初创公司尽调
尽调报告 AI search infrastructure / agent retrieval infrastructure Series C 2026-06-23

Exa

Exa 尽调报告

Exa 看起来是一家真实且有差异化的智能体搜索基础设施公司,但公开经济指标和当前 $2.2 billion 价格仍太不透明,尚不足以支持立即投资。

封面要素

成立时间 01
2021 [CO001]
总部 02
San Francisco [CO001]
最新估值 03
2200 USD M [CV001]
累计融资 04
357 USD M [CV015]
公司数(公司口径) 05
5000 companies+ [CV003]
开发者数(公司口径) 06
400000 developers+ [CV003]

公司概况

Exa 是一家位于旧金山的私营 AI 搜索基础设施公司,由 Will Bryk 和 Jeff Wang 于 2021 年创立。公司把自己定位为面向 AI agent 的定制搜索引擎,并已从搜索扩展到内容抽取、答案生成、公司搜索、监控和 agent 化研究流程。公开证据支持其真实采用和强融资动能,包括以 $2.2B 估值完成 $250M Series C,但经审计财务、治理深度、留存和精确运营规模仍披露有限。

官网
exa.ai
创始人
Will Bryk, Jeff Wang
创立地点
San Francisco
总部
San Francisco
产品
Web 搜索、内容抽取、有依据的答案、公司搜索、监控、Websets 以及 agent 式研究 API,底层依托 Exa 自有的爬取、索引、嵌入、向量和编排栈。
客户
AI 原生开发者、编码助手、研究工具、GTM 软件,以及把实时公开 Web 检索嵌入 agent 工作流的企业团队。
商业模式
搜索、内容、答案和 agent 工作流按用量计费;企业合同增加更高限额、协商账单和零数据留存等安全控制。
阶段
Series C
融资情况
最新公开融资为 2026 年 5 月以 $2.2B 估值完成的 $250M Series C,此前还有 2024 年 Series A 和 2025 年 Series B。
[CO001, CO006, CO019, CO024, CO025, CE023, CE026, CE027]

执行摘要

主要优势

  • Exa 自建抓取、索引、嵌入、向量和智能体编排栈,而不是在传统搜索外包一层壳,技术差异化看起来有实质内容。
  • 公开采用信号很强:5,000 多家公司、约 400,000 名开发者,且已嵌入 HubSpot、monday.com、OpenRouter、Cognition 等客户工作流。
  • 2026 年 5 月 $250 million Series C 和一线投资人组合显著降低近期融资风险,也给管理层扩基础设施和 go-to-market 留出空间。

主要风险

  • 公开披露仍不足以支撑承销判断:经审计收入、ARR、毛利率、NRR、烧钱速度、客户集中度和优先股堆叠条款都未披露。
  • 相比最近一次公开收入 proxy,也相比私有与公开市场可比公司,$2.2 billion 估值隐含倍数都极高。
  • 若产品控制和单位经济不能随增长成熟,隐私默认设置、仅企业版零数据留存控制以及重基础设施运营模式,会带来运营和采购风险。

未决问题

  • 董事会批准的 2025 年及当年收入、ARR 和毛利率数据,按主要产品线拆分。
  • 在声称的 5,000 多家公司基础上,前几大客户集中度、续约行为、净收入留存率,以及付费与免费转化。
  • 2026 年 Series C 前后的当前股权结构表、清算优先权、老股流动性、债务和期权池条款。
  • 准确员工数、创始人以下管理层厚度,以及敏感工作负载使用的企业 DPA 或默认留存配置。

目录

Chapter 01

01公司概览

1.1 身份、改名与运营模式

Exa 是一家位于旧金山的私营 AI 搜索基础设施公司,成立于 2021 年,最初以 Metaphor 名义推出,2024 年 1 月公开更名为 Exa。这次改名重要,因为公司叙事从「更好的搜索」转向更大的使命:为 AI 系统组织世界知识。当前产品不是面向消费者的搜索入口,而是 agent 的搜索与检索层:Exa 通过 API 和开发者工具销售 Web 搜索、内容抽取、答案生成、监控和异步研究工作流。 运营模式是按用量计费的基础设施。公开价格显示,搜索档位、页面内容检索和 agent 运行按请求收费;企业方案增加更高限额、Zero Data Retention 等协商控制。官方文档和产品页持续强调延迟、语义检索质量、结构化输出和代码搜索支持,而不是广告。这一点重要,因为它把 Exa 定位为 agent 栈里的「卖铲人」,不是媒体或 SEO 生意。官方法律页面也把总部锚定在旧金山;公司材料则描述全球招聘足迹,而不是一个规模庞大、逐一列明的办公室网络。[CO001, CO002, CO003, CO004, CO005, CO037]

FO002: Exa — 公司快照逻辑

Exa 模型里,创始人、检索基础设施、开发者分发、客户以及隐私 / 治理约束如何彼此连接。

仅表示概念流,并不意味着正式控制权,也不量化具名客户中的使用集中度。

[CO003, CO004, CO009, CO024, CO025, CO027]

1.2 创始人、领导层与治理

公开的创始人故事连贯,而且与 Exa 的产品 thesis 异常契合。Will Bryk 是联合创始人兼 CEO,拥有哈佛计算机科学和物理学背景,曾在 Cresta 做工程。Jeff Wang 在哈佛学习计算机科学和哲学,之后在 Plaid 做数据和 Web 基础设施。官方和投资人材料都把两人呈现为长期合作者,在创办 Exa 前已经一起做过搜索项目。这让公司在检索质量、系统工程和开发者工作流设计交汇的产品上具备很强的 founder-market fit。 治理图景要薄得多。公开来源明确显示 Peter Fenton 在 2025 年 9 月 Series B 后加入董事会、Sarah Wang 在 2026 年 5 月 Series C 后加入董事会,但审阅过的官方页面没有发布完整董事会名单、所有权结构或接班计划。招聘和关于页面展示了一批技术员工,却没有披露创始人之下的完整高管团队。因此,董事会控制权、管理层厚度,以及 Bryk 过高的可见度是否在招聘、融资和产品方向上形成关键人集中,仍是尽调问题。[CO006, CO007, CO008, CO016, CO020, CO023]

领导层与创始人表
人物角色背景创始人-市场匹配 / 职能覆盖关键人物依赖
Will Bryk联合创始人兼 CEOHarvard 计算机科学 + 物理;Cresta 早期工程师搜索 / 基础设施创始人匹配强;掌握公开叙事、融资和产品愿景关键——公开面孔和论点承载者
Jeff Wang联合创始人Harvard 计算机科学 + 哲学;曾在 Plaid 做数据和网络基础设施工作系统和开发者基础设施匹配强;在数据 / 网络架构上补足 Bryk高——深度产品 / 基础设施语境集中在创始人组合
Peter Fenton董事会成员(自 Series B 起)Benchmark 合伙人;投资人材料称其曾有重要搜索公司董事会经历2025 年拐点后,增加后期治理和融资模式识别能力中等——外部治理信号,不是运营依赖
Sarah Wang董事会成员(随 Series C 报道)a16z 增长投资人,关注 AI 和基础设施2026 轮后,为 agent 生态带来分发和成长期网络中等——战略网络有价值,但公开范围未完全披露

公开领导层证据偏创始人;本表包括公开点名的新增董事,但不包括完整高管组织架构图。

[CO006, CO007, CO008, CO016, CO020, CO023]

1.3 融资历史与估值路径

Exa 的资本历史像快速上楼梯。2024 年 7 月,公司宣布完成 $22M 种子轮加 Series A,其中 Series A 部分报道为 $17M,由 Lightspeed 领投,NVentures 和 Y Combinator 参投。2025 年 9 月,Exa 以 $700M 估值完成 $85M Series B,由 Benchmark 领投;这一轮也正式确立 Peter Fenton 的董事席位,并资助更大的 GPU 集群和更广泛招聘。2026 年 5 月,Exa 宣布以 $2.2B 估值完成 $250M Series C,由 Andreessen Horowitz 领投,老股东 Benchmark、Lightspeed 和 Y Combinator 也参与。 这条路径意味着截至 2026 年 5 月,已披露累计融资约 $357M;从 Series B 到 Series C 不到一年,估值上调略高于 3 倍。融资轨迹亮眼,但公开市场数据服务跟进不均:Sacra 仍只显示 $107M 融资,因为其快照停在 Series B,说明高速发展的私营公司很容易让第三方数据过时。公开来源也未披露债务、二级流动性、优先权栈、pro-rata 权利或完整股权表。因此,即便头部轮次序列表面动能强,治理和经济条款仍需要私下确认。[CO013, CO014, CO015, CO016, CO017, CO019]

利益相关方或投资人地图
利益相关方角色控制 / 经济重要性尽调问题
Lightspeed Venture PartnersSeries A 领投方;Series B 跟投方最早被点名的机构领投方,且持续供给资本确认 Series C 后的当前持股、pro-rata 权利和任何董事会观察员权利
Y Combinator加速器和多轮投资人最早公开平台支持者;强公司形成信号核验 seed 持股、SAFE 转换条款和任何信息权
NVentures (NVIDIA)Series A 和 Series B 投资人基础设施重公司中的战略算力对齐投资人检查是否存在任何商业、硬件供应或优先合作伙伴安排
Benchmark / Peter FentonSeries B 领投方和董事席位持有人在 2025 年规模化融资轮引入正式、懂搜索的董事会监督审阅 Series B 文件中的董事会同意事项、否决权和清算优先权结构
Andreessen Horowitz / Sarah WangSeries C 领投方和报道中的董事席位持有人牵头 2026 年估值重置,并很可能塑造当前治理确认当前董事会构成、保护性条款,以及进入投资组合公司的任何分发合作
开发者和 agent 生态合作伙伴客户与分发层,而非股权持有人LangChain、MCP clients、Browserbase 和具名应用构建者提高潜在转换成本衡量用量多元化程度,而不是集中在少数设计伙伴

映射最可见的公开利益相关方,不是完整 cap table;经济性、持股比例和优先权条款均为私有信息。

[CO013, CO015, CO016, CO019, CO020, CO023]

1.4 公开规模标记、客户与生态信号

最强的公开牵引力标记来自公司口径和生态,而不是经审计财务。Exa 称其为 Cursor、Cognition、HubSpot、OpenRouter、Monday.com 等提供搜索能力,并服务超过 400,000 名开发者;Series C 文章称平台现在服务超过 5,000 家公司。独立合作伙伴信号进一步说明 Exa 已嵌入当前 agent 工具链:LangChain 维护 Python 和 JavaScript 两套集成;Exa 发布自有 MCP server 指南,覆盖 Claude、Cursor、VS Code、Codex 和 Gemini CLI;Browserbase 提供基于 Exa 的求职搜索自动化模板。TechCrunch 还提到 Databricks 是用 Exa 做数据集发现的客户。 一些规模标记仍软得多。公开员工数相互冲突很大,从 Y Combinator 公司资料的 75 人,到 2026 年 5 月报道中的约 100 人,再到 Growjo 的 294 人。收入更不可靠:Sacra 估算 2025 年约 $10M,但在公司自有材料中没有找到官方经审计收入或 ARR 披露。同样,公开记录确认了旧金山总部和全球招聘姿态,却没有给出稳定二级办公室的清单。正确解读是,客户和开发者采用看起来真实,但多个封面指标在承担估值权重之前仍需私下核验。[CO024, CO025, CO026, CO027, CO028, CO030]

快照 KPI 表
指标数值 / 状态日期置信度缺口
最新估值$2.2B Series C2026-05-20标题融资价值,但没有公开优先股结构细节
已披露总融资额公开轮次隐含约 $357M2026-05-20取决于把 seed 视为约 $5M,并排除任何未披露债务或二级流动性
估计收入 / 运行收入约 $10M 收入估计2025-09仅 Sacra 估计;公司未公开经审计收入或 ARR
公司客户5,000+2026-05-20公司声称的数量;未披露 cohort 或留存拆分
开发者400,000+2026-05-20公司声称的数量;未拆分活跃与注册开发者
员工数公开引用为 75 至 2942026-06公开数据集分歧很大;需要 HRIS 或薪资确认
地点San Francisco 总部已确认;更广泛布局不清楚2026-06官方页面支持全球招聘,但未给出经核验的第二办公地点清单

混合了官方声称与第三方估计;围绕经审计收入、准确员工数和第二地点,仍有 null 质量缺口。

[CO019, CO021, CO024, CO025, CO030, CO031]
FO003: Exa — 公开证据质量记分卡

KPI 式展示:在当前公司概览材料中,公开证据哪里最强、哪里最弱。

[CO015, CO019, CO023, CO024, CO025, CO033]

1.5 里程碑与不利背景

Exa 的里程碑显示,公司以罕见速度从搜索引擎 R&D 转向 agent 基础设施:2022 年 11 月首次推出搜索引擎,2023 年初聚焦 API,2024 年 1 月更名为 Exa 并发布 Highlights,2025 年 9 月完成制度化的 Series B,2026 年 3 月发布改版 Exa Deep,2026 年 5 月完成 $2.2B 估值的 Series C。合在一起,这些事件支持一个判断:管理层反复把公司推向价值最高的 agent 工作流,而不是停留在狭窄的语义搜索利基里。 不利背景更微妙,但有实质影响。第一,Exa 自建爬取、索引和 GPU 栈,经济模型偏重基础设施;即便增长故事为正,仍可能持续烧钱。第二,Sacra 的竞争图谱突出了来自 OpenAI、Anthropic、Google、Brave、Tavily、Jina 和 Perplexity 的压力,提醒投资人这个品类正处于前沿模型和相邻搜索 API 的爆炸半径内。第三,隐私政策说明查询数据可用于改进和微调 Exa 模型,而 Zero Data Retention 被营销为企业专属功能。这不意味着不当行为,但它确实是敏感工作负载的尽调问题,也强化了检查企业合同条款的必要性,而不能默认数据隔离。[CO009, CO010, CO011, CO012, CO018, CO035]

里程碑表
日期事件类型金额 / 估值 / 状态参与方含义
2021公司以 Metaphor 名义成立创立私有公司成立Will Bryk、Jeff Wang在 ChatGPT 重塑市场之前,启动从零做搜索的努力
2021Y Combinator 支持和早期 seed 资本成为公司形成故事的一部分融资后续轮次数学隐含约 $5M seedY Combinator 和早期支持者在 Series A 之前提供第一批机构支持
2022-11首个 Exa / Metaphor 搜索引擎发布产品公开产品发布创始团队在 API 中心转向之前,建立核心检索引擎
2023-early公司公开重点转向 AI 搜索 API产品面向 AI 的首个网络搜索 API创始团队和早期开发者客户将 Exa 重新指向 agent 和 LLM 工作流
2024-01-25Metaphor 更名为 Exa,并推出 Highlights治理品牌重置加新功能Exa 团队让使命更清晰,并加入抽取式检索工具
2024-07-16Seed + Series A 公布融资总计 $22M;$17M Series ALightspeed、NVentures、Y Combinator资助模型开发、招聘和早期市场扩张
2025-09-03Series B 公布;Peter Fenton 加入董事会融资$85M,估值 $700MBenchmark、Lightspeed、NVentures、Y Combinator 等投资方标志首个重大估值跃升和治理专业化
2025-09Websets 作为招聘和市场研究 traction 产品被重点展示产品面向人的结构化搜索用例获得 tractionExa、Lightspeed、客户显示从纯 API 使用延伸到结构化搜索工作流
2026-03-04改版 Exa Deep 发布产品结构化输出、grounded citations、更低定价Exa 产品和研究团队从检索扩展到更深 agent 任务的多步骤综合
2026-05-20Series C 公布;报道称 Sarah Wang 将加入董事会融资$250M,估值 $2.2Ba16z、Benchmark、Lightspeed、Y Combinator 等投资方不到一年估值翻三倍,并资助下一代模型 / 基础设施扩张
2026-06隐私政策确认查询数据训练表述,同时企业页面营销 Zero Data Retention反向政策限制仍有效Exa Labs对假定默认保留限制的敏感工作负载,形成尽调问题

时间线仅基于公开记录;内部治理、未披露客户赢单和未公布产品实验可能不在本时间线内。

[CO002, CO010, CO011, CO012, CO013, CO015]
FO001: Exa — 资本与产品拐点时间线

高层次梳理 Exa 从创立、改名、产品发布到融资推进的历程,覆盖至 2026 年 6 月报告生成日。

[CO002, CO010, CO011, CO012, CO013, CO015]

1.6 展示材料

Chapter 02

02市场分析

2.1 市场边界与相邻支出

Exa 所在市场比泛泛的「搜索」甚至泛泛的「AI 基础设施」更窄。其自有材料描述的是面向 agent 的公开 Web 检索栈:搜索、爬取、抽取、带引用答案,以及通过 MCP 分发进开发者工具。这意味着相关支出,是开发者和企业 AI 团队用来把 agent 连接到新鲜公开 Web 信息,并把结果转成可用上下文的钱。几个相邻资金池很诱人,但容易误导。Glean 和 Elasticsearch 卖内部知识与私有数据检索;Algolia 卖电商和站内搜索优化;它们都可作为替代品或比较点,但不能直接衡量 Exa 的核心收入池。Google Custom Search 仍是传统替代品,但 Google 称其已关闭新客户入口,并将在 2027 年 sunset。因此,正确边界是面向 agent 的 AI 原生公开 Web 检索基础设施;内部企业搜索、广告、浏览器订阅和通用向量数据库支出应被排除或作为相邻项,而不是计入核心 TAM。[CM001, CM002, CM003, CM004, CM005, CM006]

市场定义表
细分 / 类别纳入支出排除支出买家 / 付款方相关性
公共网络搜索 API 层为 agent 工作流查询公共网络,提供排序结果、摘要片段和可抽取元数据消费者搜索广告和浏览器订阅收入Agent 构建者、编码工具、AI 产品团队核心市场
事实锚定和答案层带引用答案、抽取、研究运行,以及面向 agents 的更高阶检索工作流与检索无关的独立 LLM 推理支出AI 平台、产品和研究团队核心市场
协议和集成层围绕公共网络检索的 MCP 连接器、框架工具、IDE 集成和部署打包不交付检索本身的通用集成中间件开发者工具团队和企业 AI 平台负责人重要赋能层
企业内部检索跨私有文档、SaaS 系统和权限感知知识库的安全搜索作为服务的鲜活公共网络事实锚定企业 IT 和知识管理负责人相邻但排除
通用搜索 / 向量平台用于自建检索栈的更广泛搜索、分析、向量 DB 或站内搜索平台除非单独打包,否则不计入托管公共网络检索收入平台工程和数据基础设施负责人相邻性和替代压力

边界放在面向 agents 的付费公共网络检索基础设施;相邻内部搜索和泛化检索平台只用于防止 TAM 膨胀。

[CM001, CM003, CM004, CM005, CM006, CM007]
FM001: 面向智能体的公开网页检索价值链

展示从公开网页检索原语,到智能体分发和企业采购界面的商业链条。

[CM002, CM003, CM021, CM022, CM023, CM024]

2.2 规模测算视角与需求信号

公开证据支持多个规模测算视角,但没有一个干净的 TAM。审阅材料中最佳相邻市场研究是 Grand View Research 的企业搜索报告:该报告测算 2023 年市场为 $4.9B,并到 2030 年增长至 $8.9B;这适合作为外层相邻市场,不适合作为 Exa 的 TAM,因为它主要计算内部知识检索和安全企业搜索。直接品类证据反而来自供应商自身的运营数据和单位经济。Exa 定价覆盖每千次搜索类请求 $7 至 $15、每千页内容抽取 $1,以及每次 agent 运行最高 $2。Tavily 和 Brave 都发布了低摩擦入门定价、大规模用量或索引口径;Google 则展示了较老 programmable-search 基线的样子。合在一起,这些视角证明 agent 工作流内的公开 Web 检索存在真实且增长的收入池,但在没有内部合同、用量或 cohort 数据前,分析师无法为 Exa 隔离出一个中立的独立 SAM。[CM009, CM010, CM011, CM012, CM013, CM014]

TAM / SAM / 规模测算视角表
视角数值 / 信号覆盖内容置信度限制
相邻企业搜索市场2023 年 $4.867B;2030 年 $8.852B;8.9% CAGR内部企业搜索和信息访问的外层支出包络高估 Exa,因为多数支出是内部知识检索,不是公共网络 agent 事实锚定
Exa 定价和规模视角每 1k 搜索式请求 $7-$15;每 1k 页面 $1;超过 400k 开发者;超过 5k 公司直接供应商单位经济性和当前采用声称中低自我报告,缺少用量组合或留存数据,无法转换为独立市场收入
Tavily 用量视角300M 月请求;2M+ 开发者;每 credit $0.008竞争证据显示该品类已有开发者需求和用量规模自我报告,且使用不同计费单位
Brave 索引和定价视角30B+ 页面;100M 日更新;每 1k answers $4 或每 1k search requests $5规模、索引成本和品类定价基准的证据供应商声称,并受不同端点打包差异影响
传统可编程搜索视角每天 100 次免费查询;每 1k 查询 $5;服务不再面向新客户开放,并于 2027-01-01 sunset许多团队会迁出的现状基线传统基准,不是前瞻 TAM

多个视角支持需求,但没有一个能隔离 AI-native 公共网络检索基础设施的中性独立 SAM;本表主要输出是有边界的测算逻辑,而非单一 TAM 标题。

[CM009, CM010, CM011, CM012, CM013, CM014]
查询定价与打包基准
供应商 / 产品单价单位打包信号对 Exa 的含义
Exa Search$7每 1,000 次请求面向 agent 循环的原始公共网络搜索锚定低层查询经济性
Exa Deep Search / 深度推理搜索$12 / $15每 1,000 次请求更高价值检索被打包为更深工作流表明编排越深,付费意愿越高
Exa Agent$0.012-$2.00每次运行异步研究和富集工作流定价成果打包能把支出推到简单搜索之上
Brave Search / Answers$5 / 1,000 次搜索请求;$4 / 1,000 次回答 + tokens按请求计费,另加 tokens原始检索和有依据回答分别定价支撑与 Exa 类似的多层定价栈
Tavily PAYG$0.008按 credit 计费面向开发者的入门定价,用 credit 抽象低门槛上手会推高价格透明度压力
Google 自定义搜索 JSON API$5每 1,000 次查询仅面向现有客户的传统可编程搜索基准可作历史参照,但不是好的前瞻代理指标

各厂商计费单位不能直接横比;本表看的是定价结构和工作流定位, 不是干净的同口径毛利率基准。

[CM009, CM013, CM014, CM030, CM032]

2.3 买方、用户、付款方与采用路径

审阅到的工作流证据显示,买方基础由软件团队主导,而不是终端用户搜索部门。OpenAI 的 agent 工具和 MCP 生态都默认开发者、产品团队和企业 AI 平台组会把 Web 检索嵌入更大的 agent 系统。Exa 的 MCP server、Brave 的 skills 仓库、LangChain 的 Tavily 集成,都强化同一模式:第一用户通常是工程师或 agent 构建者,而不是采购主导的知识管理团队。付款方位于用户上一层。早期支出可能来自开发者主导的产品或实验预算;规模化部署更可能落在 CTO、平台、产品或集中式企业 AI 负责人手中,他们关心治理、SLA 和多工具集成。采用通常从编码辅助或研究自动化等狭窄工作流开始,通过更好的 grounding 和更少手动浏览证明价值;等治理、权限和可靠性标准满足后,才扩展到更广的企业合同。[CM019, CM020, CM021, CM022, CM023, CM024]

细分 / 买家地图
细分买家用户付款方工作流预算负责人采用触发器
编码助手供应商产品和工程领导层Agent 工程师和开发者体验团队产品 / 平台预算用实时文档、代码库、changelog 和引用为代码生成做 groundingVP Engineering 或开发者平台负责人在延迟敏感循环中需要当前网络和代码语境
Agent 应用构建者创始人、CTO 或应用 AI 负责人应用工程师和 prompt / agent 设计者核心产品或应用 AI 预算把实时网络检索、抽取和带引用答案加入任务型 agentsCTO 或 AI 产品负责人需要比静态 RAG 或封闭模型记忆更新鲜的答案
企业 AI 平台团队平台或 AI 转型领导层把获批工具集成进企业工作流的内部构建者集中 AI / 平台预算在多个内部 agents 间标准化检索、治理和供应商合同CTO、CIO 或企业 AI 负责人试点用量大到治理和 SLA 变得重要
研究和运营自动化团队知识、研究或运营负责人由 agents 辅助的分析师和运营人员带 AI 覆盖的职能运营预算自动化公共网络研究、监控和答案综合业务单元负责人加 AI 赞助人需要压缩手工浏览和综合时间
内部搜索在位者 / 构建者IT 或知识管理负责人搜索内部系统的员工知识管理或 IT 预算比较采购公共网络检索 API,还是扩展现有内部搜索栈CIO、IT 搜索负责人或知识负责人公共网络问题开始出现在现有企业 AI 助手里

各细分区分检索循环的使用者和最终付款方;规模化部署通常会从开发者主导用量转入集中平台或产品预算。

[CM019, CM020, CM021, CM022, CM023, CM024]
FM002: 从检索试点到生产合同的采用路径

映射智能体构建者通常如何从窄范围试点,推进到受治理的企业部署。

[CM019, CM020, CM025, CM027, CM028, CM029]

2.4 增长驱动、约束与保留的矛盾

最强增长驱动来自生态层面。OpenAI 正把 Web 搜索和 agent 编排变成标准构件,MCP 正降低外部工具集成税,Exa 和 Tavily 等供应商正从原始搜索端点向更丰富的研究工作流上移。与此同时,这仍是一个难以测算和承保的市场。不同供应商的消耗单位差异极大,治理和零数据留存要求会拖慢采购,内部检索栈在许多企业里仍争夺同一个预算负责人。最重要的是,公开材料保留了不应抹平的直接矛盾。Exa 称自己在每个延迟和价格点都是最高质量搜索 API,也是世界最快搜索 API;Brave 称它能替代 Exa 和 Tavily 这类较小索引竞争对手,并且其 grounding 可胜过前沿答案引擎;Tavily 称自家搜索是市场最快。这些矛盾是证据,不是噪音:它们说明尽调应优先看中立工作流 benchmark 和私有 cohort 经济,而不是供应商营销叙事。[CM027, CM028, CM029, CM031, CM032, CM033]

增长驱动与采用约束表
驱动 / 约束方向时点含义尽调问题
主要模型提供商内置 agent 工具正向当前让网页检索成为新 agent 的默认功能,也放大 Exa 这类厂商的分发机会哪些买方工作流最能把实验转成付费生产使用?
MCP 标准化正向当前降低跨客户端集成成本,并借助 IDE、框架和企业连接器拓宽渠道新增管线中有多少来自 MCP 驱动,多少来自直接 API 采用?
向上进入带引用的回答和研究工作流正向当前让厂商在每个任务上拿到比单纯原始搜索更高的价值收入中有多少来自更高阶工作流,多少来自商品化搜索调用?
内部搜索和自建栈竞争负向当前预算归属会变模糊,因为部分买方可以扩展 Glean、Elastic 或内部工具,而不是购买专门的公共网页检索厂商哪些客户群会选择专门的网页检索,而不是扩展内部栈?
安全、治理和采购要求负向当前至中期企业买方在大规模推出前,可能要求 ZDR、权限、可观测性和合同条款Exa 最大客户里,哪些安全和治理能力已成入场门槛?
厂商对质量和速度的自我宣称相互冲突负向当前第三方基准和队列经济性比营销叙事更重要哪个中立基准或客户证据最能预测胜率和留存?

驱动和约束绑定采用时点,而不是抽象市场规模;每一行都用于判断需求何时转成可持续预算。

[CM027, CM028, CM029, CM030, CM031, CM033]
保留证据的矛盾表
议题说法 A说法 B矛盾为何重要当前处理
质量领先Exa 称自己在每个延迟和价格点上都是质量最高的搜索 APIBrave 称其 API 可以替代 Exa 这类索引规模较小的竞争对手,而且更好的依据化胜过前沿回答引擎已审阅语料中没有中立第三方基准能解决这一说法保留为未解决的竞争矛盾
延迟领先Exa 称其打造了全球最快的搜索 API,延迟低于 200msTavily 称其 /search 端点 p50 为 180ms,是市场最快两种说法都来自自报,口径也不同作为特定工作流的尽调项处理,而不是既定事实
市场规模标题企业搜索研究显示相邻市场有数十亿美元规模公共网页检索品类本身更窄,也缺少独立 TAM 研究若把更宽口径报告当作 Exa 的 TAM,会抬高估值框架用多个受约束视角,而不是单一 TAM 标题
商业打包部分厂商按原始搜索请求变现其他厂商主推带引用的回答、研究运行或 credit 抽象计费单位不同,比较和预算都会变乱按工作流和捕获价值比较,不只看单次调用标价

矛盾被保留下来,而不是被抹平;本章把它们视为市场尚不成熟、且需要中立工作流基准的证据。

[CM016, CM017, CM018, CM032, CM034, CM035]

2.5 展示材料

Chapter 03

03竞争对手

3.1 格局:直接 API 同行、有 grounding 能力的巨头、企业搜索替代品与内部自建

Exa 的直接同行集合很窄,但非常真实。Tavily、Brave Search API、SerpAPI 和 Perplexity Sonar 都向开发者销售把当前 Web 上下文带进 AI 产品的方式,而且四者现在都从普通结果扩展到更丰富的 agent 或答案工作流。Tavily 把搜索、抽取、研究和爬取打包进一个 API;Brave 销售独立索引、答案生成和 OpenAI 兼容表面;SerpAPI 竞争点在引擎广度和 wrapper 便利性;Perplexity 则把搜索与答案生成和来源控制打包。 替代品集合比直接同行更宽。Google 的传统 Custom Search API 在收缩,但 Google 同时通过 Grounding、Agent Search 和更广的 Gemini 平台工具抬高门槛。Glean、Algolia 和 Elastic 从企业边界内部切入,而不是从实时公开 Web 索引切入;当预算负责人更重视权限、自有数据或既有部署足迹而不是开放 Web recall 时,它们就是可信替代品。内部自建同样可行,因为 LangChain 这类编排层已经让团队可替换模型和检索后端,不必永远绑定一家垂直整合供应商。[CP001, CP008, CP011, CP012, CP013, CP014]

竞争对手画像表
竞争对手类别规模 / 融资代理指标目标客群差异化局限
Exa直接网页搜索 API + agent 基础设施50M+ 公司索引、公开 SDK/MCP 分发、按量 API 定价AI agent、编码助手、GTM / 市场研究构建者一套栈覆盖实时网页搜索、contents、公司搜索、代码感知工作流和 agent 工具捆绑能力弱于 Google 或 Glean,公开信任能力差距正在缩小
Tavily直接同业首页称月请求量 300M+、开发者 2M+、99.99% 可用性、180 ms p50构建搜索、抽取和研究 agent 的开发者统一搜索 + 抽取 + 研究 + 爬取定位企业胜率和差异化信任范围的公开证据较薄
Brave Search API拥有独立索引的直接同业30B+ 页面独立索引、每日 100M+ 页面更新、50 QPS 搜索AI 搜索、训练数据、RAG、重引用助手独立索引、低公开入门价、OpenAI 兼容回答层、ZDR 声明仍是横向网页索引产品,没有 Exa 的公司搜索垂直能力
SerpAPI直接同业 / 封装器免费至 $275 自助档位,另有多个特定引擎 API 和企业计划希望覆盖广泛引擎、但不想运营 SERP 基础设施的开发者覆盖 Google、Maps、News、Scholar、Shopping、Amazon 等当专有召回质量是关键差异点时,封装器经济性和索引依赖会很重要
Perplexity Sonar带回答层的直接同业原始搜索 $5 / 1k,另加 Sonar 模型定价;TechCrunch 称 Zoom 是早期用户希望在一个界面获得搜索、回答生成和来源控制的开发者一套栈包含 Search API、Sonar、Agent API、MCP 和 OpenAI 兼容信任和原始搜索经济性仍与 Exa 直接竞争,并不能让其免于替代
Google Grounding / Custom Search既有平台捆绑传统 Custom Search 正停止接纳新客户,Grounding 则在 Gemini 栈内扩张已在 Google 上标准化的云和企业买方世界知识依据化、Agent Search、受监管网页依据化和捆绑能力Custom Search 属于传统产品,Grounding 是更宽的平台采购,不是轻量替换
Glean企业搜索替代品首页强调 35+ 模型,以及重点部署案例中 <2 年达 93% 采用率关注内部知识、权限和采用的大型企业权限感知搜索、agent 治理和深层内部系统上下文不是公共网页原生检索产品
Algolia自有内容替代品按请求计价的相关性栈,包含 NeuralSearch、规则、分析、商品运营和爬取摄取优化自有资产上产品或内容发现的团队面向第一方内容的强相关性工具和业务用户控制没有独立公共网页索引,也没有 agent 原生研究楔子
Elastic可自建平台替代品开源部署选项、350+ 集成、99.95% 托管 SLA、高阶 Agent Builder希望完全掌控检索、数据和部署足迹的团队混合检索、向量搜索、多云托管和规模化 agent 功能比采用现成网页搜索 API 需要更多构建工作
用 LangChain + 所选搜索 API 内部自建现状 / 潜在进入路径除所选组件外没有固定厂商加价;编排已经横跨主要模型提供商已拥有 agent 基础设施的强工程团队路由灵活性最高,也更容易在搜索厂商间多家部署集成负担最高,开箱即用的信任、支持和质量保证最弱

各行覆盖截至 2026-06-23 公开证据中,Exa 买方可见的最重要直接同业、既有厂商、替代品和内部自建选项。

[CP001, CP008, CP012, CP013, CP014, CP016]
FP001: 竞争定位图

Exa 在网页原生差异化上位置较高,但在捆绑能力上低于 Google 和 Glean;按现代开发者姿态看,Brave 和 Perplexity 是最接近的直接同业。

分数是综合判断,锚定保留下来的公开产品、定价和信任证据,而不是供应商发布的基准。

[CP012, CP016, CP021, CP024, CP028, CP032]

3.2 能力、定价、GTM 与信任:Exa 的差异化在哪里,平价化又在哪里上升

Exa 的公开产品表面对一家 AI 原生搜索供应商来说异常连贯。搜索、内容、公司 enrichment、MCP 分发和 SDK primitives 都指向同一个购买故事:一家供应商能让 agent 在实时 Web 上完成 grounding,同时暴露面向代码、公司发现和深度研究的结构化工作流。这比单纯搜索 API pitch 更强。不过,一旦纳入定价和信任,比较就变紧。Brave 的答案栈公开、低摩擦且兼容 OpenAI;Perplexity 明示原始搜索和 agent 工具定价;SerpAPI 提供简单自助阶梯;Google 仍列出商品化传统搜索定价,同时把严肃买家迁向 grounding bundles。 信任也不是垄断。Exa 凭借 SOC 2、零留存选项和受限 HIPAA-safe 检索路径,拥有有意义的企业证明。但 Brave、SerpAPI、Glean、Google 和 Elastic 都发布了各自的控制表面、认证或 uptime 承诺。因此采购不像二元的安全审查,而更像在 recall 质量、bundle 契合度、合规范围和运营成本之间做矩阵权衡。结果是,Exa 仍可凭产品连贯性和垂直扩展突出,但团队越来越能在多个可接受替代品之间比较信任、价格和分发,而不是在严肃企业评估中把 Exa 当成唯一的 AI 原生 Web 检索供应商。[CP002, CP003, CP004, CP005, CP006, CP007]

功能 / 能力矩阵
购买标准ExaTavilyBraveSerpAPIPerplexityGoogle / 企业替代品
实时公共网页检索深度
结构化研究 / agent 工作流
代码或公司垂直专长
自有数据 / 权限感知检索
明确的企业信任控制
标准 / 兼容界面
捆绑或装机基础能力

单元格是基于保留的公开产品、定价和安全页面作出的有序判断;“强”不代表所有用例的基准表现都相当。

[CP001, CP006, CP008, CP014, CP017, CP024]
定价 / 打包比较
厂商公开入门价 / 单位合同模式公开包含项公开未知项含义
ExaSearch 请求 $7 / 1k;Deep Search $12;Deep-Reasoning $15;Agent 起价 $0.012 / 请求按量 credits + 企业定制Search、contents、monitors、answer 和 agent 工具实际企业折扣和承诺消费下限未公开对开发者足够透明,但不明显是最便宜的直接同业
Tavily每月免费 1,000 credits;PAYG 每 credit $0.008基于 credit 的自助 + 企业定制Search、research 及相关 API 使用各端点 credit 到工作负载的换算没有在一张标准化卡片中完全公开容易试用,降低早期团队切换成本
Brave Search API每月 $5 免费 credits;Answers $4 / 1k 请求 + token 费用自助 + 企业定制独立索引、回答层、引用、OpenAI SDK 兼容每个搜索端点或存储权利的精确标准化成本因计划而异低公开入门价加大 Exa 的比价压力
SerpAPI每月免费 250 次搜索;$25 / 1k;$75 / 5k;$150 / 15k;$275 / 30k月度自助计划 + 企业定制引擎覆盖广度、ZeroTrace、SLA 支持的企业升级仅看定价,无法明确与独立索引质量比较当覆盖广度比召回质量更重要时,简单自助阶梯有吸引力
PerplexitySearch API 请求 $5 / 1k;Sonar 和 Agent API 按 tokens 与工具另行定价基于 token 的 API 定价 + 直接提供商成本转嫁原始搜索、Sonar 回答模型、Agent API 工具、MCP、OpenAI 兼容企业折扣和每个已回答任务的实际混合成本仍不透明适合买方把搜索和回答生成放在一起采购
Google Custom Search仅限现有客户,最高 10k / 天,$5 / 1k 查询传统按量付费 API每天 100 次免费查询和标准监控不对新客户开放,并将在 2027 年关闭传统低价不如 Google 更宽的依据化捆绑重要
Algolia / Elastic / Glean多为定制或按工作负载定价;Algolia 公开请求和爬取费用,Glean 为定制销售主导合同或云计量自有数据相关性、分析、治理或托管平台功能企业部署没有同口径公开成本卡这些替代品拼的是更宽工作流或数据栈 ROI,不是原始搜索调用价格

各厂商公开定价单位并不完全标准化;本表比较开发者或采购团队在定制谈判前实际能看到的内容。

[CP002, CP003, CP004, CP015, CP016, CP017]
FP002: 功能广度 / 能力图

Exa 最清楚的领先点,是把实时网页检索与公司、开发者工作流扩展结合起来;替代品则在自有数据和捆绑驱动用例上占优。

这张图把很宽的市场压缩成买方关心的维度,并把替代品强项与直接同业强项分开标注,而不是强行套进一张假的通用功能清单。

[CP006, CP014, CP017, CP024, CP028, CP029]

3.3 切换成本、多供应商并用与护城河耐久性:Exa 有楔子,但不是封闭赛场

公开信息里支持 Exa 护城河的最佳论据,不是孤立的搜索 API,而是实时 Web 检索、代码感知搜索、结构化公司搜索,以及通过 SDK 和远程 MCP 进行 agent 原生分发的组合。开发者想要更少供应商、更快做出生产级 agent 时,这套组合会有价值。它也让 Exa 有可信空间扩展到 GTM enrichment、编码和合规敏感检索等相邻垂直工作流。 问题在于,现代 AI 栈让多供应商并用异常便宜。Google Grounding 可以坐在任意搜索 API 之上,LangChain 抽象核心模型供应商选择,Brave 和 Perplexity 都提供降低重写成本的兼容层。竞争对手也在上移:Tavily 现在销售研究工作流,Perplexity 销售 Sonar 加 agent 工具,巨头可以把搜索采购藏进更大的企业 bundle。即便 SerpAPI 也会改变比较方式,因为团队可保留 wrapper 层,同时替换底层引擎或垂直 SERP 端点。这意味着,只有当 Exa 的索引质量和垂直扩展显著好于同行时,它的护城河才显得耐久。如果买家越来越把搜索看作可替换组件,那么价格压缩、bundle 压力、内部自建可选性和基于 wrapper 的第二供应源,都会成为承保风险,而不是边缘情形。[CP020, CP027, CP028, CP034, CP035, CP037]

护城河耐久性 / 竞争风险登记表
护城河说法威胁严重性缓释 / 尽调问题
专有网页索引质量Brave、Perplexity 和 Google 都把搜索与实时回答工作流配在一起,SerpAPI 则拓宽引擎访问要求在目标工作负载上提供 Exa、Brave、Tavily、Perplexity 和 SerpAPI 的独立相关性与延迟基准
Agent 原生工作流楔子Tavily 和 Perplexity 都在从原始搜索走向研究或 agent 编排界面验证 Exa 胜出有多常是因为工作流深度,而不是基础搜索质量
企业信任差异化Brave、SerpAPI、Glean、Google 和 Elastic 都发布合规、SLA 或治理声明中高按信任要求拆解已赢单交易,并要求审计范围对比,而不只是认证 logo
通过 SDK 和 MCP 做开发者分发OpenAI 兼容和 any-search-API 抽象让改写比传统 SaaS 更轻按 SDK 或 MCP 进入路径衡量队列留存,并识别第二供应商采用从哪里开始
公司搜索等垂直扩展Google、Glean、Elastic 和内部自建可以满足许多自有数据或企业图谱用例,而不必匹配 Exa 的精确产品形态测试公司搜索是否在 GTM 和市场研究工作流之外显著改变胜率
价格透明度和自助动线Brave、Perplexity、SerpAPI 和传统 Google 定价都为买方提供参考价格先拿到实际定价、折扣和毛利率数据,再假设 Exa 能守住溢价经济性
既有平台捆绑Google Grounding 和企业搜索平台可以把检索藏进更大的 AI、云或工作流预算跟踪 Exa 何时被作为单独项目评估,何时被买方已批准的捆绑方案替代

严重性是基于公开替代路径证据作出的承销判断,不是公司披露的内部风险登记。

[CP020, CP027, CP028, CP035, CP036, CP037]
FP003: 护城河 / 就绪度 KPIs

Exa 的承销图景里,差异化产品界面最强,捆绑暴露和多家并用阻力最弱。

KPI 数值是承销和尽调优先级的综合判断,不是公司报告的运营指标。

[CP035, CP036, CP037, CP039, CP040, CP041]

3.4 展示材料

Chapter 04

04财务

4.1 收入模式与端点变现

Exa 变现的是一整套 API primitives,而不是席位授权。公开定价表面显示,Search、Deep Search、Deep-Reasoning Search、Contents、Answer 和 Agent runs 都直接收费,Monitors 和 Websets 则把同一用量逻辑延伸到经常性或重 enrichment 的工作流。自助用户预付 credits,并随着请求、页面和 agent 工具调用发生而消耗;企业客户被推向协商定价、发票账单、更高结果数、自定义 QPS,以及零数据留存等安全承诺。因此商业问题不是 Exa 有没有收费方式,而是标价需求在扣除算力、爬取和支持成本后,能否转成持久的实际毛利。公开客户案例支持客户愿意把 Exa 深度嵌入 agent 化产品,但没有揭示企业合同规模、折扣,或高用量平台交易与长尾开发者支出的 mix。[CI001, CI002, CI003, CI004, CI005, CI006]

收入流表
收入流机制单位当前价值 / 状态收入质量尽调问题
Search API核心检索端点,按用量售卖每 1k 次请求基础 $7;超过 10 条结果后,额外结果收 $1标价清晰;实际收益率未知提供头部客群每 1k 次搜索的混合实际 ARPU
Deep / Deep-Reasoning Search更高投入的综合搜索模式每 1k 次请求deep $12;deep-reasoning $15高价用量 SKU;模型工作更多,毛利率未知披露 deep 查询占比,以及每次请求的计算成本
Contents / Answer检索加页面提取 / 有出处答案每 1k 个页面或每 1k 次请求每 1k 个页面 $1;Answer 每 1k 次请求 $5有用的附加产品;抓取和推理成本可能波动展示附加率、每次查询平均页面数和毛利率
Agent异步深度研究和增强每次运行,另加 ACU 和搜索工具调用每次运行 $0.012-$2.00,另加 $0.10/ACU 和 $0.005/search工作负载潜在价值高,但对计算成本高度敏感按 effort 层级分享平均 ACU、搜索调用和实际毛利率
Monitors / Websets周期性搜索、列表构建和增强定时运行 / 异步任务Monitors 在公开页面标价;Websets 定位为增强工作流靠周期性和 GTM 用例扩大钱包份额披露留存、重复运行量和合同打包方式
Enterprise overlays发票计费、自定义 QPS、ZDR、支持、SLA自定义合同通过销售谈判,而不是公开计算器定价更高付费意愿有可能成立;定价不透明度仍高拆出企业 ARR、平均合同额和自定义支持负担

标价和打包方式公开;实际净价、折扣和产品组合未公开。

[CI001, CI002, CI003, CI004, CI005, CI008]
定价 / 变现表
产品价格 / 合同模式标价 vs 实际价格折扣 / 未知项来源
Search按用量计费,$7/1k 次请求;超过 10 条结果后额外结果 $1标价公开;实际价格未披露仅企业版提到批量折扣exa.ai/pricing 价格页
Deep Search按用量计费,$12/1k 次请求标价公开未公开客户级折扣表exa.ai/pricing 价格页
Deep-Reasoning Search按用量计费,$15/1k 次请求标价公开未公开客户级折扣表exa.ai/pricing 价格页
Contents按用量计费,每种内容类型 $1/1k 个页面标价公开实际页面组合和抓取频率未知exa.ai/pricing + Contents 文档
Agent每次运行 + ACU + 搜索工具调用标价公开;实际混合价格取决于 effort 和工具使用未公开按 effort 模式拆分的实际数据exa.ai/pricing + Agent 文档
Self-serve billing预付点数,自动充值在线全产品化未披露公开分成或转售渠道计费文档
Enterprise发票计费、自定义价格、SLA、自定义 QPS、ZDR谈判定价,未披露折扣、最低消费和服务负担未知定价 + 计费 + 速率限制

该表只记录标价和打包方式;不应把它误当成实际收入或毛利率。

[CI001, CI002, CI003, CI004, CI005, CI006]
FI001: 收入模型桥接

在计算和抓取成本之前,Exa 如何把开发者和企业活动转成按使用量计费的收入流。

流程反映公开的变现机制,不代表经审计的收入确认政策或贡献利润率。

[CI001, CI004, CI006, CI008, CI012, CI013]

4.2 成本结构代理指标与单位经济

公开来源对基础设施野心异常清楚,对实际利润率则异常稀薄。Exa 拥有爬取、索引、嵌入、向量数据库和查询层,管理层称系统追踪超过 500B 个 URL。公司还披露了一个 $5M 的 144-H200 集群,几乎连续运行;加上更早的 80-A100 机群,已安装 GPU 总数达到 224。向量数据库文章称,这种设计在承载数十亿向量、低于 100 毫秒检索和超过 500 QPS 的同时,把云向量数据库成本削减约 10 倍。这些是有意义的效率信号,但它们没有揭示毛利率,因为交付仍取决于实时爬取新鲜度、内容广度、reranking 和 agent 算力强度。正确解读是,Exa 很可能拥有很强的内部成本优化人才,但在管理层披露端点级贡献利润率,或至少披露混合毛利率区间之前,利润率路径仍是尽调事项。[CI010, CI026, CI027, CI028, CI029, CI030]

单位经济表
指标数值 / 公开代理指标置信度重要性尽调要求
2025 收入估算$10M(Sacra 估算)已审阅材料中唯一公开收入锚点提供经审计或董事会口径的 2025 收入和 2026 运行收入率
毛利率null基础设施重的 API 业务,核心承销卡点就是毛利提供混合毛利率和端点级毛利率历史
已部署训练 / 检索算力$5M 144-H200 集群;加上较旧 A100 集群合计 224 块 GPU暗示固定成本基础重、容量野心大提供折旧政策、租赁条款和利用率
向量 DB 效率说法成本比报价的云端向量 DB 替代方案低 10x说明内部成本优化,不说明绝对毛利率提供每 1k 次成功搜索的实际基础设施成本
搜索吞吐代理指标向量 DB >500 QPS;公开 /search 默认限制 10 QPS暗示内部容量可超过公开自助默认值按层级提供持续生产 QPS、延迟和单位成本
Token 效率代理指标Agent 基准测试中,文本提取减少 >20x,token 最多减少 94%更少 token 可改善下游经济性和客户 ROI提供实测客户节省额,以及相对 API 价格的捕获率
销售效率 / 留存null未公开 CAC、回本周期、NRR 或流失率,GTM 质量不透明提供按渠道拆分的 CAC、回本周期、总留存和 NRR

Null 表示未披露的私有指标;公开代理指标强调成本驱动因素和优化说法,而不是实际盈利能力。

[CI010, CI021, CI027, CI029, CI030, CI031]
FI002: 单位经济桥接

公开成本代理显示,在任何搜索或智能体工作负载转成毛利之前,Exa 可能把钱花在哪里。

该图为定性判断,因为公开来源披露了基础设施规模和优化说法,但没有披露毛利率结果。

[CI005, CI026, CI028, CI029, CI031, CI032]
FI004: 资本强度 / 现金流图

Exa 能复用核心基础设施的地方,收入潜力最强;计算和支持强度超过实际定价的地方,现金流风险最高。

[CI009, CI010, CI028, CI029, CI036, CI037]

4.3 牵引力可见度与资本充足性

需求证明方向性强,但仍不完整。Exa 称其服务超过 5,000 家公司和 400,000 多名开发者;OpenRouter 案例研究报告,其 Exa 支持路径累计产生 73M 次搜索查询。收入方面,最佳公开数字是 Sacra 对 2025 年 $10M 的估算;此前公司曾表示 2024 年收入在几个月内增长三倍。资本可见度好于运营可见度:Exa 公开宣布 2026 年 5 月以 $2.2B 估值完成 $250M Series C,此前披露 2024 年完成 $22M 种子轮和 Series A。管理层把新资金用于模型训练、把基础设施扩展到每秒数十万次搜索,以及 GTM 建设。仍缺失的是从这笔融资通向可持续经济性的偿付能力桥梁:公开信息没有现金余额、烧钱、runway、债务或 covenant 图景。[CI014, CI015, CI016, CI021, CI022, CI034]

资本充足性表
字段公开数值 / 状态置信度影响尽调要求
账面现金null无法把 Series C 规模换算成实际流动性垫提供交割前后的季末现金
月度烧钱null基础设施重,且未公开跑道桥提供总烧钱、净烧钱,以及 capex 与 opex 拆分
跑道月数null新资金降低风险,但无法公开计算跑道提供董事会基准 / 下行情景下的跑道测算
计划资金用途$250M Series C 指定用于下一代模型、基础设施扩容和 GTM 扩张资金用于进攻,而不是解释当前单位经济提供资本配置计划和里程碑闸门
下一轮触发条件未公开披露下一轮融资是否取决于收入、算力或地域扩张仍未知提供下一轮股权或债务融资的触发指标
债务 / 项目融资义务未发现公开披露资产负债表风险仍不透明提供全部债务、租赁、供应商融资和硬件承诺
此前股权融资基础2024 年披露 $22M;Sacra 估算截至 2025 年、Series C 之前累计融资 $107M说明在规模大得多的 2026 轮之前,资本结构已经快速膨胀核对全部融资轮、老股交易和期权池刷新

分轮融资历史放在 Company Overview;该表只隔离评估未来充足性所需的融资事实。

[CI034, CI035, CI036, CI037, CI038, CI044]
FI003: 财务估计区间

少数公开财务锚点集中在融资和单一第三方收入估计上,而不是盈利能力或 runway。

公开区间收敛成点估计,因为没有来源披露收入、burn 或毛利率的低 / 中 / 高情景。

[CI021, CI034, CI035, CI044]

4.4 财务结论与承保视角

Exa 看起来商业上真实、技术上差异化、当前资本充足,但仍无法基于公开基本面承保。正面 case 很直接:按用量变现与 agent 采用一致,客户证明显示嵌入式工作流价值,Series C 显著延长公司在下一轮融资前扩大规模的能力。阻塞项在于,把产品兴奋转成可投资经济性的每个核心承保指标仍缺席:实际收入 mix、毛利率、CAC、回本周期、NRR、集中度、折扣、烧钱、runway 和债务义务。公开备案没有填补缺口;Delaware 公开搜索除非购买底层记录,否则只提供基本实体细节。实际结论是,Exa 值得继续尽调,但任何带价格的决策都应等管理层 data room,把用量增长与贡献利润率和流动性对齐。在那之前,正确的 IC 姿态是 research-more,而不是 buy 或 avoid,因为缺失数据指向盈利质量,而不是明显需求失败。[CI023, CI024, CI037, CI038, CI039, CI040]

公开财务缺口表
缺失的私有指标对判断的影响精确尽调路径
企业版 vs 自助收入组合没有组合数据,就无法判断收入质量和集中度要求按产品、合同类型和前 20 大账户拆分收入
按端点拆分的毛利率无法判断用量增长是增厚利润还是稀释毛利要求按 Search、Contents、Agent 和自定义企业工作负载拆分 COGS 桥
现金、烧钱、跑道资本充足性无法转换成存活月数或投资节奏要求最新月度现金瀑布和 12-18 个月计划
CAC、回本周期、NRR、流失率销售效率和持久性仍不透明要求 cohort 留存、扩张和全成本销售 / 营销效率
折扣和合同条款如果企业折扣很深,标价可能高估变现质量要求标准 MSA 条款、按 cohort 拆分的平均折扣和支持负担
债务、硬件承诺和供应商融资基础设施负债可能在头条股权故事之外要求债务明细、硬件采购承诺和任何保底金额

每一行都是卡口型尽调事项,不是小的加分项;缺少这些信息,定价承销决策只能依赖代理指标。

[CI023, CI024, CI037, CI038, CI039, CI040]

4.5 展示材料

Chapter 05

05产品与技术

5.1 产品表面:Exa 销售分层检索与研究栈,不是单一端点

Exa 的公开定位已从「面向 AI 的搜索引擎」扩展为面向 agent 构建者的更广运营栈。官方 quickstart 和搜索指南呈现四个核心 primitives——/search、/contents、/answer 和 /research——再把它们延伸到 Agent runs、Monitors、Websets、MCP 分发,以及公司搜索等结构化垂直场景。按客户工作流看,这意味着 Exa 可以是快速助手的检索层、CRM 或 GTM 软件的 enrichment 层、结构化情报任务的研究层,也可以是偏好 MCP 或 SDK 介入而不是直接 HTTP 集成的助手工具层。客户例子重要,因为它们显示这些表面已经进入类似生产的工作流:OpenRouter 把 Exa 支持的 Web 搜索作为共享模型市场基础设施,monday.com 用自然语言搜索加结构化输出来铺排 CRM prospecting,Cognition 公开称 Exa 支撑 Devin 的搜索能力。这种广度让 Exa 的产品定义比通用搜索 API 供应商更清晰,但也意味着产品尽调应把它看作多模块平台,其价值取决于编排、分发和搜索后工作流契合度,而不只是结果相关性。[CE001, CE002, CE003, CE005, CE007, CE008]

产品模块 / 资产矩阵
模块或资产主要用户状态 / 成熟度差异化尽调缺口
Search APIAgent / app 开发者GA;已记录从 instant 到 deep-reasoning 的六种搜索模式语义检索,带延迟 / 质量层级、结构化输出、过滤器和类别入口需要 Exa 自有评测之外的独立召回率 / 精确率基准测试
Contents APIAgent / app 开发者GA;已记录 text、highlights、summaries 和子页面抓取高 token 效率的高亮片段加新鲜度控制,降低下游上下文成本需要更清晰公开抓取广度、提取失败率和重 JS 边缘场景限制
Company SearchGTM、研究、金融产品构建者GA;每周更新的 50M+ 公司索引,带结构化实体Exa 从普通网页搜索推进到结构化公司元数据和增强需要公开证明覆盖质量、实体解析准确率和陈旧记录处理能力
Agent / Deep Search需要多步综合的团队2026 年已发布并持续扩展并行搜索 agents、结构化 JSON 输出、引用和按投入定价需要公开长期任务吞吐、失败率和治理细节
Monitors / Websets运营、研究、增强、GTM 团队GA;已记录周期性和异步工作流把搜索变成持久管线,带条件验证、增强、导入和 webhook需要更多公开的审批、审核队列和企业治理控制
MCP + SDK 分发AI assistant 构建者和平台团队GA,且已在 GitHub、PyPI、npm 和托管 MCP 上主动打包低摩擦接入 Cursor、Claude、Codex、VS Code 和自定义客户端需要明确版本保证、向后兼容性和 MCP 端点租户隔离

各行把面向买家的模块与其下方共享平台分开;成熟度反映公开文档可用性,而不是私有企业部署深度。

[CE001, CE002, CE003, CE005, CE007, CE009]
工作流 / 用例表
用户任务当前流程Exa 方案可衡量或声称收益局限
让快速 assistant 接入实时网络信息模型回答时效性问题时过期,或依赖通用浏览器插件使用 /search,配 instant/fast 模式和高亮片段公开延迟档位,instant 模式低至约 250 ms,并返回有出处片段未公开按领域对比现有网页插件的答案质量基准测试
跑带结构化输出的深度网页研究团队自己编排多次搜索、抓取、综合和 JSON 格式化使用 Deep / Agent,配 outputSchema 和 citationsExa 称一次 API 调用可替代深度研究任务的复杂编排供应商托管编排降低了中间失败处理透明度
构建 GTM 或投资实体清单分析师拼接电子表格、手工搜索和增强供应商使用 Company Search 或 Websets,配条件验证和增强结构化公司实体和已验证 webset 条目减少手工筛选步骤长尾实体覆盖和新鲜度未公开经过独立基准测试
监控变化中的主题或竞争对手团队手工重跑搜索,或自己搭 cron 任务和 webhook 管线使用 Monitors 做定时搜索,并用 webhook 去重投递周期性运行只浮现新内容,并可输出结构化结果公开文档未暴露企业工作流治理或审核控制
准备 CRM 潜客挖掘和增强销售团队手工研究账户和购买群体monday.com 案例研究用 Exa search 把自然语言 ICP 转成结构化 CRM 记录案例研究称销售代表打开的是预先准备好的工作,而不是手工建列表客户证明由公司托管,未披露错误率或提升指标
把网页搜索嵌入多模型 assistant团队为每个模型家族构建特定供应商插件OpenRouter 案例研究和 MCP 分发让搜索跨越多个模型 / 客户端界面OpenRouter 报告迄今已有 73M 次 Exa-backed 搜索,并支持服务端工具调用可移植性如果检索质量压缩到同等水平,可移植性也会降低迁出 Exa 的成本

收益来自公开产品描述和案例研究;多数未披露经审计 ROI 指标,因此运营提升应视为方向性信号。

[CE003, CE004, CE006, CE013, CE017, CE019]
FE002: 客户工作流 / 运营流

构建者如何借助 Exa 的公开模块,从用户请求走到有出处支撑的结构化行动。

[CE001, CE004, CE005, CE008, CE009, CE013]

5.2 架构与运营模式:自定义检索基础设施加面向 agent 的编排

最重要的技术问题是,Exa 只是包装第三方搜索,还是拥有检索栈的实质部分。公开证据指向后者。Exa 的文档和博客描述了基于 embeddings 的搜索、query-aware highlights、结构化垂直索引,以及自建向量数据库;后者为数十亿向量、元数据过滤、低于 100 毫秒检索和超过 500 QPS 优化。同一批材料还描述了激进压缩、聚类、reranking 和自定义查询语言阶段,这与简单把查询转发给 SERP provider 在性质上不同。基础设施披露强化了这个故事:Exa 称其运营 80-A100 集群和 144-H200 Exacluster,使用 Kubernetes 编排并配合 Pulumi 和 Ansible,依赖 NVIDIA operators 做 GPU 和网络管理,并用 Alluxio 加 S3-backed cache semantics 支持大规模训练数据访问。更高层产品也映射这套架构。Exa Deep 和 Exa Agent 被呈现为并行搜索和 subagent 系统,把检索转成结构化综合;LangChain 案例研究则描述 planner-task-observer 模式和 JSON-first 输出。实际含义是,Exa 的运营模式越来越像「检索平台加研究编排」;技术上有差异化,但也比薄 API wrapper 更吃基础设施。[CE003, CE004, CE006, CE023, CE026, CE027]

技术 / 运营架构表
层或组件角色依赖风险
搜索接口层接收自然语言查询、过滤器、输出 schema 和类别选择托管 Exa API、语言 SDK 和客户代码质量API 抽象强,但客户仍暴露在托管服务可用性和合同条款之下
内容处理层从 URL 或搜索结果返回 text、highlights、summaries 和子页面抓取爬虫覆盖、缓存新鲜度政策和提取管线选择公开文档解释控制项,但未说明全球提取成功率或区域抓取拓扑
结构化垂直层把搜索结果映射为公司或其他带元数据的类型化实体领域分类器、实体解析和垂直专用索引长尾公司和非英语表面的覆盖 / 质量仍难从外部判断
Agentic 编排层把任务拆成搜索、子 agent、增强和有出处输出Exa 托管的运行编排,加底层模型 / 工具路由在 Exa 自有示例之外,长任务正确性和预算控制部分是黑箱
检索核心计算 embeddings、搜索向量、应用元数据过滤器并重排结果自定义向量 DB、倒排索引、聚类、重排管线架构看起来有差异化,但可能资本开支重;没有内部指标就很难审计
训练和基础设施层训练检索模型,运行大规模索引工作负载Exacluster、Kubernetes、NVIDIA operators、Alluxio、S3-backed 缓存语义沉重的 GPU / 数据中心足迹可以强化护城河,也会增加固定成本和运营复杂度

该表把已记录架构和明确尽调风险混在一起;公开材料在运营或基准测试级披露处止步。

[CE004, CE005, CE007, CE009, CE026, CE027]
FE001: 产品架构图

从面向开发者的接入点,到检索基础设施和训练运营,分层看 Exa。

[CE002, CE007, CE009, CE014, CE026, CE027]
FE004: 产品成熟度 / 能力图

从公开证据看,Exa 核心模块成熟度不一;接入界面已经较成熟,但企业级运营证明更薄。

[CE007, CE009, CE012, CE013, CE014, CE015]

5.3 部署、集成与可靠性:采用路径强,公开企业运营细节较薄

Exa 很容易嵌入现代 agent 栈。文档索引暴露 Python 和 JavaScript SDK、MCP 设置、编码 agent 参考,以及 Snowflake、LangChain、Browserbase、OpenAI、Anthropic、自动化工具和语音栈的集成页面。MCP 表面尤其重要,因为它让 Exa 可以作为托管远程服务器或 npm package 分发,而不是要求每个客户自建和维护定制 tool wrapper。Snowflake 展示了另一条路径:Exa 可通过 External Access 和 stored procedures 进入受治理的数据仓库;Browserbase 则展示 Exa 作为浏览器自动化的上游发现层。开发者信号可信,而不只是名义存在。exa-mcp-server、exa-py 和 exa-js 的公开 GitHub 仓库都在 2026 年 6 月更新,其中 MCP 仓库远大于语言 SDK 仓库;PyPI 和 npm packaging 显示持续发布活动,以及 langchain-exa 等生态扩展。可靠性证据更窄。Exa 公开状态页在访问时显示 Websets 和 Exa MCP operational,uptime 为 100%,但审阅到的公开材料没有提供完整 SLA、已发布 error-budget discipline、区域部署矩阵或 self-host/VPC 运营选项。因此采用摩擦看起来低,但企业运营尽调仍部分依赖销售和 data room,而不是可从公开文档自助完成。[CE014, CE015, CE016, CE017, CE018, CE020]

路线图 / 发布 / 开发阶段表
日期或阶段功能或里程碑状态含义来源
Dec 2024定制 Web 规模向量数据库架构公开披露已发布 / 已记录说明 Exa 把检索核心当作自有基础设施,而不是外包管道Exa 向量数据库博客
May 2025Exacluster 及配套 MLOps 栈公开披露已发布 / 运行中显示公司愿意投入自有训练与索引能力Exa Exacluster 博客
Mar 2026改版 Exa Deep 上线,支持结构化输出和字段级溯源已发布深度搜索进一步走向智能体式综合,而不只是普通检索Exa Deep 发布文章
Mar 2026 包状态langchain-exa 1.1.0 上传至 PyPI已发布生态包把 Exa 带进第一方 SDK 之外的智能体框架工作流PyPI 包页面
Jun 2026Exa Agent 作为单一 API 上线,覆盖深度研究、名单构建和数据补全已发布更高层编排正在成为一线 SKU,而不只是文档里的模式Exa Agent 发布文章
当前文档索引变更日志、coding-agent 参考、MCP 和多条集成路径均已公开列出产品面活跃,但路线图信息偏少有公开发布沟通,但正式前瞻路线图仍稀疏文档索引

表中使用发布日期或当前包 / 文档状态,因为在已审阅材料中,Exa 没有发布详细的前瞻里程碑路线图。

[CE015, CE037, CE038, CE041, CE042, CE043]
FE003: 关键依赖图

会实质影响 Exa 产品交付和采用路径的外部及基础设施依赖。

[CE014, CE016, CE017, CE018, CE027, CE028]

5.4 信任、隐私、合规与技术结论:有意义的控制存在,但范围受限且有条件

Exa 的信任故事好过「一家有搜索 API 的 AI 初创公司」,但买家需要仔细读限定条件。安全文档称 Exa 获得 SOC 2 Type II 认证,并引导买家访问 trust center;企业方案可解锁零数据留存和 HIPAA 支持。HIPAA 文档的具体性很有用:合规模式不是覆盖每条产品路径的总开关,而是面向 eligible teams 的门控模式,只在受限 /search 和 /contents flows 上工作,拒绝 livecrawl 和 summary-heavy paths,并在请求需要非 HIPAA-safe processors 时 fail closed。这种精确性是正面信号,因为它说明 Exa 在把控制映射到精确请求路径,而不是把合规营销成泛平台覆盖。隐私政策构成反向砝码。它明确警告用户不要把个人信息作为查询数据提交,说明查询数据可用于改进产品并训练或微调模型,并把 business-offering processor data 放进客户协议。对企业承保来说,这意味着 Exa 可以支持敏感工作流,但只有买家处于正确方案、配置正确,并且认可这些模式之外的默认查询处理符合内部政策时才成立。技术上,Exa 看起来有差异化;流程上,信任和部署仍需要合同级尽调。[CE032, CE033, CE034, CE035, CE040, CE043]

信任 / 质量 / 合规表
控制或信号状态范围缺口
SOC 2 Type II已记录安全文档引用了组织级安全控制框架公开材料本身未列明控制例外、审计日期或面向客户的 SLA 条款
Zero Data Retention企业计划可用面向合格买家的企业商业 / 安全选项非企业查询路径的默认行为仍需政策审查
HIPAA mode仅已启用的企业团队可用仅限符合条件的 /search 和 /contents 请求,且必须走合规处理器路径并执行 fail-closed 规则不是覆盖整个平台的合规声明;livecrawl、摘要和深层路径不在范围内
查询数据隐私政策已有文档用户被告知不要提交个人信息,但查询数据可能用于改进产品并训练 / 微调模型企业需要在合同里确认其具体 SKU 和工作流的数据留存 / 训练处理方式
公开状态页可见运营信号抓取时 Websets 和 Exa MCP 处于运行状态,显示 100% 正常运行时间快照状态弱于公开的 SLA/SLO 历史或事故复盘纪律
信任中心 / 企业文档可通过 Exa 信任中心和销售流程获取支持围绕 DPA、安全文档和企业控制项展开尽调关键细节仍锁在销售或门户访问之后,而不是公开自助文档

状态条目描述公开文档里能看到的控制面,并明确标出买方仍需做合同级或门户级尽调的环节。

[CE032, CE033, CE034, CE035, CE043]

5.5 展示材料

Chapter 06

06客户

6.1 客户基础分层

Exa 的公开客户集合足够宽,可以按买方、用户、付款方和待完成任务分层,而不是只贴一个「开发者工具」标签。第一类是开发者原生 agent 和编码产品——Cursor、Cognition、OpenRouter、CodeRabbit——日常用户是开发者或 AI agent,经济买方通常在工程或产品,付款方是基础设施或软件预算。第二类是 GTM 和 CRM 自动化——monday.com、HubSpot、11x 和 Obvious——终端用户是 SDR、AE、RevOps 或代表他们行动的 AI agent,预算权在收入运营、销售技术或产品领导层。第三类是研究和专业知识工作——科学发现里的 Anara、诉讼情报里的 WhyHow——信息质量比单纯点击式搜索更重要。第四类是 StackAI 这类企业 agent 平台,Exa 嵌入受治理的多步骤企业工作流,如尽职调查和 RFP。地理分布由 Exa 直接披露的少于客户披露的,但 HubSpot 覆盖 135 个国家的 299,000 名客户、monday.com 全球 250,000+ 客户,意味着即使 Exa 没有发布干净的细分或区域收入拆分,它已位于全球分布的软件产品内部。[CU001, CU002, CU003, CU004, CU005, CU006]

客户分层表
分层示例客户买方 / 付款方主要用户核心用例战略价值 / 缺口
开发者原生智能体和编码工具Cursor、Cognition、OpenRouter、CodeRabbit 等工程 / 产品预算开发者和编码智能体为代码、文档和实时 Web 上下文提供依据战略价值高;除 Cursor 细节仍薄外,生产证明较强
CRM 和 GTM 平台HubSpot 与 monday.com产品 / RevOps / CRM 预算销售代表、RevOps 用户、AI 智能体人员 / 公司搜索、潜客挖掘、数据补全、路由说明 Exa 可以卖到开发工具之外;Exa 未按分层披露胜率或 ACV
销售自动化和市场情报11x, Obvious销售科技 / GTM 运营预算外呼智能体和研究团队信号发现、名单构建、数据补全工作流证明不错;复购支出如何跨客户放大仍不清楚
研究和科学工作流Anara研究产品预算科学家、学生、研究团队论文发现、引用支持、研究智能体对信任敏感型检索是有用证明;分层规模未披露
法律情报工作流WhyHow产品 / 法律科技预算诉讼情报智能体跨网页和备案文件检测弱信号展示差异化用例;受监管留存行为仍缺少公开证明
企业智能体平台StackAI企业平台预算企业运营人员和 AI 智能体尽调、竞争情报、RFP 回复、市场研究企业触达可信;合同期限和集中度仍未披露

客户分层来自公开案例研究和具名证明的推断;Exa 没有正式披露按分层划分的收入或地域结构。

[CU001, CU002, CU003, CU004, CU005, CU006]
FU001: 客户旅程图

Exa 通常先进入开发者实验,再嵌入工作流基础设施。

旅程综合客户故事、定价、账单和集成界面而来;并非每个客户都会按顺序走完所有阶段。

[CU002, CU003, CU004, CU022, CU028, CU029]

6.2 采用轨迹

公开采用曲线很陡,即便分母质量仍由公司定义。2024 年 7 月,Exa 称已有数千家公司和开发者集成产品,并提到 Databricks 是早期 AI 研究团队案例。到 2026 年 5 月 Series C 文章,Exa 称使用平台的公司数已增至超过 5,000 家,产品为超过 400,000 名开发者提供搜索能力。具名客户信号在这段时期也更具体:公开引用从宽泛的「AI applications」语言,转向明确提及 Cursor、Cognition、HubSpot、OpenRouter 和 monday.com。客户侧结果数据存在的地方,用量深度最强。OpenRouter 报告 Exa 支持的搜索查询从去年 2.36M+ 增至累计 73M。11x 称 Exa 已在生产中支撑数百万次搜索、Webset items 和 enrichment cells。CodeRabbit 报告切换到 Exa 后,在保持或提升输出质量的同时,Web 搜索量减少约 70-75%。合在一起,这条轨迹支持真实生产采用,但 Exa 仍未披露活跃与注册开发者数量、付费与免费公司数量,或按 cohort 的扩张。[CU010, CU011, CU012, CU013, CU014, CU015]

客户增长 / 采用轨迹表
指标数值日期来源置信度含义缺失分母
已集成公司和开发者数千2024-07-16Business Wire 新闻稿当前客户名单形成前,商业采用已经真实存在未区分付费与免费用户
使用 Exa 的公司5,000+2026-05-20Exa Series C 文章Logo 广度如今已有意义未披露活跃公司数或付费账户数
覆盖开发者400,000+2026-05-20Exa Series C 文章自下而上扩张拥有较大漏斗入口注册开发者与活跃开发者未拆分
具名客户名单Cursor、Cognition、HubSpot、OpenRouter、monday.com 等2026-05-20Exa / Lightspeed / PYMNTS公开证明已从泛泛表述转向具名头部客户未披露按垂直行业划分的胜率或收入贡献
OpenRouter 由 Exa 支撑的搜索量2.36M+ -> 73M2025 至 2026OpenRouter 案例研究单一客户内部重复使用信号强未披露这些查询对应的收入分成
11x 生产使用量数百万次搜索、Webset 条目和补全单元格202611x 案例研究暗示 GTM 智能体中已有规模化嵌入使用无合同规模或留存数据
CodeRabbit Web 搜索效率搜索量降低 70-75%;每日 7-10% 的 PR 使用 Web 搜索2026CodeRabbit 案例研究Exa 能在保持质量的同时减少查询负载披露样本期只有四天

指标混合了公司自称数据和客户案例数据;采用广度较强,但活跃使用、变现和 cohort 转化未披露。

[CU010, CU011, CU012, CU013, CU014, CU015]
FU002: 采用 / 部署流

公开可见的采用路径:先触达大量开发者,再进入具名的生产工作流嵌入。

已披露处采用真实的漏斗顶端数量;后续阶段节点为概念性表述,因为 Exa 不公布转化率或队列数量。

[CU010, CU011, CU012, CU014, CU015, CU029]

6.3 具名客户证明

对一家私营基础设施 API 公司来说,Exa 的具名客户证明异常丰富,但不同 logo 的证据质量差异很大。HubSpot 拥有最强的企业工作流证明:具名 AI 负责人称 Exa 在速度、价格和 enrichment 覆盖上击败原生模型搜索和其他替代品,案例研究把用量连接到 Breeze Assistant、agents 和 Monitors 工作流。monday.com 对 CRM lead-generation 和 routing 提供了同样具体的工作流证明。OpenRouter 给出最佳量化规模证明:累计 73M 次查询,并清楚解释 Exa 为什么在模型路由市场中重要。Cognition 的证据细节较薄,但战略上重要,因为其创始人称 Exa 支撑 Devin 的所有部分。二级引用扩大了证明集:11x、CodeRabbit、Obvious、StackAI、WhyHow 和 Anara 显示 Exa 用于 GTM 情报、代码审查、市场地图、企业 agent 栈、诉讼情报和科学研究。Cursor 是一个 marquee logo,但 Exa 的公开证明最弱:公司、Lightspeed 和 PYMNTS 都称其为客户,却没有找到单独的 Exa 案例研究或量化部署细节。因此,Cursor 作为具名 proof point 真实存在,但不如 HubSpot、monday.com、OpenRouter 或二级案例研究集合强。[CU018, CU019, CU020, CU021, CU022, CU023]

具名客户证明表
客户分层部署 / 用例生产 vs 试点结果 / 证明局限
Cursor编码智能体公开具名的 Exa 客户 / 品类头部具名生产客户,但细节薄Exa、Lightspeed 和 PYMNTS 均具名提及未找到独立 Exa 案例研究或量化工作流结果
HubSpotCRM / GTM 平台Breeze Assistant 和智能体使用人员 / 公司搜索以及 Monitors生产具名 AI 负责人称,Exa 在速度、价格和覆盖面上胜过原生及替代搜索无公开合同金额或留存指标
CognitionAI 工程 / 编码智能体覆盖 Devin 的 Web 搜索能力生产创始人引述:Exa 支撑 Devin 的所有部分公开案例研究中没有量化吞吐或成本数据
OpenRouter模型路由平台面向 400+ 模型的服务端 Web 搜索生产累计 73M 次由 Exa 支撑的查询,且工作流解释清晰Exa 是一个引擎选项,而不是明确的独家基础设施
monday.comCRM / 销售工作流线索智能体、数据补全、资格判定、路由、招聘生产直接工作流描述关联 250,000+ monday.com 客户无专门归因于 Exa 的公开 ROI 或支出数据
11xGTM 智能体平台面向外呼 AI 工作者的账户研究和补全信号生产数百万次搜索、Webset 条目和补全单元格披露了使用规模,但没有合同或留存细节
CodeRabbitAI 代码审查对文档、包和发布说明做外部验证生产搜索量降低 70-75%,质量持平或更好披露的 PR 使用指标样本期较短
Obvious市场情报相似客群和联系人发现生产一次补全调用替代 15-20 次人工迭代;成本从数十万美元降到几美元由 Exa 发布;未找到客户侧独立 ROI 确认

公开具名证明较强,且多为生产级,但不同 logo 的证据质量不均;Cursor 仍停留在 logo 级证明,而 OpenRouter、HubSpot、monday.com、11x、CodeRabbit 和 Obvious 有更丰富的工作流细节。

[CU018, CU019, CU020, CU021, CU022, CU023]
FU003: 客户证明矩阵

具名客户的证据质量差异很大,从仅有标识露出到量化生产结果都有。

矩阵划分的是公开证据质量,不是客户价值;「独立佐证」指 Exa 之外的第二个域名确认了合作关系。

[CU018, CU020, CU021, CU022, CU025, CU040]

6.4 留存与耐久性

耐久性是公开记录变薄的地方。没有审阅来源披露净收入留存、总留存、流失、续约率或合同期限,因此留存故事只能从产品嵌入和工作流行为推断,而不是直接测量。正面的结构性信号存在。Exa 不只是单一端点:客户可从 Search 起步,再扩展到 Contents、Monitors、Agent,以及协商账单和更高用量支持等企业叠层。LangChain 的专用 Exa 集成降低开发者采用摩擦;客户故事也反复把 Exa 描述为核心运营循环的一部分,而不是一次性实验。但负面信号也真实存在。OpenRouter 的架构显示,搜索可以是更大模型路由产品中的可替换引擎。Humai 和 MakerStack 都认为,当语义质量重要时 Exa 最好,但在简单或高用量查询上比替代品更贵、更复杂;两者都描述了 routing 或 multi-provider 模式,让多供应商并用常态化。ChatForest 把风险再推高一层,认为如果模型供应商内置搜索变得「足够好」,整个品类可能被压缩。结果是一个可信但证据不足的耐久性 case:Exa 看起来嵌入重要客户工作流,但公开留存披露缺失,切换成本也显然不是绝对的。[CU023, CU024, CU028, CU029, CU030, CU031]

留存 / 重复使用 / 满意度表
维度公开数值分层置信度含义尽调请求
净收入留存所有分层没有公开 NRR,扩张质量无法验证索取按 cohort 和分层拆分的董事会级 NRR
总留存 / 流失所有分层Logo 增长可能掩盖简单或低价值用例中的流失索取 GRR、流失和 logo 留存历史
合同期限 / 续约企业账户企业耐久性无法靠公开信息承销索取平均合同期限和续约节奏
重复使用信号73M 次 OpenRouter 查询;数百万次 11x 动作平台客户嵌入之后,至少部分客户有高频重复使用拆分前 20 大账户的重复使用
满意度信号具名引述正面但偏轶事具名案例研究客户客户证明支撑产品价值,但不是调查级满意度提供 NPS / CSAT 或客户访谈
切换 / 多供应商并用风险较高开发者和智能体平台路由模式和替代供应商评测暗示独家锁定较弱展示单一供应商部署与多供应商部署占比

公开留存证据大多是代理指标;null 表示已审阅材料没有披露该指标,不代表数值为零。

[CU023, CU024, CU029, CU030, CU031, CU032]
FU004: 持久性 / 切换风险矩阵

在工作流更深的企业产品里,Exa 看起来比简单搜索路由场景更有粘性。

单元格是综合客户案例研究和独立定价 / 竞争评测后的分析判断;Exa 不公布量化队列留存数据。

[CU024, CU031, CU032, CU033, CU034, CU035]

6.5 扩张与集中度风险

Exa 的扩张路径直观:自助搜索可增长到内容抽取、监控、异步研究,最终进入协商企业合同。客户 mix 本身就能看到这条路径。HubSpot 把 Search 和 Monitors 叠进 CRM agents;OpenRouter 把 Exa 变成覆盖数百个模型的共享基础设施;StackAI 把 Exa 嵌入受治理的企业 agent 部署;11x 和 Obvious 等 GTM 客户则从发现走向 enrichment 和工作流自动化。集中度问题难得多。公开具名引用偏向 AI 原生软件和 agent 构建者,这意味着存在真实风险:少数旗舰 logo 对路线图、声誉,也可能对收入 mix 的影响,超过 5,000 家公司这个标题数字所暗示的程度。公开来源没有披露 top-customer 收入占比、细分 mix 或直接与伙伴来源 bookings。生态也是双刃剑:LangChain 和其他开发者表面降低采用摩擦,也让客户更容易 benchmark 或替代多个搜索供应商。因此,Exa 的客户章节支持 land-and-expand thesis,但还不支持无集中度 thesis。[CU027, CU028, CU029, CU030, CU038, CU039]

扩张和集中度风险表
扩张 / 集中度因素证据影响尽调路径
搜索 -> 更宽产品扩张Search、Contents、Monitors、Agent 和企业计费均已公开打包支持在成功账户内先落地、再扩张索取按 ARR cohort 拆分的产品挂载率和端点组合
企业工作流深度HubSpot、monday.com、StackAI 和 WhyHow 将 Exa 嵌入多步工作流Exa 一旦进入运营闭环,切换成本会上升按分层索取实施时间和替换成本
开发者飞轮400,000+ 开发者加上 LangChain 集成,降低采用摩擦可为未来企业扩张埋种子索取自助开发者向企业付费转化的数据
具名 logo 集中公开客户名单偏向 AI 原生软件和智能体构建者少数旗舰 logo 可能驱动市场感知,也可能贡献支出披露前 10 大客户收入占比和分层集中度
生态 / 伙伴依赖LangChain 和模型路由栈让 Exa 易于采用,也易于拿来与同业基准比较获客收益伴随替代风险提供直接渠道 vs 生态渠道 bookings,以及按渠道拆分的流失
品类压缩风险独立评测警告,原生模型供应商搜索可能吸收独立 AI 搜索价值即便使用量保持高位,也可能压迫定价和留存展示对捆绑搜索选项的胜负数据

扩张逻辑可信,但公开集中度披露缺失,生态依赖也有双刃剑效应。

[CU027, CU028, CU029, CU030, CU035, CU036]

6.6 展示材料

Chapter 07

07风险

7.1 风险格局排序

Exa 的风险栈不是单一的生死缺陷,而是四个相互咬合的问题。第一,公司的默认隐私和 IP 姿态比很多企业买家预期更宽松:公开隐私政策写明,查询数据可用于改进产品和微调模型;公开安全文档则把 Zero Data Retention 和 HIPAA 控制限定在企业版,覆盖面也窄于完整产品。第二,公司的核心优势来自自建爬取、索引、向量数据库和 GPU 机群,技术上亮眼,但经济和运营负担都重。第三,公开客户证据很强,却偏向前沿 AI 构建者和 agent 工具;即便 Exa 声称已有 5,000 多家公司、40 多万名开发者,依赖和集中度疑虑仍在。第四,围绕 AI 训练、检索和网络级内容使用的监管与版权边界,变化速度快过公司的公开法律披露。 实际含义是,Exa 更像高上行空间、高经营杠杆的基础设施。公司确有缓释资产——SOC 2 Type II、企业隐私控制、信任中心,以及多次聘请外部法律顾问——但这些缓释并未完全内置为自助产品默认项,公开记录也仍缺少 DPA 细节、事件指标、收入集中度、利润率披露和领导层纵深可见度。投资人要问的不是 Exa 有没有产品市场信号;它有。真正的问题是,隐私控制、法律卫生、资本纪律和依赖管理,是否能像采用率和估值一样快地成熟。[CR047, CR048, CR049, CR050, CR051, CR052]

监管 / 法律风险登记表
规则 / 问题司法辖区状态可能性严重性缓释措施剩余暴露尽调路径
默认查询数据使用和仅企业版 ZDR美国 / 全球活跃产品政策问题SOC 2 Type II、信任中心、仅企业版 ZDR 和 HIPAA 模式默认自助行为仍允许将查询数据用于训练 / 改进;敏感工作负载可能需要合同例外获取已签署 DPA、留存时间表,以及逐客户默认日志设置
AI 版权和训练数据争议美国全行业诉讼和政策辩论活跃条款要求输入合法;已聘请外部版权 / 隐私顾问Web 规模训练、抓取和 RAG 的法律状态仍未稳定,可能需要新的授权或过滤做法审查训练数据来源政策、robots / opt-out 处理,以及外部顾问备忘录
EU AI Act 加 GDPR 义务EU法律已生效;合规边界正在收紧信任中心、安全文档和企业控制支持采购就绪公开证据没有足够细地展示 AI 治理文档、DPIA 或欧盟传输架构索取欧盟合规矩阵、DPIA、SCC/DPF 立场和事件响应手册
公开 DPA / 处理方条款可见度美国 / 欧盟部分中高安全页面将买方引向信任中心和 DPA 材料公开抓取的材料没有披露生效 DPA 文本或留存附件,合同范围仍无法核验拉取实时 DPA、留存附件、子处理方清单和企业模板订单表
合同审计 / 终止 / 竞争性使用条款美国已在公开条款中生效企业附加条款可以覆盖公开条款如果后备选项薄弱,深度嵌入的客户仍可能面对供应商的不对称议价力对比公开条款与企业 MSA、SLA、终止和可迁移性条款

各行按剩余投资重要性排序,依据公开法律、监管和官方政策证据,而不是未经核验的私下陈述。

[CR002, CR003, CR004, CR005, CR006, CR007]
FR001: 风险热力图

剩余严重度集中在隐私默认设置、资本强度,以及依赖较重的基础设施运营。

定性排序综合引用的公开证据而来,不来自私有运营指标。

[CR047, CR048, CR049, CR050, CR051, CR052]

7.2 监管、隐私与知识产权风险

最尖锐的风险,是 Exa 默认产品行为与敏感企业工作负载预期之间的缺口。隐私政策明确写明,查询数据会用于改进产品并微调支撑服务的模型;政策也写明,Exa 不会主动监控查询文本中是否含个人信息。这一组合用于通用开发者搜索尚可管理,但并不天然契合那些期待默认日志最小化、机密提示词隔离或明确仅作为处理者处理数据的客户。Exa 自己的安全文档给出了缓释路径:SOC 2 Type II 控制、信任中心,以及 Zero Data Retention、HIPAA 模式等企业专属选项。问题在于,这些控制不是基线。HIPAA 模式只面向符合条件的企业团队,只支持 `/search` 和 `/contents`,拒绝实时检索和大量摘要路径;一旦请求需要非合规处理,系统会失败关闭。换句话说,产品可以为受监管用例做得更安全,但路径依赖合同,也会限制功能,而不是普遍默认开启。 相邻的 IP 和网络爬取风险也真实存在。Exa 的条款把合法用户输入责任推给客户,禁止用服务侵犯 IP 或合同权利,保留审计和终止权,并授予 Exa 对用户输入和输出的广泛许可,用于运营和改进服务。这不能证明存在问题,但意味着双方都在用合同分配法律不确定性。美国版权局 2025 年 Part 3 报告称,已有数十起 AI 版权诉讼待决,并把训练数据收集、RAG、合理使用和许可视为未决问题。与此同时,EU AI Act 保留 GDPR 和其他个人数据义务,也提高了围绕 AI 系统的文档和治理预期。Exa 多次使用外部隐私、网络安全和版权顾问,是正面信号;但公开记录仍无法让投资人核验实际生效的 DPA、训练数据授权立场或按司法辖区划分的合规项目。[CR001, CR002, CR003, CR004, CR005, CR006]

运营 / 质量 / 安全风险登记表
故障模式发生概率严重性缓释成熟度剩余暴露未解决缺口
核心搜索路径默认服务不稳定或延迟飙升中——公开状态页和分层服务面已存在2026 年 6 月事件记录显示确有宕机;按产品披露的公开可用率 / SLA 历史缺失需要 12 个月事件日志、MTTR 和 SLA 抵扣政策
计费或权益配置错误阻断客户工作负载中高中——有文档化的额度和自动充值控制余额耗尽后请求会停止;企业发票条款的公开说明不够深入关键账户需要企业计费兜底和支出控制
算力容量或采购瓶颈中——已部署规模可观的自有 GPU 集群224 块 GPU 足迹和接近 24/7 的使用强度意味着利用率高,也带来后续采购风险需要 GPU 供应合同、云端弹性扩容方案和容量余量政策
查询走向长尾并更智能体化后,质量下降中——自研向量 DB 和重排序栈在智能体规模下,搜索质量和覆盖度仍要和成本、延迟做取舍需要基准测试方法、误报 / 漏报率和回滚流程
受监管场景里,实时检索带来安全 / 合规摩擦中高中——受限路径已有 HIPAA 模式和 ZDR最有差异化的实时 / 深度功能无法原样用于合规模式需要客户案例,证明高价值实时工作流可在合规框架下使用

运营登记表混合了公开事件历史、产品控制和基础设施披露;私下 SRE 指标仍不可得。

[CR007, CR008, CR009, CR010, CR011, CR012]
FR002: 风险传导图

法律 / 隐私问题和基础设施依赖会直接传导到客户采用、利润率和估值风险。

[CR008, CR010, CR015, CR017, CR026, CR033]

7.3 运营可靠性与依赖风险

Exa 的基础设施故事同时是护城河,也是风险放大器。公司称其购入一套 $5 million 的 H200 集群,如今运行合计 224 块 GPU 的机群,几乎 24/7 运转,并自建专有向量数据库,在数十亿向量中实现低于 100 毫秒延迟和超过 500 QPS 的搜索。这些技术指标很亮眼,但也意味着业务背着有意义的固定成本承诺、利用率风险,以及横跨爬取、索引、重排、存储、计费和合作伙伴 API 的非平凡故障模式。公开状态历史已经显示,“default” 服务在 2026 年 6 月 14 日发生过多次短暂停机。计费文档也暴露出客户层面的失败路径:预付额度用完后,请求会被阻断。速率限制文档把自助搜索封顶在 10 QPS,并把更高吞吐用户推向企业合同。运营上这说得通,但也意味着可靠性不只取决于核心正常运行时间,还取决于权益管理、速率整形和客户账户配置。 依赖关系又加了一层。TechCrunch 报道称,Exa 一边运营自有 GPU 集群,一边把产品托管在 AWS;公开生态还显示,其进一步依赖托管 MCP 基础设施、支付通道,以及 LangChain、Browserbase 等合作伙伴集成。公开客户故事清楚表明,Exa 已嵌入 OpenRouter、Cognition 和 monday.com 的关键 agent 工作流。一旦 Exa 进入这些自动化链条,即便短暂停机或延迟尖峰,也会传导为下游产品故障。因此,正确的尽调视角不是 Exa 能否演示强搜索质量——它可以——而是它是否具备 SRE 成熟度、冗余、云 / GPU 采购纪律和合同 SLA,能按融资叙事暗示的规模,表现得像任务关键型基础设施供应商。[CR011, CR012, CR013, CR014, CR015, CR016]

合作伙伴 / 依赖风险登记表
依赖交易对手角色集中度故障场景严重性缓释剩余暴露
云托管AWS与自有集群一起承载产品工作负载高,但未完全量化AWS 宕机、调价或架构约束损害服务或利润率自有 GPU 和存储栈降低部分上游模型依赖公开冗余 / 多云立场未披露
GPU 硬件和数据中心供应NVIDIA / 硬件供应商 / 数据中心运营商训练和检索容量供应延迟或成本通胀拖慢扩张,或压缩毛利率已为集群扩张融资;现有 224 块 GPU 足迹采购合同、租赁条款和备用采购来源未公开
智能体和 IDE 分发Claude、Cursor、VS Code、Codex、Gemini CLI、MCP 客户端流量和开发者采用入口中高连接器政策或协议变化削弱分发,或推高支持成本中高托管 MCP 端点和广泛生态支持Exa 不控制第三方客户端用户体验或默认位置
智能体平台客户OpenRouter / Cognition / monday.com 及类似设计伙伴使用量、标杆背书和工作流嵌入Unknown少数大型 AI 原生账户主导增长,或同步流失多个工作流里的具名证据未公开集中度、留存或 ACV
支付和开票Stripe / 企业应付账款流程额度购买和账户连续性支付失败或支出控制错配中断工作负载自动充值和企业发票计费路径公开文档没有展示任务关键型计费兜底条款

交易对手来自公开文档、客户故事和新闻报道的推断;集中度百分比大多未披露。

[CR011, CR013, CR014, CR017, CR019, CR020]
人员 / 执行风险登记表
角色 / 职能依赖或缺口发生概率严重性缓释尽调路径
创始人 / CEO / 产品愿景公开叙事和战略责任仍集中在 Bryk/Wang 身上强创始人—市场匹配;产品转向和融资中反复跑出执行力索取接班计划、已下放的运营责任和董事会监督图谱
安全 / 隐私运营控制已存在,但公开实施细节比买方需要的更窄SOC 2、信任中心、HIPAA 模式和外部隐私法律顾问审查内部安全组织、隐私负责人、审计节奏和事件演练
企业 GTM 和账户管理需要把 AI 原生动能转化为多元化企业收入中高企业功能、发票计费和信任中心资产检查管线结构、头部账户依赖和续约责任归属
基础设施 / SRE 深度任务关键型可靠性依赖的团队,在公开记录中不可见公开状态页和自有栈工程人才索取基础设施、SRE、容量规划和值班领导层组织图

公开证据显示创始人能力强,但联合创始人层以下的梯队可见度有限。

[CR007, CR008, CR015, CR016, CR021, CR026]
FR003: 依赖图

Exa 处在技术栈中心,这套栈依赖云、GPU、协议,以及少数可见的智能体平台客户。

客户侧节点代表公开证明点,不代表已确认的收入集中度百分比。

[CR017, CR019, CR020, CR021, CR022, CR023]

7.4 客户、财务模型与执行风险

公开采用故事很有吸引力,但耐久性仍披露不足。Exa 点名的客户证据——OpenRouter、Cognition、monday.com、Databricks、Cursor,以及投资人强调的其他 AI 原生账户——显示产品切中了 agent 栈需求,却也让公开证据偏向一个相对狭窄的前沿 AI 和自动化买家群体。Exa 称其现在服务 5,000 多家公司和超过 40 万名开发者,但公开材料没有披露头部客户集中度、续约率、NRR、流失、按垂直行业划分的部署广度,或来自少数设计伙伴的收入占比。这很重要,因为公司的进入市场和产品叙事仍围绕恰恰最容易受到 OpenAI、Anthropic、Google 和大型应用在位者快速平台替代的客户。 财务模型风险由此直接延伸。Sacra 估计 2025 年收入约 $10 million,而这家公司公开运行昂贵的 GPU 和爬取基础设施;同时公司估值已抬到 $2.2 billion,并引导投资人想象每秒数十万次搜索。如果 Exa 能把前沿用量转化为耐久的企业合同,这套经济模型可以成立;否则,公司可能掉进熟悉的基础设施陷阱:用量增长快过毛利。执行层面还叠加创始人集中度。可见的公开领导层记录仍高度集中在 Will Bryk 和 Jeff Wang,管理梯队披露有限,也没有公开继任计划。这对早期基础设施公司并不罕见,但在这里更重要,因为 Exa 同时在扩大产品范围、基础设施强度、法律复杂度和企业销售动作。[CR022, CR023, CR024, CR025, CR028, CR031]

缓释和退出标准表
风险可监控触发项阈值 / 事件行动含义
隐私 / 查询数据风险默认留存和训练条款敏感企业层级没有默认仅处理方 / ZDR合同控制被证明前,不承保受监管工作负载的上行
可靠性风险状态事件和 SLA 指标反复出现客户可见宕机,或没有可信 MTTR/SLO 方案按企业就绪度薄弱的基础设施重新定价
资本强度风险毛利率、GPU 利用率和下一轮资本开支计划集群支出和云成本上升时,利润率停滞假设更多稀释和更低终局倍数
客户集中风险前 10 大收入占比和按客群批次的 NRR头部账户主导使用量,或前沿 AI 客群批次明显流失下调增长耐久性假设,重看估值
平台替代风险相对 OpenAI、Google、Anthropic 和捆绑套件的输赢数据现有平台吸收搜索用例的速度快于 Exa 增加差异化工作流将投资论点从品类领导者调为小众组件供应商
关键人风险领导梯队和接班准备度创始人以下没有可信深度,或创始人离任管理连续性去风险前,暂停信心

退出标准被定义为可观察事件或缺失控制,一旦出现会直接损害承保投资论点。

[CR007, CR008, CR010, CR012, CR013, CR014]

7.5 缓释措施、监控指标与论点破裂触发器

Exa 风险画像中令人鼓舞的一点是,许多最高严重度问题可以监控。公司已经暴露出一些正确的控制面:公开状态页、明确速率限制、企业安全文档、信任中心,以及足够坦诚、能看出默认项薄弱处的隐私政策。关键是把这些披露转化为尽调要求和触发器,而不是把它们当成安慰性声明。在承销一轮成长期融资前,投资人应坚持拿到已签署 DPA 和企业条款、按产品面划分的正常运行时间 / SLA 历史、云和 GPU 供应商集中度、头部客户收入占比、毛利率桥,以及能证明创始人以下领导层纵深的组织架构图。 论点会在三类事件任一发生时破裂。第一,隐私和数据使用控制跟不上企业采用——例如,受监管客户需要 Exa 最深的实时检索功能,却只能通过关闭这些功能来获得合规行为。第二,公司的基础设施杠杆反噬——持续事故、GPU / 云瓶颈,或利润率压缩迫使收入基数成熟前继续融资。第三,客户证据无法从少数前沿 AI 工作流中多元化,同时 incumbents 把搜索更原生地嵌入产品。如果这些事件没有发生,Exa 仍可能作为 agent 栈核心层继续复利;如果发生,估值下行会很陡,因为当前故事已经预设其成为品类领导者。[CR007, CR008, CR010, CR012, CR013, CR014]

7.6 图表

Chapter 08

08估值

8.1 投资建议与价格纪律

Exa 作为公司看起来可投,但按当前价格还未到可投。记录中的正面部分,对于一家私有 AI 搜索基础设施初创公司来说异常强:$250 million Series C、$2.2 billion 估值、蓝筹投资财团、超过 5,000 家公司客户、约 40 万名开发者,以及公开客户引用,显示真实工作流嵌入,而不是模糊的 logo 幻灯片。问题是,公开运营记录仍落后于融资记录。Sacra 最后一次公开收入估计仅约 $10 million(2025 年),公开来源没有披露毛利率、烧钱、留存或集中度,也没有任何已审阅来源披露优先股结构或二级市场悬顶。 在这个背景下,正确建议是跟踪,而不是买入。最新一轮价格要求外部投资人付费购买一个未来:Exa 迅速放大到远大得多的收入基数,保有足够产品边缘来抵御捆绑压力,并在计算密集成本结构下证明类似软件的经济性。公开证据尚未跨过这条线。纪律性投资人应把 $2.2 billion 当成标题参考,而非公允价值;只有当私下尽调显示收入、留存和利润率质量显著领先公开代理集合时,才应加码。[CV001, CV003, CV005, CV014, CV016, CV027]

建议摘要表
维度当前判断原因决策含义
建议跟踪公司质量有吸引力,但公开运营记录太薄,不足以支持按价投资。保持实时尽调覆盖;不要锚定 Series C 价格。
信心判断方向清楚,但依赖不完整的公开经济性证据。只有在审查私下收入、利润率和留存后才上调。
风险评级算力强度、竞争和缺失的资本结构条款带来不对称下行。任何新资金决策前,都要求更深尽调。
估值立场昂贵约为上一份公开收入估算的 220x,高于当前公开和私有可比公司组。需要更低入场价格,或好得多的私下数据。
决策含义仅列入观察名单Exa 值得密切跟踪,因为公司可能跑赢当前疑虑。只有在尽调或价格明显向投资者有利方向移动时才接触。

建议刻意对价格敏感:表格把公司质量支撑和入场价格支撑分开。

[CV001, CV005, CV016, CV027, CV031, CV039]
FV001: 建议逻辑

建议沿着一条链条展开:产品和融资支撑强,但价格支撑下的判断更弱。

这是决策链,不是法律意义上的股本结构流程图。

[CV001, CV003, CV016, CV031, CV039]

8.2 投资论点、反论点与融资背景

投资论点很直接。Exa 已经搭出一个实时网络检索栈,客户确实把它嵌入接近生产的 AI 工作流。公司在开发者、客户和投资人三端都有公开证据;其变现靠用量,而不是概念阶段的席位定价;同时它所在品类仍能吸引高溢价私募资本。如果管理层能把 search、contents、answer 和 agent 工作流转化为耐久的扩张收入,同时让延迟和召回领先更广泛的云捆绑,公司可能长进当前估值。 反论点同样清楚。Exa 自身披露指向资本密集型架构,私有估值也已经计入很大一部分执行故事。公开客户验证了效用,但没有验证收入耐久性。隐私政策和未披露经济性表明,企业就绪度仍依赖闭门尽调。与此同时,投资人对 AI 搜索的热情也在资助 ARR 披露规模大得多的同业。这让融资背景有支撑,但并不让人安心:这一轮降低了近期生存风险,却没有告诉新投资人,他们买到的是保护充分的上行轮领导者,还是一个不透明、只是受益于热门市场的股权结构表。[CV006, CV007, CV008, CV009, CV011, CV012]

正方 / 反方论点表
立场论点证据锚点什么会改变判断
正方真实客户嵌入说明 Exa 是基础设施,不是演示型产品。OpenRouter、monday.com 和 Cognition 案例,加上 5,000+ 公司声明。独立留存、集中度和扩张数据会增强信心。
正方蓝筹投资人让近期融资风险保持低位。a16z、Lightspeed、Benchmark、YC 和 $250M Series C。股权结构表透明度和募资用途纪律会提高舒适度。
正方如果搜索仍有差异化,按用量计价可随智能体采用而放大。公开定价,以及 search、contents、answer 和智能体的产品广度文档。按产品线证明毛利率后,模型会更容易承保。
反方当前价格已经计入大量未来成功。约为上一份公开收入估算的 220x,也高于同业组。如果当前收入基数明显更大,就能中和这条反对意见。
反方基础设施强度可能压住长期经济性。GPU 集群和向量数据库披露,加上分析师对成本基础的警示。端点级贡献利润率数据可以反驳这一担忧。
反方随着同业和捆绑套件成熟,搜索预算可能更容易被替换。Perplexity、Glean,以及品类内公开定价透明度。中性基准测试胜率和耐久 NRR 会降低商品化担忧。

各行说明当前价格要成立必须满足什么,以及哪些具体尽调能证伪任一方。

[CV003, CV006, CV007, CV008, CV009, CV012]
FV004: 投资 KPI

可供投委会使用的评分卡:Exa 在产品和融资支撑上得分高,在披露支撑的价格吸引力上得分低。

[CV028, CV030, CV031, CV032, CV039]

8.3 情景与可比公司集合

可比集合比轮次标题更能说明问题。按公开和半公开代理看,Exa 的价格高于一大批披露更充分的业务。Perplexity 2025 年私募估值隐含约 100x ARR,且披露收入规模大得多。Glean 2025 年融资隐含约 72x ARR,到 2026 年已报告 $300 million run rate。公开软件基础设施公司定价低得多:Elastic 约为 3.5x 市值 / 收入,Datadog 约为 23x。Exa 按最后公开收入估计约 220x 定价,因此即便放在火热的私有 AI 市场里,也不显得明显便宜。 这会把情景分析推向入场纪律。牛市情景中,Exa 复合增长足够快,能沿着通往九位数收入的路径、维持高溢价倍数,从而证明当前估值合理。基准情景中,增长仍强,但倍数正常化让相对最新轮次的上行空间有限。熊市情景中,搜索在 Exa 证明利润率质量前变得更容易替代,下一轮融资持平或下行。当前轮次可以成立,但前提是未来运营证据要迅速追上当前投资人热情。[CV017, CV018, CV019, CV020, CV021, CV022]

牛 / 基准 / 熊情景表
情景运营假设估值 / 回报逻辑概率信号关键风险
收入扩张到约 $120M-$160M,扩张强劲,企业控制更干净,利润率改善可信。按约 18x-25x 远期收入,价值落在 $2.2B-$4.0B,才终于支撑或超过今天的标记。需要私下数据证明收入质量已经远超公开估算。执行失误、捆绑压力或利润率转化薄弱都可能打破该情景。
基准增长仍强,但收入更接近约 $60M-$90M,软件经济性扎实但并非顶尖。按约 12x-16x,价值落在 $0.7B-$1.4B,意味着相对今天价格上行有限或为负。最符合当前公开证据:好公司,经济性不完整,倍数回归正常。本轮买方可能为最终质量付了过高价格。
收入扩张缓慢,留存参差不齐,或随着搜索更可替代,毛利率令人失望。按约 $25M-$40M 收入的 6x-10x,价值落在 $0.2B-$0.4B,并指向平轮 / 下轮融资。如果管理层无法在下一轮尽调周期证明经济性,该情景会更可能发生。下轮风险、士气受损和资本结构复杂度都会上升。

区间带有判断性,但锚定可比公司组,而不是 DCF 的虚假精确。

[CV016, CV023, CV025, CV027, CV034, CV035]
可比估值表
可比公司规模指标倍数 / 估值 / 状态相关性局限
Exa(2026 Series C)Sacra 估算 2025 年收入约 $10M按 $2.2B 私有估值,对应约 220x 隐含投后 / 估算收入直接审查的价格。收入代理来自第三方,优先权结构未披露。
Perplexity(2025 私募轮)TechCrunch 称 ARR 约 $200M;Sacra 后续估算 2026 年年化收入为 $500M$20B 估值;按约 2025 ARR 代理约 100x最接近的 AI 搜索增长可比公司,投资人兴趣清晰。消费者 / 产品组合和规模都远大于 Exa 公开记录。
Glean(2025 Series F)本轮时 ARR >$100M;到 2026 年 5 月年化收入 $300M$7.2B 估值;按披露 ARR 里程碑约 72x可作为搜索加智能体工作流热度的私有可比公司。企业搜索产品比 Exa 更偏内部数据。
Datadog(公开)FY2025 收入约 $3.43B2026 年 6 月市值 / 收入约 23x高质量公开基础设施软件的溢价倍数天花板。规模远大于 Exa,平台也更宽。
Elastic(公开)FY2026 收入约 $1.74B2026 年 6 月市值 / 收入约 3.5x低倍数搜索 / 检索基准,有公开申报文件。成熟上市公司,产品组合和增长画像与 Exa 差异很大。

私有公司行使用已披露估值和 ARR/收入里程碑;公开公司行使用 2026 年 6 月市值和最新 SEC 申报收入。

[CV016, CV019, CV022, CV023, CV024, CV025]
FV002: 估值敏感性

只有当 Exa 做到大得多的收入规模、且仍守住高溢价倍数时,当前价格才显得合理。

数值只是基于 $2.2B 标记价和从可比公司组中选取的倍数区间做简单代数计算。

[CV016, CV019, CV022, CV023, CV025, CV026]
FV003: 估值 / 回报区间

情景区间明显偏斜:当前价格主要只在牛市情景下说得通。

区间反映与收入规模和常态化倍数区间挂钩的情景逻辑,不是 DCF。

[CV034, CV035, CV036, CV037]

8.4 最终尽调问题与论点破裂触发器

公司质量与价格质量之间的缺口仍可弥合,但只能靠私下证据。尽调议程应聚焦四个问题:2025 年退出、进入 2026 年时的收入和 ARR 到底是多少;在爬取、推理和支持成本之后,毛利率是否改善;客户集中度和留存是否支撑高溢价估值假设;新投资人实际会坐在资本结构的哪个位置。没有这些答案,IC 承销的就会是叙事,而不是经济性。 论点破裂点很具体。如果管理层无法展示可信的近期收入路径,来缩小与当前估值的差距,价格就应被视为过高。如果增长真实但利润率质量薄弱,这一轮仍可能破坏价值。如果下一轮融资或员工流动性事件把定价重置到低于 Series C 标记,那会说明公开采用证据不足以撑住私募市场。在尽调补齐这些缺口之前,审慎姿态是保持接近、把公司留在观察名单上,等待更好的证据或更好的入场点。[CV014, CV031, CV038, CV039, CV040, CV041]

投资论点破裂和退出触发项表
触发项阈值对投资论点的传导行动含义
收入证明达不到公允价值门槛管理层无法展示一条可信路径,让收入足够快地进入至少高双位数百万美元,从而支撑当前标记。打破“经营证明会追上融资动能”的论点。除非价格大幅下调,否则后退。
单位经济性仍不透明或偏弱深入尽调后,毛利率和贡献利润率仍偏薄或恶化。把产品赢家变成资本黑洞。不承保当前价格。
留存 / 集中度令人失望头部客户主导支出,或扩张弱于增长叙事暗示。让按用量变现的耐久性低于牛市情景假设。重新评级为回避,除非价格补偿风险。
搜索变得更容易替换中性基准测试显示,相对同业或捆绑替代品几乎没有持久质量优势。同时压低增长预期和估值倍数。按大宗商品化 API 供应商重新看待这家公司。
融资重估或资本结构出现反向意外下一轮融资、二级交易或内部估值低于 Series C,或显示出很强的下行保护。说明当前名义估值夸大了普通股价值。暂停推进,按新条款而非旧名义估值重新承销。

触发项刻意锚定可衡量的运营或融资事件,而不是泛泛的不安。

[CV031, CV032, CV034, CV037, CV040, CV042]
最终尽调问题表
主题缺失证据重要性负责人或尽调路径
收入桥接董事会或管理层按产品和客户 cohort 拆分,从 2025 年实际收入桥接到当前 2026 年收入 run rate。这是估值争议背后最核心的缺失输入。任何 IC 投票前,先索取 CFO 材料或 data room 导出。
毛利率与单位经济混合口径和产品级毛利率,包括抓取、推理和支持成本。没有这些,就无法给增长质量定价。索取财务 / 基础设施成本材料,并与使用量指标对账。
留存与集中度NRR、毛留存、logo 集中度和前 20 大客户收入占比。只有客户能持续扩张且集中度可控,高倍数才站得住。索取 cohort 表和头部账户清单。
资本结构与优先权清算优先权、参与权、二级交易、期权池,以及任何债务或硬件承诺。如果保护条款很重,名义估值可能夸大普通股价值。索取 cap table、股份购买协议和债务清单。
竞争证明中立基准下的胜率、流失原因和第二供应商使用情况。需要判断 Exa 到底有差异化,还是只是卡在热门品类的早期位置。做客户访谈,并把该工作负载与同业供应商跑基准。

这些问题是闸门项,不是锦上添花;每一项都可能改变估值立场。

[CV014, CV029, CV031, CV038, CV042]

免责声明

本报告的元判断基于运行日期可获得的公开来源和公司提供来源。它不构成投资、法律、会计或税务建议;私营公司条款或业绩可能与公开记录存在重大差异。

证据索引

结论
编号陈述可信度来源
CO001 Exa is a San Francisco-based private AI search infrastructure company founded in 2021. SO013, SO014, SO023
CO002 The company was originally called Metaphor and publicly renamed itself Exa on 2024-01-25 to better reflect its mission. SO008, SO023
CO003 Exa's current product is an AI-native web search and retrieval API designed for agents, coding tools, and research workflows rather than ad-supported human browsing. SO001, SO010, SO020
CO004 Exa monetizes primarily through usage-based API pricing, while enterprise plans add custom limits, security terms, and Zero Data Retention. SO003, SO022
CO005 Exa Labs Inc. lists 430 Shotwell Street, San Francisco, California 94110 as its contact address in the privacy policy. SO004
CO006 Will Bryk is Exa's co-founder and CEO, and Jeff Wang is the other named co-founder in current public materials. SO002, SO013, SO014
CO007 Bryk studied computer science and physics at Harvard and previously worked as an engineer at Cresta before starting Exa. SO002, SO013
CO008 Wang studied computer science and philosophy at Harvard and previously built data and web infrastructure at Plaid. SO002, SO013, SO016
CO009 Exa says its crawlers track more than 500 billion URLs and that it trains search models on its own GPU cluster. SO002, SO007
CO010 Exa says it launched its first search engine in November 2022. SO006
CO011 By early 2023 Exa had pivoted its public focus toward being a web search API for AI products and agents. SO007, SO014
CO012 The January 2024 rebrand to Exa was paired with the launch of Highlights for extracting semantically relevant webpage passages. SO008
CO013 Exa announced $22 million of combined seed and Series A financing on 2024-07-16, with a $17 million Series A led by Lightspeed and participation from NVentures and Y Combinator. SO005, SO014, SO015
CO014 The public 2024 financing narrative implies roughly $5 million of prior seed backing before the $17 million Series A. SO014, SO007, SO028
CO015 Exa's September 2025 Series B raised $85 million at a $700 million valuation, led by Benchmark with participation from Lightspeed, Y Combinator, and NVentures. SO006, SO017, SO018
CO016 Benchmark partner Peter Fenton joined Exa's board as part of the Series B financing. SO006, SO018
CO017 Series B proceeds were allocated to a 5x larger GPU cluster and broader hiring across engineering, go-to-market, and operations. SO006, SO022
CO018 Exa launched a revamped Exa Deep endpoint on 2026-03-04 with structured outputs, field-level grounding, and lower pricing than the prior Deep tier. SO009, SO003
CO019 Exa's May 2026 Series C raised $250 million at a $2.2 billion valuation led by Andreessen Horowitz, with Benchmark, Lightspeed, and Y Combinator also participating. SO007, SO019, SO021
CO020 Multiple public Series C reports say a16z general partner Sarah Wang joined Exa's board in conjunction with the 2026 financing. SO019, SO021
CO021 Adding the publicly disclosed seed, Series A, Series B, and Series C amounts implies approximately $357 million of lifetime capital raised by May 2026. SO007, SO021, SO028
CO022 Sacra still showed approximately $107 million of total funding because its snapshot stopped at the September 2025 Series B and therefore lagged the 2026 Series C. SO022
CO023 Public governance disclosure remains sparse beyond the Peter Fenton and Sarah Wang board-seat reports, and the reviewed official pages do not publish a full board, cap table, or voting-control summary. SO002, SO007, SO013
CO024 Exa says it powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com, and more than 400,000 developers. SO002, SO007
CO025 Exa said on its Series C announcement that the number of companies using its platform had grown to over 5,000. SO007, SO019
CO026 Public customer-proof includes Databricks using Exa for dataset discovery and Browserbase publishing an Exa-based automation template for AI job search. SO014, SO025
CO027 LangChain maintains both Python and JavaScript integrations for Exa, indicating active framework support in the agent developer ecosystem. SO026, SO027
CO028 Exa publishes an MCP server integration guide covering Claude, Cursor, VS Code, Codex, Gemini CLI, Windsurf, and other coding-agent clients. SO011
CO029 Exa's careers page says the company is fully in-person with limited exceptions and is willing to sponsor visas for exceptional candidates. SO012
CO030 Y Combinator's Exa company profile listed 75 employees in San Francisco as of the June 2026 retrieval. SO013
CO031 The Silicon Valley Post reported around 100 employees at Exa in May 2026. SO019
CO032 Growjo estimated 294 employees and 87% year-over-year employee growth, materially above the lower public profile counts. SO024
CO033 Because public headcount figures range from 75 to 294, Exa's current staffing level should be treated as a low-confidence diligence item rather than a settled KPI. SO013, SO019, SO024
CO034 The public record clearly supports San Francisco as Exa's headquarters and a worldwide hiring footprint, but it does not provide a consistently verified list of secondary offices. SO002, SO004, SO023
CO035 Exa's privacy policy says query data is used to improve products, including training and fine-tuning models, while also warning users not to submit personal information in open-text query fields. SO004
CO036 Zero Data Retention is presented as an enterprise feature rather than the default behavior for all Exa users. SO003, SO004
CO037 Exa's public product stack now spans search, contents, answers, monitors, deep search, async agents, and Websets-style structured collection workflows. SO003, SO010
CO038 Sacra identifies OpenAI, Anthropic, Google, Brave, Tavily, Jina AI, and Perplexity as relevant competitive or substitute pressures on Exa's position. SO022
CO039 Exa's core cost structure is capital intensive because it operates its own crawling, indexing, and GPU-heavy retrieval infrastructure. SO007, SO022
CO040 Bryk and Wang's search and infrastructure backgrounds create strong founder-product fit, but they also concentrate product vision, recruiting magnetism, and fundraising narrative in a very small founding group. SO002, SO014, SO016
CO041 Lightspeed said Websets became a human-facing product by September 2025 and was gaining traction in recruiting, lead generation, and market research use cases. SO018
CO042 Exa publicly markets sub-200ms search and 20x token reduction from its extraction stack for latency-sensitive agent workflows. SO007, SO010
CO043 Sacra estimated Exa at roughly $10 million of revenue in 2025, but no official audited revenue or ARR disclosure was found in the reviewed company materials. SO022
CO044 The public milestone pattern is search engine launch in 2022, AI API pivot in 2023, rename and Highlights in 2024, Websets/Series B scale-up in 2025, and Deep plus Series C expansion in 2026. SO006, SO008, SO009, SO018, SO019
CM001 Exa positions the category as one API for web search, crawling, and research agents rather than a generic search product. SM001
CM002 Exa's documentation splits the workflow into search, answer, and MCP layers, indicating that the product sells live-web retrieval infrastructure for agents. SM003, SM005, SM022, SM032
CM003 Included spend in this market is usage-based search, extraction, answer, and research-run consumption that helps agents retrieve and ground public-web information. SM001, SM002, SM022, SM024, SM032
CM004 Glean's positioning around internal knowledge, enterprise systems, and permission-aware answers shows that enterprise knowledge search is an adjacent but excluded budget from Exa's public-web retrieval category. SM027
CM005 Elastic sells a broad private-data retrieval platform spanning structured data, vectors, analytics, and agent workflows, which is adjacent infrastructure rather than public-web search as a service. SM018
CM006 Algolia's NeuralSearch and merchandising stack target site search and conversion workflows, making commerce search another adjacency that should be excluded from Exa's market boundary. SM029
CM007 Google Custom Search JSON API remains a status-quo programmable substitute, but Google says it is unavailable to new customers and will discontinue on 2027-01-01. SM012
CM008 The real substitutes for Exa are legacy programmable search APIs, competing public-web retrieval APIs, and in-house stacks assembled from broader search and vector tools. SM006, SM012, SM018
CM009 Exa prices search at $7 per 1,000 requests, deep search at $12, deep-reasoning search at $15, contents at $1 per 1,000 pages, and agent requests from $0.012 to $2.00 per run. SM002
CM010 Exa says it already powers search for over 400,000 developers and has grown to over 5,000 companies using Exa. SM038
CM011 Exa says its crawlers track over 500 billion URLs and that it is scaling infrastructure to handle hundreds of thousands of searches per second. SM038
CM012 Tavily claims 300 million monthly requests handled, 2 million-plus developers, 99.99% uptime, and 180 ms p50 search latency. SM009
CM013 Tavily prices its service with 1,000 free monthly credits and $0.008 per credit pay-as-you-go, signaling low-friction developer entry pricing. SM010
CM014 Brave markets Answers at $4 per 1,000 requests plus $5 per million tokens and says all plans include $5 in free monthly credits. SM006, SM033
CM015 Brave says its independent index covers more than 30 billion pages and more than 100 million daily page updates. SM006
CM016 Grand View Research sizes the adjacent enterprise search market at $4.867 billion in 2023 and $8.852 billion by 2030, but that measure is dominated by internal information access use cases. SM020
CM017 The reviewed public corpus does not provide a standalone third-party TAM or SAM for AI-native public-web retrieval infrastructure for agents. SM002, SM006, SM009, SM012, SM020
CM018 Using enterprise-search analyst TAM as Exa's direct TAM would overstate the opportunity because those reports include intranet, document, and private-data retrieval budgets that Exa does not directly monetize. SM020, SM027, SM018
CM019 OpenAI says developers and enterprises are now building agents with built-in web search, file search, computer use, and orchestration SDKs. SM015
CM020 Anthropic says MCP exists to connect AI assistants to repositories, business tools, and development environments, and cites early adoption by companies such as Block and Apollo plus dev-tool vendors. SM035
CM021 The MCP specification says the protocol is supported across Claude, ChatGPT, VS Code, Cursor, and other clients, reducing the cost of plugging retrieval tools into agent workflows. SM017
CM022 Exa's MCP server README shows direct install paths for Cursor, VS Code, Claude, Codex, Gemini CLI, and other agent clients, indicating that developer-tool distribution is a primary go-to-market channel. SM005
CM023 Brave's skills repository targets Claude Code, Cursor, GitHub Copilot, Codex, Gemini CLI, VS Code, Windsurf, OpenClaw, and other coding agents, reinforcing coding-assistant vendors as a live buyer segment. SM008
CM024 LangChain exposes Tavily Search, Extract, Crawl, and Map as tools, showing how agent frameworks can become distribution wedges for retrieval vendors. SM014
CM025 Glean's emphasis on permission-aware enterprise context and lower token usage shows that enterprises often separate internal knowledge budgets from public-web retrieval budgets even when the end-user workflow looks similar. SM027
CM026 Exa's Answer endpoint packages cited answers rather than raw ranked results, implying that some buyers will pay for outcome-level research APIs instead of only low-level search calls. SM032
CM027 OpenAI's agent tooling makes web search a default building block for new agents, which should expand the addressable demand for retrieval vendors that integrate cleanly into agent stacks. SM015
CM028 MCP standardization expands the number of client surfaces where retrieval vendors can plug in without bespoke integrations, improving distribution leverage. SM017, SM035, SM005
CM029 Tavily's research endpoint and Exa's agent-run pricing both show category expansion from simple search APIs toward multi-step research workflows. SM002, SM024
CM030 Google's planned retirement of Custom Search JSON API for new customers creates a migration tailwind toward newer agent-native retrieval vendors. SM012
CM031 Because buyers can still solve parts of the job with Glean, Elastic, or other internal retrieval stacks, budget ownership for public-web retrieval infrastructure is often ambiguous at first deployment. SM018, SM027
CM032 Category pricing is heterogeneous across per-query, per-page, per-token, per-credit, and per-run units, which weakens apples-to-apples TAM estimation and procurement comparison. SM002, SM006, SM010
CM033 Security and governance requirements can lengthen adoption because Brave emphasizes SOC 2 and Zero Data Retention while Glean emphasizes permissions and enterprise governance controls. SM006, SM027, SM033
CM034 Exa says customers choose it because it is the highest quality search API at every latency and price point. SM038
CM035 Brave says its API can replace smaller-index competitors like Exa and Tavily and that higher-quality grounding can let open-weight models beat ChatGPT, Perplexity, and Google AI Mode. SM006, SM033
CM036 Tavily says its /search endpoint is the fastest on the market at 180 ms p50. SM009
CM037 Exa says it built the fastest search API in the world at sub-200ms and cut text extraction token counts by more than 20x. SM038
CM038 The most plausible near-term buyer segments are coding-assistant vendors, agent/application developers, enterprise AI platform teams, and research or operations automation teams. SM005, SM008, SM015, SM017
CM039 Users are usually engineers building agent loops, while payers are more likely platform, product, CTO, or enterprise AI leaders rather than the end users of the final agent. SM015, SM017, SM027
CM040 Adoption commonly starts with a developer integration in a framework, IDE, or MCP client and later expands into governance, SLAs, and larger contract packaging once usage scales. SM005, SM008, SM014, SM017, SM025
CM041 Because this market sells infrastructure into agent loops, the commercially important variables are relevance quality, latency, coverage, extraction quality, and workflow fit rather than consumer search share or advertising reach. SM001, SM002, SM006, SM015
CM042 The market boundary should exclude browser subscriptions, search advertising, and standalone vector-database spend because those budgets do not directly buy public-web grounding for agents. SM006, SM018, SM027
CP001 Exa sells a web-native API stack for search, crawling, and research agents. SP001, SP003
CP002 Exa public list pricing starts at $7 per 1,000 Search requests, $12 per 1,000 Deep Search requests, $15 per 1,000 Deep-Reasoning Search requests, $1 per 1,000 Contents pages, and $5 per 1,000 Monitors requests. SP002
CP003 Exa Agent fixed-effort pricing runs from $0.012 to $1.00 per request with separate charges for search tool calls and contact enrichment. SP002
CP004 Exa gives new accounts $10 in onboarding credits and $7 in monthly credits when a payment method is on file. SP005
CP005 Exa default limits are 10 QPS on /search and 100 QPS on /contents. SP006
CP006 Exa markets SOC 2 Type II certification plus optional zero-data-retention and HIPAA enterprise arrangements. SP007, SP008
CP007 Exa HIPAA mode is limited to compliant cached-retrieval workflows and can reject live-retrieval search paths. SP008
CP008 Exa Company Search indexes more than 50 million companies and updates weekly. SP004
CP009 Exa distributes through remote MCP across Claude, Cursor, VS Code, Codex, Gemini CLI, and other assistants. SP010
CP010 Exa’s Python SDK exposes search, answer, contents, streaming, and agent-run primitives that reduce integration friction. SP011
CP011 Exa’s 2024 launch post positioned the company as neural search that filters by meaning rather than by keyword matching. SP009
CP012 Tavily packages search, extraction, research, and web crawling in one API. SP012
CP013 Tavily claims 300M+ monthly requests, 2M+ developers, 99.99% uptime, and 180 ms p50 search latency. SP012
CP014 Tavily’s research endpoint supports mini, pro, and auto modes plus JSON-schema-structured output. SP014
CP015 Tavily public pricing spans a free 1,000-credit monthly tier and pay-as-you-go usage at $0.008 per credit, with enterprise custom terms. SP013
CP016 Brave Search API markets an independent web index with $5 in free monthly credits and 50 QPS search capacity. SP015
CP017 Brave’s Answers endpoint is OpenAI SDK compatible and priced at $4 per 1,000 requests plus token charges. SP015
CP018 Brave says its index covers over 30 billion pages and receives over 100 million page updates every day. SP015
CP019 Brave advertises SOC 2 Type II attestation and zero data retention for enterprise deployments. SP015
CP020 Brave explicitly says its API can replace smaller-index options like Exa or Tavily and higher-latency options like SerpAPI. SP015
CP021 SerpAPI’s breadth is distribution-based rather than index-based, with APIs across Google, Bing, Amazon, Maps, News, Scholar, Shopping, and many other vertical surfaces. SP016
CP022 SerpAPI public plans run from free 250 searches per month to $25 for 1,000, $75 for 5,000, $150 for 15,000, and $275 for 30,000 searches, with enterprise custom tiers above that. SP016
CP023 SerpAPI markets ZeroTrace Mode, SOC 2 Type II, SOC 3, ISO 27001, and a 99.95% SLA as enterprise trust features. SP016, SP017
CP024 Perplexity exposes an Agent API, Search API, Sonar API, MCP Server, and OpenAI compatibility in one developer stack. SP019
CP025 Perplexity prices raw Search API requests at $5 per 1,000 and its Agent API web_search tool at $0.005 per call while passing third-party model rates through with no markup. SP018
CP026 TechCrunch reported that Sonar launched with base and Pro tiers, customizable sources, and Zoom as an early enterprise user. SP020
CP027 Google’s Custom Search JSON API is unavailable to new customers and scheduled for discontinuation on 2027-01-01. SP021
CP028 Google Grounding can connect Gemini to Google Search, Agent Search, RAG Engine, Elasticsearch, any external search API, regulated web grounding, and parallel web search. SP022
CP029 Algolia competes as owned-content search infrastructure with NeuralSearch, ranking controls, analytics, merchandising, and crawl-based ingestion rather than a public-web index. SP023
CP030 Elastic competes as a buildable search platform with full-text, vector, and hybrid retrieval, more than 350 integrations, and open-source deployment options. SP024
CP031 Elastic’s hosted service adds multi-cloud management, a 99.95% monthly uptime SLA, and higher-tier AI Assistant and Agent Builder features. SP025
CP032 Glean competes as permission-aware enterprise search and agent infrastructure across internal systems rather than as a general web index. SP026
CP033 Glean claims support for 35+ LLM models, 30% lower token usage, and 93% enterprise adoption in under two years on highlighted deployments. SP026
CP034 LangChain lowers internal-build friction by giving teams a configurable agent harness that already supports OpenAI, Anthropic, Google, and other model providers. SP027
CP035 Exa’s strongest direct differentiation is the combination of live web retrieval, code-aware search, structured company enrichment, and agent-native tooling in one vendor stack. SP001, SP003, SP004, SP010
CP036 Exa is not obviously the low-price leader in public list pricing because Brave’s Answers API and Perplexity’s raw Search API are both priced around $4–$5 per 1,000 requests while Google Custom Search also lists $5 per 1,000. SP015, SP018, SP021
CP037 Multi-homing is structurally easy because Google Grounding accepts any search API, LangChain abstracts model providers, and Brave and Perplexity both advertise compatibility layers that reduce integration rewrites. SP015, SP019, SP022, SP027
CP038 Usage-based contracts and stateless API surfaces make direct-peer switching costs lower than classic enterprise workflow-software lock-in. SP002, SP013, SP015, SP018
CP039 Exa’s trust moat is real but not unique because Brave, SerpAPI, Glean, and Google all make explicit security or compliance claims that a procurement team can compare directly. SP007, SP015, SP017, SP022, SP026
CP040 Incumbent substitutes win when the buyer’s primary need is internal permissioning or owned-data relevance rather than web-native recall. SP022, SP023, SP024, SP026
CP041 Commodity pressure is rising because Tavily and Perplexity both extend from raw search into research or agent workflows, shrinking the novelty of Exa’s agentic positioning. SP012, SP014, SP018, SP019, SP020
CP042 Google’s shift from a shrinking Custom Search product to a broader grounding platform raises the competitive bar from search API to search plus enterprise AI stack. SP021, SP022
CP043 Exa’s MCP and SDK distribution give it a lower-friction developer GTM path than enterprise substitutes that rely on heavier platform rollouts or custom sales. SP010, SP011, SP026
CP044 SerpAPI’s many engine connectors can reduce a buyer’s dependency on any single web index even if they increase dependence on a wrapper vendor. SP016
CI001 Exa prices standard Search at $7 per 1,000 requests with up to 10 results included and $1 per 1,000 additional results above 10. SI002
CI002 Exa prices Deep Search at $12 per 1,000 requests and Deep-Reasoning Search at $15 per 1,000 requests. SI002
CI003 Exa prices the Contents endpoint at $1 per 1,000 pages per content type and the Answer endpoint at $5 per 1,000 requests. SI002
CI004 Exa Agent is priced separately from core search with a quoted range of $0.012 to $2.00 per run depending on effort. SI002, SI005
CI005 Agent runs also meter usage through $0.10 Agent Compute Units and $0.005 search tool calls, making Exa Agent a two-part compute plus tool-retrieval product. SI002, SI005
CI006 Exa bills self-serve customers through a prepaid credit balance that blocks requests when the balance runs out. SI003
CI007 New accounts receive $10 of onboarding credits and funded accounts receive an additional $7 of monthly credits that do not roll over. SI003
CI008 Enterprise customers can negotiate postpaid invoice billing rather than relying solely on prepaid credits. SI003
CI009 Exa's pricing page positions enterprise contracts around custom pricing, volume discounts, higher result counts, custom QPS, SLAs, and zero-data-retention features. SI002
CI010 The default public rate limit is 10 QPS for /search, 100 QPS for /contents, and 15 concurrent tasks for the deprecated /research/v1 workflow. SI004
CI011 Enterprise buyers are explicitly told to contact sales when they need higher rate limits than the documented defaults. SI004, SI002
CI012 Monitors run recurring Exa searches on a schedule and deliver de-duplicated results to webhooks, creating a recurring intelligence workload rather than a one-shot query. SI006
CI013 Websets monetizes higher-value enrichment workflows by verifying results against criteria and extracting structured fields such as funding amount and contact information. SI009
CI014 Exa says more than 5,000 companies were using the product by May 2026. SI011, SI022
CI015 Exa says it serves over 400,000 developers. SI001, SI011
CI016 OpenRouter says usage through its Exa-backed search path reached 73 million search queries to date from 2.36 million plus in the prior year. SI016
CI017 monday.com uses Exa inside AI lead agents for prospecting, enrichment, qualification, and routing inside monday CRM. SI017
CI018 Cognition states that Exa powers all parts of Devin's web search capability. SI018
CI019 LangChain maintains a dedicated langchain-exa package and native Exa retriever/tool examples, lowering developer onboarding friction for self-serve adoption. SI024
CI020 Y Combinator describes Exa as web search rebuilt for LLMs, reinforcing an API-first developer go-to-market identity rather than a seat-based SaaS workflow. SI023
CI021 Sacra estimates Exa reached $10 million of revenue in 2025, up roughly 1,010 percent year over year from an estimated $0.9 million in 2024. SI021
CI022 Exa's July 2024 funding announcement said revenue had tripled in the preceding few months without disclosing the starting base. SI012, SI019
CI023 No reviewed public source discloses Exa ARR, net revenue retention, or customer concentration. SI011, SI021, SI025
CI024 The public record does not separate self-serve credit revenue from enterprise contract revenue. SI002, SI003, SI011, SI021
CI025 The self-serve documentation shows prepaid credits and standard QPS while the enterprise documentation emphasizes negotiated billing, custom QPS, and security terms. SI002, SI003, SI004
CI026 Exa says its crawlers track over 500 billion URLs. SI011
CI027 Exa says it built the fastest search API in the world with sub-200 millisecond latency and text extraction that cuts LLM token counts by over 20 times. SI011, SI007
CI028 The Contents API can live-crawl, respect cache freshness windows, and crawl linked subpages, so delivery costs can expand with freshness and breadth settings. SI008
CI029 Exa recently purchased a $5 million 144-H200 GPU cluster that runs nearly 24/7 to support retrieval-model training and embedding workloads. SI013
CI030 Exa's combined old and new clusters total 224 GPUs, roughly 26.4 TB of GPU RAM, and about 350 TB of local NVMe storage. SI013
CI031 Exa built a proprietary vector database that targets billions of vectors, under-100 millisecond search, and more than 500 queries per second at reasonable cost. SI014
CI032 Exa says its custom vector database cut costs roughly 10 times versus quoted cloud vector database services. SI014
CI033 Exa says it owns the crawling, indexing, embedding, vector-database, and query stack instead of wrapping another search engine. SI011, SI014
CI034 Exa raised $250 million in a May 2026 Series C at a $2.2 billion valuation led by a16z with Benchmark, Lightspeed, and YC participating. SI011, SI020, SI022
CI035 Exa raised $22 million of seed and Series A capital in July 2024 led by Lightspeed with NVentures and YC participating. SI012, SI019
CI036 Management says the 2026 capital will fund next-generation models, infrastructure scaled to hundreds of thousands of searches per second, and a larger go-to-market organization. SI011, SI022
CI037 The public materials reviewed do not disclose current cash on hand, monthly burn, or post-Series-C runway. SI011, SI020, SI021, SI025
CI038 No reviewed public source discloses debt facilities, project finance, or other non-equity obligations for Exa. SI011, SI020, SI025, SI026
CI039 No reviewed public source provides gross margin by endpoint or a blended gross margin figure for Exa. SI011, SI021, SI025
CI040 No reviewed public source provides CAC, payback, NRR, or discount-rate data sufficient to assess sales efficiency. SI011, SI021, SI025
CI041 Exa's Snowflake integration requires a paid Snowflake account with External Access support, a dedicated warehouse, and an Exa API key, implying enterprise-style deployment overhead. SI010
CI042 AIbase described the Exa and OpenRouter partnership as bringing real-time web search to over 400 language models via a shared integration path. SI027
CI043 Sacra characterizes Exa's business model as entirely usage-based, with monetization tied to search queries, content retrieval, and AI-generated answers rather than seat licenses. SI021
CI044 Sacra estimates Exa had raised about $107 million and was valued at $700 million after the 2025 Series B, before the 2026 Series C. SI021
CI045 The SEC public search portal is the standard public entry point for filer history, yet the reviewed source set did not surface filing-grade financial statements for Exa. SI025, SI011
CI046 Delaware's free entity search exposes only basic formation details while charged copies are required for underlying filed documents, limiting what outsiders can verify without paid records. SI026
CI047 Fresh capital substantially reduces near-term solvency risk, but the lack of public burn and margin data prevents a firm runway conclusion. SI020, SI011, SI021
CI048 Because Exa owns crawling, vector retrieval, and model-serving infrastructure, its unit economics depend on realized usage margins rather than on list pricing alone. SI002, SI008, SI013, SI014, SI021
CI049 Exa's capital intensity is unusually high for an API startup because it funds both model-training clusters and web-scale indexing infrastructure. SI011, SI013, SI014, SI021
CI050 Exa is not underwriteable on public fundamentals today because the sources show strong demand and capitalization but not realized margins, sales efficiency, or filing-grade financial statements. SI011, SI021, SI025, SI026
CE001 Exa positions its product as one API for search, crawling, and research agents. SE001
CE002 Exa’s getting-started documentation presents /search, /contents, /answer, and /research as the core first-party API surfaces. SE002
CE003 Exa documents six search modes—auto, instant, fast, deep-lite, deep, and deep-reasoning—with published latency-quality tradeoffs. SE003
CE004 Exa says search supports output_schema and system_prompt across search types for grounded text or structured extraction. SE003, SE024
CE005 Contents API can return full text, highlights, summaries, and crawled subpages with configurable freshness controls such as maxAgeHours. SE004
CE006 Exa markets highlights as token-efficient extracts for LLMs and claims about 10x token efficiency versus fuller webpage context. SE003, SE001
CE007 Exa’s public retrieval surface extends beyond generic web search into typed categories including company, people, research paper, news, personal site, and financial report. SE001, SE003
CE008 Company Search indexes 50M+ companies, updates weekly, and returns structured entity metadata such as workforce, headquarters, financials, and traffic fields. SE008
CE009 Agent API creates asynchronous runs that can search, read, reason, enrich rows, and return either natural-language or structured outputs with citations. SE005, SE024
CE010 Agent runs are created, polled, streamed, and cancelled through dedicated run and event endpoints and can terminate by completion, failure, cancellation, or timeout. SE005
CE011 Exa prices Agent runs through effort-tier request charges plus Agent Compute Units and per-search tool-call fees. SE005
CE012 Monitors run Exa searches on recurring schedules, deduplicate new results across runs, and deliver outputs to webhooks. SE006
CE013 Websets turns search into an asynchronous pipeline that verifies criteria, enriches each match, supports imports, and can keep datasets fresh with monitors and webhooks. SE007
CE014 Exa MCP is open source, exposes search and fetch tools, and can be consumed either from a hosted remote endpoint or via the exa-mcp-server npm package. SE009, SE021, SE026
CE015 The docs index shows Exa packaging MCP, SDKs, integrations, changelog entries, and multiple coding-agent references around the core API. SE013
CE016 Exa’s integrations directory spans frameworks, automation tools, voice platforms, data platforms, and model ecosystems rather than a single integration path. SE015
CE017 Exa’s Snowflake integration uses External Access and stored procedures so search and contents can run inside Snowflake SQL and Cortex Agent workflows. SE016
CE018 Browserbase’s Exa template uses Exa for company and job discovery before Stagehand and Playwright automate downstream extraction and application steps. SE020
CE019 OpenRouter’s Exa customer story says OpenRouter has processed 73 million Exa-backed search queries to date. SE017
CE020 OpenRouter moved from plugin-based search toward server-side tool calls where models can decide when to search while Exa supplies highlighted, cited web context. SE017, SE027, SE031
CE021 monday.com says it uses Exa Search to convert natural-language prospect descriptions into structured CRM enrichment and downstream agent actions. SE018
CE022 Cognition publicly states that Exa powers all parts of Devin’s web search capability. SE019
CE023 Exa describes itself as a custom search engine built for AIs rather than a general-purpose wrapper around legacy search APIs. SE001, SE003
CE024 The public search materials claim category-specific indexes of 1B+ people, 100M+ research papers, and 50M+ companies. SE003, SE008
CE025 Company Search and Websets position Exa for GTM intelligence, investment research, and market-mapping workflows in addition to chat-time grounding. SE007, SE008
CE026 Exa’s differentiation story depends on combining generic retrieval with structured vertical indexes and orchestration products rather than selling raw search alone. SE001, SE005, SE007, SE008
CE027 Exa disclosed an internal training and indexing stack with 80 A100 GPUs plus 144 H200 GPUs, Kubernetes orchestration, NVIDIA operators, and Alluxio-backed storage caching. SE032
CE028 Exa says its custom vector database is designed to search billions of vectors in under 100 milliseconds while supporting filters and more than 500 QPS. SE033
CE029 The vector-database writeup says Exa uses Matryoshka-trained embeddings, truncation to 256 dimensions, binary quantization, clustering, reranking, and inverted indexes, and claims roughly 10x lower cost than cloud vector-database quotes. SE033
CE030 Exa Deep is described as a revamped agentic search endpoint that uses optimized query expansion, multiple parallel search agents, and field-level grounded structured outputs. SE035
CE031 Exa Agent is described as a subagent-based research API that splits large tasks into subtasks and uses model fusion plus token-efficient highlights with reported token reductions of up to 94 percent. SE034
CE032 Exa’s security documentation says the company is SOC 2 Type II certified and offers enterprise zero data retention and HIPAA solutions through its trust and sales motion. SE010, SE011
CE033 HIPAA mode is enterprise-gated and limited to eligible /search and /contents requests, while live retrieval, summaries, and deep paths are rejected for compliant workflows. SE010, SE011
CE034 Exa’s privacy policy says query data is not intended for personal information but may still be used to improve products and train or fine-tune models unless a customer agreement governs business-offering data. SE012
CE035 Exa’s public status page showed 100 percent uptime and operational status for Websets and Exa MCP at the time of review. SE014
CE036 As of late June 2026, the public GitHub API showed exa-mcp-server with 4,607 stars and 349 forks, versus 217 stars for exa-py and 131 stars for exa-js. SE021, SE022, SE023
CE037 Public package distribution includes the official exa-py SDK and a separate langchain-exa package, with langchain-exa version 1.1.0 uploaded on 2026-03-26. SE024, SE025
CE038 The npm package page shows exa-mcp-server version 3.2.1 with 47 published versions, indicating active MCP packaging iteration. SE026
CE039 LangChain’s Exa case study describes a planner-task-observer research architecture that returns structured JSON and runs in roughly 15 seconds to 3 minutes depending on complexity. SE028
CE040 Anthropic’s MCP announcement and the MCP introduction frame Exa’s MCP strategy as participation in an open cross-client standard instead of a proprietary assistant-specific integration. SE029, SE030
CE041 Exa Deep was launched publicly on 2026-03-04 as a faster, cheaper, structured-output search product oriented to agentic research use cases. SE035
CE042 Exa Agent was launched publicly on 2026-06-16 as a single API for deep research, list-building, and entity-enrichment workloads. SE034
CE043 The reviewed public materials expose changelog and release artifacts but do not publish a detailed forward roadmap, explicit public SLA, or self-serve deployment topology matrix. SE013, SE014, SE010
CU001 Exa's public customer base spans developer-native agent tools, CRM and GTM platforms, research products, and enterprise-agent software. SU001, SU016
CU002 monday.com uses Exa Search to power lead-generation and enrichment workflows inside monday CRM. SU005
CU003 HubSpot uses Exa to connect Breeze Assistant and agents to live people and company intelligence across CRM workflows. SU002
CU004 OpenRouter uses Exa to provide web search infrastructure across a marketplace of hundreds of models. SU004
CU005 Cognition says Exa powers all parts of Devin. SU003
CU006 StackAI uses Exa inside enterprise AI-agent workflows including due diligence, competitive intelligence, RFP response, and market research. SU010
CU007 WhyHow uses Exa to score millions of newly published pages daily for litigation-intelligence signals. SU009
CU008 Anara uses Exa for paper discovery, citation support, and research agents for scientists, students, and research teams. SU011
CU009 11x and Obvious show that Exa can be paid for out of sales, RevOps, or business-intelligence budgets rather than only engineering budgets. SU006, SU008
CU010 In July 2024 Exa said thousands of companies and developers had already integrated the product. SU015
CU011 By May 2026 Exa said the number of companies using the platform had grown to over 5,000. SU012, SU014
CU012 Exa says its product powers search for more than 400,000 developers. SU012, SU013
CU013 Public named-customer references became far more specific by 2026, explicitly naming Cursor, Cognition, HubSpot, OpenRouter, and monday.com. SU012, SU014
CU014 OpenRouter's Exa-backed search volume grew from 2.36M+ queries last year to 73M cumulative queries to date. SU004
CU015 11x says Exa has powered millions of searches, Webset items, and enrichment cells in production. SU006
CU016 CodeRabbit says switching to Exa reduced web-search volume by about 70-75% while preserving or improving review quality. SU007
CU017 CodeRabbit's public sample showed web search ran for only 7-10% of PRs on a given day, implying Exa is used for targeted external verification rather than every task. SU007
CU018 Cursor is publicly named by Exa, Lightspeed, and PYMNTS as a customer or category leader using Exa. SU012, SU013, SU014
CU019 No standalone Cursor-specific Exa case study or quantified Exa workflow outcome was found in the reviewed source set. SU001, SU012, SU013, SU014
CU020 HubSpot says Exa was better than native model search and other alternatives on speed, cost, and enrichment coverage. SU002, SU001
CU021 HubSpot uses Exa's Monitors API to track companies and trigger workflows when public-web signals change. SU002, SU024
CU022 monday.com uses Exa to automate prospecting, enrichment, qualification, routing, and internal recruiting workflows. SU005
CU023 OpenRouter's workflow shows Exa can be embedded as a reusable, server-side search tool across supported models. SU004
CU024 OpenRouter's architecture also makes Exa valuable but potentially replaceable because the same tool definition can travel across different models and engine options. SU004
CU025 Obvious says Exa found companies missing from other tools, replaced 15-20 manual research iterations with one enriched call, and cut research cost from hundreds of thousands to a few dollars. SU008
CU026 StackAI positions Exa inside governed enterprise agent deployments, broadening proof beyond startup self-serve usage. SU010
CU027 Exa's named-customer set spans GTM agents, coding agents, research assistants, legal intelligence, and enterprise workflow agents. SU001, SU006, SU007, SU009, SU010, SU011
CU028 LangChain maintains a dedicated langchain-exa package and Exa tool wrappers, lowering developer adoption friction. SU017
CU029 Exa's public surfaces support expansion from Search into Contents, Monitors, Agent, and enterprise contract features. SU016, SU022, SU023, SU024, SU025
CU030 Exa's billing docs confirm a path from self-serve pricing to negotiated enterprise contracts. SU022, SU023
CU031 No public NRR, GRR, churn, renewal-rate, or contract-length disclosure was found in the reviewed materials. SU016, SU018, SU019, SU020
CU032 Customer durability must therefore be inferred from workflow embed and product attach rather than measured retention statistics. SU016, SU022, SU023, SU025
CU033 Humai concludes Exa is strongest on semantic and research queries but more complex and expensive than simpler tools for basic applications. SU018
CU034 MakerStack says usage-based add-ons can muddy bills and that lower-cost providers may fit simple, high-volume search better. SU019
CU035 Both Humai and MakerStack describe multi-provider routing patterns, implying multi-homing is common in AI search stacks. SU018, SU019
CU036 ChatForest argues that Exa's category could compress if native search inside model providers becomes good enough. SU020
CU037 ChatForest describes Exa's named customers as production integrations rather than experimental accounts. SU020
CU038 Public named references skew heavily toward AI-native software and agent builders, suggesting sector concentration may be higher than the 5,000-company headline implies. SU001, SU012, SU013
CU039 Public sources do not disclose top-customer revenue share, segment mix, or direct-versus-partner sourced bookings. SU016, SU020
CU040 Exa's overall customer evidence is strong on named production references and workflow breadth but weak on retention, concentration, and Cursor-specific depth. SU001, SU012, SU019, SU020
CU041 HubSpot serves more than 299,000 customers across 135 countries and monday.com serves more than 250,000 customers worldwide, implying Exa is embedded in globally distributed software products. SU002, SU005
CU042 Lightspeed says Exa now helps power category leaders like Cursor and Cognition and that companies setting the pace in AI are reaching for Exa first. SU013
CU043 Business Wire's 2024 press release cited Databricks as an early AI research-team user, showing adoption before the current named-customer roster. SU015
CU044 Anara says showing the expected papers in the right workflows increases user trust, supporting value beyond simple search recall. SU011
CU045 Exa does not publish a clean customer-geometry split by geography, size, or vertical, so those cuts must be inferred from individual customer footprints. SU001, SU002, SU005
CU046 Proof quality differs materially by logo: HubSpot and OpenRouter have workflow-specific evidence, while Cursor is still mostly corroborated by named mentions rather than a dedicated case study. SU001, SU002, SU004, SU012, SU013, SU014
CR001 Exa’s privacy policy applies to the website, search engine, API, and websets, but says customer data processed on behalf of business customers is governed by separate customer agreements. SR001
CR002 Exa’s privacy policy says query data is used to improve products and technology, including training and fine-tuning models that power the services. SR001
CR003 The privacy policy says users should not submit personal information in query fields and that Exa does not actively monitor query data for personal information. SR001
CR004 Exa’s public terms grant the company a broad, perpetual, and irrevocable license to use, store, publish, distribute, and modify user input and output to provide and improve services and comply with legal obligations. SR002
CR005 The public terms require customers to have all rights necessary for submitted input and prohibit directing the service to generate output that violates intellectual-property rights, contractual restrictions, or law. SR002
CR006 The public terms reserve Exa’s right to audit API usage for compliance and to terminate API use rights at any time. SR002
CR007 Exa’s security page says the company is SOC 2 Type II certified and points enterprise buyers to the trust center for SOC 2 reports, DPA materials, and security documentation. SR003, SR005
CR008 Exa’s published security docs make Zero Data Retention and HIPAA compliance enterprise-plan features rather than default self-serve behavior. SR003, SR004
CR009 HIPAA mode is only available for eligible enterprise teams and only on the /search and /contents endpoints. SR004
CR010 HIPAA mode includes Zero Data Retention but fails closed when a request would require live retrieval, summaries, or deep/auto search paths. SR004
CR011 Exa exposes a public status page that separately reports operational state for Websets and Exa MCP. SR006
CR012 Exa’s public incident history shows repeated automated “Default is down” and recovery events on June 14, 2026. SR007
CR013 Exa’s billing docs say API requests are blocked when prepaid balances run out unless customers top up or enable auto recharge. SR008
CR014 Exa documents default self-serve rate limits of 10 QPS for /search and 100 QPS for /contents, with higher needs routed to enterprise sales. SR009
CR015 Exa disclosed a $5 million H200 cluster and a combined fleet of 224 GPUs with 350 TB of local NVMe running nearly 24/7. SR010, SR027
CR016 Exa said late-2024 inference requirements were roughly 100x larger than before and the ML team was 5x larger, evidencing scaling pressure. SR010
CR017 Exa’s vector-database post says the system searches billions of vectors under 100 milliseconds and supports more than 500 QPS at reasonable cost. SR011
CR018 Exa said its proprietary vector database cut cost roughly 10x versus quoted cloud vector database services. SR011
CR019 Exa’s hosted MCP endpoint is distributed across Claude, Cursor, VS Code, Codex, Gemini CLI, and other assistant clients. SR012
CR020 LangChain’s integration page exposes Exa’s live crawling, summaries, and tool use through a major agent-development ecosystem. SR013
CR021 Browserbase’s template uses Exa to discover companies and job postings inside an automated agentic workflow. SR014
CR022 OpenRouter says Exa powered its initial web-search launch and has served 73 million cumulative OpenRouter search queries to date. SR015
CR023 Cognition publicly says Exa powers all parts of Devin’s web-search capability. SR016
CR024 monday.com says its lead agents depend on Exa Search to turn natural-language prospect criteria into structured CRM results. SR017
CR025 Exa’s Series C announcement and a16z’s investment note say the company serves more than 5,000 companies and more than 400,000 developers. SR018, SR026
CR026 Exa says Series C proceeds will fund next-generation model training and systems scaled for hundreds of thousands of searches per second. SR018, SR023
CR027 TechCrunch reported that Exa hosted its product on AWS even while operating its own GPU cluster. SR022
CR028 TechCrunch identified Databricks as a named customer and noted that Exa combined a free tier with paid tiers. SR022
CR029 Latham’s Series B note says Exa’s financing involved dedicated data privacy, cybersecurity, and copyright counsel. SR025
CR030 Latham’s Series A and Series B financing notes show Exa repeatedly using outside counsel as the capital stack and legal surface expanded. SR024, SR025
CR031 a16z’s investment note frames Exa’s core search challenge as a three-way trade-off among freshness, latency, and cost. SR026
CR032 a16z quotes a customer saying search would happen 100% of the time if not constrained by GPUs, latency, and cost. SR026
CR033 Sacra estimates Exa generated about $10 million of revenue in 2025 after raising roughly $107 million through Series B. SR027
CR034 Sacra describes Exa’s cost structure as including substantial GPU infrastructure expense plus crawling and data costs. SR027
CR035 Sacra lists OpenAI, Anthropic, Google, Brave, Tavily, Jina, and Perplexity as direct competitive alternatives around search or adjacent AI infrastructure. SR027
CR036 The EU AI Act creates a harmonized AI framework while explicitly preserving existing EU personal-data obligations such as GDPR. SR028
CR037 The U.S. Copyright Office’s 2025 Part 3 report says dozens of lawsuits are pending over the use of copyrighted works in generative-AI development. SR029
CR038 Exa’s privacy policy says the company receives information from data partners and publicly accessible sources to provide and improve products, generate query responses, and train models. SR001
CR039 The privacy policy says Exa discloses information to vendors supporting cloud storage, security, analytics, payments, and other services. SR001
CR040 The public Y Combinator profile visibly centers on the two founders, leaving broader management depth opaque in the reviewed public set. SR021
CR041 PYMNTS framed Exa’s May 2026 fundraise against major AI-search moves by Google, underscoring incumbent platform pressure. SR023
CR042 The status page architecture separates Websets and Exa MCP from the default service, implying multiple production surfaces that can fail independently. SR006
CR043 Because HIPAA mode only supports cached, non-generative retrieval and fails closed on live/deep paths, Exa’s most differentiated retrieval features are harder to deploy unchanged in regulated workflows. SR004
CR056 Exa’s freshness docs say cached content is the default and `maxAgeHours` controls when the system livecrawls fresh pages, making freshness a configurable cost/latency trade-off rather than a universal default. SR031
CR044 Exa’s terms prohibit using the service to develop a competitive product or service. SR002
CR045 Exa’s privacy policy says professional advisors such as auditors, law firms, and accounting firms may receive information for legal and regulatory compliance purposes. SR001
CR046 TechCrunch said Exa’s early paid use cases included helping companies find training data, making data provenance and licensing more sensitive than in simple link retrieval alone. SR022
CR047 Because Exa’s privacy policy permits training on query data while ZDR and HIPAA are sold as enterprise add-ons, default privacy posture is materially weaker than the hardened posture implied by the trust narrative. SR001, SR003, SR004
CR048 Because public terms grant broad use rights over user input/output and shift lawful-use obligations to customers, privacy and IP uncertainty is being managed contractually as much as technically. SR001, SR002, SR029
CR049 Because the public customer set is dominated by OpenRouter, Cognition, monday.com, and agent-tooling ecosystems, the visible proof base is skewed toward frontier-AI workflows rather than a clearly diversified enterprise cohort. SR015, SR016, SR017, SR026
CR050 Because Exa runs its own GPU fleet, depends on AWS-hosted product components, and plans to scale much further, margin and uptime depend on continued capital access and disciplined capacity planning. SR010, SR018, SR022, SR027
CR051 Because the EU AI Act preserves data-protection obligations and the U.S. Copyright Office describes active litigation over AI training, Exa operates inside an evolving rather than settled compliance perimeter. SR028, SR029
CR052 Because public evidence shows outages, hard rate limits, and request blocking when credits run out, customers treating Exa as infrastructure must design around both vendor-level and account-level failure modes. SR007, SR008, SR009
CR053 Because public sources do not disclose top-customer share, retention, or contract duration, customer durability cannot be underwritten from the current public record. SR018, SR027
CR054 Because public leadership visibility remains founder-centric, key-person risk stays elevated until bench depth and succession are demonstrated in diligence. SR021, SR022
CR055 Because Latham’s Series B note names privacy, cybersecurity, and copyright counsel, the company itself appears to recognize that its legal surface now extends well beyond standard startup financing work. SR025
CV001 Exa announced a $250 million Series C in May 2026 at a $2.2 billion valuation. SV001, SV012, SV019, SV020
CV002 Exa said the Series C will fund new models, stronger GTM, and infrastructure scaled for hundreds of thousands of searches per second. SV001, SV018, SV020
CV003 Exa said it serves more than 5,000 companies and about 400,000 developers. SV001, SV020
CV004 Exa's July 2024 financing totaled $22 million, including a $17 million Series A led by Lightspeed with NVentures and Y Combinator participating. SV002, SV013, SV017
CV005 Sacra estimated that Exa reached about $10 million of revenue by September 2025 and was valued at $700 million in its Series B. SV015
CV006 Exa's public monetization is usage-based across search, contents, answer, and agent workloads rather than seat-based subscriptions. SV003
CV007 OpenRouter said its Exa-backed path has handled roughly 73 million cumulative search queries. SV007
CV008 monday.com publicly presents Exa as a web-search layer inside enterprise AI agents. SV008
CV009 Cognition publicly presents Exa as part of Devin's live-search stack and says its internal evaluation outperformed alternatives. SV009
CV010 Silicon Valley Investclub argues that Exa's revenue base is still small versus its private-market valuation and that the category can commoditize. SV016
CV011 Exa's privacy policy says prompts, queries, and outputs may be used to improve and fine-tune its models unless customers obtain stronger contractual controls. SV004
CV012 Blue-chip investor support is real: Andreessen, Benchmark, Lightspeed, and Y Combinator all show up in the recent financing path. SV001, SV010, SV011, SV019
CV013 Lightspeed said Exa revenue tripled in the months before the Series B and tied the round to a 5x cluster expansion. SV011
CV014 The reviewed public record does not disclose liquidation preferences, secondary liquidity, option-pool refreshes, debt, or other downside-protection terms tied to the latest round. SV001, SV012, SV015, SV019
CV015 Summing the publicly disclosed 2024, 2025, and 2026 rounds implies roughly $357 million of lifetime equity raised. SV001, SV011, SV015
CV016 At the $2.2 billion Series C mark, Exa trades at roughly 220x Sacra's $10 million 2025 revenue estimate. SV001, SV015
CV017 Elastic's market capitalization was about $6.15 billion in June 2026. SV022
CV018 Elastic reported fiscal 2026 revenue of about $1.739 billion in SEC filing-derived data. SV026, SV031
CV019 Elastic's market-cap-to-revenue ratio was about 3.5x on June 2026 market cap and fiscal 2026 revenue. SV022, SV026
CV020 Datadog's market capitalization was about $78.79 billion in June 2026. SV023
CV021 Datadog reported fiscal 2025 revenue of about $3.427 billion in SEC filing-derived data. SV028, SV032
CV022 Datadog's market-cap-to-revenue ratio was about 23x on June 2026 market cap and fiscal 2025 revenue. SV023, SV028
CV023 TechCrunch reported that Perplexity raised $200 million at a $20 billion valuation in September 2025 with ARR approaching $200 million. SV021, SV024
CV024 Sacra estimated that Perplexity reached roughly $500 million of annualized revenue in April 2026 while still carrying a $20 billion private valuation. SV024
CV025 TechCrunch reported that Glean raised $150 million at a $7.2 billion valuation in June 2025 after surpassing $100 million of ARR. SV029
CV026 TechCrunch reported that Glean reached $300 million of annualized revenue by May 2026 while the last disclosed valuation remained $7.2 billion. SV029, SV030
CV027 On public proxies, Exa's current valuation is richer than Perplexity (~100x), Glean (~72x), Datadog (~23x), and Elastic (~3.5x). SV001, SV015, SV021, SV024, SV029, SV022, SV023, SV026, SV028
CV028 Exa's strongest public proof comes from adoption breadth and customer embedment rather than audited financial disclosure. SV001, SV007, SV008, SV009, SV015
CV029 Public customer case studies show product embedment but still do not reveal concentration, retention, or realized contract economics. SV007, SV008, SV009
CV030 Exa's own infrastructure disclosures point to a GPU- and crawl-intensive cost base, making gross-margin quality a central underwriting issue. SV005, SV006, SV015
CV031 The public record still lacks filing-grade visibility on gross margin, burn, cash, NRR, discounting, and concentration. SV001, SV012, SV015, SV019
CV032 Price transparency is increasing across AI-search vendors, which raises the risk that search becomes a more swappable budget line and compresses multiples. SV003, SV021, SV024, SV029, SV030
CV033 The private comp set shows investors paying up for AI-search leaders, but those peers also disclose materially more revenue scale than Exa's public record does. SV021, SV024, SV029, SV030
CV034 For the current $2.2 billion price to look fair, Exa likely needs either much higher present revenue than public proxies show or a rapid path into nine-figure revenue with durable margins. SV001, SV015, SV022, SV023, SV024, SV026, SV028, SV029
CV035 The best bull case is that Exa becomes default retrieval infrastructure for agent builders and compounds into a substantially larger revenue base before public multiples compress. SV001, SV003, SV007, SV010, SV011, SV020
CV036 The base case is that Exa keeps growing but ultimately settles into strong infrastructure-software valuation bands that leave limited upside from the latest mark. SV022, SV023, SV026, SV028, SV029, SV030
CV037 The bear case is that search and retrieval get bundled or commoditized before Exa proves margin quality and retention, leading to a flat or down financing outcome. SV015, SV016, SV021, SV029, SV030
CV038 Entry discipline should focus on revenue quality, gross margin, retention, and cap-stack terms before anchoring on the headline Series C valuation. SV001, SV012, SV015, SV019
CV039 The current evidence supports tracking Exa rather than underwriting the latest price because operating proof still lags financing momentum. SV001, SV015, SV021, SV029, SV030
CV040 Thesis-break triggers include a slower-than-expected march to material revenue, failure to prove healthy unit economics, or a financing reset below the 2026 mark. SV001, SV015, SV021, SV029
CV041 SEC submissions and XBRL companyfacts make Elastic and Datadog cleaner external price-discovery anchors than private AI-search startups. SV025, SV026, SV027, SV028
CV042 The financing context is strong enough to keep Exa out of near-term distress, but not transparent enough to tell new investors where they would sit in the preference stack. SV010, SV011, SV012, SV014, SV019
来源
编号出版方标题引文
SO001 Exa Labs Exa homepage
SO002 Exa Labs Exa about page
SO003 Exa Labs Exa pricing page
SO004 Exa Labs Exa Privacy Policy Query Data is used to improve our products and technology, including by training and fine-tuning models that power our Services
SO005 Exa Labs Exa Announces Series A Funding for AI Search Technology Development
SO006 Exa Labs Exa Raises $85M to Build the Search Engine for AIs
SO007 Exa Labs Exa Raises $250M Series C to Build the Search Engine for AIs
SO008 Exa Labs Announcing Exa: The AI Search Engine with Semantic Search Technology
SO009 Exa Labs Introducing Exa Deep: An Agent for Every Search
SO010 Exa Labs Exa Search API documentation
SO011 Exa Labs Web Search MCP documentation
SO012 Exa Labs Exa careers page
SO013 Y Combinator Exa: Web search rebuilt for LLMs
SO014 TechCrunch Exa raises $17M from Lightspeed, Nvidia, Y Combinator to build a Google for AIs
SO015 Latham & Watkins Latham & Watkins Advises Exa in US$17 Million Series A Financing From Lightspeed and NVIDIA
SO016 Lightspeed Venture Partners Exa: Redesigning Search for AI
SO017 Latham & Watkins Latham & Watkins Advises Exa in US$85 Million Series B Financing With Benchmark and Lightspeed
SO018 Lightspeed Venture Partners Search, Perfected for AI: Why We're Doubling Down on Exa
SO019 The Silicon Valley Post Exa Raises $250M at a $2.2B Valuation
SO020 Andreessen Horowitz Investing in Exa
SO021 Techno Trenz Exa Secures $250M to Scale AI Search Platform
SO022 Sacra Exa revenue, valuation & funding Infrastructure costs: The company's GPU-intensive architecture imposes high fixed costs that require increasing usage volume to sustain.
SO023 Craft Exa Company Profile
SO024 Growjo Exa: Revenue, Competitors, Alternatives
SO025 Browserbase AI job application automation with Exa
SO026 LangChain Exa search integration (Python)
SO027 LangChain ExaSearchResults integration (JavaScript)
SO028 Silicon Valley Investclub Exa
SM001 Exa Labs Exa Web search, built for AI agents One API for search, crawling, and research agents
SM002 Exa Labs API Pricing | Exa Search $7 / 1k requests; Deep Search $12 / 1k; Deep-Reasoning Search $15 / 1k; Contents $1 / 1k pages.
SM003 Exa Docs Welcome to Exa - Exa Exa finds the exact content you’re looking for on the web, with four core functionalities: /search, /contents, /answer, /research.
SM005 GitHub GitHub - exa-labs/exa-mcp-server: Exa MCP for web search and web crawling! Connect AI assistants to Exa's search capabilities: web search, code search, and company research.
SM006 Brave Brave Search API | Brave The Brave Search API can replace several competing options that may have smaller indexes (Tavily or Exa).
SM007 Brave Brave Search - API Web Search provides access to our comprehensive index of web pages.
SM008 GitHub GitHub - brave/brave-search-skills: Official skills for using Brave Search API with AI coding agents. Official skills for using Brave Search API with AI coding agents.
SM009 Tavily Tavily 300M+ monthly requests handled.
SM010 Tavily Tavily 1,000 API credits / month; $0.008 / credit.
SM011 Tavily Docs Tavily Search - Tavily Docs Execute a search query using Tavily Search.
SM012 Google for Developers Custom Search JSON API  |  Google for Developers This API is not available for new customers.
SM014 LangChain Tavily integrations - Docs by LangChain The langchain-tavily package exposes Tavily’s Search, Extract, Crawl, and Map endpoints as LangChain tools.
SM015 OpenAI New tools for building agents Built-in tools including web search, file search, and computer use.
SM017 Model Context Protocol What is the Model Context Protocol (MCP)? - Model Context Protocol MCP is an open-source standard for connecting AI applications to external systems.
SM018 Elastic Elasticsearch: The Official Distributed Search & Analytics Engine | Elastic Elasticsearch is an open source, distributed search and analytics engine built for speed, scale, and AI applications.
SM020 Grand View Research Enterprise Search Market Size, Share & Growth Report, 2030 The global enterprise search market size was estimated at USD 4,867.2 million in 2023 and is projected to reach USD 8,851.7 million by 2030.
SM022 Exa Docs Search - Exa The search endpoint lets you search the web and extract contents from the results.
SM024 Tavily Docs Create Research Task - Tavily Docs Tavily Research performs comprehensive research on a given topic by conducting multiple searches, analyzing sources, and generating a detailed research report.
SM025 Brave Guides | Brave This guide covers the steps required to enable Brave Search as a tool to be used in the Claude desktop app using the Model Context Protocol (MCP).
SM027 Glean Glean – Work AI that Works | Agents, Assistant & Search Glean connects knowledge, systems, and context so AI can actually work.
SM029 Algolia Pricing NeuralSearch combines semantic search with keyword search.
SM031 Tavily {"message":"Alive"} {"message":"Alive"}
SM032 Exa Docs Answer - Exa The Answer endpoint returns an answer with citations.
SM033 Brave Brave launches most powerful search API for AI to date | Brave Brave's high-quality grounding data allows cheaper open-weight LLMs to beat ChatGPT, Google AI Mode, and Perplexity.
SM035 Anthropic Introducing the Model Context Protocol MCP provides a universal, open standard for connecting AI systems with data sources.
SM038 Exa Labs Exa Raises $250M Series C to Build the Search Engine for AIs Exa already powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com and over 400,000 developers.
SP001 Exa Exa
SP002 Exa API Pricing | Exa
SP003 Exa Search - Exa
SP004 Exa Company Search
SP005 Exa Billing - Exa
SP006 Exa Rate Limits - Exa
SP007 Exa Enterprise Documentation & Security
SP008 Exa HIPAA - Exa
SP009 Exa Announcing Exa
SP010 Exa Labs Exa MCP Server
SP011 Exa Labs Exa Python SDK
SP012 Tavily Tavily
SP013 Tavily Tavily pricing
SP014 Tavily Research endpoint
SP015 Brave Brave Search API | Brave The Brave Search API can replace several competing options that may have smaller indexes (Tavily or Exa), higher latencies (SerpAPI or Serper), or more limited access (Google and Bing).
SP016 SerpApi SerpApi: Google Search API
SP017 SerpApi Security
SP018 Perplexity Pricing - Perplexity
SP019 Perplexity Overview - Perplexity
SP020 TechCrunch Perplexity launches Sonar, an API for AI search
SP021 Google for Developers Custom Search JSON API
SP022 Google Cloud Grounding overview
SP023 Algolia Pricing
SP024 Elastic Elasticsearch
SP025 Elastic Elastic Cloud Hosted pricing
SP026 Glean Glean – Work AI that Works | Agents, Assistant & Search
SP027 LangChain LangChain overview
SI001 Exa Exa We power search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com and over 400,000 developers.
SI002 Exa API Pricing | Exa
SI003 Exa Billing - Exa
SI004 Exa Rate Limits - Exa
SI005 Exa Overview - Exa
SI006 Exa Monitors - Exa
SI007 Exa Exa Search API - Exa
SI008 Exa Contents API - Exa
SI009 Exa Websets - Exa
SI010 Exa Snowflake - Exa
SI011 Exa Exa Raises $250M Series C to Build the Search Engine for AIs Exa just raised $250M at a $2.2B valuation led by a16z to power all agents with the highest quality web search.
SI012 Exa Exa Announces Series A Funding for AI Search Technology Development
SI013 Exa The Exacluster: Powering Our Neural Network Search Engine Exa recently purchased a 5 million dollar GPU cluster to train retrieval models over the web.
SI014 Exa How We Built a Web-Scale Vector Database for Our Neural Network Search Engine
SI015 Exa Introducing Exa Agent
SI016 Exa How OpenRouter gives 400+ models agentic web search with Exa
SI017 Exa monday.com x Exa Case Study
SI018 Exa Cognition x Exa Case Study
SI019 Business Wire Exa Raises $22MM to Build the Search Engine for AI
SI020 Latham & Watkins Latham & Watkins Advises Exa in US$250 Million Series C at US$2.2 Billion Valuation
SI021 Sacra Exa revenue, valuation & funding Infrastructure costs: The company's GPU-intensive architecture imposes high fixed costs that require increasing usage volume to sustain.
SI022 PYMNTS Exa Raises $250 Million for AI-Powered Search Infrastructure | PYMNTS.com
SI023 Y Combinator Exa: Web search rebuilt for LLMs | Y Combinator
SI024 LangChain Exa integrations - Docs by LangChain
SI025 U.S. Securities and Exchange Commission SEC.gov | Search Filings
SI026 Delaware Division of Corporations Division of Corporations - Filing
SI027 AIbase Over 400 AI Models Unleash Web Search! Exa Joins Forces with OpenRouter to Ignite RAG Revolution
SE001 Exa Exa | Web Search API, AI Search Engine, & Website Crawler One API for search, crawling, and research agents.
SE002 Exa Welcome to Exa
SE003 Exa Exa Search API
SE004 Exa Contents API
SE005 Exa Overview - Agent API
SE006 Exa Monitors
SE007 Exa Websets
SE008 Exa Company Search
SE009 Exa Exa MCP
SE010 Exa Enterprise Documentation & Security
SE011 Exa HIPAA
SE012 Exa Privacy Policy
SE013 Exa Exa docs index (llms.txt)
SE014 Exa Exa Status
SE015 Exa Exa Integrations — Find the Right Tool for Your Stack
SE016 Exa Snowflake - Exa
SE017 Exa OpenRouter customer story
SE018 Exa monday.com customer story
SE019 Exa Cognition customer story
SE020 Browserbase Exa + Browserbase template
SE021 GitHub GitHub API - exa-mcp-server
SE022 GitHub GitHub API - exa-py
SE023 GitHub GitHub API - exa-js
SE024 PyPI exa-py · PyPI
SE025 PyPI langchain-exa · PyPI
SE026 npm exa-mcp-server · npm
SE027 OpenRouter Web search plugin docs
SE028 LangChain How Exa built a Web Research Multi-Agent System with LangGraph
SE029 Anthropic Model Context Protocol
SE030 Model Context Protocol Introduction
SE031 AIbase Exa and OpenRouter web search article
SE032 Exa We Needed a Bigger Hammer
SE033 Exa Building a Web-Scale Vector Database
SE034 Exa Exa Agent launch post
SE035 Exa Introducing Exa Deep: An Agent for Every Search
SU001 Exa Customer Stories | How top teams amplify their products with web search | Exa Exa is necessary for agents on HubSpot to supplement internal data with high quality, real-time people and company search.
SU002 Exa HubSpot × Exa Case Study We found that other alternatives were slower, more expensive, and did not have comprehensive enrichment coverage.
SU003 Exa Cognition x Exa Case Study Exa powers all parts of Devin.
SU004 Exa How OpenRouter gives 400+ models agentic web search with Exa Last year, OpenRouter had powered 2.36M+ search queries with Exa. Today, that number is 73M search queries to date.
SU005 Exa monday.com x Exa Case Study monday.com is a global software company... for more than 250,000 customers worldwide.
SU006 Exa How 11x finds novel GTM signals with Exa Across 11x's production usage to date, Exa has powered millions of searches, Webset items, and enrichment cells.
SU007 Exa CodeRabbit × Exa Case Study Switching to Exa Search cut CodeRabbit's web-search volume by about 70-75% while preserving, and sometimes improving, review quality.
SU008 Exa Obvious × Exa Case Study One enriched call replaces 15-20+ manual iterations with comparable quality.
SU009 Exa Legal Tech Case Study WhyHow built an AI-powered litigation intelligence engine that scores millions of newly published pages daily.
SU010 Exa StackAI × Exa Case Study StackAI leverages Exa's Search, Answer, and Websets APIs to power enterprise AI agents with real-time web intelligence.
SU011 Exa Anara × Exa Case Study Scientists trust our product further when the relevant papers they expect to see are available to them in the right workflows.
SU012 Exa AI agents will search the web more than humans this year Exa already powers search for Cursor, Cognition, HubSpot, OpenRouter, Monday.com and over 400,000 developers.
SU013 Lightspeed Venture Partners Tripling Down on Exa to build the Search Engine for AI Cursor, Cognition, HubSpot, Gamma, OpenRouter, and over 400,000 developers — the companies setting the pace in AI are all reaching for Exa first.
SU014 PYMNTS Exa Raises $250 Million for AI-Powered Search Infrastructure Since launching its AI-focused API in early 2023, Exa's customer base has grown to more than 5,000 companies, including Cursor, Cognition, HubSpot, OpenRouter, and Monday.com.
SU015 Business Wire Exa Raises $22MM to Build the Search Engine for AI So far, thousands of companies and developers have integrated Exa... to AI research teams at companies like Databricks.
SU016 Sacra Exa Companies often begin with basic search integration and later adopt advanced features such as multi-agent research, real-time content processing, and enterprise-grade privacy controls.
SU017 LangChain Exa for LangChain
SU018 Humai AI Search APIs Compared: Tavily vs Exa vs Perplexity Exa is the most technically impressive API I've tested... But it's also more complex and expensive, which makes it overkill for simple applications.
SU019 MakerStack Exa Review (2026) - MakerStack Usage-based add-ons can muddy your bill.
SU020 ChatForest Every AI Agent Needs to Search the Web. Exa Just Raised $250 Million to Be the One They Call. These are not experimental accounts — these are production integrations inside tools that millions of people use daily.
SU021 Cursor Cursor · Customers By February 2025, every Coinbase engineer had utilized Cursor.
SU022 Exa Pricing
SU023 Exa Billing API guide Requests are billed according to the rates on exa.ai/pricing, or per your enterprise contract if you have one.
SU024 Exa Monitors API guide
SU025 Exa Introducing Exa Agent: frontier web research at a fraction of the cost
SR001 Exa Labs Privacy Policy - Exa Query Data is used to improve our products and technology, including by training and fine-tuning models that power our Services.
SR002 Exa Labs Exa Labs Terms of Service You grant us a nonexclusive, royalty-free, transferable, sub-licensable, worldwide, perpetual and irrevocable license to access, use, host, cache, store, reproduce, transmit, display, publish, distribute, and modify any User Input and Output.
SR003 Exa Labs Enterprise Documentation & Security
SR004 Exa Labs HIPAA HIPAA mode includes Zero Data Retention behavior for those requests: Exa does not persist PHI, and the request follows a compliant processor path that only uses approved subprocessors.
SR005 Vanta / Exa Labs exa.ai Trust Center
SR006 Exa Labs Exa Status
SR007 Exa Labs Exa Status History - June 2026 Default is down at the moment. This incident was automatically created by Instatus monitoring.
SR008 Exa Labs Billing overview
SR009 Exa Labs Rate Limits
SR010 Exa Labs We Needed a Bigger Hammer Exa recently purchased a 5 million dollar GPU cluster to train retrieval models over the web.
SR011 Exa Labs Building web-scale vector DB We cut costs 10x compared to quotes from cloud vector database services.
SR012 exa-labs exa-mcp-server
SR013 LangChain Exa integration
SR014 Browserbase Exa Browserbase template
SR015 Exa Labs How OpenRouter uses Exa to power search Last year, OpenRouter had powered 2.36M+ search queries with Exa. Today, that number is 73M search queries to date.
SR016 Exa Labs How Cognition uses Exa to power web search capability across their agentic products Exa powers all parts of Devin.
SR017 Exa Labs How monday.com uses Exa Search for lead agents
SR018 Exa Labs Announcing Series C
SR019 Exa Labs Series A
SR020 Exa Labs Documentation Index (llms.txt)
SR021 Y Combinator Exa | Y Combinator
SR022 TechCrunch Exa raises $17M from Lightspeed, Nvidia, Y Combinator to build search for AI
SR023 PYMNTS Exa Raises $250 Million for AI-Powered Search Infrastructure
SR024 Latham & Watkins Latham & Watkins advises Exa in US$17 million Series A financing
SR025 Latham & Watkins Latham & Watkins advises Exa in US$85 million Series B financing
SR026 Andreessen Horowitz Investing in Exa If the underlying data is stale, incomplete, or incorrect, everything downstream breaks.
SR027 Sacra Exa
SR028 European Union Regulation (EU) 2024/1689 (Artificial Intelligence Act)
SR029 U.S. Copyright Office Copyright and Artificial Intelligence Part 3: Generative AI Training
SR030 Delaware Division of Corporations Entity Search
SR031 Exa Docs Content Freshness - Exa
SV001 Exa Labs Exa Raises $250M Series C to Build the Search Engine for AIs
SV002 Exa Labs Exa Announces Series A Funding for AI Search Technology Development
SV003 Exa Labs API Pricing | Exa
SV004 Exa Labs Exa Privacy Policy
SV005 Exa Labs The Exacluster: Powering Our Neural Network Search Engine
SV006 Exa Labs How We Built a Web-Scale Vector Database for Our Neural Network Search Engine
SV007 Exa Labs How OpenRouter gives 400+ models agentic web search with Exa
SV008 Exa Labs monday.com x Exa Case Study
SV009 Exa Labs Cognition x Exa Case Study
SV010 Andreessen Horowitz Investing in Exa
SV011 Lightspeed Venture Partners Search, Perfected for AI: Why We're Doubling Down on Exa
SV012 Latham & Watkins Latham & Watkins Advises Exa in US$250 Million Series C at US$2.2 Billion Valuation
SV013 Business Wire Exa Raises $22MM to Build the Search Engine for AI
SV014 Y Combinator Exa: Web search rebuilt for LLMs | Y Combinator
SV015 Sacra Exa revenue, valuation & funding
SV016 Silicon Valley Investclub Exa | Silicon Valley Investclub
SV017 TechCrunch Exa raises $17M from Lightspeed, Nvidia, Y Combinator to build a Google for AIs
SV018 PYMNTS Exa Raises $250 Million for AI-Powered Search Infrastructure
SV019 Gunderson Dettmer Andreessen Horowitz Leads Exa’s $250 Million Series C at $2.2 Billion Valuation
SV020 SiliconANGLE Exa Labs raises $250M at $2.2B valuation for its AI search tools
SV021 TechCrunch Perplexity reportedly raised $200M at $20B valuation
SV022 CompaniesMarketCap Elastic (ESTC) market capitalization
SV023 CompaniesMarketCap Datadog (DDOG) market capitalization
SV024 Sacra Perplexity revenue, valuation & funding
SV025 U.S. Securities and Exchange Commission Elastic submissions metadata (CIK 0001707753)
SV026 U.S. Securities and Exchange Commission Elastic XBRL company facts (CIK 0001707753)
SV027 U.S. Securities and Exchange Commission Datadog submissions metadata (CIK 0001561550)
SV028 U.S. Securities and Exchange Commission Datadog XBRL company facts (CIK 0001561550)
SV029 TechCrunch Enterprise AI startup Glean lands a $7.2B valuation
SV030 TechCrunch Glean's top line crosses $300M as AI budget cutting becomes its major selling point
SV031 U.S. Securities and Exchange Commission Elastic Form 10-K for fiscal year ended April 30, 2026
SV032 U.S. Securities and Exchange Commission Datadog Form 10-K for fiscal year ended December 31, 2025