | Databricks generated $1.6 billion of revenue for the year ended January 31, 2024. | 中 | SV012 |
| CV021 | TechCrunch reported that Databricks reached a $134 billion valuation at more than $4.8 billion of run-rate revenue in December 2025, implying roughly a 27.9x run-rate revenue multiple. | 中 | SV013 |
| CV022 | Harvey announced a $200 million financing at an $11 billion valuation and said more than 100,000 lawyers across 1,300 organizations use the platform. | 中 | SV025 |
| CV023 | Harvey shows that vertical-AI application companies can command double-digit-billion valuations, but only after much deeper workflow embed and customer scale than Reka has publicly disclosed. | 中 | SV025, SV002 |
| CV024 | Snowflake reported $4.684 billion of FY2026 revenue and had a June 2026 market capitalization of about $80.51 billion, implying roughly a 17.2x revenue multiple. | 高 | SV014, SV015, SV016, SV017 |
| CV025 | NVIDIA had FY2026 revenue of about $215.938 billion and a June 2026 market capitalization of about $5.103 trillion, implying roughly a 23.6x revenue multiple. | 中 | SV018, SV019, SV020 |
| CV026 | C3.ai reported $389.1 million of FY2025 revenue and had a June 2026 market capitalization of about $1.49 billion, implying roughly a 3.8x revenue multiple. | 高 | SV021, SV022 |
| CV027 | Palantir had about $4.475 billion of FY2025 revenue and a June 2026 market capitalization of about $307.98 billion, implying roughly a 68.8x revenue multiple. | 中 | SV023, SV024 |
| CV028 | Reka's implied ~91.7x trailing multiple is above the public multiples observed for Snowflake, NVIDIA, Palantir, and C3.ai in the retained sources. | 中 | SV002, SV014, SV015, SV016, SV018, SV019, SV021, SV022, SV023, SV024 |
| CV029 | At a $1 billion valuation, Reka would trade at 25x on $40 million of revenue, 20x on $50 million, and 15x on roughly $66.7 million. | 中 | SV002, SV004, SV005 |
| CV030 | The bull case requires Reka to reach roughly $60 million-$75 million of forward revenue within the next 12-24 months while sustaining software-like gross margins and turning partner access into repeatable enterprise sales. | 中 | SV001, SV002, SV026 |
| CV031 | On an 18x-22x revenue multiple, that bull case supports roughly a $1.1 billion-$1.65 billion valuation range. | 中 | SV004, SV009, SV013 |
| CV032 | A base case of roughly $35 million-$50 million of forward revenue on 14x-18x revenue supports about a $0.5 billion-$0.9 billion valuation range. | 中 | SV004, SV005, SV014, SV015, SV021, SV022 |
| CV033 | A bear case of roughly $20 million-$30 million of forward revenue on 10x-14x revenue supports only about a $0.2 billion-$0.4 billion valuation range. | 中 | SV004, SV005, SV021, SV022 |
| CV034 | Because the base case remains below the current round while the bull case requires unusually strong execution, the public evidence supports a monitor / price-sensitive stance instead of a straightforward buy recommendation. | 中 | SV002, SV004, SV005, SV027 |
| CV035 | The strongest drivers that would justify the current valuation are proof of a $30 million-$40 million forward run-rate, channel conversion through Snowflake, and gross margins above roughly 65%. | 中 | SV026, SV014, SV005 |
| CV036 | The strongest drivers that would undercut the valuation are compute-heavy gross margins, partner concentration, customer concentration, and a weak cap-table position for new money. | 中 | SV005, SV010, SV028, SV029 |
| CV037 | Public sources reviewed for this chapter still do not disclose audited ARR, net retention, burn, top-customer concentration, channel economics, or the 2025 preference stack. | 中 | SV001, SV002, SV003, SV027 |
| CV038 | Those gaps matter because structured AI financings and compute-linked contracts can make headline valuations look cleaner than the underlying economics. | 中 | SV005, SV010 |
| CV039 | If management can show run-rate revenue above $30 million-$40 million, net retention above about 120%, gross margins above 65%, and clean 1x non-participating preferences, the current round becomes materially easier to defend. | 中 | SV005, SV014, SV026 |
| CV040 | If management cannot show those items, investors should either negotiate materially better entry discipline or wait, because the current public evidence does not justify paying for perfect execution. | 中 | SV004, SV005, SV027 |
| CV041 | Strategic interest from Snowflake and the 2024 acquisition-talk reporting make a future strategic sale plausible, but they do not establish a hard floor above the current round. | 中 | SV026, SV028, SV029 |
| CV042 | A near-term IPO looks unlikely on public evidence because Reka's visible revenue base remains a small fraction of even the smallest public AI/software anchors reviewed here. | 中 | SV002, SV014, SV021, SV024 |
| CV043 | GetLatka's tracker estimates the latest round involved about 11% sold, which is not aggressive dilution by late-stage standards but says nothing about the preference stack. | 低 | SV002 |
| CV044 | NVIDIA and Snowflake backing reduce signaling risk, but strategic investors can also make price discovery less clean because they may value product access or ecosystem leverage more than a pure financial investor would. | 中 | SV001, SV010, SV026 |
| CV045 | Public valuation support is therefore stronger as a strategic-option story than as a trailing-fundamentals story. | 中 | SV001, SV002, SV005, SV026 |
| CV046 | The scenario tree implies asymmetric public-data risk/reward from a $1 billion entry: the bull case offers only moderate upside support while the base and bear cases both sit below the round. | 中 | SV002, SV004, SV005, SV014, SV026 |
| CV047 | The disciplined next step is diligence, not conviction: stay close to the company, but do not treat the public record as sufficient support for an immediate positive price call. | 中 | SV002, SV004, SV005, SV027 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Reka AI | Reka | We're building models and infrastructure for the physical AI era. |
| SO002 | Reka API | Reka API Documentation | |
| SO003 | Reka AI | Reka Vision | |
| SO004 | Reka AI | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | Reka Vision ... is used by companies such as Shutterstock [and] Turing Video. |
| SO005 | Reka AI | Reka Flash 3.1 and Reka Quant | Reka Flash 3.1 improves by 10 points on LiveCodeBench v5 from Reka Flash 3. |
| SO006 | Reka AI | Reka Research: Knowledge Made Accessible | Enterprises also have the options to deploy Reka Research on-premise, in their private cloud, or use through our API at scale. |
| SO007 | Reka AI | Reka and Turing Partner to Pioneer Agentic Video Surveillance Platform | Turing is a leader in security surveillance solutions with over 13,000+ site deployments and 10M+ daily events processed. |
| SO008 | Reka AI | Reka Vision: Intelligence Made Visible | |
| SO009 | Reka AI | Adding Reka Vision Without Replacing VMS: What Actually Works | You've read about Reka Vision cutting case resolution time by 65%, reducing false alarms by 95%. |
| SO010 | Reka AI | Defence & Security - Sovereign Multimodal AI | Reka supports qualified defence and security programmes with air-gapped deployments, on-premise infrastructure, and models built for mission-critical visual intelligence. |
| SO011 | Snowflake | Snowflake invests in Reka, Further Expanding LLM Capabilities in the Data Cloud | Today we’re excited to announce our investment and partnership with Reka. |
| SO012 | Snowflake | Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex | Today we are excited to announce we’re furthering our partnership with Reka to support its suite of highly capable multimodal models in Snowflake Cortex. |
| SO013 | arXiv | Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models | Core performs competitively to GPT4-V ... and on video question answering ... Core outperforms Gemini Ultra. |
| SO014 | Business Wire | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | |
| SO015 | Investing.com / Reuters syndication | Reka AI raises $110 million, valuation tops $1 billion | The company ... tripling its valuation to over $1 billion. |
| SO016 | CRN | Snowflake Eyes Reka AI Buy For $1B To Boost Generative AI, LLMs | Snowflake is seeking to acquire AI startup company Reka AI for over $1 billion. |
| SO017 | MarketScreener | Snowflake Reportedly in Talks to Purchase Reka AI for over $1 Billion | |
| SO018 | AK&M | Snowflake and Reka AI have stopped negotiations on a $1.0 billion deal | Snowflake and Reka AI have stopped negotiations on a $1.0 billion deal. |
| SO019 | Shutterstock Investor Relations | Reka Announces Partnership with Shutterstock | Reka will also add Shutterstock to its growing roster of customers as the company retains Reka to further enhance the value of the metadata supporting their image and video library. |
| SO020 | GetLatka | Reka AI Revenue 2025: $10.9M ARR, $1B Valuation | |
| SO021 | Tracxn | Reka company profile | |
| SO022 | Reka API | API Pricing | Reka Research ... costs $25 per 1000 queries regardless of the number of tokens and the steps taken. |
| SO023 | Reka API | Vision API Pricing | Enterprise: custom arrangement ... No rate limits. |
| SO024 | Reka API | Vision API | |
| SO025 | Reka API | Reka Research | |
| SM001 | Gartner | Gartner predicts 80% of enterprise software and applications will be multimodal by 2030 | Eighty percent of enterprise software and applications will be multimodal by 2030, up from less than 10% in 2024. |
| SM002 | Gartner | Gartner forecasts worldwide AI spending to grow 47% in 2026 | Worldwide spending on AI is forecast to total $2.59 trillion in 2026. |
| SM003 | IDC | IDC FutureScape 2026 | IDC’s FutureScape 2026 reveals how AI is scaling from pilots to enterprise transformation. |
| SM004 | ResearchAndMarkets | Multimodal AI Market Report | Global Multimodal AI Market, Segmentation by Type ... Offering ... Data Modality ... Vertical ... Historic and Forecast. |
| SM005 | The Business Research Company | Foundation Artificial Intelligence (AI) Models Market Report | The main model types ... are language models, vision models, multimodal models, speech recognition, speech generation, and code generation models. |
| SM006 | Control Risks | The AI compute contest | In 2026, getting access to compute will require diplomacy as much as money. |
| SM007 | ARK Invest | The State Of AI Infrastructure: Demand, Costs, And Custom Silicon | Enterprise demand is also growing at a torrid pace. As measured by OpenRouter ... token demand has risen 28x since December 2024. |
| SM008 | Artificial Analysis | LLM leaderboard models | Comparison and ranking the performance of over 100 AI models across key metrics including intelligence, price, performance and speed. |
| SM009 | NVIDIA | Leading Inference Providers Achieve Lowest Token Cost With Open Source Models on NVIDIA Blackwell | These providers ... are using the NVIDIA Blackwell platform, which helps them reduce cost per token by up to 10x compared with the NVIDIA Hopper platform. |
| SM010 | Snowflake | Snowflake, AWS & NVIDIA Blackwell Power Enterprise AI | By embedding NVIDIA Blackwell-class compute into Snowflake architecture, customers can build powerful AI models, agents and applications ... within a governed security perimeter. |
| SM011 | Reka AI | Reka Vision | Purposefully engineered for enterprises, creators, and developers who need state-of-the-art multimodal AI. |
| SM012 | Reka AI | Reka Research: Knowledge Made Accessible | Reka Research can synthesize information from multiple sources in a multi-hop manner. |
| SM013 | Reka AI | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | This investment will ... scale Reka’s multimodal platforms, aiming for wider enterprise adoption. |
| SM014 | Reka AI | Reka and Turing Partner to Pioneer Agentic Video Surveillance Platform | Turing launched Guardian AI, an agentic video surveillance software. |
| SM015 | Snowflake | Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex | This will allow our customers to seamlessly unlock value from more types of data with the power of multimodal AI in the same environment where their data lives. |
| SM016 | Reka API | Vision API | Vision API offers video upload and management, semantic search, Q&A, clip generation, and metadata tagging. |
| SM017 | Reka API | Reka Research | Reka Research is best suited for answering factual questions that require accessing up to 20 sources. |
| SM018 | Reka API | API Pricing | Pay as you go. Get started with no upfront costs. |
| SM019 | OpenAI | OpenAI API Pricing | OpenAI publishes multimodal and realtime pricing tiers across text, image, and audio. |
| SM020 | Google DeepMind | Gemini 3.5 | Gemini emphasizes advanced multimodal understanding, long horizon tasks, and multi-step problem solving. |
| SM021 | Google AI for Developers | Gemini Developer API pricing | Start building free of charge with generous limits, then scale up with paid and enterprise pricing. |
| SM022 | Cohere | Command Models: AI-Powered Solutions for Enterprise | Secure, production-ready efficiency for agentic intelligence. |
| SM023 | Mistral | Models Overview | A list of all our available models, helping you explore their capabilities, performance, trade-offs, and more. |
| SM024 | Google DeepMind | Gemma | Our most advanced open models help developers create AI applications that run wherever users need them. |
| SM025 | Meta | Industry Leading, Open-Source AI | Llama | Industry Leading, Open-Source AI | Llama |
| SP001 | Reka API | API Pricing | Reka Flash ... $0.80 ... $2.00. |
| SP002 | Reka AI | Reka Flash 3.1 and Reka Quant | A multimodal version of Reka Flash 3.1 serves as a base model for our core products Reka Research and Reka Vision. |
| SP003 | Reka API | Vision API | Vision API offers video and image management, QA, semantic search, clip generation, metadata tagging, and more. |
| SP004 | Reka API | Reka Research | Reka Research can browse the web and private documents to answer complex questions. |
| SP005 | Artificial Analysis | LLM leaderboard models | Comparison and ranking the performance of over 100 AI models. |
| SP006 | OpenAI | OpenAI API Pricing | Power applications across text, image, and audio with models built for real-time interaction. |
| SP007 | OpenAI | Hello GPT-4o | GPT-4o accepts as input any combination of text, audio, image, and video. |
| SP008 | Anthropic | Plans & Pricing | Claude | Access to Research ... enterprise search across your organization. |
| SP009 | Anthropic | Introducing Claude 3.5 Sonnet | The model costs $3 per million input tokens and $15 per million output tokens, with a 200K token context window. |
| SP010 | Anthropic | Introducing the next generation of Claude | The Claude 3 models have sophisticated vision capabilities on par with other leading models. |
| SP011 | Google DeepMind | Gemini 3.5 | Advanced multimodal understanding ... long horizon tasks ... multi-step problem-solving. |
| SP012 | Google AI for Developers | Gemini Developer API pricing | Start building free of charge ... then scale up with paid ... enterprise. |
| SP013 | Mistral | Mistral Studio | One platform to build, deploy, and govern agentic AI systems—all with enterprise privacy, security, and full ownership of your data. |
| SP014 | Mistral | Models Overview | A list of all our available models. |
| SP015 | Cohere | Command Models: AI-Powered Solutions for Enterprise | Secure, production-ready efficiency for agentic intelligence. |
| SP016 | Cohere | Pricing | Secure and Scalable Enterprise AI | Move from proof of concept into production with our enterprise-ready AI solutions — private, secure, and built to work with your existing systems. |
| SP017 | Aleph Alpha | Aleph Alpha | Unsere SLLMs laufen kompromisslos auf europäischer Infrastruktur. |
| SP018 | Google DeepMind | Gemma | Our most advanced open models help developers create AI applications that run wherever users need them. |
| SP019 | Meta | Industry Leading, Open-Source AI | Llama | Industry Leading, Open-Source AI | Llama |
| SP020 | Snowflake | Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex | This will allow our customers to seamlessly unlock value from more types of data. |
| SP021 | Snowflake | Snowflake invests in Reka, Further Expanding LLM Capabilities in the Data Cloud | Through this partnership, Snowflake customers will be able to bring Reka's leading multimodal assistant to their data. |
| SP022 | Tracxn | Reka company profile | The company has 302 active competitors. |
| SP023 | GetLatka | Reka AI Revenue 2025: $10.9M ARR, $1B Valuation | How Reka AI CEO Dani Yogatama grew to $10.9M revenue with a 60 person team in 2025. |
| SP024 | Reka API | Vision API Pricing | Enterprise ... no rate limits. |
| SP025 | Snowflake | Snowflake, AWS & NVIDIA Blackwell Power Enterprise AI | Snowflake ... is addressing these challenges by unifying the AI lifecycle within the Snowflake AI Data Cloud. |
| SP026 | OpenAI | ChatGPT Enterprise | Deploy enterprise-grade ChatGPT—powered by OpenAI’s industry-leading models, products, and expertise, and connected to your company’s data. |
| SI001 | Reka API | API Pricing | Pay as you go. Get started with no upfront costs. You only pay for what you use. |
| SI002 | Reka API | Vision API Pricing | Enterprise: Monthly invoicing (billed at end of month) ... Bulk discounts available ... No rate limits. |
| SI003 | Reka API | Reka Research | Reka Research can browse the web and private documents to answer complex questions. |
| SI004 | Reka API | Vision API | Vision API offers video and image management, QA, semantic search, clip generation, metadata tagging, and more. |
| SI005 | Reka AI | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | The company's focus on efficient training and serving infrastructure has enabled it to develop market-leading models at a fraction of the cost. |
| SI006 | Business Wire | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | This investment will significantly accelerate Reka's technical development efforts. The funding will also scale Reka's multimodal platforms, aiming for wider enterprise adoption. |
| SI007 | Snowflake | Snowflake Invests in Reka, Further Expanding LLM Capabilities in Data Cloud | Through this partnership, Snowflake customers will be able to bring Reka's leading multimodal assistant to their data, with the ability to run and fine-tune it all within their Snowflake account. |
| SI008 | Snowflake | Multimodal LLM in Snowflake with Reka | We're furthering our partnership with Reka to support its suite of highly capable multimodal models in Snowflake Cortex. |
| SI009 | Snowflake Documentation | Multimodal AI in Snowflake Cortex AI Functions | Cortex AI Functions support multimodal analysis across documents, images, audio, and video, enabling end-to-end media understanding and processing pipelines directly inside Snowflake. |
| SI010 | Snowflake Documentation | Snowflake Cortex AI model capabilities and regional availability | To achieve the best performance per credit, choose a model that's a good match for the content size and complexity of your task. |
| SI011 | Snowflake | Cortex AI | Build gen AI applications directly in SQL or via APIs, analyze multimodal data and build agents — all within Snowflake's secure perimeter. |
| SI012 | Snowflake | Getting Started with Multimodal Analysis on Snowflake Cortex | You'll learn how to build an end-to-end application for multimodal analysis using AI models through Snowflake Cortex AI. |
| SI013 | Snowflake | Optimizing the AI Data Cloud with NVIDIA Blackwell to Secure Proprietary IP and Accelerate Full-Lifecycle AI Development | Snowflake moves beyond offering faster instances to deliver a cohesive platform ... while reducing the total cost of ownership (TCO) through improved operational velocity. |
| SI014 | Investing.com | Reka AI raises $110 million, valuation tops $1 billion | Reka AI has secured $110 million in a new funding round from investors including Nvidia and Snowflake, tripling its valuation to over $1 billion. |
| SI015 | Tech Funding News | Reka rockets to unicorn status with $110M round, leading the multimodal AI generation | Reka plans to use the new capital to expand the reach of its multimodal platforms, continue technical development, and hire more engineering talent. |
| SI016 | AIM Media House | Foundation model startup Reka just tripled its worth with $110 million in funding | Its headcount has already grown to 50, and the company says it will invest further in infrastructure to support broader enterprise adoption. |
| SI017 | CRN | Snowflake Eyes Reka AI Buy For $1B To Boost Generative AI, LLMs | Snowflake is seeking to acquire AI startup company Reka AI for over $1 billion. |
| SI018 | MarketScreener | Snowflake Reportedly in Talks to Purchase Reka AI for over $1 Billion | The company also lets customers use third-party AI models, such as those from Reka, on their data within Snowflake. |
| SI019 | Shutterstock Investor Relations | Reka Announces Partnership with Shutterstock | Reka will also add Shutterstock to its growing roster of customers as the company retains Reka to further enhance the value of the metadata supporting their image and video library. |
| SI020 | GetLatka | Reka AI Revenue 2025: $10.9M ARR, $1B Valuation | In 2025, Reka AI's revenue reached $10.9M. |
| SI021 | Tracxn | Reka - 2026 Company Profile, Team, Funding & Competitors | Reka has raised a total funding of $168M over 2 rounds ... latest funding round was a Series B round on Jul 22, 2025 for $110M. |
| SI022 | NVIDIA | How Inference Providers Use Blackwell to Reduce Cost Per Token | NVIDIA Blackwell ... helps them reduce cost per token by up to 10x compared with the NVIDIA Hopper platform. |
| SI023 | Control Risks | The AI Compute Contest | In 2026, getting access to compute will require diplomacy as much as money. |
| SI024 | ARK Invest | The State of AI Infrastructure: Demand, Costs, and Custom Silicon | AI training costs have been falling 75% per year. Inference costs are falling faster. |
| SI025 | Reka API | Reka API Documentation | Use our models via the API to build scalable production workloads. |
| SI026 | U.S. Securities and Exchange Commission | Snowflake Inc. Annual Report (Form 10-K) | |
| SI027 | World Economic Forum | Is power grid connectivity the strategic bottleneck for AI? | Investment in AI data centres is growing faster than power grids can keep up, making grid connectivity a constraint. |
| SE001 | Reka AI | Quickstart | |
| SE002 | Reka AI | Chat API overview | |
| SE003 | Reka AI | Chat with image, video, and audio | |
| SE004 | Reka AI | Function calling | |
| SE005 | Reka AI | Models | |
| SE006 | Reka AI | Video Management | |
| SE007 | Reka AI | Video Search | |
| SE008 | Reka AI | Video Q&A | |
| SE009 | Reka AI | Highlight Clip Generation | |
| SE010 | Reka AI | Metadata Tagging | |
| SE011 | Reka AI | Image Search | |
| SE012 | Reka AI | MCP Server | |
| SE013 | Reka AI | Errors | |
| SE014 | Reka AI | Vision rate limits | |
| SE015 | Reka AI | Reka Edge | Physical AI at the Edge | |
| SE016 | Reka AI | Reka Labs | Where Multimodal Reasoning Is Built | |
| SE017 | Reka AI | Reka Edge: Frontier-Level Edge Intelligence for Physical AI | |
| SE018 | Reka AI | Reka Speech: High Throughput Speech Transcription and Translation Model with Timestamps | |
| SE019 | Reka AI | Introducing Parallel Thinking for Reka Research | |
| SE020 | Reka AI | Introducing Research-Eval: A Benchmark for Search-Augmented LLMs | |
| SE021 | Reka AI | Research at Reka: Reasoning | |
| SE022 | Reka AI | Reinforcement Learning for Reka Flash 3.1 | |
| SE023 | Reka AI | Reka Quantization Technology | |
| SE024 | Reka AI | Privacy Policy | |
| SE025 | arXiv | Reka Core, Flash, and Edge: A Series of Powerful Multimodal Language Models | |
| SE026 | Artificial Analysis | Reka Flash - Intelligence, Performance & Price Analysis | |
| SE027 | GitHub | reka-ai organization | |
| SE028 | GitHub | reka-ai/n8n-nodes-reka | |
| SE029 | Hugging Face | RekaAI/reka-flash-3.1 | |
| SE030 | Oracle | Oracle Unveils New Defense Ecosystem Members | |
| SE031 | NVIDIA | Build a Video Search and Summarization (VSS) Agent Blueprint by NVIDIA | |
| SE032 | Reka AI | Using NVIDIA AI Blueprint for Video Search and Summarization with Reka Vision Agent | |
| SE033 | Snowflake | Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex | |
| SE034 | OpenAI | ChatGPT Enterprise | |
| SE035 | Google DeepMind | Gemini 3.5 | |
| SE036 | AIPedia | Reka AI Review: Edge 2, Pricing & Physical AI (June 2026) | |
| SU001 | Reka | Reka | |
| SU002 | Reka | Reka Vision | |
| SU003 | Reka | Defence & Security - Sovereign Multimodal AI | Reka | |
| SU004 | Reka | Quickstart | Reka API | |
| SU005 | Reka | API Pricing | Reka API | |
| SU006 | Reka | FAQs | Reka API | |
| SU007 | Reka | Reka Vision overview | Reka API | |
| SU008 | Reka | Rate Limits | Reka Vision API | |
| SU009 | Reka | MCP Server | Reka Vision API | |
| SU010 | Reka | Reka Announces Partnership with Shutterstock | Reka will also add Shutterstock to its growing roster of customers as the company retains Reka to further enhance the value of the metadata supporting their image and video library. |
| SU011 | Shutterstock | Reka Announces Partnership with Shutterstock | Shutterstock expects to leverage Reka’s leading multimodal models to augment details and enhance the metadata attached to its library of digital assets. |
| SU012 | Shutterstock | How Reka Uses Shutterstock Data to Create State-of-the-Art Multimodal AI Models | In return, the AI company would enhance the metadata of Shutterstock’s 550 million assets across images and video. |
| SU013 | Benzinga | Shutterstock Expands AI Horizons: New Partnership with Reka AI to Enhance Digital Asset Metadata - Apple | |
| SU014 | Photutorial | Reka.ai partners with Shutterstock to enhance AI and metadata capabilities | |
| SU015 | Snowflake | Snowflake invests in Reka, Further Expanding LLM capabilities in the Data Cloud | Through this partnership, Snowflake customers will be able to bring Reka's leading multimodal assistant to their data, with the ability to run and fine-tune it all within their Snowflake account. |
| SU016 | Snowflake | Snowflake Brings Gen AI to Images, Video and More With Multimodal Language Models from Reka in Snowflake Cortex | This will allow our customers to seamlessly unlock value from more types of data with the power of multimodal AI in the same environment where their data lives. |
| SU017 | Snowflake | Cortex AI Functions: Multimodal | Snowflake Documentation | |
| SU018 | Snowflake | Getting Started with Multimodal Analysis on Snowflake Cortex AI | |
| SU019 | Snowflake | Snowflake Cortex AI | |
| SU020 | Reka | Reka and Turing Partner to Pioneer Agentic Video Surveillance Platform | Guardian AI was built on top of Turing’s platform and Reka Vision. |
| SU021 | Turing AI | General Agent | Turing AI | |
| SU022 | Reka | How Reka Vision Is Powering the Future of AI-Driven Security | The Ohio Police Department is using the solution to augment investigations. |
| SU023 | reka-ai | GitHub - reka-ai/n8n-nodes-reka: Official n8n nodes to use Reka's AI in our workflows | |
| SU024 | reka-ai | GitHub - reka-ai/reka-mcp: Reka AI's MCP server | |
| SU025 | n8n | Generate AI video clips from YouTube using Reka Vision API and Gmail | n8n workflow template | |
| SU026 | Oracle | Oracle Unveils New Defense Ecosystem Members | |
| SU027 | eesel AI | Reka AI pricing: A complete 2025 overview | Using a foundational model API from someone like Reka might seem like a good deal, but it comes with a lot of hidden work and headaches. |
| SU028 | Comparateur-IA | Reka AI — Multimodal Models for Text, Image, Audio & Video | |
| SU029 | Reka | Adding Reka Vision Without Replacing VMS: What Actually Works | Frontier intelligence scales linearly. Success begins with a high-value subset: the 20% of cameras that cover 80% of your security surface area. |
| SR001 | Reka | Terms of Use - Reka | If you make a paid request ... Reka will not use Your Content for model training unless you have explicitly opted in. |
| SR002 | Reka | Business Terms - Reka | |
| SR003 | European Commission | General-purpose AI obligations under the AI Act | Obligations for all providers of GPAI models: Draw up technical documentation, implement a copyright policy, publish a summary of the model's training content. |
| SR004 | European Union | Regulation (EU) 2024/1689 (Artificial Intelligence Act) | |
| SR005 | U.S. Copyright Office | Copyright and Artificial Intelligence | On May 9, 2025, the Office released a pre-publication version of Part 3 ... A final version of Part 3 will be published in the future, without any substantive changes expected in the analysis or conclusions. |
| SR006 | Library of Congress Copyright Blog | Inside the Copyright Office’s Report, Copyright and Artificial Intelligence, Part 2: Copyrightability | The outputs of generative AI can be protected by copyright only where a human author has determined sufficient expressive elements. |
| SR007 | Information Commissioner’s Office | Artificial intelligence | AI and data protection risk toolkit ... practical support for organisations assessing the risks to individual rights and freedoms caused by their own AI systems. |
| SR008 | California Department of Justice | California Consumer Privacy Act (CCPA) | The right to delete ... the right to opt-out ... the right to correct inaccurate personal information. |
| SR009 | European Data Protection Board | EDPB opinion on AI models: GDPR principles support responsible AI | The opinion looks at when and how AI models can be considered anonymous, whether legitimate interest can be used, and what happens if an AI model is developed using personal data that was processed unlawfully. |
| SR010 | NIST | AI Risk Management Framework | The NIST AI Risk Management Framework is intended for voluntary use and to improve the ability to incorporate trustworthiness considerations into the design, development, use, and evaluation of AI products, services, and systems. |
| SR011 | Cloud Security Alliance / Enkrypt AI | Multimodal AI Faces New Safety Threats | These two models are 60 times more prone to generate child sexual exploitation material-related textual responses ... and 18-40 times more likely to produce dangerous CBRN information. |
| SR012 | MIT AI Risk Repository | MIT AI Risk Repository | |
| SR013 | International Energy Agency | Data centre electricity use surged in 2025, even with tightening bottlenecks driving a scramble for solutions | Electricity demand from data centres soared by 17% in 2025 ... electricity consumption from data centres is set to double by 2030, and power use from those focused on AI is poised to triple. |
| SR014 | RAND | How Much Power Will AI Systems Need? | AI data centers could require 68 gigawatts of power globally by 2027 ... If U.S. companies cannot find adequate power, they may build data centers abroad. |
| SR015 | U.S. Department of Energy | Recommendations on Powering Artificial Intelligence and Data Center Infrastructure | Connection requests for hyperscale facilities of 300-1000MW or larger with lead times of 1-3 years are stretching the capacity of local grids. |
| SR016 | BloombergNEF | AI and the Power Grid: Where the Rubber Meets the Road | Data-center power demand hits 106 gigawatts by 2035 ... in PJM, BNEF forecasts data center capacity could 31GW by 2030. |
| SR017 | Shaping Tomorrow | AI Infrastructure Race: Navigating Critical Risks and Opportunities | AI Infrastructure Concentration Risk: $650-700B hyperscaler capex creates systemic dependency on four vendors. |
| SR018 | Senior Executive AI Think Tank | Competing in AI When Infrastructure Is Controlled by Hyperscalers | Startups should leverage open-source models like Mistral or Llama ... use modular, API-driven architectures that stay cloud-agnostic. |
| SR019 | OpenAI | Enterprise privacy at OpenAI | Our commitments provide you with ownership and control over your business data (inputs and outputs ...) and support for your compliance needs. |
| SR020 | Anthropic / Claude | Plans & Pricing | Claude by Anthropic | Enterprise ... Single sign-on (SSO) ... Compliance API ... HIPAA-ready offering ... No model training on your content by default. |
| SR021 | Mistral AI | Mistral Studio | Your AI production platform. | One platform to build, deploy, and govern agentic AI systems—all with enterprise privacy, security, and full ownership of your data. |
| SR022 | Mistral AI | Mistral Small 3.1 | Mistral Small 3.1 is released under an Apache 2.0 license ... up to 128k tokens ... 150 tokens per second ... can run on a single RTX 4090. |
| SR023 | Cohere | Cohere Command Models: AI-Powered Solutions for Enterprise | Deploy securely, whether through private deployments or in a hyperscaler VPC. |
| SR024 | Cohere | AI Security and Data Protection | Cohere | Opt out of model training at any time — your data stays yours ... Deploy through your virtual private cloud (VPC), on-premises setup, or dedicated, Cohere-managed Model Vault. |
| SR025 | Meta | Industry Leading, Open-Source AI | Llama | |
| SR026 | Google DeepMind | Gemma | Our most advanced open models help developers create AI applications that run wherever users need them — from cloud servers to laptops and even phones. |
| SR027 | SiliconANGLE | Multimodal AI startup Reka AI raises $110M at $1B valuation | Reka AI ... raised $110 million in fresh funding backed by Nvidia and Snowflake. |
| SR028 | GetLatka | Reka AI Revenue 2025: $10.9M ARR, $1B Valuation | Reka AI employs approximately 60 people as of 2026 ... In 2025, Reka AI's revenue reached $10.9M. |
| SR029 | Tracxn | Reka | Reka has 64 employees as of May 26 ... Reka has raised $168M in funding ... with a current valuation of $1B. |
| SR030 | Nudge Security | Is Reka AI Safe? Learn if Reka AI Is Legit | The following security profile for Reka AI includes ... security certifications, supply chain details, privacy policy, terms of service, GDPR compliance, breach history, and more. |
| SR031 | Reka | Latest changes - Reka | April 17th 2025: Access Reka Flash 3 for free on Space chat ... March 10th 2025: Added new reasoning model, Reka Flash 3, to the API. |
| SV001 | Reka | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms | Reka ... announced it has secured a $110 million investment. This funding is backed by new and existing investors including NVIDIA and Snowflake. |
| SV002 | GetLatka | Reka AI Revenue 2025: $10.9M ARR, $1B Valuation | In 2025, Reka AI's revenue reached $10.9M. |
| SV003 | Tracxn | Reka | Reka has raised $168M in funding. |
| SV004 | Bessemer Venture Partners | State of the Cloud 2024 | the private sector has rebounded and arguably bubbled up again, largely on the back of AI Cloud. |
| SV005 | Equidam | AI Startup Valuation: Revenue Multiples, 2025 Insights, Trends | For AI companies with their unique cost structures and technical risks, this crude approach isn’t just inadequate—it’s dangerous. |
| SV006 | Sacra | Cohere revenue, funding & news | Sacra estimates that Cohere hit $240 million in annual recurring revenue (ARR) in 2025. |
| SV007 | BetaKit | Cohere reportedly soars past revenue target, with $240-million USD ARR | Cohere reportedly hit $240 million USD in annual recurring revenue (ARR) last year. |
| SV008 | TechCrunch | Mistral is rumored to be raising €3B at €20B valuation | The funding round would value the company at around €20 billion. |
| SV009 | Sacra | Glean revenue, funding & news | Sacra estimates Glean hit $300M in annual recurring revenue (ARR) in May 2026. |
| SV010 | Semafor | Why hot AI startup Anthropic wanted a lower valuation | It also aimed to peg its worth somewhere between $15 billion and $20 billion. |
| SV011 | TechCrunch | Lidl owner and Bosch Ventures co-lead $500M Series B into German AI startup Aleph Alpha | German AI startup Aleph Alpha has raised a Series B funding round of $500 million. |
| SV012 | TechCrunch | Databricks keeps marching forward with $1.6B in revenue | For the year ending January 31, 2024, the late-stage startup pulled in $1.6 billion. |
| SV013 | TechCrunch | Databricks raises $4B at $134B valuation as its AI business heats up | Databricks ... has just raised more than $4 billion in a Series L funding round at a $134 billion valuation. |
| SV014 | Snowflake | Snowflake Reports Financial Results for the Fourth Quarter and Full-Year of Fiscal 2026 | Revenue of $1.28 billion in the fourth quarter ... Snowflake annual revenue for 2026 was $4.684B. |
| SV015 | CompaniesMarketCap | Snowflake (SNOW) - Market capitalization | As of June 2026 Snowflake has a market cap of $80.51 Billion USD. |
| SV016 | Macrotrends | Snowflake Revenue 2020-2026 | SNOW | Snowflake annual revenue for 2026 was $4.684B. |
| SV017 | U.S. Securities and Exchange Commission | Snowflake, Inc. Annual Report (FY2026 XBRL viewer) | |
| SV018 | CompaniesMarketCap | NVIDIA (NVDA) - Market capitalization | As of June 2026 NVIDIA has a market cap of $5.103 Trillion USD. |
| SV019 | Macrotrends | NVIDIA Revenue 2012-2026 | NVDA | NVIDIA annual revenue for 2026 was $215.938B. |
| SV020 | NVIDIA | NVIDIA Corporation - Financial Reports | |
| SV021 | C3 AI | C3 AI Announces Record Fiscal Fourth Quarter and Full Fiscal Year 2025 Financial Results | $389.1 million, an increase of 25% compared to $310.6 million one year ago. |
| SV022 | CompaniesMarketCap | C3 AI (AI) - Market capitalization | As of June 2026 C3 AI has a market cap of $1.49 Billion USD. |
| SV023 | CompaniesMarketCap | Palantir (PLTR) - Market capitalization | As of June 2026 Palantir has a market cap of $307.98 Billion USD. |
| SV024 | Macrotrends | Palantir Technologies Revenue 2019-2026 | PLTR | Palantir Technologies annual revenue for 2025 was $4.475B. |
| SV025 | Harvey | Harvey Raises at $11 Billion Valuation to Scale Agents Across Law Firms and Enterprises | The round values Harvey at $11 billion. |
| SV026 | Snowflake | Snowflake invests in Reka, Further Expanding LLM Capabilities in the Data Cloud | Through this partnership, Snowflake customers will be able to bring Reka's leading multimodal assistant to their data. |
| SV027 | Investing.com / Reuters syndication | Reka AI raises $110 million, valuation tops $1 billion | The company ... tripling its valuation to over $1 billion. |
| SV028 | CRN | Snowflake Eyes Reka AI Buy For $1B To Boost Generative AI, LLMs | Snowflake is seeking to acquire AI startup company Reka AI for over $1 billion. |
| SV029 | MarketScreener | Snowflake Reportedly in Talks to Purchase Reka AI for over $1 Billion | |
| SV030 | Business Wire | Reka Secures $110 Million to Accelerate Adoption of Its Multimodal AI Platforms |