Recursive
Recursive 尽职调查报告
Recursive 在战略上重要,按公司年龄看技术可信度也高,但当前 $4.65 billion 估值缺少公开证据支撑,尚不足以下买入结论。
封面要素
公司概况
Recursive 是一家成立于 2025 年的 AI 研究实验室,公开叙事围绕「构建能递归改进 AI 的 AI」,并最终自动化科学研究。 公开证据支持其在 2026 年 5 月走出隐身,获得超过 $650 million Series A 轮资本、估值 $4.65 billion;随后 2026 年 7 月 的技术发布让研究计划更具体。公开证据尚未支持的是常规商业画像:没有公开收入、没有具名客户、没有公开定价,治理细节也很少。
- 产品
- 公开可见资产描述的是内部自动化 AI 研究系统和相关基准工件,而不是已经完整打包的企业软件产品。
- 客户
- 早期买家很可能是成熟研究机构、前沿模型开发团队和技术型企业团队,但公开材料没有披露具名客户。
- 商业模式
- 尚未公开披露;最可能的路径是未来将自动化 AI 研究工作流或相关企业软件变现,而不是已经跑通的软件收入引擎。
- 阶段
- Series A
- 融资情况
- 2026 年 5 月走出隐身,获得超过 $650 million Series A 轮资本,估值 $4.65 billion。
执行摘要
主要优势
- 非同寻常的融资通道和高质量投资人,让 Recursive 有更大空间招人、购买算力,并继续推进前沿研究。
- 2026 年 7 月的技术发布明显提高了信心:融资叙事背后确实有一套自动化研究系统。
- 2026 年 AI 支出规模庞大,AI for science 顺风仍在;如果 Recursive 把技术证明转成产品化工作流,仍有机会跑出赢家。
主要风险
- 没有公开收入、定价或具名客户披露,当前估值不像正常软件投资那样有支撑。
- 既有平台已把治理、支出控制和工作流 AI 打包;Recursive 尚未展示公开控制面,就已面临捆绑压力。
- 股权结构、优先权和治理条款不明,可能让下行结果比私募 headline 估值暗示的更差。
未决问题
- 当前收入、烧钱速度、算力承诺和现金跑道均未公开披露。
- 完整股权结构表、清算优先权和投资人权利包仍未公开。
- 公开资料没有显示具名客户、试点或可进入采购流程的产品控制面。
目录
01公司概况
1.1 身份、版图与研究假设
今天的 Recursive 更应被理解为前沿 AI 研究实验室,而不是一家已有出货产品目录的商业软件供应商。它目前的官方线上入口是 recursive.com 域名,公司在这里把目标表述为构建能递归改进自身的 AI,用以自动化知识发现,并首先自动化 AI 科学本身。 同一批官方材料强调安全,也把公司定位为追求开放式算法,而不只是扩展常规基础模型栈。公开足迹信号不大,但彼此一致: 首页和 X 资料都指向 San Francisco 与 London,Tech.eu 还补充了 London 注册说明。组合起来看,这像是一家拥有英国法律根基 和美国运营存在的跨大西洋实验室。同样重要的是概览中明显缺席的内容:网站没有展示定价、客户标识或公开产品注册入口, 独立发布报道也明确称,公司走出隐身时尚未发布产品。因此,本章层面的判断应是:研究优先的身份、异常充足的资本支持, 但商业化证据仍然有限。[CO001, CO002, CO003, CO015, CO024, CO025]
| 指标 | 数值 / 状态 | 日期 | 置信度 | 缺口 / 注意事项 |
|---|---|---|---|---|
| 官方网站 | recursive.com | 当前 | 高 | 第三方追踪站有时仍列出 recursive.ai |
| 成立时间 | 2025 | 2025 | 中 | 本章未审阅公司注册文件包 |
| 阶段 | Series A 轮 | 2026-05-13 | 高 | 未披露后续轮次 |
| 最近披露融资额 | $650M+ | 2026-05-13 | 高 | 部分来源四舍五入为正好 $650M |
| 最近披露估值 | $4.65B | 2026-05-13 | 高 | 公开来源未发现后续老股估值标记 |
| 投资人财团 | GV、Greycroft、Nvidia、AMD Ventures 投资财团 | 2026-05-13 | 高 | 控制条款和确切持股未披露 |
| 员工数信号 | 25+ 且 <30 | 2026-05 to 2026-07 | 中 | 区间来自官方和发布报道推断,不是精确员工名册 |
| 办公地点 | San Francisco 和 London | 当前 | 中 | London 注册说明来自 Tech.eu,不是官方法律文件 |
| 产品状态 | 未公开披露已发布产品 | 2026-05 to 2026-07 | 中 | 研究产物已公开,但商业化包装缺失 |
| 公开技术里程碑 | 自动化 AI 研究结果和 GitHub 产物已发布 | 2026-07 | 中 | 结果由公司发布,本报告未独立复现 |
| 收入 / 客户指标 | 未公开披露 | 2026-07-20 | 高 | 后续财务和客户章节的重大缺口 |
这张快照把已交叉验证的融资事实,与公司发布的技术和团队披露放在一起。员工数和商业化仍是区间判断, 不是精确数字。
[CO001, CO003, CO004, CO005, CO006, CO007]Recursive 的公开叙事把顶尖研究履历、战略资本和自我改进 AI 研究循环连了起来,但商业化证明仍滞后。
[CO002, CO005, CO006, CO007, CO011, CO012]头部指标显示,Recursive 在资本和定位上异常强,但经审计的运营证明偏弱。
员工数用有来源支撑的区间表示,而不是精确数字,因为公开来源只给出团队规模范围。
[CO003, CO005, CO008, CO009, CO010, CO015]1.2 创始人、融资信号与治理可见度
创始信号是公开市场和私募投资者这么早就关注 Recursive 的主要原因。官方文案称,联合创始人曾创建 Salesforce 和 Uber 的 AI 实验室,并领导过 OpenAI、DeepMind、Google Brain 和 Meta 的团队;外部报道则稳定指向 Richard Socher 和 Tim Rocktäschel,范围更广的报道还加入 Yuandong Tian 以及 Josh Tobin、Jeff Clune、Tim Shi、Alexey Dosovitskiy、 Caiming Xiong 等更大圈层。这样的广度对技术可信度是正向信号,但也带来真实的尽调褶皱:当前官网没有公布完整领导层或董事会名单, 公开来源无法完全校准确切的联合创始人名单或治理结构。不过,融资头条清晰得多。Wilson Sonsini、Tech.eu 和 TechCrunch 都支持同一核心事实:Recursive 于 2026 年 5 月 13 日走出隐身,完成超过 $650 million 的 Series A 轮,估值 $4.65 billion, 由 GV 和 Greycroft 领投,Nvidia 与 AMD Ventures 参投。对一家成立于 2025 年的实验室来说,这是异常强的资本和信号结果, 但不能把它误读为控制条款、董事席位或老股交易的披露深度。[CO004, CO005, CO006, CO007, CO008, CO011]
| 人物 / 群体 | 公开角色或状态 | 背景 / 证据 | 创始人—市场匹配或报道覆盖 | 关键人 / 尽调备注 |
|---|---|---|---|---|
| Richard Socher | CEO 兼联合创始人 | Tech.eu、TNW、OfficeChai 和 Foundra 点名 | 直接把实验室与 Salesforce AI、MetaMind 和 You.com 经验连接起来 | 关键公开门面;确切股权分配和董事会角色未披露 |
| Tim Rocktäschel 本人 | 联合创始人兼研究负责人 | Tech.eu 和 OfficeChai 点名 | 开放式研究,以及 DeepMind / UCL 背景信号强 | 角色深度由外部描述,不在官网呈现 |
| Yuandong Tian | 外部发布报道中的联合创始人 | SCMP 和 TNW 点名 | 增加 Meta FAIR 和优化领域可信度 | 官网未列出其姓名 |
| Josh Tobin / Jeff Clune / Tim Shi 等人 | 更广泛发布报道中反复出现 | OfficeChai 和 Lab Index 将其与创始团队关联 | 扩展机器人、开放式 AI 和 OpenAI 履历 | 需要管理层确认组织架构和确切头衔 |
| Alexey Dosovitskiy / Caiming Xiong | 仅由 TNW / Foundra 点名 | 出现在更广泛的八人名单报道中 | 若确认,将显示 Transformer 和系统能力深度 | 名单一致性仍需尽调 |
| 官方网站领导层展示 | 没有列名的完整团队页面 | 首页描述过往机构,但没有完整名单 | 支撑投资逻辑和招聘故事,不支撑治理透明度 | 公开名单不透明本身就是概览风险 |
这里有意只做部分覆盖,因为官网没有列举完整团队,外部来源对发布时正式点名的联合创始人人数也不一致。
[CO011, CO012, CO013, CO014, CO029, CO030]| 利益相关方 | 角色 | 控制权或经济重要性 | 重要性证据 | 尽调问题 |
|---|---|---|---|---|
| GV | 领投方 | Series A 的锚定风险投资支持方 | 多家融资来源称其为领投方 | 确认董事席位、信息权和持股比例 |
| Greycroft | 联合领投方 | 显示 VC 支持面较广,而非单一机构押注 | 法律和新闻报道中与 GV 一同被点名 | 确认 Greycroft 是否也持有治理权 |
| Nvidia | 战略参投方 | 作为算力供应方和市场信号可能重要 | Tech.eu、WSGR、TNW 和 Europe Alternatives 点名 | 厘清任何优先访问、联合项目或非财务商业条款 |
| AMD Ventures | 战略参投方 | 增加第二家芯片厂商信号和潜在供应多元化 | 融资报道和法律公告中被点名 | 厘清参投是否纯财务性质,或与硬件合作绑定 |
| 创始团队 | 经济和技术控制中心 | 公开叙事显示,团队本身是被押注的核心资产 | Foundra 明确写到,团队清晰可辨带来溢价 | 索要完整股权结构表、归属安排、投票控制和继任结构 |
公开材料识别了本轮领投方和战略芯片投资人,但没有披露确切持股、董事会构成或优先股堆叠条款。
[CO005, CO006, CO007, CO029, CO038]1.3 技术里程碑与运营就绪度
2026 年 7 月,Recursive 发布首篇技术文章和配套 GitHub 仓库,缩短了雄心与证据之间的距离。这些材料重要,是因为公司从纯承诺 迈向了至少一个可检查的研究里程碑。文章描述了一个自动化 AI 研究循环:提出想法、实现想法、运行实验、验证结果,然后选择下一步尝试。 Recursive 随后声称在三个基准上取得 SOTA 结果,覆盖固定预算小模型训练、以速度为目标的小模型训练和 GPU kernel 优化。 具体数字仍由公司发布,尚未独立复现,因此应视为有意义但尚未完全承保的证据。即便如此,公开工件集也强于普通隐身页:公司仓库中有 基准专属文件夹,与 Karpathy 的 autoresearch、NanoGPT speedrun 等公开基准生态相连,并明确称融资会帮助扩展算力基础设施, 迈向首个 Level 1 自主训练系统。这个组合让命题更可读,但仍未证明商业产品或持久护城河。[CO016, CO017, CO018, CO019, CO020, CO021]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2025 | Recursive 成立 | 创立 | 公司成立 | 创始团队 | 锚定从成立到巨额融资之间异常短的时间线 |
| 2026-05-13 | 退出隐身并宣布 Series A | 融资 | $650M+,估值 $4.65B | Recursive 与 GV、Greycroft、Nvidia、AMD Ventures | 为本报告建立核心估值和阶段锚点 |
| 2026-05-13 | 披露战略芯片投资人 | 合作 | Nvidia 和 AMD Ventures 参投 | Recursive + 芯片生态 | 显示发布时已有异常强的算力供应信号 |
| 2026-05-13 | 报道 London 注册和双办公室布局 | 治理 | 在 London 注册;办公室位于 London 和 San Francisco | Tech.eu 与 Recursive | 增加司法辖区和运营地点背景 |
| 2026-05 | X 账号公开出现 | 规模 | 2026 年 5 月加入 | Recursive | 标记首个轻量级公开分发入口 |
| 2026-05-14 | 反向发布报道强调无产品且员工少于 30 人 | 反向 | 没有已发布产品;员工少于 30 人 | TNW | 凸显成熟度与估值之间的张力 |
| 2026 年中目标 | 披露公开发布目标 | 产品 | 目标为 2026 年中 | Recursive / 发布报道 | 显示公司有意从隐身叙事走向公开产品化 |
| 2026-07 | 首篇自动化 AI 研究结果文章和 GitHub 产物发布 | 产品 | 基准测试结果和开源产物已发布 | Recursive | 增加可检查技术证据,提高尽调质量 |
| 2026-07-20 | 截至运行日期,本章来源未浮现公开监管里程碑 | 监管 | 未发现已披露的备案、批准或执法事件 | 已审阅的公开来源 | 当前监管故事是披露缺席,而不是活跃事件 |
行使用里程碑公开日期,不一定是内部成交或完成日期。
[CO004, CO005, CO007, CO016, CO018, CO025]公开时间线显示,Recursive 从 2025 年创立到 $4.65B Series A,再到首批已发布技术产物,推进速度很快。
[CO004, CO005, CO006, CO007, CO015, CO016]1.4 概览层面的风险与后续章节仍需证明的事项
本章的主要反向结论并不是 Recursive 缺少人才或雄心,而是价格、成熟度和披露严重错位。TNW 和 Foundra 在这一点上说得很明确: 公司只有几个月大、员工不到 30 人、没有发布产品、没有披露收入,却拿到了 $4.65 billion 估值。这并不代表该轮融资不理性—— 精英前沿 AI 融资已反复奖励团队质量和战略位置——但这意味着后续尽调必须证明的东西远超本章所能证明的范围。开放问题直白且重要: 准确员工数、董事会构成、股权结构表条款、客户或合作伙伴集中度、商业化路线图,以及技术里程碑能否转化为真实产品,而不只是更强的招聘 和融资叙事。甚至公司别名和创始人数量在公开记录中也不完全干净。因此,恰当的概览判断是:强创始人—市场匹配和异常强的资本触达, 由较薄的治理可见度和几乎没有经审计的运营证明所抵消。[CO009, CO010, CO014, CO015, CO029, CO030]
1.5 展示材料
02市场分析
2.1 市场边界窄于 AI 宏观叙事,但宽于单一研究工具
Recursive 并不争夺整个 AI 经济。它自己的材料把公司定位在自动化 AI 研究本身,这意味着相关市场边界位于前沿模型开发工具、 智能体研究自动化,以及帮助团队锚定事实、评估并运营 AI 工作的企业或平台系统的交叉处。这个范围宽于小众基准工具,但窄于外界有时 贴在完整 AI 机会上的数万亿美元口径。因此,纳入的支出至少要覆盖三类:基础设施和模型开发平台、为 AI 工作流提供事实锚定的企业检索 与知识系统,以及开发者、研究人员或受监管知识工作者使用的高价值智能体工具。排除的支出应包括通用消费者聊天订阅、非 AI SaaS、 以及与 AI 工作负载无关的广义云支出。替代品集合已经拥挤。买家可以转向模型 API、RAG 栈、企业搜索产品,或更低成本的搜索和检索 API, 完全不必购买独立研究自动化平台。这就是为什么市场定义如此重要:Recursive 追的是一个有价值的问题,但今天还没有清晰预算线。[CM001, CM002, CM003, CM004, CM025, CM029]
| 细分市场 / 品类 | 纳入支出 | 排除支出 | 买家 / 付款方 | 对 Recursive 的意义 |
|---|---|---|---|---|
| 前沿 AI 研究自动化 | 自动化实验设计、评估和模型改进的工具 | 通用企业聊天和基础消费者使用 | 研究负责人、CTO、AI 平台预算 | 最接近 Recursive 投资逻辑的概念归属 |
| AI 基础设施和模型开发平台 | 用于 AI 系统的算力、模型服务、编排和评估栈 | 非 AI 云或商品化 IT 运维 | 超大规模云厂商、平台团队、前沿实验室 | 设定 Recursive 必须置身其中的资本和集成环境 |
| 企业 RAG 与知识系统 | 事实锚定、检索、企业搜索和知识库系统 | 没有 AI 工作流的一般文档存储 | CIO、平台、知识管理预算 | 当前企业端已货币化高价值 AI 推理工作流的代理指标 |
| 智能体开发者与编码工具 | 模型 API、编码智能体、路由、支出控制和工作流自动化 | 没有 AI 执行或路由的传统开发工具 | 工程负责人、产品工程预算 | 显示 Recursive 可能需要匹配的相邻预算线和购买标准 |
| 低成本检索替代品 | 搜索 API 和开放权重模型路线 | 没有 API 层的定制研究自动化 | 开发者、成本敏感型运营方 | 对任何缺乏差异化的检索层形成替代压力 |
边界有意把研究自动化与更广义的 AI 经济分开。Recursive 最接近第一行, 但在独立品类完全形成前,必须卖进相邻预算桶。
[CM001, CM002, CM003, CM004, CM025, CM029]Recursive 位于 AI 宏观总市场之下的好几层:其可能的变现路径始于研究自动化和高阶企业工作流,而不是所有 AI 支出。
[CM002, CM005, CM006, CM008, CM009, CM017]Recursive 可能必须先从研究证明走到预算匹配,才有资格拿到稳定的软件预算科目。
[CM001, CM015, CM020, CM021, CM033, CM036]2.2 自上而下的支出预测巨大,但自下而上的买家相关性紧得多
自上而下的数字无疑很大。Gartner 2026 年 5 月预测,全球 AI 支出当年约为 $2.596 trillion,其中 AI 基础设施单独超过 $1.43 trillion,AI 软件约 $453 billion。Gartner 1 月发布的数字已经高达 $2.52 trillion,因此即便官方预测在同一年内也继续上修。 IDC 给出了更接近运营买家思维的用例视角。在 IDC 指南中,AI Infrastructure Provisioning 是最大单一用例,2024 年达到 $30.3 billion,预计 2028 年达到 $47 billion;客服和欺诈分析工作流也已经各自占据数百亿美元中高段支出。这些数字重要, 因为它们同时说明两点。第一,AI 宏观市场真实存在且仍在扩张。第二,对 Recursive 这样的产品前公司来说,立即相关的部分小得多: 由研究机构、超大规模云厂商平台团队,以及愿意为昂贵知识工作自动化付费的成熟企业掌控的预算。因此,Recursive 可信的可服务市场(SAM) 足够大,值得关注,但仍远低于 AI 宏观总量头条。[CM005, CM006, CM007, CM008, CM009, CM017]
| 发布方 / 视角 | 年份 | 地理范围 / 范围 | 数值 | 方法 | 置信度 | 限制 |
|---|---|---|---|---|---|---|
| Gartner AI 总支出 | 2026 | 全球 | $2.596T | 覆盖多个 AI 细分市场的自上而下市场预测 | 中 | 远宽于 Recursive 眼下可触达市场 |
| Gartner AI 基础设施 | 2026 | 全球 | $1.432T | AI 总支出中的细分市场 | 中 | 主要是供应商和超大规模云厂商支出,不是初创公司可直接触达的收入池 |
| Gartner AI 软件 | 2026 | 全球 | $453.2B | AI 总支出中的软件细分市场 | 中 | 包含研究自动化之外的广泛 AI 软件品类 |
| Gartner AI 模型 | 2026 | 全球 | $32.6B | AI 总支出中的模型细分市场 | 中 | 相比一家产品前实验室,更适合模型提供商 |
| IDC AI 基础设施供给 | 2024 to 2028 | 全球 | $30.3B(2024);到 2028 年 $47B | 用例支出指南 | 中 | 该视角只是一个用例,不是整个市场 |
| IDC AI 赋能客户服务 / 自助服务 | 2024 | 全球 | $16.7B | 用例支出指南 | 中 | 强势 AI 预算品类,但不是 Recursive 的核心滩头阵地 |
| McKinsey 企业技术预算 | 2026 年背景 | 全球样本公司 | AI 最高吃掉三分之一变革预算 | 企业预算分配调查与分析 | 中 | 预算占比,不是市场规模 |
| Recursive 近期 SAM | 2026 年背景 | 前沿实验室 + 高阶企业团队 | 远小于 AI 总支出 | 基于买家匹配度和产品成熟度、受证据约束的推断 | 低 | 没有公开收入或产品数据,无法建模更窄区间 |
这张表保留了彼此不兼容但对决策有用的视角:宏观支出、基础设施 / 用例支出和预算占比信号。 不应把它们压成一个看似精确的 TAM 数字。
[CM005, CM006, CM007, CM008, CM009, CM010]不同视角会给出非常不同的数字:从巨大的宏观预测,到小得多的用例池和预算占比信号。
[CM005, CM006, CM007, CM008, CM017]2.3 买家地图分裂在前沿构建者、企业平台所有者和预算敏感型开发者之间
这个市场的买家、用户和付款方并不相同。在前沿实验室和超大规模云厂商中,经济买家很可能是掌控算力、模型和工具预算的研究或平台高管; 用户是研究人员和工程师;付款方则是更大的基础设施或 AI 组织。在企业里,检索和智能体系统常由 CIO、平台或知识管理预算购买, 但由产品团队、分析师和开发者使用。云和模型既有厂商进一步强化这种结构。OpenAI、Anthropic、AWS、Google Cloud 和 Azure 都把模型访问、搜索、事实锚定和智能体控制的某种组合卖进集中预算,而不是纯个人刷卡工作流。这让采用路径分叉。一条路是与治理、 隐私或集成要求相连的自上而下平台采购;另一条路是开发者自下而上试验,等工作流证明 ROI 后再把预算往上拉。Recursive 若要商业化, 很可能两者都需要:在精英构建者那里有研究可信度,也要有一套包装叙事,让企业或平台所有者明白产品为何属于现有预算,而不是一个理想化 的科学项目。[CM023, CM024, CM025, CM026, CM027, CM029]
| 细分市场 | 买家 | 用户 | 付款方 / 预算所有者 | 工作流 | 采用触发条件 |
|---|---|---|---|---|---|
| 前沿 AI 实验室 | 研究副总裁或 CEO | 研究员和研究工程师 | 中央 AI 研发预算 | 自动化实验循环和模型改进 | 明确的基准提升或算力杠杆 |
| 超大规模云厂商 AI 平台团队 | 总经理或平台负责人 | 应用科学家和平台工程师 | 云 / AI 平台预算 | 把智能体、知识锚定或评估系统嵌入托管服务 | 需要让平台功能做出差异化,或压低模型成本 |
| 企业知识平台 | CIO 或知识平台负责人 | 分析师、运营人员、产品团队 | IT 或转型预算 | 锚定内部数据,自动化高价值推理任务 | 可衡量的生产力或合规收益 |
| 开发者工作流买方 | 工程经理或 CTO | 开发者 | 工程工具预算 | 把路由、代码辅助和研究自动化接入软件工作流 | 迭代更快或 token 成本更低 |
| 受监管知识团队 | 业务单元负责人加 IT | 法务、财务、研究或合规用户 | 职能预算加 IT 支持 | 在治理约束下,自动化高成本、文档密集型推理 | 可审计性、引用或领域专属性 |
买方、用户和付款方常常不是同一批人。终端用户市场成形之前,Recursive 很可能先从技术能力强的买方切入。
[CM023, CM024, CM025, CM026, CM029, CM032]买方复杂度是这个市场的核心特征:不同细分市场购买理由不同,但都关心 ROI、控制和集成。
[CM023, CM025, CM029, CM032, CM036, CM037]2.4 需求驱动真实存在,但市场也在惩罚浪费和无差异支出
最强的结构性驱动在于,企业和平台显然想要更多智能体自动化、事实锚定检索和 AI 驱动的工作流加速。Deloitte 称,员工 AI 访问率在 2025 年跃升 50%,且至少 40% 项目已进生产的公司数量将在六个月内翻倍。与此同时,McKinsey 显示 AI 正挤入变革预算并增加运行成本, CNBC 则记录了明确的支出挤压:账单跑赢 ROI 时,买家会削减或改道使用量。这些事实让市场有吸引力,但也不宽容。买家会为能减少人力 或释放差异化产出的 AI 买单;对无纪律的 token 支出、简单任务过度使用前沿模型,以及治理薄弱的智能体部署,他们越来越怀疑。 这就是 Recursive 的核心市场判断。顺风异常强:基础设施投资、大型企业兴趣、快速演进的智能体平台。约束同样真实:算力强度、 治理成熟度不足、模型路由压力,以及既有厂商捆绑相邻能力。Recursive 进入的是一个需求真实存在的市场,但清晰经济包装会比宏大 理论承诺更快拿到回报。[CM010, CM011, CM012, CM013, CM014, CM015]
| 驱动因素 / 约束 | 方向 | 时间 | 含义 | 尽调问题 |
|---|---|---|---|---|
| 员工 AI 使用权限提高 50% | 正向 | 近期 | 扩大 AI 原生工作流可触达的用户基础 | 衡量使用权限能否转化为付费、持久使用 |
| 生产环境项目预计翻倍 | 正向 | 近期 | 表明买方正在越过试点阶段 | 跟踪 Recursive 目标细分市场中的生产环境占比 |
| AI 占变革预算最高三分之一 | 正负混合 | 当前 | 预算存在,但挤出风险上升 | 确认 Recursive 产品会替代或补充哪些预算 |
| 支出收紧与 ROI 审查 | 负向 | 当前 | 缺乏差异化的模型或 token 支出承压 | 证明价值可衡量,并高于更便宜的替代品 |
| 自主智能体治理成熟的组织仅占五分之一 | 负向 | 当前至中期 | 控制顾虑可能拖慢智能体部署 | 明确监督、可审计性和安全失效设计 |
| 模型路由与开放权重替代品 | 负向 | 当前 | 通用前沿模型使用场景的定价权被压缩 | 说明 Recursive 为什么能抓住独特价值,而不是只做薄封装 |
| 既有平台捆绑 | 负向 | 当前 | 云和模型平台能吸收相邻能力 | 说明哪些能力无法被捆绑带走 |
| 基础设施与智能体平台建设 | 正向 | 中期 | 部署更容易、预算池更大,有利于新品类出现 | 观察 Recursive 能否借力既有平台,而不是正面硬打 |
市场有吸引力,因为需求真实且广泛;但预算更讲纪律后,变现门槛也在抬高。
[CM010, CM012, CM013, CM015, CM020, CM021]03竞争格局
3.1 竞争格局:直接实验室、既有平台与替代品
Recursive 并不处在一个干净的单一品类市场。它的官方材料把公司定位在改进 AI 的 AI 和自动化知识发现上,听起来像前沿研究实验室命题; 但它最终必须进入的实际预算池,已经由解决相邻问题的公司共同占据。直接前沿竞争来自那些用 AI 加速模型开发、推理和知识工作、 同时已经出货可变现产品的实验室和平台。OpenAI 与 Anthropic 把前沿能力打包进企业计划,附带安全控制、编程智能体和用量分析。 Google、AWS 和 Microsoft 各自把 AI 搜索、智能体和模型平台推入既有云或生产力关系。Glean 与 Perplexity 从检索和企业搜索侧 攻击知识工作问题,而不是从实验室侧切入。Together AI 和 DeepSeek 压低基础设施与模型成本;Llama 4 等开放权重替代方案降低了 买家为无差异模型访问付费的意愿。与此同时,Ricursive Intelligence 和 Sakana AI 等其他前沿实验室说明,投资者也在资助围绕 自我改进系统、芯片共同演化和研究自动化的平行命题。因此,Recursive 进入的是一个可信替代方案已经横跨研究、工具、企业工作流自动化 和内部自建的市场。[CP001, CP002, CP003, CP004, CP005, CP006]
| 竞争对手 | 类别 | 目前销售内容 | 分发证据 | 对 Recursive 的影响 | 从 Recursive 视角看主要局限 |
|---|---|---|---|---|---|
| OpenAI | 前沿模型与企业平台 | Business 和 Enterprise 套餐、前沿模型、Codex 智能体 | 企业工作区、代码工作流、广泛合作伙伴覆盖 | 设定企业级 AI 与智能体的标杆 | 若 Recursive 能证明价值,OpenAI 的通用定位会给专门化研究自动化留下空间 |
| Anthropic | 前沿模型与代码平台 | Claude 模型、Claude Code、Team 和 Enterprise 套餐 | 技术用户采用度高,管理控制也强 | 争夺重视安全的技术买方和代码智能体预算 | 消费端分发弱于 OpenAI 或 Google |
| Google Cloud / Gemini | 既有平台 | Agent Search 和 Gemini 企业工具 | 既有云与 Workspace 采购 | 能把搜索、模型和企业上下文捆绑销售 | 产品面很宽,有时显得分散 |
| AWS Bedrock | 既有模型平台 | 托管接入领先 AI 实验室的模型和智能体 | AWS 存量客户基础 | 买方可组合模型选择,不必押注单一实验室 | 更像平台层,而不是有差异化的应用体验 |
| Azure AI Search | 既有检索平台 | 企业搜索与 RAG 基础设施 | Microsoft 企业客户关系 | 争夺 Recursive 可能切入的企业知识工作流 | 仅有搜索层不等于自主研究 |
| Glean | 企业工作 AI 平台 | 企业上下文、智能体、连接器、权限 | 单租户云、感知权限的部署 | 已经把自动化卖进知识工作预算 | 前沿研究定位弱于 Recursive |
| Perplexity Enterprise 企业版 | AI 答案引擎 / 搜索替代品 | 企业搜索与答案工作流 | 引用优先的答案体验 | 争夺研究型用户行为和知识发现入口 | 平台与基础设施深度较窄 |
| Ricursive Intelligence | 相邻前沿 AI 实验室 | AI 与芯片协同进化平台逻辑 | 高关注度 Series A 与创始人履历 | 表明投资人已经在投递归改进的变体 | 即期领域不同:芯片设计,而不是自动化 AI 研究 |
各行强调当前商业现实,而不是愿景层面的相似性。Recursive 仍处于产品前阶段,因此比较重点放在相邻玩家已经卖进重叠技术预算或企业预算的内容。
[CP003, CP004, CP006, CP012, CP013, CP015]这张序数图用两个有证据支撑的轴,摆放 Recursive 及相邻竞争者:商业化准备度和投资逻辑专门化程度。
轴分数是作者根据公开产品界面和定位语言给出的序数估计,不是第三方数字基准。
[CP001, CP003, CP006, CP008, CP013, CP017]3.2 直接前沿与研究实验室同类
最接近的概念同类不是传统 SaaS 竞争对手,而是 AI 实验室和研究密集型平台。它们能可信地宣称比一家产品前初创公司更快提升模型能力、 开发者生产率或基础设施经济性。OpenAI Enterprise 明确营销前沿模型和智能体,并把 Codex 加进去,帮助团队把目标变成完成的工作。 Anthropic 在 Team 和 Enterprise 计划中销售 Claude Code,并附带细粒度支出上限和分析。Ricursive Intelligence 尤其相关, 因为它采用了相似的递归框架:闭合 AI 与为其供能的芯片之间的反馈环,并在上线不到两个月后达到 $4 billion 估值。Sakana AI 把自己定位为围绕自然启发智能构建的前沿实验室,并在日本拥有地域差异化。Together AI 并不被框定为纯实验室,但其面向研究优化的云、 99% 可用性 SLA、预训练和推理栈,会争夺 Recursive 未来也需要的同一批技术买家注意力。直接同类给出的关键教训是,大多数相邻玩家 已经拿出了 API、平台、托管云或企业界面。Recursive 还没有。在做到之前,它的竞争主要是人才、投资人注意力和未来买家心智, 而不是当前钱包份额。[CP012, CP013, CP014, CP015, CP016, CP017]
| 公司 | 研究自动化逻辑 | 企业工作流控制 | 定价可见性 | 分发面 | 竞争含义 |
|---|---|---|---|---|---|
| Recursive | 高 | 低 | None | 低 | 逻辑清晰,但没有公开商业化界面 |
| OpenAI | 中 | 高 | 高 | 高 | 树立集成式前沿产品标杆 |
| Anthropic | 中 | 高 | 高 | 中高 | 技术用户与代码智能体领域的强标杆 |
| Google Cloud / Gemini | 中低 | 高 | 中 | 高 | 既有平台替代品 |
| Glean | 低 | 高 | 中低 | 中 | 企业上下文护城河已商业化 |
| DeepSeek / 开放权重 | 低 | 低 | 高 | 中 | 价格压缩与内部自建威胁 |
序数评分概括的是已审阅公开证据,而非经审计的性能测试。Recursive 在投资逻辑新颖性上得分高,但公开工作流、定价和支持披露得分低。
[CP012, CP013, CP014, CP015, CP016, CP017]Recursive 的逻辑聚焦度异常高,但在商业化功能广度和买方工具上明显落后于同行。
单元格是分析师基于已审阅公开页面和披露给出的序数判断,不是实验室基准测量。
[CP003, CP006, CP008, CP010, CP011, CP012]3.3 企业平台与替代品压力
如果 Recursive 最终把产品商业化到企业知识工作或研究自动化,最难的竞争可能来自那些从不自称递归自我改进公司的既有厂商。 Google Agent Search、AWS Bedrock 和 Azure AI Search 已经把事实锚定搜索、模型访问和智能体基础设施卖进现有采购通道。 Glean 销售的是建立在企业语境、连接器、权限和运行时控制之上的横向 AI 平台。Perplexity Enterprise 则从以引用为中心的答案引擎 切入同一广义问题空间。这些产品重要,是因为它们把抽象研究命题转化成 CIO 今天就能购买的东西。价格纪律进一步强化替代压力。 CNBC 报道称,企业正从堆 token 转向效率,并越来越多探索模型路由,而不是默认把每个任务交给最昂贵的前沿模型。 Anthropic 和 OpenAI 都已用支出上限、分析和业务控制来回应。这对 Recursive 直接不利。一家年轻实验室没有分发界面、运营历史 或定价页,就必须证明的不只是技术有效,还要证明它胜过企业工作流里已经安装好的、捆绑式、有治理且懂预算的替代方案。[CP022, CP023, CP024, CP025, CP026, CP027]
| 公司 | 公开包装 | 定价可见性 | 管理 / 买方控制 | 这对 Recursive 意味着什么 |
|---|---|---|---|---|
| Recursive | 官网 + 研究文章 + GitHub 产物 | None | 未找到公开定价、分析或企业控制页面 | 采购清晰度上还无法竞争 |
| OpenAI | Business 和 Enterprise 工作区 | Business 价格公开;Enterprise 定制 | SSO、分析、EKM、SCIM、支持 | 团队若想找单一托管供应商,OpenAI 是难打的既有玩家 |
| Anthropic | API + Team / Enterprise + Claude Code 包装 | API 价格公开;企业高级席位和控制 | 支出上限、席位管理、使用分析 | 技术与代码买方的强标杆 |
| AWS Bedrock | 托管模型平台 | 云式定价与多模型访问 | 企业云治理 | 让买方推迟选择单一模型供应商 |
| Azure AI Search | 搜索 / 检索平台 | 云平台定价模型 | 既有 Microsoft 治理足迹 | 争夺基于企业知识的工作流 |
| Together AI | AI 原生云与预留推理 | 基于 token 和预留吞吐量的定位 | 强调 99% 正常运行时间 SLA | 在应用层护城河成形前,先拿下重视基础设施的买方 |
本表比较的是包装与采购清晰度,而不是绝对功能质量。Recursive 目前是该组中商业选项最不清晰的一家,因为公开记录仍以研究产物为中心。
[CP003, CP006, CP007, CP008, CP009, CP011]简要判断:当前公开证据在多大程度上支撑 Recursive 已经具备持久竞争位置。
[CP003, CP006, CP008, CP010, CP011, CP012]3.4 切换成本、护城河耐久度与反向证据
Recursive 的竞争上行在于,少有竞争对手围绕公司精确命题组织起来:用自动化 AI 研究来改进 AI 本身,再把这套打法扩展到更广泛科学。 下行在于,它可能变现进入的大多数相邻市场,在模型层技术锁定效应弱,在工作流层既有厂商力量强。CNBC 对模型路由的报道很直接: 即便更便宜的模型可以完成任务,约 95% 的企业 AI 使用仍停留在昂贵前沿模型上,这意味着路由成熟后价格压缩应会加剧。DeepSeek 以及 Llama 4 等开放权重家族进一步压低通用访问价值。一旦新工作流证明有价值,OpenAI、Anthropic 和云平台也能快速吸收功能。 关于 Recursive 自身,最不利的公开证据仍与公司概览一致:TNW 描述的是一家没有公开产品、披露很薄、却坐在 $4.65 billion 估值上的公司。 从竞争角度看,这意味着 Recursive 当前叙事护城河强、商业护城河弱。要补上这个缺口,它需要明显更强的研究产出、专有数据和工作流嵌入, 或者能把研究引擎转成可重复购买动作的分发伙伴。[CP032, CP033, CP034, CP035, CP036, CP037]
| 威胁向量 | 证据 | 当前强度 | 为什么对 Recursive 不利 | 哪些因素能抵消 |
|---|---|---|---|---|
| 模型路由与预算控制 | CNBC 称企业正从 tokenmaxxing 转向效率和路由 | 高 | 降低买方为无差异模型使用支付溢价的意愿 | 展示 Recursive 能实质改善结果的工作流,而不只是提升模型质量 |
| 开放权重 / 低成本替代品 | Llama 4 和 DeepSeek 让低成本模型访问更可行 | 高 | 让通用前沿能力的复制成本更低 | 专有研究闭环、数据或工作流嵌入 |
| 既有供应商采购捆绑 | Google、AWS 和 Azure 把 AI 卖进既有合同 | 高 | 买方无需采用新创业公司,也能解决相邻问题 | 找到一个楔子:既有玩家没有交付,或无法可信地优先投入 |
| 企业上下文平台 | Glean 已经销售连接器、权限和企业上下文 | 中高 | 在 Recursive 进入前,先拿下知识工作自动化预算 | 证明 Recursive 能产出更高价值的自主研究结果 |
| 产品前阶段可信度缺口 | TNW 强调其没有产品、没有收入 | 高 | 如果产品证明滞后,叙事护城河可能消失 | 公开产品发布、客户证明和反复赢得基准测试 |
威胁强度是基于已引用公开证据的分析师判断,不是公司披露的排名。
[CP022, CP023, CP028, CP029, CP032, CP034]3.5 展示材料
04财务情况
4.1 收入模型与当前变现状态
Recursive 目前没有像一家软件公司通常那样披露公开收入模型。官网呈现的是研究使命,2026 年 7 月技术文章呈现的是基准证据, 但二者都没有发布定价表、API 注册入口、客户分层结构或支持的购买动作列表。Tech.eu、TNW、OfficeChai、Tech Funding News 和 Foundra 的发布报道都强化了同一种经济解读:市场投资这家公司,是看团队质量、命题和未来潜力,而不是看得到的当前现金生成。 这个区别重要,因为它意味着公开财务分析必须把公司未来可能变现的东西,与公司今天正在变现的东西分开。最可能的未来收入流是 与自动化研究工作流、模型开发工具或专业科学自动化绑定的企业软件或平台访问。但这些是推断出来的未来路径,不是已披露的当前业务线。 相比之下,OpenAI 和 Anthropic 已经通过公开商业定价、企业控制、用量分析和支出控制,提供了达到变现级包装的具体参照。 Recursive 尚未公开展示同等内容。财务后果很简单:现在还没有公开方法检验价格实现、买家支付意愿或需求可重复性。[CI001, CI002, CI003, CI007, CI008, CI009]
| 潜在收入来源 | 当前公开证据 | 当前状态 | 重要性 | 尽调问题 |
|---|---|---|---|---|
| 自动化 AI 研究平台访问 | 公司逻辑和技术文章暗示,这可能是核心产品路径 | 仅为推断的未来收入来源 | 与公司使命和差异化最一致 | 索取路线图,说明包装、买方和定价计划 |
| 企业工作流 / 知识自动化 | Glean、Azure、Google 等已给出相邻市场证据 | Recursive 没有公开产品 | 如果 Recursive 扩展到内部 AI 研究之外,预算池可能很大 | 澄清企业知识工作是否真的是进入市场目标 |
| 开发者或 API 访问 | 未找到公开 API、文档或价目表 | 尚未公开发布 | 这会是最清晰的经常性软件变现路径 | 索取 API 路线图、控制平面和用量 / 定价模型 |
| 科学或实验室合作 | 使命表述显示,未来可能扩展到科学发现 | 未公开披露 | 在大规模软件包装前,可能先带来定制化高价值合同 | 索取试点、研究伙伴和商业化条款清单 |
| 当前收入 | 已审阅来源未披露公开收入 | 未知 / 可能不重大或未披露 | 决定估值由业务基本面支撑,还是纯期权价值支撑 | 按收入来源提供 ARR、已确认收入和客户集中度 |
本表把可行的未来变现路径与当前变现的公开证明分开。只有最后一行是在描述当前公开记录。
[CI001, CI002, CI003, CI007, CI023, CI024]| 公司 / 界面 | 公开价格或定价姿态 | 公开可见的买方控制 | 对 Recursive 的含义 |
|---|---|---|---|
| Recursive | 未找到公开价格页面 | 未找到公开支出控制或管理分析 | 外部还无法检验变现能力 |
| OpenAI Business | 年付每用户每月 20 USD;企业定制 | 使用分析、预算、支出控制、SSO | 表明前沿能力可以多快被包装成采购就绪的软件 |
| Anthropic | 公开 API 定价,加企业高级控制 | 支出上限、席位管理、使用分析 | 为技术买方支付意愿设定标杆 |
| 云平台(AWS / Azure / Google) | 云平台式或企业定制定价 | 既有治理和采购控制 | 捆绑替代品可能压低新独立供应商的空间 |
| 精打细算的 2026 年买方 | 越来越要求效率和路由 | 支出控制和分析已在走向基本门槛 | 相比 2023 年那批 AI 卖方,Recursive 很可能面对更讲纪律的采购环境 |
同行定价不是 Recursive 价格的代理;它只是 2026 年企业级变现包装应有形态的标杆。
[CI008, CI009, CI010, CI011, CI012, CI013]Recursive 的公开记录仍是从研究论点走向未来变现,而不是从当前产品走向已确认收入。
[CI001, CI002, CI003, CI008, CI014, CI015]4.2 成本结构与单位经济含义
虽然 Recursive 没有披露烧钱速度或毛利率,现有证据指向的是算力密集型成本结构,而不是轻 SaaS 画像。公司自己的技术材料突出三条 依赖昂贵前沿硬件和专业工程的基准轨道。nanoGPT speedrun 结果在 8x H100 上测得。nanochat autoresearch 结果在单个 Modal B200 GPU 上跨 10 个随机种子运行。已发布的 SOL-ExecBench 样本显示,在 B200 上测量了 235 个 GPU-kernel 实现中的 10 个。这些事实不能让外部人 算出烧钱速度,但确实说明公司工作在昂贵基础设施边缘,而不是做商品化 CPU 受限实验。Series A 轮募资用途也指向同一方向:多个公开来源称, 融资将帮助锁定大规模算力基础设施,并支持首个 Level 1 自主训练系统。此外,市场环境并不宽容。McKinsey 称 AI 正消耗变革预算, CNBC 报道称买家正从堆 token 转向效率、模型路由和支出控制。这意味着即便 Recursive 以后把软件卖进企业 AI 预算, 也可能在尚未形成规模或采购杠杆前,就遇到成本敏感买家。因此,公开单位经济分析只能给出方向性结论:算力、研究人力和基础设施访问 很可能主导成本底座,而变现仍未定义。[CI004, CI005, CI012, CI013, CI014, CI015]
| 成本驱动因素 | 公开证据 | 可能重要性 | 仍缺什么 | 财务含义 |
|---|---|---|---|---|
| 前沿 GPU 算力 | 8xH100 nanoGPT 竞速结果和基于 B200 的基准测试工作 | 极高 | 签约算力规模、预留容量、综合成本 | 产品收入起来之前,烧钱速度很可能主要由它拉动 |
| 研究工程人力 | 公开团队规模约 25 至 30 人,加上顶尖创始人背景 | 高 | 薪酬结构、招聘计划、研究与产品人员占比 | 人才成本可能集中在资深人员上 |
| 基准实验 | 10 个 seed 的 autoresearch 评估和 235 个 kernel 项目意味着反复实验 | 高 | 实验节奏和失败运行开销 | 烧钱速度可能随迭代频率放大,而不只是随员工数放大 |
| 商业支持 / GTM | 未披露公开销售或支持组织 | Unknown | 销售计划、支持承诺、客户成功模式 | 上线前可能保持低位,之后快速上升 |
| 数据 / 基础设施工具 | 已发布资产依赖上游仓库和专用工具 | 中 | 供应商合同、托管组合、软件许可 | 开源杠杆可能有帮助,但消不掉算力成本 |
本表刻意不编造美元估算,而是基于已披露工作负载画像,对重要性作方向性排序。
[CI014, CI015, CI016, CI017, CI018, CI025]成本底座先呈现为技术成本,而不是商业成本。
[CI006, CI014, CI015, CI016, CI017, CI018]公开市场为同业提供的价格锚点,比 Recursive 自身更清晰。
[CI008, CI010, CI012, CI013, CI014, CI015]4.3 资本充足性与融资依赖
Recursive 的融资头条足够亮眼,容易遮住仍然缺失的大量承保输入。对一家成立于 2025 年、在公开报道中约 25 到 30 人、产品前的 AI 实验室来说,以 $4.65 billion 估值拿到超过 $650 million Series A 轮资本,是很罕见的结果。这轮融资几乎肯定足以让公司积极招聘、 购买算力,并在没有立即融资压力的情况下继续运营。但「现在大概率够用」不等于「足以支撑战略」。公开来源没有披露当前现金余额、 月度烧钱速度、计划招聘节奏、预留算力承诺、债务义务或优先股堆叠细节。它们也没有说明公司预期 Series A 轮是支撑到商业发布、 支撑到技术里程碑,还是仅仅支撑到更强的下一轮。公司自己的叙事暗示,这次融资是为一个雄心勃勃的研究步骤买单,而不是常规商业化扩张。 这会增加融资依赖,因为研究里程碑可能在收入追上来之前很久就消耗资本。没有公开债务或老股交易披露,本身并不令人安心; 它只说明公开记录不完整。因此,尽调中的财务问题不是 Recursive 今天有没有资本——它显然有——而是相对于命题背后的隐藏算力需求 和商业化路线图,它的资本是否足够。[CI004, CI005, CI006, CI019, CI020, CI021]
| 指标 | 公开支持的数值 / 状态 | 它说明什么 | 它不能说明什么 |
|---|---|---|---|
| Series A 融资额 | 超过 650M USD | 短期资本获取能力强 | 剩余现金、烧钱速度或里程碑覆盖度 |
| 估值 | 4.65B USD | 投资者给了极高的战略期权价值 | 财务基本面是否撑得住这个价格 |
| 成立年份 | 2025 | 相对融资轮规模,公司仍很早期 | 上线前运营体系搭建到什么程度 |
| 团队规模信号 | 上线报道中为 25+ 且少于 30 人 | 相对融资额,团队很精简 | 真实薪资支出、外包负荷和本轮后招聘节奏 |
| 算力雄心 | 公开提到 Level 1 自主训练系统和更大算力集群 | 资金很可能投向昂贵的技术扩展 | 确切算力承诺和资金可支撑多久 |
资本充足性按已披露的雄心评估,而不只是看融资额本身。
[CI004, CI005, CI006, CI021, CI022, CI029]| 缺失数据 | 为什么重要 | 严重性 | 尽调路径 |
|---|---|---|---|
| 收入 / ARR / 确认收入 | 需要判断估值反映的是业务牵引,还是只有投资逻辑 | 关键 | 按收入流和期间索取收入桥接表 |
| 烧钱速度和现金跑道 | 需要理解下一轮融资依赖的时间点 | 关键 | 索取月度烧钱速度、现金余额和现金跑道模型 |
| 算力合同和预留容量义务 | 很可能是最大的隐藏成本项 | 关键 | 索取 GPU / 云合同、最低采购承诺和供应商组合 |
| 毛利率 / 服务成本路径 | 需要评估其经济模型更像软件公司还是研究实验室 | 重要 | 索取试点 P&L 或建模后的单位经济 |
| 股权结构、优先权和治理经济条款 | 会影响投资者回报,不止名义估值 | 重要 | 索取融资文件、董事会权利和优先权摘要 |
这些缺口是把本章从描述性分析推进到可承销判断所需的最低私有数据集。
[CI019, CI020, CI030, CI036, CI037, CI040]能看见融资能力,看不见现金流基本面。
[CI002, CI004, CI006, CI019, CI020, CI022]4.4 财务结论与关键尽调阻滞项
公开记录支持一个狭窄但重要的财务结论。Recursive 在其成熟度阶段展现出异常强的资本触达和高质量投资人财团。它也在 2026 年 7 月 展示了足够的技术严肃性,走出了纯隐身叙事。它尚未公开证明的是收入质量、变现纪律或软件式经济性。没有公开年经常性收入(ARR)、没有公开合同证据、 没有披露毛利率、没有发布定价,也没有销售效率代理指标。因此,本章不能支持经典的「强单位经济」或「高效增长」判断。最多只能支持 研究实验室资本形成判断。这对后续估值章节很重要:在 2026 年前沿 AI 市场里,$4.65 billion 标记或许有战略可理解性,但没有收入 和成本披露,就无法用常规财务基础证明。决定性尽调阻滞项很直白:实际烧钱速度、算力承诺、股权结构表条款、招聘计划、首个产品路线图, 以及证明未来任何产品都能在买家已经要求支出控制和路由效率的市场里捕获价值。在这些信息出现前,财务姿态应描述为资金充足, 但尚未获得财务承保。[CI001, CI002, CI009, CI012, CI021, CI022]
4.5 展示材料
05产品与技术
5.1 用工作流定义产品
Recursive 目前应被理解为一家研究系统公司,而不是应用公司。公司的公开材料没有描述打磨成熟的终端用户应用、开发者 API 或商业工作流套件。相反,它们描述的是一个自动化 AI 研究若干环节的系统。官方命题很明确:构建能递归改进 AI 的 AI, 先推进 AI 科学,再把这套打法用于更广泛的科学发现。2026 年 7 月文章和 GitHub 仓库让这个命题更具体,暴露了三组资产包: nanoGPT speedrun 解法、nanochat autoresearch 测试框架,以及 SOL-ExecBench 工作中的 GPU-kernel 样本输出。用客户工作流语言说, 公开交付物不是「和模型聊天」,而是「用一个系统更快提出、实现、测试并优化研究想法,速度超过纯人工循环」。这个区别很重要, 因为它把核心产品框定为知识生产和系统优化引擎。弱点也同样清楚:工作流引擎还不是商业包装。公司仍缺少面向外部买家的公开引导流程、 定价、权限、支持承诺或明确产品边界。[CE001, CE002, CE003, CE004, CE005, CE006]
| 资产或模块 | 公开证据 | 当前角色 | 买方 / 用户实际会接触什么 | 当前成熟度 |
|---|---|---|---|---|
| 官网 / 论点层 | recursive.com | 使命和定位 | 叙事和高层研究论点 | 概念 / 公司展示面 |
| 技术文章 | 2026 年 7 月文章 | 解释循环和基准结果 | 文档和证据层 | 已有文档的原型 |
| 仓库根目录 | GitHub 顶层仓库 | 打包已发布资产 | 代码入口和 README | 可检查资产 |
| nanoGPT speedrun 包 | GitHub 子目录 | 训练速度优化基准 | 基准包,而不是产品 UI | 可跑基准 |
| nanochat autoresearch 包 | GitHub 子目录 | 小模型训练的自主研究框架 | 研究框架和评估脚本 | 可跑基准 |
| SOL-ExecBench 样例包 | GitHub 子目录 | kernel 优化示例 | 示例性代码样本集 | 部分发布 |
矩阵划分的是今天公开可见的内容,而不是私下可能存在的资产。这些模块目前都还不是传统意义上的定价产品界面。
[CE001, CE003, CE004, CE005, CE006, CE011]| 工作流步骤 | 公开描述的行为 | 证据 | 为什么重要 |
|---|---|---|---|
| 想法生成 | 系统提出可尝试的想法 | 文章 | 显示循环从假设形成开始,而不只是执行 |
| 实现 | 系统编写或修改代码 / 方法 | 文章 + 仓库 | 让系统成为主动研究操作者 |
| 实验执行 | 系统运行有基准的实验 | 文章 + 基准包 | 把推理连接到可衡量结果 |
| 验证 | 系统评估改动是否改善结果 | 文章 + 基准统计 | 形成选择机制,而不是纯生成 |
| 选择 / 下一步 | 系统选择下一步尝试什么 | 文章 | 把单次实验变成复利循环 |
这是 Recursive 公开描述的运行工作流中最清晰的一版。
[CE002, CE008, CE012, CE013]Recursive 的公开架构从研究论点延伸到特定基准测试的工件包。
[CE001, CE002, CE003, CE004, CE005, CE008]公开用户旅程仍主要是研究操作员的工作流,而不是传统终端用户的软件旅程。
[CE002, CE012, CE013, CE014, CE015, CE016]5.2 架构、基准与运营模型
从公开工件可以读出 Recursive 的运营模型。文章描述了一个循环:提出想法、实现想法、运行实验、验证结果,然后选择下一步尝试。 仓库把这个循环打包成面向具体基准的形态。在 nanoGPT speedrun 中,系统把 GPT-2-small 训练优化到 8x H100 上平均 77.3 秒, 同时通过目标验证阈值并击败同硬件基线。在 nanochat autoresearch 中,Recursive 发布多个解法,在单个 Modal B200 GPU 上跨 10 个 seeds 评估五分钟,其中包括从 Karpathy 基线优化出的最佳解法。在 SOL-ExecBench 中,Recursive 分享了 235 个 kernel 实现中的 10 个在 B200 上的评分,同时将多数保留为私有,以避免影响排行榜。仓库也清楚显示,Recursive 的工作是组合式的, 不是从零开始:它建立在 KellerJordan 的 modded-nanogpt、Karpathy 的 autoresearch 和 Karpathy 的 nanochat 之上,并在 Apache-2.0 打包中保留上游声明。这个公开架构意味着一种实验室工作流,混合了自主实验、基准测试框架、人工选择的基线和高端硬件。 它更接近内部研究平台,而不是公共应用栈。[CE011, CE012, CE013, CE014, CE015, CE016]
| 层 | 公开证据 | 依赖 | 技术含义 | 局限 |
|---|---|---|---|---|
| 基准框架 | nanoGPT、nanochat、SOL-ExecBench 包 | 基准生态和精选基线 | 让 Recursive 能在公认任务上证明进展 | 如果没有后续产品化,可能夸大通用性 |
| 自主实验循环 | 文章描述提出 / 实现 / 运行 / 验证 / 选择流程 | 内部编排未完全开源 | 如果能稳定复利,就是核心差异化 | 内循环实现细节仍部分不透明 |
| 前沿算力 | 8x H100 和 B200 相关表述 | NVIDIA 级 GPU 和支持 GPU 的平台 | 支撑高速实验和系统工作 | 抬高基础设施依赖和成本 |
| 上游开源基线 | modded-nanogpt、autoresearch、nanochat 开源基线 | 开源社区工作 | 加快进展,也提升可比性 | 削弱全栈独特性的主张 |
| 部分自研层 | 225 个未发布 kernel 仍属私有 | 公司持有的资产 | 可能包含今天未公开的重要技术诀窍 | 无法外部审计 |
架构由已发布资产及其上游引用推断而来。
[CE011, CE012, CE014, CE015, CE016, CE017]Recursive 的公开栈依赖上游基线、前沿 GPU 和基准测试生态。
[CE003, CE004, CE014, CE015, CE016, CE017]5.3 依赖、信任控制与已知限制
Recursive 最强的公开技术证据是技术产出;最弱的是信任和部署就绪度。公开来源显示出清晰依赖:已发布工作绑定 H100 和 B200 等前沿 GPU、 SOL-ExecBench 等基准生态,以及 nanochat、modded-nanogpt 等上游开源基线。这些依赖并非天然负面——现代研究系统就是这样搭出来的—— 但也意味着公司当前护城河并非基于端到端封闭基础栈。信任控制是公开记录薄弱之处。首页强调安全,公司法律页面显示标准网站隐私和条款, 但没有公开模型卡、企业安全页面、审计文档或部署控制文档,无法与更后期 AI 供应商发布的材料相比。公开工件包也刻意保持部分披露。 Recursive 明确保留大多数 kernel 实现,以免污染基准;这可以理解,但也意味着外部人无法完整检查技术资产底座。最后,所审查来源没有显示 公开客户部署、可用性承诺或支持模型。公司已经展示了真实工程信号,但尚未展示公开生产就绪度。[CE025, CE026, CE027, CE028, CE029, CE030]
| 维度 | 公开证据 | 当前判断 | 缺口或注意事项 |
|---|---|---|---|
| 安全姿态 | 官网措辞强调安全 | 正向意图信号 | 未发现公开模型卡或安全框架 |
| 隐私 / 法务基础 | 已有隐私政策和条款页面 | 存在标准网站治理 | 不等同于企业 AI 合规 |
| 质量证据 | 基准包带有统计结果和可复现代码 | 对研究可信度支撑强 | 仍由公司发布,范围限于基准 |
| 部署控制 | 未发现公开管理、权限或支持文档 | 公开部署就绪信号弱 | 可能私下存在,但未披露 |
| 可审计性 | 只有部分资产公开 | 好坏参半 | 大部分 kernel 清单仍是私有 |
公开记录中,信任证据明显薄于技术证据。
[CE025, CE026, CE027, CE028, CE029, CE030]Recursive 在论点清晰度和基准测试证据上最强,在公开部署就绪度上最弱。
单元格为分析师根据审阅过的公开记录给出的序数判断。
[CE001, CE003, CE004, CE011, CE025, CE029]5.4 路线图、发布节奏与成熟度判断
公开路线图仍以里程碑为中心,而不是以产品为中心。发布报道称,Series A 轮将资助更大算力基础设施和首个 Level 1 自主训练系统, 并把公开发布目标设在 2026 年中。到运行日为止,真正已经出货且最清晰可见的里程碑,是 2026 年 7 月文章和 GitHub 发布。 这很重要,因为它把公司从无法证伪的隐身叙事,升级为可检查的技术叙事。但它仍未解决最大的成熟度问题:外部用户的产品边界是什么? 当前公开材料至少支持五个成熟度结论。第一,公司有连贯命题。第二,它展示了多资产技术栈,而不是单一基准技巧。第三,它能以可复现仓库 和许可合规打包研究结果。第四,公开产出仍像研究工件,不像部署级软件。第五,未来路线图很可能既依赖改进原始研究循环,也同样依赖将引擎 运营化,并包上信任、控制和买家工作流。因此,技术章节对技术可信度持正面看法,对产品成熟度保持谨慎。[CE036, CE037, CE038, CE039, CE040]
| 里程碑 | 日期或期间 | 公开证据 | 证明什么 | 不能证明什么 |
|---|---|---|---|---|
| 退出隐身和 Series A | 2026-05-13 | 发布报道 | 投资者和公司在投资逻辑上达成一致 | 本身不能说明产品成熟度 |
| 公开发布目标 | 报道中的 2026 年中目标 | Tech Funding News / Europe Alternatives 一类报道 | 管理层有意走出隐身状态 | 直到 7 月资产发布前,目标是否达成仍不清楚 |
| 首篇技术文章和仓库 | 2026-07 | 官方文章 + GitHub | 公开技术实质和打包纪律 | 仍不是定价产品发布 |
| 基准广度 | 截至 2026 年 7 月 | 披露三条不同路线 | 系统不只是一次性结果 | 广度仍以基准为中心 |
| 下一道成熟度门槛 | 运行日期之后推断 | 基于当前证据的分析师判断 | 需要打包、控制和买方工作流定义 | 公开来源尚无法证明 |
最后一行是分析性的成熟度判断,不是公司发布的里程碑。
[CE003, CE004, CE006, CE011, CE017, CE036]5.5 展示材料
06客户情况
6.1 客户分层与可能的买家原型
Recursive 的公开材料没有列出客户,但确实暗示了一个狭窄的第一批买家画像。一家构建自动化 AI 研究和知识发现循环的公司, 不太可能从主流 SMB 用户或休闲消费者起步。最可能的第一批买家是先进研究机构、前沿模型开发团队、超大规模云厂商内部平台组, 以及已经购买昂贵 AI 工具的高技术企业知识团队。已发布工件的性质间接支持这种分层:基准测试框架、GPU kernel 工作和模型训练优化 与技术组织相关,而不是面向普通办公生产力买家。在后续阶段,公司更广泛的科学发现叙事可能把买家集合扩展到研究密集型垂直领域, 但公开记录尚未显示这件事发生。因此,客户视角是面向未来的:如果当前研究系统被包装成真实产品,谁会合乎逻辑地付费?答案是一小群 预算深、技术上能容忍不完美早期产品的成熟买家。这与研究实验室商业化路径一致,但也意味着早期集中度风险和较长证明周期。[CU001, CU002, CU003, CU004, CU005, CU006]
| 客群 | 为什么契合投资逻辑 | 今天的公开支撑 | 当前置信度 | 主要注意事项 |
|---|---|---|---|---|
| 前沿实验室 / 模型团队 | 直接重视自动化研究和系统优化 | 基准资产显示强间接契合 | 中 | 未披露具名客户 |
| 超大规模云厂商平台团队 | 可在内部使用研究自动化和 kernel 优化 | GPU 和基准重点显示中等间接契合 | 中 | 会面临买还是自建的压力 |
| 企业知识 / 研发团队 | 后续可能重视自动化知识发现 | 更广泛的科学叙事显示中低契合 | 中低 | 尚无面向它们的公开产品打包 |
| 科研机构 / 实验室 | 使命延展到更广泛的科学发现 | 官方论点显示概念契合 | 中低 | 没有公开垂直产品证据 |
| 通用企业用户 | 理论市场很大 | 当前契合度低 | 低 | 当前公开资产太技术化,打包不足 |
本表映射的是可能的买方画像,不是已确认客户。
[CU001, CU003, CU004, CU005, CU006, CU007]早期客户路径很可能从技术好奇开始,最后收窄到少数高阶试点买家。
[CU003, CU004, CU005, CU010, CU013, CU014]6.2 采用轨迹与公开证明信号
Recursive 最好的公开采用证据不是客户案例,而是围绕 2026 年 7 月技术发布的一组弱但真实的社区信号。在 GitHub 上,公开仓库的 network 页面显示 174 stars 和 15 forks,还有一条外部用户在 2026 年 6 月 11 日提出的开放 issue,讨论 autoresearch 超参数优化的先前工作。pulls 页面没有开放或关闭的 pull request,releases 和 tags 页面则显示尚无 release。 这些细节重要,因为它们揭示外部用户今天如何接触公司:把它当作研究工件生产者,而不是打包软件供应商。缺少 release 尤其重要。 这说明仓库仍被当作代码和文档消费,而不是旨在易部署的版本化软件。X 资料提供公开公司身份和分发界面,也形成另一个轻量证明信号, 但同样不是客户证明。本章观点是,Recursive 已经从研究和开发者社区获得一些早期技术关注,但公开证据没有显示生产用户、试点、合同 或可引用付费账户。[CU010, CU011, CU012, CU013, CU014, CU015]
| 信号 | 公开证据 | 日期 | 说明什么 | 为什么不足 |
|---|---|---|---|---|
| 退出隐身 / 产品缺口 | 上线报道称尚未发布产品 | 2026-05 | 上线时客户牵引很可能很少 | 没有采用指标 |
| 公开 X 账号 | 公司 X 主页处于活跃状态 | 2026-05 之后 | 公开身份和轻量漏斗顶部认知 | 不是客户转化证据 |
| 技术文章 + 仓库发布 | 官方发布已上线 | 2026-07 | 有意义的技术认知事件 | 仍不是面向客户的发布 |
| GitHub 星标和 fork | network 页面显示 174 个 star 和 15 个 fork | 2026-07-20 观察到 | 开发者 / 研究者有外部兴趣 | 不是收入或生产使用 |
| 外部 issue 活动 | 一条来自外部用户的未关闭 issue | 2026-06-11 | 有一些外部社区互动 | 单个 issue 太弱,无法推断产品市场契合 |
这些是关注度和社区信号,不是经典收入采用指标。
[CU010, CU011, CU012, CU013, CU014, CU015]| 实体或证明类型 | 公开状态 | 证据质量 | 实际能说什么 | 关键缺口 |
|---|---|---|---|---|
| 具名付费客户 | 未披露 | None | 在已审阅来源中未发现具名付费客户 | 需要直接客户访谈或合同 |
| 具名试点客户 | 未披露 | None | 未发现公开披露的具名试点或 LOI | 需要试点清单和范围 |
| GitHub 社区用户 | 部分 | 低 | 外部用户已 fork、点星并评论该仓库 | 不等同于产品客户 |
| X / 公开受众 | 部分 | 低 | 公司在 X 上有公开分发阵地 | 受众不等于买方 |
| 技术读者 / 基准测试观察者 | 部分 | 低-中 | 文章和仓库可能触达了研究导向读者 | 没有转化证据 |
本表穷尽了本章审查过的公开证明类型。
[CU013, CU014, CU015, CU016, CU017, CU018]公开漏斗从认知到真实客户证据迅速收窄。
数值是序数阶段评分,不是实测转化率。它们展示了从认知到客户证据之间的证据缺口。
[CU010, CU013, CU014, CU015, CU016, CU017]公开证据对技术兴趣的支撑强于对经济采用的支撑。
单元格是对证据质量的序数分析师判断,不是实测得分。
[CU003, CU010, CU013, CU014, CU016, CU018]6.3 留存、重复使用与集中度风险
Recursive 没有公开留存数据。所审查来源没有披露净留存率(NRR)、总留存率(GRR)、流失、续约、席位扩张或合同期限指标。这意味着常规客户质量测试 无法执行。唯一可用的重复使用线索偏活动,而不是偏收入:有仓库互动、一些外部 issue 活动,以及可见的持续 commit 历史。这些显示 持续技术维护或外部好奇心,但没有显示投资者通常想看到的重复价值捕获。如果第一批付费买家出现,集中度风险很可能很高。一个扎根于 自动化 AI 研究的产品,逻辑上会从少数前沿实验室、超大规模云厂商或精英技术团队开始,而不是从广泛横向需求开始。这会让第一批客户群 既有声望也脆弱。市场语境也不宽容。买家越来越想要支出控制、路由效率和采购就绪治理。Google、AWS、Azure、Glean、OpenAI、 Anthropic 和 Perplexity 等平台供应商已经服务相邻工作流。因此,Recursive 面临双重门槛:先证明真实客户需求超出技术兴趣, 再证明这些客户会留下来并扩张,而不是把系统当作有趣的基准工件。[CU020, CU021, CU022, CU023, CU024, CU025]
| 指标或代理指标 | 公开状态 | 解读 | 局限 |
|---|---|---|---|
| NRR / GRR / 流失 | 未披露 | 没有公开的经常性收入持续性证据 | 关键缺口 |
| 续约 / 合同期限 | 未披露 | 没有公开的合同持续性证据 | 关键缺口 |
| 仓库星标 | 观察到 174 | 兴趣或收藏的弱信号 | 不是使用强度 |
| 仓库 fork | 观察到 15 | 动手试用的弱信号 | 不是客户留存 |
| 未关闭 issue / PR / 版本发布 | 1 个未关闭 issue,0 个 PR,无版本发布 | 有一些外部互动,但产品包装成熟度低 | 不是队列或满意度指标 |
因为没有公开的真实留存数据,GitHub 活动只作为很弱的外部兴趣代理指标。
[CU020, CU021, CU022, CU023, CU024, CU034]| 风险 | 为何重要 | 当前判断 | 降低风险的条件 |
|---|---|---|---|
| 首批买方范围小 | 首批客户很可能是精英技术团队 | 高 | 展示更宽的 ICP 和销售管线深度 |
| 验证周期长 | 研究型产品买方需要时间验证工作流价值 | 高 | 提供试点转生产的转化证据 |
| 相邻既有厂商替代 | 买方已有 OpenAI、Anthropic、Glean、Google、AWS、Azure 等选项 | 高 | 证明研究结果显著更好 |
| 公开产品化薄弱 | 没有版本发布、支持承诺或客户文档 | 高 | 交付产品文档、发布版本并提供部署指南 |
| 将社区关注误当作客户牵引 | 星标和 fork 可能夸大真实采用 | 中-高 | 披露真实用户、试点和重复使用 |
这些是当前公开证据显示的主要集中度和扩张约束。
[CU025, CU026, CU027, CU028, CU029, CU030]公开材料里唯一的重复使用证据来自社区活动,不是客户经济性。
[CU012, CU016, CU017, CU018, CU019, CU020]6.4 客户结论与后续尽调仍需补齐的内容
正确的客户结论不是「没有市场」,而是「尚无公开客户证明」。Recursive 的命题与真实技术买家问题相吻合,尤其是那些试图加速模型开发、 系统优化或研究迭代的组织。公司也已在 2026 年 7 月公开做出足够动作,吸引技术熟练观察者关注。但关注不是采用,采用也不是变现。 最大缺口很直白:具名客户引用、试点或生产状态、实际部署工作流、重复使用证据,以及任何扩张经济性迹象。即便是基本包装信号——release、 版本管理、客户文档、支持承诺——在公开材料里也很薄或缺席。如果公司的第一批客户确实是少数精英技术买家,后续尽调应预期早期客户基础 集中,且证明周期偏企业式、偏长。在私人证据显示相反情况之前,Recursive 应被视为一家买家逻辑可信、但公开采用证明非常有限的公司。[CU003, CU015, CU020, CU025, CU026, CU028]
6.5 展示材料
07风险
7.1 法律、监管与披露风险
Recursive 的公开法律和披露姿态有序但很薄。网站有标准隐私和条款页面,7 月仓库发布也采用 Apache-2.0 打包,并为 MIT 许可组件 保留上游声明。这是正面信号:公司似乎重视自己选择发布代码的基础法律卫生。更深的风险在于公开记录没有披露的内容。 没有公开董事会名单、没有股权结构表摘要、没有融资条款细节、没有客户合同证据,也没有公开产品治理文档,让外部人能够在研究引擎变成可部署 产品后评估责任边界。网站安全语言停留在高层,没有运营化。已发布仓库也把大部分技术资产底座留在私有状态,从基准角度可以理解, 但限制了外部可审计性。简言之,所审查记录没有浮现明显进行中的诉讼或执法事件,但存在实质透明度风险:投资者和未来企业买家必须凭信任, 而不是凭披露证据,承保大量重要法律和治理变量。[CR001, CR002, CR003, CR004, CR005, CR006]
| 风险 | 公开证据 | 当前严重性 | 为何重要 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 治理不透明 | 已审查来源中没有公开董事会或股权结构表细节 | 高 | 投资者无法评估控制权或下行保护 | 要求提供董事会名单、融资文件和优先权摘要 |
| 产品治理不透明 | 有安全表述,但没有公开的运营治理文件 | 高 | 部署责任难以尽调 | 要求提供安全框架、模型卡和部署政策 |
| 知识产权 / 许可规范性 | Apache-2.0 打包中可见上游声明 | 中 | 是正面信号,但发布范围有限 | 审查完整来源链和内部知识产权归属 |
| 可审计性缺口 | 大多数内核库存仍未公开 | 中-高 | 外部评审无法完整检查护城河 | 要求独立技术审计 |
| 监管就绪度缺口 | 未发现公开的企业合规或控制文档 | 中-高 | 之后可能拖慢受监管客户采用 | 要求安全、隐私和合规路线图 |
本登记表聚焦公开记录中可见的法律与披露界面;本章审查的来源未发现正在进行的公开执法行动。
[CR001, CR002, CR003, CR004, CR005, CR006]Recursive 当前最实质的风险,集中在执行、封装和依赖:发生可能性高,影响也高。
可能性和影响是基于公开证据做出的序数尽调判断,不是精算概率。
[CR001, CR003, CR010, CR013, CR018, CR026]7.2 运营、质量与安全风险
Recursive 目前公开的产品形态仍是研究软件,因此运营风险画像很具体。公司已经证明自己能产出基准测试产物, 但公开资料看不到运行时间承诺、支持体系、部署流程或面向客户的控制。已发布仓库也只是部分开放: 10 个内核实现公开,225 个仍为私有,外部无法完整判断最强内部能力能否泛化或规模化。技术栈依赖 H100、B200 等前沿 GPU,也依赖基准测试驱动的工作流;成本和供应敏感度因此都被抬高。运营上, Recursive 容易受硬件到位延迟、基础设施费用上升,或基准表现转化为可用产品质量的路径慢于预期影响。 安全和质量风险也受开源与社区动态牵动。GitHub 页面显示有一定外部互动,但没有公开发布节奏、打包版本、 PR 历史,issue 线索也有限。年轻研究项目出现这种情况并不异常,但也凸显公开运营形态仍非常早期。 技术可能是真的;生产运营的公开证据仍很少。[CR010, CR011, CR012, CR013, CR014, CR015]
| 风险 | 公开证据 | 当前严重性 | 为何重要 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 基准到产品的转化风险 | 公开证明以基准为中心,而非以部署为中心 | 高 | 出色研究结果仍可能转不成可用软件 | 要求产品路线图和试点反馈 |
| 前沿 GPU 依赖 | 已发布工作明确依赖 H100 和 B200 | 高 | 供应和成本冲击会显著拖慢进展 | 要求供应商策略和算力应急计划 |
| 部分发布风险 | 235 个内核仅 10 个公开 | 中-高 | 最强能力可能无法由外部验证 | 要求抽样审计未发布资产 |
| 运营不成熟 | 没有公开版本发布、支持承诺或可用性承诺 | 高 | 外部用户无法把当前产物当作生产软件 | 要求发布流程、支持模式和 SLA |
| 安全 / 部署控制缺口 | 未发现公开的企业控制界面 | 高 | 限制严肃客户的信任 | 要求管理控制、权限模型和日志方案 |
严重性反映了技术产物质量与运营就绪度之间的公开缺口。
[CR010, CR011, CR012, CR013, CR014, CR015]若干可见风险会相互传导,而不是彼此隔离。
[CR010, CR013, CR014, CR022, CR023, CR028]7.3 合作伙伴、依赖与市场结构风险
Recursive 的依赖栈集中在几处关键节点。公开发布显示,公司依赖 nanochat、autoresearch、 modded-nanogpt 等上游开源基线,依赖 SOL-ExecBench 等基准生态,也依赖前沿 NVIDIA 级 GPU。 这些依赖对先进 AI 实验室都不罕见,但叠在一起意味着 Recursive 还没有公开证明自己拥有完全自给的技术栈。 商业环境又加了一层依赖风险。买家越来越看重支出控制和路由效率,而 OpenAI、Anthropic、Google、 AWS、Azure、Glean、Perplexity 等在位厂商已经覆盖许多相邻工作流。因此 Recursive 不只依赖供应商和硬件, 还依赖自己能否赶在在位厂商把相关功能打包进现有产品前进入市场。可能的首批客户也少且集中, 合作伙伴集中和客户集中风险会同时出现。商业化如果依赖少数实验室、单一超大规模云厂商关系, 或一小群技术设计伙伴,谈判筹码可能在客户手里,而不是 Recursive 手里。[CR020, CR021, CR022, CR023, CR024, CR025]
| 依赖 | 公开证据 | 当前严重性 | 为何重要 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 开源上游基线 | 明确沿用了 nanochat、autoresearch 和 modded-nanogpt 谱系 | 中 | 能加速工作,但削弱全栈独立性的说法 | 澄清哪些才是真正自研 |
| 基准生态 | SOL-ExecBench 和基准测试框架支撑证明 | 中 | 成功可能过拟合基准声誉,而不是买方结果 | 展示非基准产品结果 |
| NVIDIA 级硬件 | 已发布工作依赖 H100 和 B200 级别 | 高 | 造成基础设施集中度和成本敏感性 | 提供多元化算力和供应商计划 |
| 既有平台 | OpenAI、Anthropic、Google、AWS、Azure、Glean 和 Perplexity 已经服务相邻工作流 | 高 | 可通过打包吃掉部分机会 | 证明独特工作流或结果护城河 |
| 首批客户集中 | 首批买方很可能只有少数精英技术团队 | 高 | 议价杠杆与收入集中度可能很极端 | 展示销售管线广度和多细分市场需求 |
风险不在于这些依赖存在,而在于公开记录还没有显示 Recursive 如何摆脱它们。
[CR020, CR021, CR022, CR023, CR024, CR025]Recursive 的公开风险面由上游代码、基准测试、硬件和既有厂商依赖共同塑形。
[CR020, CR021, CR022, CR024, CR025, CR026]7.4 人员、执行与终止标准
最后一组风险是执行。Recursive 的公开故事建立在创始人履历、投资人信号和可信技术里程碑之上—— 但公开视角下,公司仍然小、年轻,且尚未形成产品。即便不像某些同行那样依赖单一明星创始人, 这种组合也带来关键人和节奏排序风险。执行必须在多个维度同时跑通:自动化研究循环要持续产出更强结果; 公司要把这些结果运营化为可用产品;而且必须赶在买家认定现有平台厂商已经足够好之前完成。公开董事会和治理细节缺席, 又放大了这个问题,因为外部看不到公司如何在研究纯度、产品化和商业化之间取舍。因此,正确的终止标准并不抽象。 如果公司不能在下一个合理里程碑窗口内展示公开产品形态、具名买家、部署控制,或持续技术领先, 估值风险会急剧升高。反过来,如果它能把 2026 年 7 月的技术证据与真实包装和客户证明配起来, 现有几项风险会很快收窄。现阶段,这组风险应视为高,但仍可通过私下尽调主动压低。[CR030, CR031, CR032, CR033, CR034, CR035]
| 风险 | 公开证据 | 当前严重性 | 为何重要 | 缓释措施 / 尽调问题 |
|---|---|---|---|---|
| 小团队执行风险 | 公开员工数信号仍约为 25 到 30 | 高 | 几次执行失误就可能被放大 | 要求组织架构图和招聘计划 |
| 创始人 / 领导层可见度缺口 | 公开来源更强调履历,而不是当前运营结构 | 中-高 | 难以评估管理带宽和角色清晰度 | 要求领导层名单和决策权安排 |
| 产品化顺序风险 | 有技术发布,但没有公开产品界面 | 高 | 研究进展可能跑在商业化包装前面 | 要求从仓库到产品的里程碑计划 |
| 商业化风险 | 还没有具名客户或定价 | 高 | 即便研究很强,高估值公司也可能错失市场 | 要求试点、买方反馈和产品包装路线图 |
| 治理监督缺口 | 董事会和控制权未公开 | 中-高 | 难以评估风险取舍如何管理 | 要求治理材料 |
相比估值,公司还处早期,执行风险因此被放大。
[CR030, CR031, CR032, CR033, CR034, CR035]| 风险簇 | 降低风险的条件 | 终止触发 / 投资逻辑破裂 | 当前状态 |
|---|---|---|---|
| 技术可信度 | 更多公开基准胜利,加上更广泛的资产披露 | 公开结果停滞,或无法泛化到当前展示之外 | 2026 年 7 月发布已部分降低风险 |
| 产品成熟度 | 公开产品界面、控制能力、版本发布和部署文档 | 下一个主要里程碑窗口后仍没有产品包装或控制能力 | 风险尚未降低 |
| 客户证明 | 具名试点、部署和重复使用 | 产品包装后仍没有具名客户证明 | 风险尚未降低 |
| 资本充足性 | 算力计划、烧钱可见度和融资条款透明度 | 公开产品牵引出现前就需要新资本 | 风险尚未降低 |
| 竞争护城河 | 自研工作流优势或嵌入式数据引力证据 | Recursive 找到切入点前,既有厂商复制了功能集 | 风险尚未降低 |
终止标准是尽调分析阈值,不是公司发布的里程碑。
[CR036, CR037, CR038, CR039, CR040, CR041]7.5 图表
08估值
8.1 建议与当前价格纪律
正确起点不是 Recursive 是否令人印象深刻,而是公开证据是否支持今天支付 $4.65 billion。质量上, 公司确实有真亮点。多项公开来源指向同一融资事实:超过 $650 million 的 Series A 资本、 $4.65 billion 估值,以及包括 GV、Greycroft 和 NVIDIA 在内的投资财团。2026 年 7 月的文章和仓库发布, 也让故事比单纯隐身融资更具体。但这些事实都没有补上核心估值缺口。公开来源仍未披露收入、定价、客户、 毛利率、留存或优先股堆叠。官方呈现仍是研究使命和技术证明集,而不是可进入采购的产品。 这意味着新投资人不是在 $4.65 billion 买入一家已验证的软件公司;他们是在押注一家资金异常充足的研究实验室未来商业化的期权。 价格合适时,这仍可能有吸引力,但这不同于买入一家已经变现的 AI 平台。因此,有纪律的建议是继续研究 / 观察, 而不是买入;纪律来自价格敏感性,不是因为不相信技术野心。[CV001, CV002, CV003, CV004, CV005, CV006]
| 决策字段 | 当前看法 | 决策含义 |
|---|---|---|
| 建议 | 继续研究 / 观察 | 保持跟踪,但不要把 2026 年 5 月价格视为已由公开证据支撑。 |
| 信心 | 中 | 公开记录对融资很强,对变现、治理和股权结构条款很弱。 |
| 风险评级 | 高 | 商业化、融资结构和平台打包风险都可能迅速压缩股权故事。 |
| 估值立场 | 偏高 | $4.65B 估值标记在叙事上说得通,但公开业务基本面并不能按常规支撑。 |
| 入场纪律 | 要求更好证据或更好价格 | 产品包装、买方和治理细节是主要解锁点。 |
本建议明确对价格敏感:它评估的是当前公开估值语境,而不是孤立评价团队质量。
[CV001, CV005, CV007, CV008, CV039, CV040]| 估值视角 | 公开证据 | 支撑什么 | 不支撑什么 |
|---|---|---|---|
| 标志性轮次 | 2026 年 5 月,Series A 融资超过 $650M,估值 $4.65B | 异常强的投资者信心和融资能力 | 客户、收入或利润率已经支撑该价格 |
| 创始人 / 团队信号 | 公开报道强调创始人履历和高规格投资团 | 高质量人才叙事 | 商业化里程碑执行 |
| 技术信号 | 2026 年 7 月文章和仓库给出面向基准的证明 | 提高对研究项目真实性的信心 | 企业产品包装或需求已经存在 |
| 官方产品界面 | 网站和文章仍以使命和研究为中心 | 解释公司想构建什么 | 未显示定价、买方工作流或管理控制 |
| 估值纪律结论 | 公开数据主要支撑期权价值 | 如果执行跑通,潜在上行仍大 | 当前估值下安全边际很薄 |
本表区分可见估值驱动因素,以及仍不透明的业务驱动因素。
[CV001, CV002, CV003, CV004, CV005, CV006]从价格锚点、证据质量和商业化缺口,推导出继续研究的建议。
这是定性投委会逻辑链,不是数学模型。
[CV001, CV004, CV011, CV026, CV039, CV040]评分卡显示,叙事和技术上行空间异常强,但公开商业化证据薄弱。
1-5 档评分综合各章证据,并有意对缺失的非公开披露扣分。
[CV001, CV004, CV005, CV006, CV007, CV011]8.2 估值支撑与商业证明
Recursive 的公开估值支撑来自三处:资本进入能力、市场叙事和技术可信度。资本进入能力从 Series A 的规模和质量上很清楚。 市场叙事也看得见:Stanford、Gartner 和 McKinsey 都在描述一个 2026 年环境, AI 投资、面向科学的 AI 和企业实验仍足够大,能够让前沿赢家获得不成比例的回报。 2026 年 7 月发布后,技术可信度明显提高。但这些都是上游估值支撑,不是变现的直接证据。 缺失的商业证明格外重要,因为 2026 年的企业 AI 买家已经不再只为模型新颖性付费。CNBC 和 McKinsey 描述的是更重效率的预算环境, 而 Microsoft、Snowflake、OpenAI 和 Anthropic 都公开宣传支出控制、治理功能、隐私边界或使用分析。 在已审阅的公开记录中,Recursive 还没有展示可比的产品包装层。这造成估值不对称: 公司可能因为野心和人才而值得溢价,但公开证据尚未显示它能否把研究优势转成可控、可采购的工作流产品, 从而支撑数十亿美元级价格。[CV004, CV005, CV006, CV011, CV012, CV013]
8.3 可比公司组与情景区间
Recursive 没有公开收入基数,传统收入倍数框架会制造伪精确。更好的方法,是把里程碑估值与公开可比公司边界结合起来。 可比表不能证明 Recursive 值 $4.65 billion,但能显示市场已经给哪些公司远高于 Recursive 的估值: Palantir、Snowflake、ServiceNow、Microsoft 和 MongoDB 的公开市值都远高于 Recursive 的标记, 同时也都通过可见的企业级包装和治理表面销售真实产品。在这个公开可比组里,Anthropic 是最相关的前沿实验室参照, 因为它 2026 年的官方融资更新把巨大估值与明确的收入运行率和企业采用表述配在一起。Recursive 还没有做到。 这个缺口比绝对估值算术更重要。按情景看,乐观情景只有在 Recursive 推出真实产品形态、 拿下具名买家并守住技术领先时,才可支撑更高估值。基准情景只有在产品化和客户证明迅速出现时, 才能支撑接近当前标记。悲观情景——商业延迟、平台打包、融资条件转弱或技术滑坡—— 很可能迫使公司以降轮方式重置,因为当前价格已经预设了大量未来证明。[CV012, CV018, CV019, CV020, CV021, CV022]
| 可比公司 | 指标 | 倍数 / 估值 / 状态 | 参考价值 | 局限 |
|---|---|---|---|---|
| Anthropic | 私募估值 + 收入背景 | 官方 2026 年 Series H 轮,投后估值 $965B,收入运行率 $47B | 最合适的前沿实验室参照,说明大额估值可以和明确商业规模并存 | 阶段晚得多、收入基数大得多,客户采用也比 Recursive 广得多 |
| Palantir | 公开市值 + 企业 AI 平台 | 2026 年 7 月市值约 $317.35B;AIP 平台可见 | 说明拥有政府 / 企业分发能力的规模化 AI 工作流厂商,公开市场可以给出什么估值 | 成熟上市公司,已有收入、分发和政府业务敞口,与 Recursive 不同 |
| Snowflake | 公开市值 + 治理型 AI 功能套件 | 2026 年 7 月市值约 $95.31B;Cortex AI 以及 RBAC / 隐私控制可见 | 适合对标企业 AI 包装方式和治理预期 | 数据云老牌厂商,已有装机基础,不是早期研究实验室 |
| ServiceNow / Microsoft 平台组合 | 公开市值 + 广泛企业 AI 捆绑 | ServiceNow 约 $102.30B;Microsoft 约 $2.899T;Copilot 把治理和分发打包进去 | 说明老牌厂商已经嵌在企业工作流里,带来捆绑压力 | 两者都不是一一对应的创业公司可比对象;它们更像竞争边界设定者,而不是估值镜像 |
| MongoDB / Databricks 技术栈类比 | 公开市值 + 私有数据平台包装背景 | MongoDB 市值约 $25.92B;Databricks 主打私有、受治理的 AI / 数据平台工作流 | 有助于理解数据平台厂商如何把 AI 包进更大的企业技术栈 | 但仍不是前沿 AI 实验室的直接可比对象;后者没有公开收入或产品界面 |
这组可比公司主要用于设定边界,不是直接倍数可比:Recursive 没有公开收入基数,公开记录里也尚未进入产品阶段。
[CV012, CV013, CV018, CV019, CV020, CV021]| 情景 | 关键假设 | 估值 / 回报逻辑 | 概率信号 |
|---|---|---|---|
| 乐观 | Recursive 推出清晰产品界面,披露具名试点或客户,守住技术领先,且前沿 AI 融资环境仍开放 | 私募估值可高于当前水平;约 $6B-$8B 的估值区间在叙事上站得住 | 概率较低,因为多个目前缺失的证据点都必须很快出现 |
| 基准 | 技术进展延续,产品化启动,但变现证据仍早,买方仍挑剔 | 当前估值可以守住,或只小幅扩张;约 $4B-$5.5B 更像宽泛持有区间 | 若公司能执行、但还没证明软件级经济性,这是最可能情景 |
| 悲观 | 商业包装滞后,老牌厂商捆绑相邻功能,融资条款收紧,或技术领先收窄 | 估值可能按降轮式重置到约 $2B-$3.5B;若优先权悬压明显,这一风险更高 | 概率偏高,因为当前估值已经折现了很多未来成功 |
| 终止 / 逻辑破裂 | 没有可见客户证明、没有可部署控制面,且下一轮融资条款转弱 | 普通股结果可能显著差于名义私募估值所暗示的水平 | 若隐藏优先权和商业化不透明同时出现,下行可能是二元的 |
这些区间是美元十亿级的定性估值包络,来自里程碑逻辑,不是 DCF 或披露收入倍数。
[CV011, CV026, CV033, CV034, CV035, CV036]Recursive 的估值最敏感于商业化证据,而不是只看 TAM 叙事。
1–5 序数条形反映基于公开证据集做出的估值敏感性判断。
[CV007, CV015, CV020, CV033, CV035, CV036]围绕当前私人市场估值标记的示意性估值包络。
这些区间是随情景变化的估值边界,不是点估计;前提是假设没有未披露的灾难性法律或技术失败。
[CV001, CV033, CV035, CV036, CV037, CV038]8.4 投资逻辑、反向逻辑与最终尽调问题
投资逻辑很直接:Recursive 是市场上资本最充足的早期 AI 实验室之一,投资人信号异常强, 所在品类如果自动化 AI 研究能够复利增长,确实有真实上行空间;现在公开技术证据也足够让故事不只是叙事。 反向逻辑同样强:当前公开记录仍更像研究项目,而不是已验证业务;当商业化、治理和融资条款不透明时, 数十亿美元级私募估值会变得脆弱。决定性尽调问题不是公司是否有前景,而是新投资人要从这个价格赚钱, 哪些条件必须成立。至少,管理层需要展示第一版商业包装、某些客户或试点证明、适合企业部署的产品控制表面、 从基准胜利到可重复购买行为的可信路径,以及不会用厚重优先权掩盖下行的股权结构条款。没有这些输入, 买家实际上是在以溢价购买期权。因此,终止触发条件和最终尽调问题非常关键: 它们定义了哪些确切证据能把判断从值得关注推进到可投资。[CV001, CV003, CV008, CV010, CV011, CV027]
| 论点 | 方向 | 改变看法的条件 |
|---|---|---|
| Recursive 拥有出色融资能力和顶级投资人,能争取时间招聘、购买算力并继续推进前沿。 | 投资逻辑 | 如果后续轮次条款转弱,或资金主要支持没有回报的研究烧钱,这种支持的意义会下降。 |
| 2026 年 7 月的技术发布提供了真实证据,证明该实验室能产出可发表的基准进展。 | 投资逻辑 | 如果这些资产转化为买方能部署、能治理的产品,投资逻辑会明显增强。 |
| 2026 年,科学 AI 和企业 AI 市场仍足够大,能够奖励差异化赢家。 | 投资逻辑 | 如果 Recursive 展示出具体可变现切入点,而不是宽泛使命叙事,看法会改善。 |
| 没有公开收入、定价或具名客户,意味着当前估值不能像软件公司那样被支撑。 | 反向逻辑 | 披露产品界面、价格模型和客户证明,会显著降低估值折扣。 |
| 既有厂商已把治理、支出控制和工作流 AI 打包进更广平台。 | 反向逻辑 | 如果 Recursive 展示出这些平台难以复制的工作流护城河,风险会下降。 |
| 未知优先权和治理条款可能把表面不变的名义估值变成较弱的普通股价值。 | 反向逻辑 | 干净股权结构表和轻优先股堆叠会改善下行保护。 |
本表把公司质量和价格吸引力拆开,因为核心尽调问题不是 Recursive 是否值得关注,而是当前入场价是否有吸引力。
[CV003, CV004, CV009, CV011, CV012, CV013]| 触发项 | 阈值 | 对投资逻辑的传导 | 行动含义 |
|---|---|---|---|
| 产品化延迟 | 到下一个重要里程碑窗口仍没有公开产品界面或部署控制 | 故事从产品可选性转为长期研究烧钱 | 从继续研究转向回避 / 等待重置 |
| 客户证明缺口 | 尽管融资可见度持续,仍没有具名试点、客户或合作伙伴工作流证据 | 削弱“技术证明正在转化为买方需求”的判断 | 要求更深折价,或等证据出现 |
| 技术滑坡 | 同行快速推进时,Recursive 没有新的重要基准或能力证明 | 压缩支撑溢价定价的感知护城河 | 下调估值区间和信心 |
| 融资条款转弱 | 下一轮平轮或降轮,且保护条款很重 | 说明私募市场支持没有头条动能暗示的那么稳 | 按新条款重新承销,不再沿用旧估值 |
| 治理意外 | 优先股堆叠、董事会控制或权利包对投资人不友好 | 普通股价值可能远低于企业价值口径的头条数字 | 暂停,直到完整建模股权结构表 |
这些是可监控的投资逻辑破裂信号,不是泛泛风险;每一项都会实质改变当前估值对新资金的含义。
[CV006, CV007, CV010, CV027, CV033, CV034]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 商业化 | 产品路线图、包装计划和定价模型 | 决定基准测试证明能否变成可变现的工作流软件 | 管理层介绍和产品演示 |
| 客户 / 试点 | 具名客户、试点或设计伙伴,以及用例 | 把叙事需求和可观察的买方采用区分开 | 客户访谈和管线复核 |
| 治理 / 控制 | 管理控制、日志、安全态势和合规路线图 | 严肃企业部署必须具备,也用于对比老牌替代方案 | 安全与合规尽调 |
| 财务 | 当前收入、烧钱速度、算力承诺和现金跑道 | 把融资故事转成可承销的业务故事 | 财务资料室审阅 |
| 股权结构表 / 优先权 | 完整证券层级、优先权、按比例认购权和特殊权利 | 理解名义估值对普通股结果意味着什么,这是必备项 | 法律尽调和瀑布模型 |
| 里程碑计划 | 未来 6-12 个月哪些具体成果应支撑下一次估值 | 定义能推动建议变化的证据路径 | 董事会材料和运营计划 |
这些尽调要求有意收窄并直接服务决策;若回答充分,建议可以上调。若不能,当前价格仍难以辩护。
[CV007, CV015, CV016, CV017, CV039, CV040]免责声明
本报告仅基于截至 2026-07-20 已审阅的公开来源,不能替代非公开的财务、法律、技术和客户尽调。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Recursive uses https://www.recursive.com as its live official website and describes itself there simply as Recursive. | 中 | SO001 |
| CO002 | Recursive says its mission is to build self-improving AI that automates knowledge discovery and ultimately improves the science of AI itself. | 中 | SO001, SO005 |
| CO003 | Recursive publicly lists San Francisco and London as its offices. | 中 | SO001, SO005 |
| CO004 | Independent profile and funding-coverage sources place Recursive's founding in 2025. | 中 | SO010, SO011, SO015 |
| CO005 | Wilson Sonsini says Recursive came out of stealth on May 13, 2026 and announced a Series A financing above $650 million at a $4.65 billion valuation. | 中 | SO012, SO006 |
| CO006 | The Series A was led by GV and Greycroft. | 中 | SO007, SO012, SO013 |
| CO007 | NVIDIA and AMD Ventures were disclosed participants in the Series A syndicate. | 高 | SO007, SO012, SO015 |
| CO008 | The disclosed round is explicitly described as a Series A. | 高 | SO012, SO006 |
| CO009 | Recursive says its team is over 25 people and still growing. | 中 | SO001, SO009 |
| CO010 | Tech.eu and TNW both describe Recursive as having fewer than 30 employees at launch. | 中 | SO007, SO008 |
| CO011 | Recursive says its co-founders previously created the AI research labs at Salesforce and Uber and led teams at OpenAI, DeepMind, Google Brain, and Meta. | 中 | SO001 |
| CO012 | Tech.eu identifies Richard Socher as CEO and co-founder and Tim Rocktäschel as a co-founder and former Google DeepMind scientist. | 中 | SO007, SO009 |
| CO013 | SCMP reports that former Meta FAIR research scientist director Yuandong Tian is one of Recursive's eight co-founders. | 中 | SO013, SO008 |
| CO014 | OfficeChai and Lab Index both associate Jeff Clune, Josh Tobin, and Tim Shi with Recursive's founding group, but the official site does not publish a full roster. | 中 | SO009, SO011, SO001 |
| CO015 | TNW says Recursive had not released a product at the time it emerged from stealth. | 中 | SO008 |
| CO016 | Tech Funding News and Europe Alternatives both say Recursive targeted a public launch in mid-2026. | 中 | SO010, SO015 |
| CO017 | Funding coverage says the new capital is intended to scale compute infrastructure and research operations. | 中 | SO010, SO008 |
| CO018 | Recursive's July 2026 article is the clearest public milestone between stealth emergence and run date because it reports concrete benchmark results from the company's automated AI research system. | 中 | SO002, SO016 |
| CO019 | Recursive says its system automates an iterative research loop that proposes ideas, implements them, runs experiments, validates results, and chooses follow-on experiments. | 中 | SO002 |
| CO020 | Recursive reported improving NanoChat autoresearch benchmark performance from 0.9372 to 0.9109 validation BPB. | 中 | SO002, SO023 |
| CO021 | Recursive reported reducing NanoGPT speedrun training time from 79.7 seconds to 77.5 seconds on its disclosed benchmark setup. | 中 | SO002, SO024 |
| CO022 | Recursive reported increasing mean SOL-ExecBench score from 0.699 to 0.754 across 235 kernels. | 中 | SO002, SO025, SO019 |
| CO023 | Recursive has publicly open-sourced research artifacts from its first automated AI research results on GitHub. | 中 | SO002, SO016 |
| CO024 | Recursive says it prioritizes safety while pursuing recursively self-improving AI. | 中 | SO001 |
| CO025 | X shows the Recursive account joined in May 2026, consistent with a company surfacing publicly around the Series A announcement. | 中 | SO005 |
| CO026 | The X profile uses the phrase self-improving superintelligence to automate knowledge discovery, reinforcing that Recursive is presenting a research-lab thesis rather than a finished product suite. | 中 | SO005 |
| CO027 | Lab Index lists Recursive Superintelligence as an alias for Recursive and shows London as the headline HQ, while the live official site emphasizes a dual San Francisco and London footprint. | 中 | SO011, SO001 |
| CO028 | Tech.eu says Recursive was incorporated in London. | 中 | SO007 |
| CO029 | Public overview sources do not disclose board composition, control rights, debt, or secondary-sale terms. | 中 | SO001, SO012, SO006 |
| CO030 | Foundra characterizes the launch as a legibility premium on an eight-person founding roster rather than a product or revenue story. | 低 | SO014 |
| CO031 | TechCrunch's July 2026 unicorn tracker places Recursive among the year's highest-profile new private AI unicorns. | 中 | SO006 |
| CO032 | Wilson Sonsini publicly disclosed that it advised Recursive on the Series A transaction. | 中 | SO012 |
| CO033 | No public customer count, revenue figure, ARR figure, or pricing page is disclosed on the official website reviewed for this chapter. | 中 | SO001 |
| CO034 | TNW frames the combination of a $4.65 billion valuation, four months of existence, and no released product as an unusually aggressive maturity-to-price relationship. | 中 | SO008 |
| CO035 | Recursive's official materials consistently describe open-ended algorithms and self-improvement of AI systems as the company's central technical thesis. | 中 | SO001, SO002 |
| CO036 | Recursive says the same playbook it is building for AI research could later extend into other scientific disciplines. | 中 | SO001 |
| CO037 | The GitHub artifact repository includes benchmark-specific subdirectories for autoresearch, NanoGPT speedrun, and SOL-ExecBench runs. | 中 | SO016, SO023, SO024, SO025 |
| CO038 | TNW argues that investment from both Nvidia and AMD implies the chipmakers view Recursive as a near-term buyer of frontier compute. | 中 | SO008, SO007 |
| CM001 | Recursive's official positioning is about AI that improves AI, which places its market closer to research automation and frontier model-development tooling than to generic consumer AI apps. | 中 | SM001, SM002 |
| CM002 | The relevant included spend for Recursive spans frontier AI research tooling, model-development infrastructure, and enterprise knowledge or agent platforms that automate complex reasoning workflows. | 中 | SM001, SM006, SM014, SM015, SM016 |
| CM003 | The excluded spend should include broad non-AI SaaS, generic cloud services unrelated to AI workloads, and consumer chat usage that does not purchase research-automation capability. | 中 | SM006, SM007, SM008 |
| CM004 | Status-quo substitutes already cover pieces of the buyer problem through enterprise search, RAG platforms, model APIs, and lower-cost search APIs. | 中 | SM012, SM014, SM015, SM016, SM020 |
| CM005 | Gartner forecasts worldwide AI spending of $2.595 trillion in 2026, up 47% year over year. | 中 | SM006 |
| CM006 | Gartner's January 2026 view put worldwide AI spending at $2.52 trillion for 2026, implying the market outlook was revised upward by May. | 中 | SM007, SM006 |
| CM007 | Gartner says AI infrastructure is the largest 2026 spending segment at $1.431 trillion and over 45% of total AI spend. | 中 | SM006 |
| CM008 | Gartner sizes 2026 AI software spending at roughly $453.2 billion. | 中 | SM006 |
| CM009 | Gartner sizes 2026 AI models spending at roughly $32.6 billion. | 中 | SM006 |
| CM010 | McKinsey says AI is gobbling up to a third of companies' change budgets while also adding to run costs. | 中 | SM008 |
| CM011 | McKinsey argues deliberate modernizers earmark at least one third of technology expenditures for change initiatives. | 中 | SM008 |
| CM012 | Deloitte says worker access to AI rose by 50% in 2025. | 中 | SM009 |
| CM013 | Deloitte says the number of companies with at least 40% of projects in production is set to double within six months. | 中 | SM009 |
| CM014 | Deloitte says only 34% of organizations are truly reimagining the business with AI rather than mainly pursuing productivity. | 中 | SM009 |
| CM015 | Deloitte says only one in five companies has a mature governance model for autonomous AI agents. | 中 | SM009 |
| CM016 | Deloitte says 42% of companies believe strategy is highly prepared for AI adoption, but operational preparedness lags. | 中 | SM009 |
| CM017 | IDC says AI Infrastructure Provisioning was a $30.3 billion use case in 2024 and is projected to reach $47 billion by 2028, representing about 30% of total AI spending. | 中 | SM010 |
| CM018 | IDC says AI-enabled Customer Service and Self Service commanded $16.7 billion of spending in 2024. | 中 | SM010 |
| CM019 | IDC says Augmented Fraud Analysis and Investigation drew more than $17 billion of investment in 2024 with a 31% five-year CAGR. | 中 | SM010 |
| CM020 | CNBC reports a growing enterprise spend crunch in which buyers are shifting from token maximization to efficiency and ROI discipline. | 中 | SM011 |
| CM021 | CNBC cites an example where Lindy switched traffic away from Anthropic to cheaper alternatives and expected millions in savings within months. | 中 | SM011 |
| CM022 | CNBC reports that some midsize companies are still waiting 12 to 18 months before making big AI spending decisions. | 中 | SM011 |
| CM023 | OpenAI's business pricing page lists a team-oriented workspace at $20 per user per month with analytics, budgeting, and spend controls. | 中 | SM012 |
| CM024 | Anthropic's pricing documentation shows a wide token-price band between premium frontier models and lower-cost sonnet-tier models. | 中 | SM013 |
| CM025 | Google positions Agent Search as an out-of-the-box RAG and enterprise search system grounded in enterprise data. | 中 | SM014 |
| CM026 | AWS says Bedrock serves more than 100,000 organizations worldwide. | 中 | SM015 |
| CM027 | AWS says Bedrock gives access to hundreds of foundation models and agent-development tooling on one platform. | 中 | SM015 |
| CM028 | AWS says prompt routing can cut costs by up to 30% and distilled models can cost up to 75% less. | 中 | SM015, SM017 |
| CM029 | Azure AI Search describes itself as an enterprise knowledge and RAG system built for end-to-end retrieval applications. | 中 | SM016 |
| CM030 | Brave Search API is a live status-quo substitute for web retrieval and search access within agentic workflows. | 中 | SM020 |
| CM031 | OpenAI's GPT-OSS release shows that open-weight frontier-class models are becoming part of the competitive landscape. | 中 | SM021 |
| CM032 | Anthropic's enterprise announcement shows that usage controls, analytics, and provisioning are becoming standard buying criteria for agentic coding tools. | 中 | SM022 |
| CM033 | GitHub's VS Code auto model selection feature shows that model routing is becoming a normal part of developer workflow procurement. | 中 | SM023 |
| CM034 | CNBC reports that roughly 95% of enterprise AI usage still runs on frontier models, according to Glean CEO Arvind Jain. | 中 | SM011, SM024 |
| CM035 | Recursive's realistic near-term addressable market is narrower than the trillions in total AI spend because the company remains pre-product and is targeting research automation first. | 中 | SM001, SM002, SM006, SM008 |
| CM036 | The most plausible initial buyers for a Recursive-like product are frontier labs, hyperscaler AI platform teams, and enterprise teams running high-value knowledge workflows. | 中 | SM001, SM014, SM015, SM016 |
| CM037 | Budget ownership in this market typically sits with CTO, CIO, platform engineering, or research leadership rather than end users themselves. | 中 | SM008, SM009, SM014, SM015, SM016 |
| CM038 | Growth drivers include agentic automation, enterprise demand for grounded retrieval, and willingness to fund infrastructure that supports differentiated AI workflows. | 中 | SM006, SM009, SM014, SM015 |
| CM039 | The main adoption constraints are compute intensity, ROI scrutiny, governance immaturity, and the ability of incumbents to bundle adjacent functionality. | 中 | SM006, SM008, SM009, SM011, SM015 |
| CM040 | For Recursive, the market gap is not lack of macro AI demand but the challenge of turning research automation into a product category that budgets already recognize. | 中 | SM001, SM003, SM008, SM011 |
| CP001 | Recursive publicly positions itself around AI that improves AI rather than around a launched application category. | 中 | SP001, SP002 |
| CP002 | That positioning places Recursive closer to automated AI research and knowledge-discovery tooling than to generic consumer chat. | 中 | SP001, SP002, SP003 |
| CP003 | As of the run date, Recursive’s public surface is a homepage, a technical article, and GitHub research artifacts rather than a priced product catalog. | 中 | SP001, SP002, SP004 |
| CP004 | TNW explicitly says Recursive had not released a product at stealth exit. | 中 | SP004 |
| CP005 | Tech.eu launch coverage frames the company as emerging from stealth with a large funding round rather than a commercial launch. | 中 | SP003 |
| CP006 | OpenAI Business publicly bundles chat, coding, analysis, workflows, connectors, SSO, analytics, and spend controls. | 中 | SP006 |
| CP007 | OpenAI Enterprise separately markets agents, Codex, governance controls, data protections, support, and SLAs. | 中 | SP007 |
| CP008 | Anthropic publicly markets Claude Code inside Team and Enterprise plans with spend caps, analytics, and admin tooling. | 中 | SP009 |
| CP009 | Public API pricing from Anthropic means buyers can benchmark price and usage without waiting for a custom sales conversation. | 中 | SP008, SP009 |
| CP010 | Google Agent Search shows that enterprise-grounded retrieval and search already have credible incumbent supply. | 中 | SP010 |
| CP011 | Amazon Bedrock positions model choice itself as a product, reducing the need for buyers to commit to a single lab too early. | 中 | SP011 |
| CP012 | Azure AI Search gives Microsoft a governed retrieval and enterprise-search surface that competes for the same broad knowledge-work budgets Recursive may target. | 中 | SP012 |
| CP013 | Glean markets enterprise context, permissions, memory, and agent runtime control as its core moat. | 中 | SP013 |
| CP014 | Perplexity Enterprise represents a citation-first answer engine substitute for users who want research-style outputs without adopting a frontier lab directly. | 中 | SP014 |
| CP015 | Llama 4 shows that open-weight alternatives continue to improve performance and efficiency, which is adverse to premium pricing for generic capability. | 中 | SP015 |
| CP016 | DeepSeek publishes per-million-token pricing, reinforcing that low-cost model access is a real outside option for technical buyers. | 中 | SP016 |
| CP017 | Ricursive Intelligence is a useful adjacent peer because it also describes a recursive-improvement thesis and reached a $4 billion Series A valuation quickly. | 中 | SP017 |
| CP018 | Presenc AI’s 2026 funding leaderboard describes capital concentrating in OpenAI, Anthropic, and xAI while mid-tier labs race to keep up. | 中 | SP018 |
| CP019 | Sakana AI’s positioning as a frontier lab in Japan shows the geography of frontier-lab competition is broadening, not narrowing. | 中 | SP019 |
| CP020 | Together AI competes for infrastructure-minded technical buyers with a research-optimized cloud, managed inference, and pre-training stack. | 中 | SP020 |
| CP021 | Recursive therefore competes for future technical-buyer mindshare against both labs and platforms that already expose APIs or managed products. | 中 | SP006, SP007, SP009, SP010, SP011, SP012, SP013, SP020 |
| CP022 | CNBC reports that companies are shifting from default frontier-model usage toward model routing. | 中 | SP021 |
| CP023 | The same CNBC reporting says roughly 95% of enterprise AI usage is still running on expensive frontier models, leaving room for future optimization pressure. | 中 | SP021 |
| CP024 | A budget-aware market is adverse for a startup that has not yet proven why its workflow should command premium spend. | 中 | SP021, SP022 |
| CP025 | OpenAI and Anthropic have both responded to budget pressure with spend controls and analytics rather than relying only on raw model quality. | 中 | SP006, SP009, SP022 |
| CP026 | Gartner says enterprises will expand their use of GenAI models embedded in existing software and agentic workflows. | 中 | SP023 |
| CP027 | That Gartner dynamic is adverse to Recursive because incumbents can distribute AI through products buyers already own. | 中 | SP023, SP010, SP011, SP012 |
| CP028 | Deloitte’s enterprise AI work supports the view that adoption is moving into production settings, raising buyer expectations for governance and support. | 中 | SP024 |
| CP029 | Stanford HAI’s economy framing supports that AI has moved into a capital-rich, high-velocity competitive environment rather than an experimental niche. | 中 | SP025, SP018 |
| CP030 | Recursive has not publicly matched peer disclosures on pricing, support, usage analytics, or enterprise admin controls in the sources reviewed here. | 中 | SP001, SP002, SP004 |
| CP031 | Because Recursive is pre-product, its direct competition today is more for talent, capital, and future demand than for reported public customer wins. | 中 | SP003, SP004, SP018 |
| CP032 | The strongest adverse public fact in the competitive story is that Recursive’s $4.65 billion valuation arrived before a public product or customer surface. | 中 | SP003, SP004, SP005 |
| CP033 | Open-weight families and low-cost APIs weaken the durability of any moat based only on access to generic model capability. | 中 | SP015, SP016, SP021 |
| CP034 | For Recursive to create switching cost, it likely needs proprietary research outputs, workflow embedding, or a hard-to-copy system advantage beyond model access. | 中 | SP001, SP002, SP013, SP021 |
| CP035 | Enterprise context and permissions are already an explicit moat claim from Glean, meaning Recursive would need to solve or partner on that layer if it enters enterprise workflows. | 中 | SP013 |
| CP036 | Public sources do not yet show a Recursive distribution channel comparable to OpenAI, Anthropic, Google, AWS, or Microsoft. | 中 | SP001, SP002, SP006, SP007, SP009, SP010, SP011, SP012 |
| CP037 | The strongest pro-Recursive competitive argument is that the self-improvement loop itself may be the differentiated product if benchmark gains keep compounding. | 中 | SP001, SP002 |
| CP038 | Ricursive Intelligence demonstrates that investors are willing to fund adjacent recursive-improvement theses even when they target different technical domains. | 中 | SP017, SP018 |
| CP039 | Together AI, AWS Bedrock, and cloud incumbents all let buyers access frontier capability without betting on a single early research startup. | 中 | SP011, SP020 |
| CP040 | Commercial moat durability for Recursive is currently unproven and remains a diligence blocker until the company shows repeatable product outcomes or workflow lock-in. | 中 | SP001, SP002, SP004, SP021 |
| CI001 | Recursive does not publish a public pricing page, API rate card, or enterprise package in the reviewed sources. | 中 | SI001, SI002 |
| CI002 | No public revenue, ARR, or customer count is disclosed in the reviewed sources as of 2026-07-20. | 中 | SI001, SI004, SI010 |
| CI003 | Recursive publicly frames its mission around AI improving AI and broader scientific discovery, which implies a future product path but not a present revenue line. | 中 | SI001, SI002 |
| CI004 | Recursive emerged from stealth in May 2026 with over $650 million raised at a $4.65 billion valuation. | 中 | SI003, SI008, SI009 |
| CI005 | Multiple public sources say the funding will help secure large compute clusters and support a first Level 1 autonomous training system. | 中 | SI005, SI006 |
| CI006 | Public reporting places the team at roughly 25 to 30 people around launch. | 中 | SI004, SI006 |
| CI007 | Because no public product package is disclosed, no public sales-efficiency metrics can be calculated from the reviewed sources. | 中 | SI001, SI002, SI004 |
| CI008 | The official web presence emphasizes mission and research evidence rather than monetization details. | 中 | SI001, SI002, SI025 |
| CI009 | Peer frontier-AI vendors now package monetization with explicit buyer controls, a contrast that makes Recursive financially harder to underwrite. | 中 | SI015, SI016, SI017, SI018 |
| CI010 | OpenAI Business lists a $20 per user per month annualized seat price and a $25 monthly-billed price. | 中 | SI015 |
| CI011 | Anthropic publishes API pricing and enterprise controls, giving buyers public monetization anchors before talking to sales. | 中 | SI017, SI018 |
| CI012 | CNBC reports that companies are shifting from frontier-model overspend toward model routing and efficiency. | 中 | SI021, SI022 |
| CI013 | McKinsey says AI is consuming large portions of technology change budgets while also adding run costs. | 中 | SI020 |
| CI014 | Recursive’s published nanoGPT speedrun benchmark uses 8x H100 hardware. | 中 | SI011, SI012, SI027 |
| CI015 | Recursive reports its best nanoGPT speedrun solution reached 77.3 seconds on 8x H100. | 中 | SI012 |
| CI016 | Recursive’s nanochat autoresearch bundle evaluates solutions for 5 minutes on a single Modal B200 GPU across 10 seeds. | 中 | SI013, SI029 |
| CI017 | Recursive’s published SOL-ExecBench bundle exposes 10 of 235 GPU-kernel implementations measured on B200. | 中 | SI014 |
| CI018 | Those benchmark disclosures imply a research program that is compute-intensive and specialized even before any public product is launched. | 中 | SI011, SI012, SI013, SI014, SI026, SI027, SI029, SI030 |
| CI019 | No public burn rate, cash balance, or runway disclosure appears in the reviewed sources. | 中 | SI001, SI003, SI004, SI010 |
| CI020 | No public gross margin, cost-to-serve, or recognized revenue disclosure appears in the reviewed sources. | 中 | SI001, SI004, SI010 |
| CI021 | A $650M+ Series A almost certainly provides meaningful near-term operating capacity, but the public record does not show whether it fully funds the roadmap. | 中 | SI004, SI005, SI006, SI008 |
| CI022 | The $4.65B valuation reflects strategic option value and founder-market fit more clearly than disclosed current cash-flow fundamentals. | 中 | SI004, SI008, SI010 |
| CI023 | No public customer references or contract terms are disclosed that would let an outsider model revenue mix or concentration. | 中 | SI001, SI004, SI010 |
| CI024 | No public retention, renewal, or expansion data is disclosed that would support recurring-revenue quality analysis. | 中 | SI001, SI004 |
| CI025 | The July 2026 article and GitHub release improve confidence in technical seriousness more than they improve confidence in present monetization. | 中 | SI002, SI011, SI012, SI013, SI014, SI028 |
| CI026 | OpenAI and Anthropic illustrate that monetization-grade enterprise AI in 2026 is packaged with controls, analytics, and support—not just model quality. | 中 | SI015, SI016, SI017, SI018 |
| CI027 | Recursive has not publicly shown equivalent monetization packaging, which lowers confidence in near-term revenue readiness. | 中 | SI001, SI002, SI015, SI016, SI017, SI018 |
| CI028 | Public cloud and platform vendors already offer procurement-ready AI surfaces, making it harder for a new vendor to monetize on narrative alone. | 中 | SI023, SI024, SI021 |
| CI029 | Recursive is a 2025-founded company, meaning the capital raised is unusually large relative to age. | 中 | SI003, SI007 |
| CI030 | The reviewed public record does not disclose debt, secondaries, or financing-term details beyond the headline Series A. | 中 | SI003, SI008, SI010 |
| CI031 | A buyer environment defined by spend controls and routing efficiency is adverse to any future premium software pricing strategy Recursive may pursue. | 中 | SI016, SI018, SI021, SI022 |
| CI032 | Gartner says AI infrastructure is the largest spending segment in 2026, which supports demand for compute but does not guarantee app-layer capture for Recursive. | 中 | SI019 |
| CI033 | Because the largest AI spend segment is infrastructure, a company like Recursive still has to prove it can capture value above the compute layer. | 中 | SI019, SI020 |
| CI034 | 2026 enterprise AI buyers increasingly expect analytics, budget controls, and predictable pricing as table stakes. | 中 | SI015, SI016, SI018, SI021, SI022 |
| CI035 | The most plausible future revenue path is some combination of platform access, workflow software, or specialized scientific automation, but no such stream is publicly launched yet. | 中 | SI001, SI002, SI023, SI024 |
| CI036 | Capital adequacy cannot be underwritten from public data without burn, cash, compute-commitment, and hiring-plan disclosure. | 中 | SI019, SI020, SI021, SI022 |
| CI037 | Public financial diligence is missing the minimum private package of cap-table terms, runway, compute contracts, and product roadmap milestones. | 中 | SI008, SI010, SI021 |
| CI038 | Recursive is well funded in headline terms but not financially underwritten in conventional operating terms. | 中 | SI004, SI008, SI010, SI019, SI020 |
| CI039 | Recursive has published only a small subset of its GPU-kernel outputs, meaning outsiders cannot fully audit the economic value of the technical asset base. | 中 | SI014 |
| CI040 | The minimum investable financial story would require proof of a real product, real buyers, a costed compute plan, and a credible path from research milestones to recurring revenue. | 中 | SI001, SI002, SI010, SI021 |
| CI041 | Recursive’s published benchmarks sit squarely on frontier NVIDIA and GPU-accelerated infrastructure references rather than commodity compute baselines. | 中 | SI027, SI029, SI030 |
| CE001 | Recursive publicly frames itself as building AI that recursively improves AI. | 中 | SE001, SE024 |
| CE002 | The company’s public thesis is first to improve the science of AI itself, then to expand into broader scientific discovery. | 中 | SE001, SE002 |
| CE003 | Recursive’s current public surface consists of a homepage, a technical article, and GitHub artifacts rather than a priced application or API. | 中 | SE001, SE002, SE003 |
| CE004 | The public artifact set spans three disclosed bundles: nanoGPT speedrun, nanochat autoresearch, and SOL-ExecBench samples. | 中 | SE003, SE004, SE005, SE008, SE010 |
| CE005 | The released stack therefore looks like a research-system asset base rather than a conventional application SKU set. | 中 | SE001, SE002, SE003, SE004 |
| CE006 | Launch coverage and the public record reviewed here do not show a public pricing page or customer onboarding path. | 中 | SE001, SE020 |
| CE007 | A user interacting with the current public Recursive stack would primarily inspect documents, repositories, and benchmark artifacts rather than use a live product interface. | 中 | SE001, SE002, SE003 |
| CE008 | Recursive’s article describes a loop that proposes ideas, implements them, runs experiments, validates results, and selects what to try next. | 中 | SE002 |
| CE009 | That loop is the closest thing the public record offers to a product definition. | 中 | SE002 |
| CE010 | No reviewed source shows that the company has yet wrapped this loop in public enterprise controls, support, or deployment guidance. | 中 | SE001, SE020, SE025 |
| CE011 | The repository root says it collects training scripts and kernel implementations discovered by Recursive’s automated AI-research system. | 中 | SE003, SE004 |
| CE012 | The loop is benchmark-centered rather than product-telemetry-centered in the public materials. | 中 | SE002, SE003, SE004 |
| CE013 | The public workflow implies that code or method changes are an explicit part of the autonomous research loop. | 中 | SE002, SE009 |
| CE014 | Recursive’s nanoGPT speedrun benchmark is framed around training GPT-2-small to the target loss on 8x H100 as fast as possible. | 中 | SE005, SE012 |
| CE015 | The published from_best nanoGPT solution reached 77.3 seconds while beating the same-hardware baseline. | 中 | SE006 |
| CE016 | Recursive’s nanochat autoresearch bundle evaluates solutions for 5 minutes on a single Modal B200 GPU across 10 seeds. | 中 | SE008 |
| CE017 | Recursive’s published SOL-ExecBench bundle contains 10 of 235 kernel implementations scored on B200. | 中 | SE010 |
| CE018 | The article and repository show that Recursive builds on upstream work such as modded-nanogpt and nanochat rather than claiming every layer was invented from scratch. | 中 | SE003, SE005, SE008, SE015, SE014 |
| CE019 | Karpathy’s autoresearch project is an explicit upstream reference point for the nanochat-style autonomous research framing. | 中 | SE012, SE013 |
| CE020 | The repository NOTICE and licensing notes preserve upstream MIT notices inside an Apache-2.0 package. | 中 | SE003, SE011 |
| CE021 | NVIDIA’s SOL-ExecBench provides an external benchmark context for the kernel-optimization work Recursive cites. | 中 | SE017, SE010 |
| CE022 | Recursive intentionally withholds most of its kernel inventory to avoid biasing the leaderboard. | 中 | SE010 |
| CE023 | The majority of the GPU-kernel asset base therefore remains private even after the July 2026 release. | 中 | SE010 |
| CE024 | Taken together, the public architecture looks like an internal research platform with benchmark adapters and selective artifact release. | 中 | SE002, SE003, SE004, SE010 |
| CE025 | Recursive’s homepage emphasizes safety as part of its mission framing. | 中 | SE001 |
| CE026 | The public Privacy Policy and Terms provide standard website legal scaffolding rather than enterprise AI governance proof. | 中 | SE025 |
| CE027 | The released work depends on frontier GPU classes such as H100 and B200. | 中 | SE005, SE008, SE010 |
| CE028 | That hardware dependence is consistent with a serious frontier-research program but raises both cost and operational dependency. | 中 | SE005, SE008, SE010 |
| CE029 | No public model card, enterprise security page, or deployment-control documentation was identified in the reviewed sources. | 中 | SE001, SE025 |
| CE030 | The public technical evidence is stronger than the public trust or deployment evidence. | 中 | SE002, SE003, SE025 |
| CE031 | Benchmark bundles provide reproducibility and measurement structure, but they remain company-selected evidence rather than third-party production audits. | 中 | SE002, SE003, SE017 |
| CE032 | The withheld 225 kernels mean outsiders cannot fully audit the breadth or defensibility of Recursive’s systems-level asset base. | 中 | SE010 |
| CE033 | The current public release therefore improves credibility but leaves meaningful uncertainty about private technical depth. | 中 | SE003, SE010 |
| CE034 | No reviewed source shows uptime commitments, support SLAs, or production operations guidance for external users. | 中 | SE001, SE020 |
| CE035 | From a buyer perspective, the product remains technically legible but operationally under-specified. | 中 | SE001, SE002, SE025 |
| CE036 | Launch coverage said the company was targeting a public launch in mid-2026 while scaling compute and a Level 1 autonomous training system. | 中 | SE019, SE021 |
| CE037 | By the run date, the clearest visible public milestone that actually shipped is the July 2026 article and repository. | 中 | SE002, SE003 |
| CE038 | That milestone materially improves underwriteability of the technical thesis even though it does not constitute a priced product launch. | 中 | SE002, SE003, SE020 |
| CE039 | Recursive has demonstrated a coherent multi-asset technical stack rather than a single isolated benchmark result. | 中 | SE004, SE005, SE008, SE010 |
| CE040 | The next maturity hurdle is productization: packaging the research engine in buyer-facing controls, support, and workflow boundaries. | 中 | SE001, SE002, SE020 |
| CU001 | Recursive does not publicly disclose named customers on its website or technical article pages. | 中 | SU001, SU002 |
| CU002 | Launch coverage in May 2026 explicitly said the company had not released a product. | 中 | SU012 |
| CU003 | The most plausible first buyers are frontier labs or model-development teams that directly value automated AI research. | 中 | SU001, SU002, SU015 |
| CU004 | Hyperscaler platform teams are also plausible buyers because the released work includes systems optimization and GPU-kernel outputs. | 中 | SU002, SU020 |
| CU005 | Research-intensive enterprise knowledge or R&D teams are a plausible later buyer set if Recursive productizes beyond internal AI research. | 中 | SU001, SU017, SU018, SU019 |
| CU006 | The broader scientific-discovery framing could eventually widen the buyer set beyond AI labs. | 中 | SU001 |
| CU007 | General enterprise users are a weak current fit because the public artifact set is highly technical and under-packaged. | 中 | SU002, SU012 |
| CU008 | The likely first-customer set is therefore small and technically sophisticated. | 中 | SU001, SU002, SU017, SU018, SU019, SU020 |
| CU009 | That buyer profile implies concentration risk even in a successful early commercialization scenario. | 中 | SU017, SU018, SU019, SU020 |
| CU010 | Recursive’s public identity and awareness surface expanded with launch coverage and its X presence in 2026. | 中 | SU003, SU011, SU013, SU014 |
| CU011 | The July 2026 article and repository are the clearest public adoption-proof event because they gave outsiders something concrete to inspect. | 中 | SU002, SU004 |
| CU012 | The repository releases page says there are no releases, which suggests the project is not yet packaged as easy-to-deploy software. | 中 | SU010 |
| CU013 | The X profile is evidence of public identity and distribution, not evidence of customer conversion. | 中 | SU003 |
| CU014 | The repository itself is a stronger signal of technical interest than the X profile because it exposes usable artifacts. | 中 | SU002, SU004 |
| CU015 | By the run date, public proof is still centered on artifact inspection rather than deployment proof. | 中 | SU002, SU004, SU012 |
| CU016 | The GitHub network page shows 174 stars and 15 forks for the public repository. | 中 | SU005 |
| CU017 | The public repository shows one open issue from an outside user and zero open pull requests. | 中 | SU007, SU008 |
| CU018 | No public source reviewed in this chapter shows a named paying customer, named pilot, or production deployment. | 中 | SU001, SU002, SU012 |
| CU019 | The strongest current proof type is community attention around the technical release, not economic adoption. | 中 | SU004, SU005, SU007, SU010 |
| CU020 | Recursive does not publicly disclose NRR, GRR, churn, renewal, or contract-length metrics. | 中 | SU001, SU002 |
| CU021 | No public revenue or seat-expansion data exists that would let an outsider evaluate repeat monetization quality. | 中 | SU001, SU012 |
| CU022 | Repository stars and forks are weak interest proxies, not retention or satisfaction metrics. | 中 | SU005 |
| CU023 | The contributors and commits pages suggest activity around the repository but do not demonstrate a paying-user cohort. | 中 | SU006, SU009 |
| CU024 | The absence of releases further weakens the case for treating repository interaction as real product retention. | 中 | SU010 |
| CU025 | If Recursive commercializes successfully, its first customer cohort is likely to be small and concentrated in elite technical organizations. | 中 | SU003, SU017, SU018, SU019, SU020 |
| CU026 | Adjacent vendors already serve many of the buyer workflows Recursive may later target. | 中 | SU017, SU018, SU019, SU020, SU023, SU024, SU025 |
| CU027 | That adjacent supply means early Recursive pilots could remain experimental rather than expand broadly if the product wedge is not strong enough. | 中 | SU017, SU018, SU019, SU020, SU023, SU024, SU025 |
| CU028 | CNBC’s 2026 reporting shows that enterprise buyers increasingly expect spend controls and routing efficiency from AI vendors. | 中 | SU021, SU022 |
| CU029 | Those buyer expectations raise the adoption bar for a new vendor that has not yet published customer-facing controls or packaging. | 中 | SU021, SU022, SU012 |
| CU030 | The biggest expansion risk is mistaking technical enthusiasm for durable product demand. | 中 | SU005, SU007, SU010 |
| CU031 | The biggest customer-quality gap is the absence of named references, deployment scope, and repeat-usage data. | 中 | SU001, SU002, SU012 |
| CU032 | Recursive should presently be described as having plausible buyer logic and very limited public adoption proof. | 中 | SU002, SU005, SU012, SU021 |
| CU033 | The tags page mirrors the absence of public releases, reinforcing that the repository is not yet packaged as versioned software for customer deployment. | 中 | SU026 |
| CU034 | The GitHub pulse page suggests repository activity can be monitored publicly, but it still does not provide buyer or usage evidence. | 中 | SU027 |
| CU035 | The community standards page shows open-source project hygiene but not customer support or commercial deployment readiness. | 中 | SU028 |
| CR001 | Recursive’s website exposes standard privacy and terms pages. | 中 | SR003, SR004 |
| CR002 | Those legal pages provide baseline web hygiene but do not answer product-governance or enterprise-control questions. | 中 | SR003, SR004 |
| CR003 | The public repository uses Apache-2.0 packaging and preserves upstream notices for derivative code. | 中 | SR017, SR018, SR019 |
| CR004 | That licensing hygiene reduces one class of IP risk in the released codebase. | 中 | SR017, SR018, SR019 |
| CR005 | No public board, cap-table, or control-rights summary appears in the reviewed sources. | 中 | SR001, SR008 |
| CR006 | No public product-governance or deployment-responsibility document was identified in the reviewed sources. | 中 | SR001, SR003, SR004 |
| CR007 | The public repository intentionally leaves most of the kernel inventory private. | 中 | SR009 |
| CR008 | That partial release limits external auditability of the technical moat. | 中 | SR009, SR020 |
| CR009 | No active public enforcement or litigation event was surfaced in the reviewed sources for this chapter. | 中 | SR001, SR007, SR008 |
| CR010 | Recursive’s current public product surface is benchmark-centric rather than deployment-centric. | 中 | SR002, SR009 |
| CR011 | The July 2026 release proves technical seriousness but not production operations readiness. | 中 | SR002, SR009 |
| CR012 | The repository’s actions, branches, labels, and milestones pages show a project surface, but not a mature packaged product operation. | 中 | SR011, SR013, SR014, SR015 |
| CR013 | The released work explicitly depends on frontier GPUs such as H100 and B200. | 中 | SR002, SR020 |
| CR014 | Frontier GPU dependence raises both cost and supply sensitivity. | 中 | SR002, SR020 |
| CR015 | No public release cadence, packaged binaries, or uptime commitments were identified for external users. | 中 | SR009, SR011 |
| CR016 | The current public operations surface is therefore too thin for a serious customer to treat as enterprise-ready software. | 中 | SR009, SR011, SR012 |
| CR017 | The public release is strong enough to reduce vaporware risk. | 中 | SR002, SR009, SR017 |
| CR018 | The public release is not strong enough to reduce customer-deployment risk. | 中 | SR002, SR009, SR012 |
| CR019 | Code-frequency and issue data show some activity but do not provide a reliability or quality metric for external customers. | 中 | SR010, SR016 |
| CR020 | Recursive’s public stack depends on upstream open-source baselines including nanochat, autoresearch, and modded-nanogpt lineage. | 中 | SR009, SR017, SR018, SR019 |
| CR021 | Recursive also depends on benchmark ecosystems such as SOL-ExecBench to frame public proof. | 中 | SR002, SR020 |
| CR022 | The company depends on NVIDIA-class GPU infrastructure for at least the released H100 and B200 benchmark work. | 中 | SR002, SR020 |
| CR023 | These dependencies are normal for frontier AI labs but still create concentration risk if supply or cost conditions worsen. | 中 | SR020 |
| CR024 | OpenAI, Anthropic, Google, AWS, Azure, Glean, and Perplexity already serve adjacent workflows that Recursive may try to commercialize into. | 中 | SR021, SR022, SR023, SR024, SR025 |
| CR025 | That incumbent supply creates bundling risk for a young startup without a public product surface. | 中 | SR021, SR022, SR023, SR024, SR025 |
| CR026 | Buyers are becoming more cost disciplined and increasingly value routing efficiency and spend controls. | 中 | SR021, SR022, SR024, SR025 |
| CR027 | A small first-customer set likely gives early buyers meaningful negotiating leverage. | 中 | SR005, SR006, SR023 |
| CR028 | The likely first buyers are a narrow technical cohort rather than a broad horizontal user base. | 中 | SR001, SR002, SR023 |
| CR029 | That concentration means customer and partner risk are tightly linked in Recursive’s early commercialization path. | 中 | SR023, SR024, SR025 |
| CR030 | Public reporting still places Recursive at roughly 25 to 30 people around launch. | 中 | SR005, SR007 |
| CR031 | A small team at a very high valuation heightens execution pressure because the organization must get multiple things right in sequence. | 中 | SR005, SR006, SR007 |
| CR032 | The company has demonstrated a research milestone but not yet a public product milestone. | 中 | SR002, SR005 |
| CR033 | Commercialization risk remains high because no named customers, pricing, or deployment controls are public. | 中 | SR001, SR005, SR006 |
| CR034 | Governance opacity compounds execution risk because outsiders cannot see how research, product, and financing trade-offs are being managed. | 中 | SR001, SR008 |
| CR035 | The strongest public execution upside is that these risks are still reducible if productization and customer proof appear quickly. | 中 | SR002, SR017 |
| CR036 | A public product surface with controls, releases, and buyer docs would materially reduce current operational risk. | 中 | SR011, SR012, SR015 |
| CR037 | Named pilots or deployments would materially reduce current customer and commercialization risk. | 中 | SR005, SR006 |
| CR038 | Compute-plan and financing-term transparency would materially reduce current capital and dependency uncertainty. | 中 | SR008, SR020 |
| CR039 | If the next major milestone arrives without public productization or customer proof, the valuation risk becomes materially sharper. | 中 | SR005, SR006, SR008 |
| CR040 | The overall risk stack is high today because multiple moderate risks reinforce each other rather than offsetting each other. | 中 | SR005, SR006, SR021, SR022 |
| CR041 | The tags page mirrors the absence of public packaged releases, reinforcing operational immaturity for external users. | 中 | SR026 |
| CR042 | The public pulse view may show activity, but activity is not a substitute for production operations evidence. | 中 | SR027 |
| CR043 | The community standards page is an open-source hygiene signal, not a customer-risk mitigation signal. | 中 | SR028 |
| CV001 | Recursive emerged from stealth in May 2026 with more than $650 million raised at a $4.65 billion valuation. | 高 | SV003, SV004, SV005, SV007 |
| CV002 | Public coverage places Recursive’s founding in 2025, making the current multibillion-dollar valuation unusually early-stage. | 中 | SV004, SV007, SV008 |
| CV003 | Recursive’s official site frames the company as building AI that recursively improves AI and ultimately automates scientific research. | 中 | SV001, SV008 |
| CV004 | The July 2026 article materially improved public confidence that Recursive has a serious automated-research system rather than only a stealth narrative. | 中 | SV002 |
| CV005 | Recursive does not publish a public pricing page or business package in the reviewed sources. | 中 | SV001, SV002 |
| CV006 | The public record reviewed for this chapter does not disclose board composition, cap-table terms, or preference-stack detail. | 中 | SV001, SV003, SV006 |
| CV007 | No public revenue, ARR, or named-customer disclosure was located for Recursive as of 2026-07-20. | 中 | SV001, SV002, SV006 |
| CV008 | The current valuation is therefore supported more by option value and investor signal than by public business fundamentals. | 中 | SV001, SV003, SV006 |
| CV009 | Multiple public sources identify GV, Greycroft, and NVIDIA among the financing backers, reinforcing the strength of the syndicate signal. | 中 | SV003, SV004, SV005 |
| CV010 | A new investor at the current mark is underwriting future commercialization rather than a publicly proven software business. | 中 | SV001, SV002, SV006, SV007 |
| CV011 | Stanford, Gartner, and McKinsey all support a 2026 environment in which AI investment and AI experimentation remain strategically important. | 中 | SV009, SV010, SV011 |
| CV012 | Stanford’s 2026 AI Index science section supports the idea that AI-for-science remains a live, high-upside category rather than a trivial niche. | 中 | SV029 |
| CV013 | CNBC and McKinsey also point to a more efficiency-focused buyer environment, implying that frontier capability alone may no longer win budgets. | 中 | SV011, SV012 |
| CV014 | Anthropic’s official Series H update paired its 2026 valuation with explicit run-rate revenue and enterprise adoption language. | 高 | SV013, SV014 |
| CV015 | Anthropic publicly markets enterprise controls such as spend caps, seat management, analytics, and a compliance API. | 中 | SV014, SV026 |
| CV016 | OpenAI publicly markets business pricing and enterprise spend controls, providing a concrete benchmark for procurement-ready AI packaging. | 中 | SV024, SV025 |
| CV017 | Microsoft 365 Copilot publicly emphasizes enterprise data protection, IT controls, reporting, and integration into existing workflow software. | 中 | SV021 |
| CV018 | Snowflake publicly emphasizes RBAC, privacy boundaries, and governance controls for its AI features. | 中 | SV018 |
| CV019 | Palantir publicly presents AIP as an artificial intelligence platform, showing that enterprise AI workflow packaging now exists at scale. | 中 | SV016 |
| CV020 | Databricks publicly markets a private, governed data-and-AI platform, further raising the packaging bar for new entrants. | 中 | SV022 |
| CV021 | AWS, Google Cloud, OpenAI, Anthropic, Microsoft, Snowflake, and Databricks all show publicly visible AI packaging surfaces that can compete with or absorb adjacent workflows. | 中 | SV014, SV018, SV021, SV022, SV024, SV027, SV028 |
| CV022 | As of July 2026, Palantir’s public market capitalization is about $317.35 billion. | 中 | SV015 |
| CV023 | As of July 2026, Snowflake’s public market capitalization is about $95.31 billion. | 中 | SV017 |
| CV024 | As of July 2026, ServiceNow’s public market capitalization is about $102.30 billion. | 中 | SV019 |
| CV025 | As of July 2026, Microsoft’s public market capitalization is about $2.899 trillion. | 中 | SV020 |
| CV026 | As of July 2026, MongoDB’s public market capitalization is about $25.92 billion. | 中 | SV023 |
| CV027 | Recursive’s $4.65 billion mark is smaller than mature public AI software valuations in absolute terms, but those companies also have real revenue, distribution, and governance surfaces. | 中 | SV015, SV017, SV019, SV020, SV021, SV018 |
| CV028 | Anthropic is the most informative frontier-lab comparator in this set because its official valuation update included explicit commercial scale that Recursive has not disclosed. | 中 | SV013, SV014, SV026 |
| CV029 | Public-comparable analysis for Recursive is boundary setting, not direct multiple comping, because there is no public revenue denominator. | 中 | SV015, SV017, SV023 |
| CV030 | A conventional DCF or revenue-multiple method would create false precision for Recursive at this stage. | 中 | SV001, SV006, SV029 |
| CV031 | The most defensible public valuation method is milestone and scenario analysis anchored by commercialization evidence rather than current revenue. | 中 | SV002, SV006, SV029 |
| CV032 | The current public record supports a high-upside category narrative, but not enough disclosed economics to call the current price cheap. | 中 | SV009, SV010, SV012, SV006 |
| CV033 | A bull case requires product packaging, customer proof, and retained technical lead rather than technical narrative alone. | 中 | SV002, SV012, SV021 |
| CV034 | A base case assumes technical progress continues and productization begins, but monetization proof remains early. | 中 | SV002, SV011, SV012 |
| CV035 | A bear case becomes plausible if commercialization lags while incumbents keep bundling adjacent AI workflows. | 中 | SV012, SV021, SV022, SV024 |
| CV036 | Bundling pressure from incumbent platforms is a real downside variable because many enterprise AI controls and workflow surfaces are already publicly available elsewhere. | 中 | SV014, SV018, SV021, SV022, SV024, SV027, SV028 |
| CV037 | A down-round style reset toward roughly $2 billion to $3.5 billion is plausible if product proof and financing support do not arrive fast enough. | 中 | SV006, SV012, SV030 |
| CV038 | Unknown preference and governance terms make downside harder to model and can reduce the value of common equity relative to the headline mark. | 中 | SV003, SV006 |
| CV039 | The price-sensitive recommendation is research-more / track rather than buy at the current valuation. | 中 | SV001, SV003, SV006, SV012 |
| CV040 | Confidence should be capped at medium because the public evidence set is much stronger on financing than on business fundamentals. | 中 | SV003, SV006, SV007 |
| CV041 | Key thesis-break triggers are no visible productization, no customer proof, technical slippage, weaker next-round terms, and adverse cap-table surprises. | 中 | SV002, SV006, SV012 |
| CV042 | The fastest ways to move the recommendation positively are packaged product, named buyers, governance-grade controls, and transparent cap-table economics. | 中 | SV014, SV021, SV024 |
| CV043 | Final diligence should focus on commercialization, customers, governance controls, financials, cap-table structure, and milestone plan. | 中 | SV006, SV014, SV021 |
| CV044 | Foundra’s adverse framing reinforces that the public record still lacks both product and revenue disclosure. | 中 | SV006 |
| CV045 | Recursive’s technical release reduces existential skepticism about the research program but does not solve the underwriting problem created by missing commercialization evidence. | 中 | SV002, SV006 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SO002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results: in fixed-budget language model training, small-model training speed, and GPU kernel optimization. |
| SO003 | Recursive | Privacy Policy - Recursive | |
| SO004 | Recursive | Terms of Use - Recursive | |
| SO005 | X | Recursive (@Recursive_SI) on X | Recursive self-improving superintelligence to automate knowledge discovery. |
| SO006 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | |
| SO007 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | The startup, which was incorporated in London and has offices in London and San Francisco, said a clear trend was emerging in AI. |
| SO008 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | Recursive Superintelligence is four months old, has fewer than 30 employees, and has not released a product. It is valued at $4.65 billion. |
| SO009 | OfficeChai | Recursive Raises $650 Million At $4.65 Billion Valuation To Create Self-Improving AI | Recursive was founded by former team leaders from OpenAI, Google DeepMind, Meta AI, Salesforce AI, and Uber AI. |
| SO010 | Tech Funding News | UK AI startup Recursive hits $4.65B valuation with $650M raise from Nvidia and GV — TFN | The company plans a public launch in mid-2026 as it scales compute infrastructure and research operations across San Francisco and London. |
| SO011 | Lab Index | Recursive | Frontier AI research lab pursuing recursive self-improvement. Emerged from stealth May 13, 2026 with a $650M round led by GV and Greycroft. |
| SO012 | Wilson Sonsini | Wilson Sonsini Advises Recursive on $650 Million Series A Funding at a $4.65 Billion Valuation | Recursive ... announced that it raised over $650 million in its Series A round at a $4.65 billion valuation. |
| SO013 | South China Morning Post | Ex-Meta Chinese star joins race for self-improving AI with US$4.6b start-up | Tian Yuandong ... launched Recursive Superintelligence alongside seven other co-founders. |
| SO014 | Foundra | The 8-Cofounder Cap Table: What Recursive Superintelligence's $650M Round Says to First-Time Founders | There is no product yet. There is no revenue. There is, however, a roster that any AI partner at a top-five firm can read in 90 seconds and underwrite. |
| SO015 | Europe Alternatives | Recursive Superintelligence raises $650M Seed | The company has stated that a public launch is targeted for mid-2026. |
| SO016 | GitHub | GitHub - recursive-org/first-steps-toward-automated-ai-research: Research artifacts from Recursive's automated AI research system | Research artifacts from Recursive's automated AI research system. |
| SO017 | GitHub | GitHub - karpathy/autoresearch: AI agents running research on single-GPU nanochat training automatically | AI agents running research on single-GPU nanochat training automatically. |
| SO018 | GitHub | GitHub - KellerJordan/modded-nanogpt: NanoGPT (124M) in 90 seconds | This repository hosts the NanoGPT speedrun. |
| SO019 | NVIDIA Research | SOL-ExecBench | GPU Kernel Performance Benchmarks by NVIDIA | GPU Kernel Performance Benchmarks by NVIDIA. |
| SO020 | arXiv | NorMuon: Making Muon more efficient and scalable | NorMuon consistently outperforms both Adam and Muon, achieving 21.74% better training efficiency than Adam. |
| SO021 | arXiv | Conditional Memory via Scalable Lookup: A New Axis of Sparsity for Large Language Models | We introduce conditional memory as a complementary sparsity axis. |
| SO022 | Ensue | Ensue | |
| SO023 | GitHub | autoresearch-h100 · recursive-org/first-steps-toward-automated-ai-research | |
| SO024 | GitHub | nanogpt-speedrun-h100 · recursive-org/first-steps-toward-automated-ai-research | |
| SO025 | GitHub | sol-execbench-b200 · recursive-org/first-steps-toward-automated-ai-research | |
| SM001 | Recursive | Recursive | |
| SM002 | Recursive | First Steps Toward Automated AI Research - Recursive | |
| SM003 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | |
| SM004 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | |
| SM005 | Stanford HAI | Economy | The 2026 AI Index Report | |
| SM006 | Gartner | Gartner Forecasts Worldwide AI Spending to Grow 47% in 2026 | |
| SM007 | Gartner | Gartner Says Worldwide AI Spending Will Total $2.5 Trillion in 2026 | |
| SM008 | McKinsey | Recalibrating technology budgets for the AI era | |
| SM009 | Deloitte | The State of AI in the Enterprise - 2026 AI report | |
| SM010 | IDC | IDC's Global Outlook on AI and Generative AI Spending - Use Case Insights | |
| SM011 | CNBC | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency | |
| SM012 | OpenAI | Business Pricing | |
| SM013 | Anthropic | Pricing | |
| SM014 | Google Cloud | Agent Search on Gemini Enterprise Agent Platform | |
| SM015 | AWS | Amazon Bedrock – Build genAI applications and agents at production scale – AWS | |
| SM016 | Microsoft Azure | Azure AI Search | Microsoft Azure | |
| SM017 | AWS | Amazon Bedrock Pricing – AWS | |
| SM018 | Microsoft Azure | Azure AI Search pricing | |
| SM019 | Google Cloud | Generative AI App Builder pricing | |
| SM020 | Brave | Search API | |
| SM021 | OpenAI | Introducing GPT-OSS | |
| SM022 | Anthropic | Claude Code on Team and Enterprise | |
| SM023 | GitHub | Auto model selection now routes based on your task in VS Code | |
| SM024 | CNBC | Model routing on AI is a problem for OpenAI and Anthropic | |
| SM025 | CNBC | Anthropic Mythos, Claude Fable 5 | |
| SP001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SP002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results. |
| SP003 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | Recursive Superintelligence emerges from stealth with $650M raise. |
| SP004 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | Recursive Superintelligence is four months old, has fewer than 30 employees, and has not released a product. |
| SP005 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | |
| SP006 | OpenAI | Business Pricing | Usage analytics, budgeting, and spend controls. |
| SP007 | OpenAI | ChatGPT Enterprise | Deploy enterprise-grade ChatGPT, powered by OpenAI’s smartest models and agents including ChatGPT Work and Codex. |
| SP008 | Anthropic | Pricing | Pricing |
| SP009 | Anthropic | Claude Code and new admin controls for business plans | Both Team and Enterprise plans include granular spend caps, self-serve seat management, and Claude Code usage analytics. |
| SP010 | Google Cloud | Agent Search on Gemini Enterprise Agent Platform | Agent Search on Gemini Enterprise Agent Platform. |
| SP011 | AWS | Amazon Bedrock – Build genAI applications and agents at production scale | OpenAI models on Amazon Bedrock expand that choice, so customers can find the right model for every use case, from every leading AI lab. |
| SP012 | Microsoft Azure | Azure AI Search | Azure Cognitive Search is now Foundry IQ ( Azure AI Search). |
| SP013 | Glean | AI Platform for Work | Glean Work AI for Enterprise | Glean builds connectors that capture enterprise signals, search that indexes your data, an Enterprise Graph that maps how everything relates, and enterprise memory that learns your processes. |
| SP014 | Perplexity | Perplexity Enterprise | Perplexity Enterprise |
| SP015 | Meta | Unmatched Performance and Efficiency | Llama 4 | Unmatched Performance and Efficiency | Llama 4 |
| SP016 | DeepSeek | Models & Pricing | DeepSeek API Docs | The prices listed below are in units of per 1M tokens. |
| SP017 | Crunchbase News | AI Lab Ricursive Intelligence Lands $300M Series A At $4B Valuation Less than Two Months After Launch | Ricursive Intelligence ... raised $300 million in a Series A round of funding at a $4 billion valuation. |
| SP018 | Presenc AI | AI Lab Funding Leaderboard 2026 | Foundation-model lab funding hit unprecedented velocity in 2026. |
| SP019 | Sakana AI | Sakana AI | Building Frontier AI in Japan |
| SP020 | Together AI | Together AI | The AI Native Cloud | Accelerate inference, model shaping and pre-training on a research-optimized platform. |
| SP021 | CNBC | Model routing is a fix for AI overspending. That’s a problem for OpenAI and Anthropic | Companies are shifting from running everything on the most powerful AI model to matching each task to the right one. |
| SP022 | CNBC | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency. |
| SP023 | Gartner | Gartner forecasts worldwide AI spending to grow 47 percent in 2026 | Enterprises will expand their use of both the GenAI models embedded in existing software applications and the new AI agents within multiple workflows. |
| SP024 | Deloitte | State of AI in the Enterprise | |
| SP025 | Stanford HAI | AI Index 2026 Economy | |
| SI001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SI002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results. |
| SI003 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | Recursive Superintelligence emerges from stealth with $650M raise. |
| SI004 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | Recursive Superintelligence is four months old, has fewer than 30 employees, and has not released a product. |
| SI005 | OfficeChai | Recursive Raises $650 Million At $4.65 Billion Valuation To Create Self-Improving AI | The funding gives the team runway to secure large compute clusters and run what they’re calling their first “Level 1” autonomous training run. |
| SI006 | Tech Funding News | UK AI startup Recursive hits $4.65B valuation with $650M raise from Nvidia and GV | The company plans a public launch in mid-2026 as it scales compute infrastructure and research operations across San Francisco and London. |
| SI007 | Lab Index | Recursive | Frontier AI research lab pursuing recursive self-improvement. |
| SI008 | Wilson Sonsini | Wilson Sonsini Advises Recursive on $650 Million Series A Funding at a $4.65 Billion Valuation | Recursive ... raised over $650 million in its Series A round at a $4.65 billion valuation. |
| SI009 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | |
| SI010 | Foundra | The 8-Cofounder Cap Table: What Recursive Superintelligence’s $650M Round Says to First-Time Founders | There is no product yet. There is no revenue. |
| SI011 | GitHub | GitHub - recursive-org/first-steps-toward-automated-ai-research | This repo collects the training scripts and kernel implementations discovered by Recursive’s automated AI-research system. |
| SI012 | GitHub | first-steps-toward-automated-ai-research/nanoGPT_speedrun | Best solution ... reached 77.3 s on 8×H100. |
| SI013 | GitHub | first-steps-toward-automated-ai-research/nanochat_autoresearch | Every solution is a single self-contained training script trained for 5 minutes on a single Modal B200 GPU. |
| SI014 | GitHub | first-steps-toward-automated-ai-research/SOL-ExecBench | 10 of the 235 GPU kernel implementations produced by Recursive’s automated AI-research system ... shared as illustrative examples. |
| SI015 | OpenAI | Business Pricing | 20 / user / month |
| SI016 | OpenAI | New usage analytics and updated spend controls for enterprises | These capabilities help companies track credit usage, understand adoption patterns, and make more informed decisions about how AI is deployed. |
| SI017 | Anthropic | Pricing | Pricing |
| SI018 | Anthropic | Claude Code and new admin controls for business plans | Both Team and Enterprise plans include granular spend caps, self-serve seat management, and Claude Code usage analytics. |
| SI019 | Gartner | Gartner forecasts worldwide AI spending to grow 47 percent in 2026 | The need for capacity will make AI infrastructure ... the largest segment of the market. |
| SI020 | McKinsey | Recalibrating technology budgets for the AI era | AI is gobbling up to a third of companies’ change budgets while adding to run costs. |
| SI021 | CNBC | Model routing is a fix for AI overspending. That’s a problem for OpenAI and Anthropic | Companies are shifting from running everything on the most powerful AI model to matching each task to the right one. |
| SI022 | CNBC | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency. |
| SI023 | AWS | Amazon Bedrock | OpenAI models on Amazon Bedrock expand that choice. |
| SI024 | Microsoft Azure | Azure AI Search | Azure Cognitive Search is now Foundry IQ ( Azure AI Search). |
| SI025 | Recursive | Privacy Policy - Recursive | |
| SI026 | Modal | Plan Pricing | Plan Pricing |
| SI027 | NVIDIA | NVIDIA H100 GPU | NVIDIA H100 GPU |
| SI028 | Ensue | Ensue | Ensue |
| SI029 | Modal | GPU acceleration | GPU acceleration |
| SI030 | NVIDIA | NVIDIA GB200 NVL72 | NVIDIA GB200 NVL72 |
| SE001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SE002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results. |
| SE003 | GitHub | GitHub - recursive-org/first-steps-toward-automated-ai-research | This repo collects the training scripts and kernel implementations discovered by Recursive’s automated AI-research system. |
| SE004 | GitHub | first-steps-toward-automated-ai-research/README.md at main | First Steps Toward Automated AI Research |
| SE005 | GitHub | first-steps-toward-automated-ai-research/nanoGPT_speedrun | train GPT-2-small tier model to ≤ 3.28 FineWeb validation loss on 8×H100, as fast as possible |
| SE006 | GitHub | first-steps-toward-automated-ai-research/nanoGPT_speedrun/from_best | Best solution ... reached 77.3 s. |
| SE007 | GitHub | first-steps-toward-automated-ai-research/nanoGPT_speedrun/from_unoptimized | Solution discovered from a weak ~15-minute baseline ... reaching ≈ 185 s. |
| SE008 | GitHub | first-steps-toward-automated-ai-research/nanochat_autoresearch | Every solution is a single self-contained training script trained for 5 minutes on a single Modal B200 GPU. |
| SE009 | GitHub | first-steps-toward-automated-ai-research/nanochat_autoresearch/solutions | optimized_from_karpathy.py |
| SE010 | GitHub | first-steps-toward-automated-ai-research/SOL-ExecBench | 10 of the 235 GPU kernel implementations produced by Recursive’s automated AI-research system |
| SE011 | GitHub | first-steps-toward-automated-ai-research/NOTICE at main | See NOTICE for the full attribution. |
| SE012 | GitHub | GitHub - karpathy/autoresearch | AI agents running research on single-GPU nanochat training automatically. |
| SE013 | GitHub | autoresearch/train.py at master | autoresearch/train.py at master |
| SE014 | GitHub | GitHub - karpathy/nanochat | The best ChatGPT that $100 can buy. |
| SE015 | GitHub | GitHub - KellerJordan/modded-nanogpt | NanoGPT (124M) in 90 seconds |
| SE016 | GitHub | modded-nanogpt/records at master | modded-nanogpt/records at master |
| SE017 | NVIDIA Research | SOL-ExecBench | SOL-ExecBench |
| SE018 | Ensue | Ensue | Ensue |
| SE019 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | Recursive Superintelligence emerges from stealth with $650M raise. |
| SE020 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | The company has not released a product. |
| SE021 | OfficeChai | Recursive Raises $650 Million At $4.65 Billion Valuation To Create Self-Improving AI | The funding gives the team runway to secure large compute clusters and run what they’re calling their first “Level 1” autonomous training run. |
| SE022 | Lab Index | Recursive | Frontier AI research lab pursuing recursive self-improvement. |
| SE023 | Wilson Sonsini | Wilson Sonsini Advises Recursive on $650 Million Series A Funding at a $4.65 Billion Valuation | Recursive ... raised over $650 million in its Series A round at a $4.65 billion valuation. |
| SE024 | X | Recursive (@Recursive_SI) on X | Recursive self-improving superintelligence to automate knowledge discovery. |
| SE025 | Recursive | Privacy Policy - Recursive | |
| SU001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SU002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results. |
| SU003 | X | Recursive (@Recursive_SI) on X | Recursive self-improving superintelligence to automate knowledge discovery. |
| SU004 | GitHub | GitHub - recursive-org/first-steps-toward-automated-ai-research | Research artifacts from Recursive’s automated AI research system. |
| SU005 | GitHub | Forks · recursive-org/first-steps-toward-automated-ai-research | Fork 15 / Star 174 |
| SU006 | GitHub | Contributors to recursive-org/first-steps-toward-automated-ai-research | Contributions per week to main, excluding merge commits |
| SU007 | GitHub | Issues · recursive-org/first-steps-toward-automated-ai-research | #1 ... opened on Jun 11, 2026 |
| SU008 | GitHub | Pull requests · recursive-org/first-steps-toward-automated-ai-research | 0 Open / 0 Closed |
| SU009 | GitHub | Commits · recursive-org/first-steps-toward-automated-ai-research | Commits · recursive-org/first-steps-toward-automated-ai-research |
| SU010 | GitHub | Releases · recursive-org/first-steps-toward-automated-ai-research | There aren’t any releases here |
| SU011 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | Recursive Superintelligence emerges from stealth with $650M raise. |
| SU012 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | The company has not released a product. |
| SU013 | OfficeChai | Recursive Raises $650 Million At $4.65 Billion Valuation To Create Self-Improving AI | The funding gives the team runway to secure large compute clusters. |
| SU014 | Tech Funding News | UK AI startup Recursive hits $4.65B valuation with $650M raise from Nvidia and GV | The company plans a public launch in mid-2026. |
| SU015 | Lab Index | Recursive | Frontier AI research lab pursuing recursive self-improvement. |
| SU016 | Wilson Sonsini | Wilson Sonsini Advises Recursive on $650 Million Series A Funding at a $4.65 Billion Valuation | Raised over $650 million in its Series A round. |
| SU017 | Glean | AI Platform for Work | Glean Work AI for Enterprise | The horizontal AI platform for enterprise superintelligence |
| SU018 | Google Cloud | Agent Search on Gemini Enterprise Agent Platform | Agent Search on Gemini Enterprise Agent Platform |
| SU019 | Microsoft Azure | Azure AI Search | Azure AI Search |
| SU020 | AWS | Amazon Bedrock | OpenAI models on Amazon Bedrock expand that choice. |
| SU021 | CNBC | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency | Users shift from tokenmaxxing to efficiency. |
| SU022 | CNBC | Model routing is a fix for AI overspending. That’s a problem for OpenAI and Anthropic | Companies are shifting from running everything on the most powerful AI model to matching each task to the right one. |
| SU023 | OpenAI | ChatGPT Enterprise | Frontier AI built for enterprise |
| SU024 | Anthropic | Claude Code and new admin controls for business plans | Granular spend caps, self-serve seat management, and Claude Code usage analytics. |
| SU025 | Perplexity | Perplexity Enterprise | Perplexity Enterprise |
| SU026 | GitHub | Tags · recursive-org/first-steps-toward-automated-ai-research | There aren’t any releases here |
| SU027 | GitHub | Pulse · recursive-org/first-steps-toward-automated-ai-research | Pulse · recursive-org/first-steps-toward-automated-ai-research |
| SU028 | GitHub | Community Standards · recursive-org/first-steps-toward-automated-ai-research | Community Standards · recursive-org/first-steps-toward-automated-ai-research |
| SR001 | Recursive | Recursive | The fastest path to superintelligence will be realized by AI that recursively improves itself. |
| SR002 | Recursive | First Steps Toward Automated AI Research - Recursive | Across three benchmarks, the system achieves state-of-the-art results. |
| SR003 | Recursive | Privacy Policy - Recursive | |
| SR004 | Recursive | Terms of Use - Recursive | |
| SR005 | The Next Web | A four-month-old startup just raised $650 million to build AI that improves itself | The company has not released a product. |
| SR006 | Foundra | The 8-Cofounder Cap Table: What Recursive Superintelligence’s $650M Round Says to First-Time Founders | There is no product yet. There is no revenue. |
| SR007 | Tech.eu | Recursive Superintelligence emerges from stealth with $650M raise | Recursive Superintelligence emerges from stealth with $650M raise. |
| SR008 | Wilson Sonsini | Wilson Sonsini Advises Recursive on $650 Million Series A Funding at a $4.65 Billion Valuation | Raised over $650 million in its Series A round at a $4.65 billion valuation. |
| SR009 | GitHub | GitHub - recursive-org/first-steps-toward-automated-ai-research | Research artifacts from Recursive’s automated AI research system. |
| SR010 | GitHub | Issues · recursive-org/first-steps-toward-automated-ai-research | #1 ... opened on Jun 11, 2026 |
| SR011 | GitHub | Actions · recursive-org/first-steps-toward-automated-ai-research | Actions · recursive-org/first-steps-toward-automated-ai-research |
| SR012 | GitHub | Security · recursive-org/first-steps-toward-automated-ai-research | Build software better, together |
| SR013 | GitHub | Branches · recursive-org/first-steps-toward-automated-ai-research | Branches · recursive-org/first-steps-toward-automated-ai-research |
| SR014 | GitHub | Labels · recursive-org/first-steps-toward-automated-ai-research | Labels · recursive-org/first-steps-toward-automated-ai-research |
| SR015 | GitHub | Milestones · recursive-org/first-steps-toward-automated-ai-research | Milestones · recursive-org/first-steps-toward-automated-ai-research |
| SR016 | GitHub | Code frequency · recursive-org/first-steps-toward-automated-ai-research | Code frequency · recursive-org/first-steps-toward-automated-ai-research |
| SR017 | GitHub | first-steps-toward-automated-ai-research/LICENSE at main | Apache License, Version 2.0 |
| SR018 | GitHub | nanoGPT_speedrun/LICENSE-modded-nanogpt | LICENSE-modded-nanogpt |
| SR019 | GitHub | nanochat_autoresearch/LICENSE-nanochat | LICENSE-nanochat |
| SR020 | NVIDIA Research | SOL-ExecBench | SOL-ExecBench |
| SR021 | CNBC | Model routing is a fix for AI overspending. That’s a problem for OpenAI and Anthropic | Companies are shifting from running everything on the most powerful AI model to matching each task to the right one. |
| SR022 | CNBC | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency | OpenAI and Anthropic face new AI reality as users shift from tokenmaxxing to efficiency. |
| SR023 | Glean | AI Platform for Work | Glean Work AI for Enterprise | The horizontal AI platform for enterprise superintelligence |
| SR024 | OpenAI | ChatGPT Enterprise | Frontier AI built for enterprise |
| SR025 | Anthropic | Claude Code and new admin controls for business plans | Granular spend caps, self-serve seat management, and Claude Code usage analytics. |
| SR026 | GitHub | Tags · recursive-org/first-steps-toward-automated-ai-research | There aren’t any releases here |
| SR027 | GitHub | Pulse · recursive-org/first-steps-toward-automated-ai-research | Pulse · recursive-org/first-steps-toward-automated-ai-research |
| SR028 | GitHub | Community Standards · recursive-org/first-steps-toward-automated-ai-research | Community Standards · recursive-org/first-steps-toward-automated-ai-research |
| SR029 | GitHub | Actions · recursive-org/first-steps-toward-automated-ai-research | Actions · recursive-org/first-steps-toward-automated-ai-research |
| SR030 | GitHub | Milestones · recursive-org/first-steps-toward-automated-ai-research | Milestones · recursive-org/first-steps-toward-automated-ai-research |
| SV001 | Recursive | Recursive | We are building AI that recursively improves AI, with the ultimate goal of automating all of scientific research. |
| SV002 | Recursive | First steps toward automated AI research | We are releasing a substantial fraction of the code and artifacts behind our first experiments in automated AI research. |
| SV003 | Wilson Sonsini | Wilson Sonsini advises Recursive on $650 million Series A funding at a $4.65 billion valuation | Recursive ... raised over $650 million in its Series A round at a $4.65 billion valuation. |
| SV004 | TechCrunch | Almost 40 new unicorns have been minted so far this year — here they are | Recursive — a startup aiming to automate all of scientific research — raised a $650 million Series A in May at a $4.65 billion valuation. |
| SV005 | OfficeChai | Recursive Raises $650 Million At $4.65 Billion Valuation To Create Self-Improving AI | Recursive says it wants to create self-improving AI. |
| SV006 | Foundra | The 8-Cofounder Cap Table: What Recursive Superintelligence’s $650M Round Says to First-Time Founders | There is no product yet. There is no revenue. |
| SV007 | South China Morning Post | Ex-Meta, Chinese star researcher joins race for self-improving AI in US$4.6b start-up | The US start-up was founded last year and emerged from stealth in May with a US$4.65 billion valuation. |
| SV008 | Lab Index | Recursive | Recursive is a research lab building AI that recursively improves AI. |
| SV009 | Stanford HAI | The 2026 AI Index Report — Economy | Private AI investment climbed meaningfully in 2025 and frontier-model economics remained concentrated among large players. |
| SV010 | Gartner | Gartner forecasts worldwide AI spending to grow 47% in 2026 | Worldwide AI spending is forecast to grow 47% in 2026. |
| SV011 | McKinsey | Recalibrating technology budgets for the AI era | AI is consuming technology budgets and forcing leaders to reallocate spend. |
| SV012 | CNBC | OpenAI and Anthropic face a new AI spending reality as users shift to efficiency | Users are shifting to efficiency as AI spending matures. |
| SV013 | Anthropic | Anthropic raises $65B in Series H funding at $965B post-money valuation | Anthropic has raised $65 billion in Series H funding ... valuing the company at $965 billion post-money. |
| SV014 | Anthropic | Claude Code and new admin controls for business plans | Team and Enterprise plans include granular spend caps, self-serve seat management, and Claude Code usage analytics. |
| SV015 | CompaniesMarketCap | Palantir (PLTR) - Market capitalization | As of July 2026 Palantir has a market cap of $317.35 Billion USD. |
| SV016 | Palantir | Palantir Artificial Intelligence Platform | Palantir Artificial Intelligence Platform |
| SV017 | CompaniesMarketCap | Snowflake (SNOW) - Market capitalization | As of July 2026 Snowflake has a market cap of $95.31 Billion USD. |
| SV018 | Snowflake | Snowflake AI and ML | You have control over your team’s use of Snowflake AI Features through familiar role-based access control. |
| SV019 | CompaniesMarketCap | ServiceNow (NOW) - Market capitalization | As of July 2026 ServiceNow has a market cap of $102.30 Billion USD. |
| SV020 | CompaniesMarketCap | Microsoft (MSFT) - Market capitalization | As of July 2026 Microsoft has a market cap of $2.899 Trillion USD. |
| SV021 | Microsoft | Microsoft 365 Copilot for Business: Enterprise AI Solutions | Copilot Chat ... includes IT controls and enterprise-grade privacy and security. |
| SV022 | Databricks | Databricks IQ: AI-Driven Analytics for Faster Data Insights | The Databricks Data Intelligence Platform allows your entire organization to use data and AI. |
| SV023 | CompaniesMarketCap | MongoDB (MDB) - Market capitalization | As of July 2026 MongoDB has a market cap of $25.92 Billion USD. |
| SV024 | OpenAI | Business Pricing | 20 / user / month |
| SV025 | OpenAI | New usage analytics and updated spend controls for enterprises | We’re introducing new usage analytics and updated spend controls for enterprises. |
| SV026 | Anthropic | Claude pricing | Claude pricing |
| SV027 | AWS | Amazon Bedrock pricing | Amazon Bedrock pricing |
| SV028 | Google Cloud | Agent Search | Agent Search |
| SV029 | Stanford HAI | Science | The 2026 AI Index Report | On end-to-end scientific research tasks, the best AI agents score roughly half of what PhD experts achieve. |
| SV030 | Presenc AI | AI Lab Funding Leaderboard 2026 | AI lab funding leaderboard 2026 |