初创公司尽调
尽调报告 AI Software / Developer Productivity Late Private (post-Series A) 2026-05-09

Cognition AI

Cognition AI 定义了自主软件工程平台这个品类,ARR 9 个月从 $1M 冲到 $73M;但 $10.2B 估值约等于 140× ARR,已经把近乎完美的执行定价进去,而这个市场的基准能力正快速商品化,关键财务指标也未披露。

封面要素

最近一轮估值 01
$10.2B (September 2025) [CV004]
累计融资 02
~$1.575B (seed + Series A + $400M round) [CV010]
估计 ARR 03
~$73M (April 2025 est.; third-party, unconfirmed) [CV030]
ARR 增长 04
$1M → $73M in 9 months (Dec 2024 – Apr 2025) [CV001]
领投方 05
Founders Fund IX (CIK 0001971631) [CV007]
核心产品 06
Devin — autonomous AI software engineer (cloud + VPC deployment) [CO002]
Windsurf 收购 07
Acquired July 2025; 350+ enterprise accounts, 250K+ DAU [CV006]
已披露客户 08
Nubank (12× efficiency, 20× cost savings), Mercedes-Benz, Cognizant [CU003]
安全事件 09
Prompt-injection vulnerability disclosed Dec 2024; patched [CR006]
基准差距 10
Devin 13.86% SWE-bench Full (2024); Claude Code 72.5% Verified (2025) [CR018]

公司概况

Cognition AI 是一家总部位于 San Francisco 的 AI 初创公司,2023 年由 Scott Wu、Steven Hao 和 Walden Yan 创立;三位创始人都是竞技编程冠军,也曾获得奥林匹克竞赛奖牌。公司的旗舰产品 Devin 是云端托管的自主 AI 软件工程师,能够端到端规划、写代码、测试和部署,无需持续人工输入。用户可通过 API、Slack、Jira 或网页界面访问 Devin,定价采用基于用量的 Agent Compute Units(ACUs)模式。July 2025,Cognition 收购 Windsurf IDE,新增 250K+ 日活跃用户和 350+ 企业账户。公司已融资约 $1.575B,最近一轮为 September 2025 的 $400M,估值 $10.2B。

官网
cognition.ai
成立时间
2023-01-01
创始人
Scott Wu, Steven Hao, Walden Yan
创立地点
San Francisco, CA
总部
San Francisco, CA
产品
Devin 是云端托管的自主 AI 软件工程师,接收来自 Slack、Jira 或网页界面的任务后,会用 Agent Compute Units(ACUs)独立规划、编写代码、测试并部署。Devin 2.0(April 2025)新增直接合并 PR 的能力,每项任务的有效价格降低 3×。Windsurf IDE(July 2025 收购)为平台带来 250K+ 日活跃用户和 350+ 企业账户。
客户
企业工程团队(1,000+ 开发者)、中型科技公司,以及个人专业开发者。已披露企业客户包括 Nubank、Mercedes-Benz、Cognizant。美国政府垂直业务于 February 2026 推出。
商业模式
基于用量的 SaaS:Team 方案 $500/month,超额部分 $2.25/ACU;Enterprise 方案需直接谈判。Windsurf IDE 增加免费转付费漏斗。Cognizant 渠道合作伙伴关系增加经销收入层。
阶段
Late Private (post-Series A)
融资情况
$21M 种子轮(2023),$175M Series A、估值 $2B(Founders Fund IX,Jan/Apr 2025),$400M 轮、估值 $10.2B(Sep 2025)。已披露累计约 $1.575B。
[CO001, CO002, CO003, CI001, CI002]

执行摘要

主要优势

  • ARR 增长极快:9 个月从 $1M 到 $73M,是一代级增长冲刺,在企业 SaaS 史上几乎找不到直接参照
  • 品类先行者:多数竞争者还没看清品类时,Devin 已把“自主 AI 软件工程师”变成企业采购项
  • Windsurf IDE 整合:250K+ DAU 与 350+ 企业账户,拼出智能体 + IDE 的独特分发护城河,单一产品竞争者难以复制
  • 标杆客户验证:Nubank 带来 12× 效率、20× 成本节省,并有公开案例,是金融服务企业 ROI 的高质量参照
  • 创始团队底色强:Scott Wu、Steven Hao、Walden Yan 都是竞赛编程冠军,具备自主编程所需的 AI/ML 工程直觉
  • Devin 自用 Devin:内部每周合并 659 个 Devin PR,形成快速模型改进反馈环

主要风险

  • 基准能力商品化:Claude Code Opus 4 的 SWE-bench Verified 为 72.5%,而 Devin Full 为 13.86%;15 个月里竞争能力提升约 5×,到 2026 年底技术差异可能被磨平
  • 安全阴影未消:2024 年 12 月曝出提示词注入漏洞;Devin 2.0 后未公开渗透测试或漏洞赏金计划;自主 PR 合并能力扩大了攻击面
  • 估值倍数过高:$10.2B 约等于估计 ARR 的 140×,比 Cursor/Anysphere 高出 7×;这个入场价下,风险调整后收益在基准情景为负,在悲观情景会很糟
  • LLM 供应商利益冲突:Anthropic 和 OpenAI 既是主要 LLM API 供应商,又通过 Claude Code 与 OpenAI Codex 直接竞争;定价权和合同条款未披露
  • 客户集中:只有 Nubank 一个量化企业案例;$73M ARR 来自数量未披露的少数账户,单一大客户流失的敏感度很高
  • EU AI Act GPAI 合规:2026 年 8 月是期限;公司未披露 GPAI 注册、欧盟 DPO 任命或透明度文件;宣布扩张欧洲时,合规基础设施尚不清楚

未决问题

  • 分队列净留存率(NRR):未披露;这是验证 $10.2B 估值下 ARR 质量最关键的单项指标,必须达到 ≥110% 才能支撑溢价
  • 分产品线毛利率:未披露;LLM 推理成本结构未知,没有这项指标就无法建模盈利路径
  • 前五大客户 ARR 集中度:未披露;Nubank 是唯一量化参照;若单一客户占比超过 30%,投资逻辑会被打穿
  • Devin 2.0 后安全审计:2025 年 4 月扩展智能体能力后,没有公开渗透测试报告或红队披露;SOC 2 Type II(2024 年 3 月)早于当前攻击面
  • Windsurf 整合后的客户留存:2025 年 7 月获得 350+ 企业账户;6 个月留存率(2026 年 1 月)未知,这是验证 IDE 护城河的关键证据
  • LLM API 合同条款:与 Anthropic/OpenAI 的量价、智能体用例限制和控制权变更条款未披露;毛利底线和供应商利益冲突无法评估

目录

Chapter 01

01公司概况

1.1 身份与商业模式

Cognition AI 是一家总部位于 San Francisco、以 AI 为底层的软件公司,正在构建可执行端到端软件工程任务的自主智能体。旗舰产品 Devin 对外定位为全球首个 AI 软件工程师:它能理解工单、规划方案、编写和调试代码、运行测试,并在极少人工介入下部署到生产环境。公司面向个人开发者和小团队提供 SaaS 订阅(Core 方案 $20/month),Team 方案为 $500/month,包含 250 个 Agent Compute Units,并向企业提供带 VPC 部署选项的定制合同。July 2025 收购 AI 原生 IDE Windsurf 后,Cognition 扩充了产品组合,可以服务更偏好交互式、IDE 中心工作流的开发者,而不只覆盖完全自主的智能体。双产品策略让 Cognition 同时切入智能体自动化板块,以及由 GitHub Copilot 和 Cursor 主导的传统代码助手市场。收入主要按用量计费:超出方案包含额度的 Agent Compute Units 约按 $2.25/ACU 定价,形成随企业任务量扩张的计量模型。公司网站和博客是面向公众的渠道;公司不发布 GAAP 财务数据,只通过投资人沟通和公司主导的新闻稿披露信息。[CO001, CO002, CO003, CO028, CO029]

1.2 创始人与领导层

Cognition AI 由三位顶尖竞技程序员共同创立;每人都是 International Olympiad in Informatics 金牌得主,公司从一开始就拥有极深的技术底座。CEO Scott Wu 三次获得 IOI 金牌,曾是 2011 MATHCOUNTS 全国冠军,并拥有 Harvard 经济学学位。在 Codeforces 竞技编程平台上,Wu 的称号为最高等级 Legendary Grandmaster,体现了持续的算法解题能力。创办 Cognition 之前,他曾联合创立 AI 职业社交产品 Lunchclub。CTO Steven Hao 同样是 IOI 金牌得主,曾在 Scale AI、DeepMind、Waymo、Nuro、Modal 和 Cursor 任职,经历横跨前沿 AI 研究与生产工程系统。CPO Walden Yan 则带来产品领导力和 IOI 级算法深度。公司创立时约 10 人的团队合计持有 10 枚 IOI 金牌,这种竞技编程人才密度在早期创业公司中极为罕见。这个画像吸引了顶级投资人与人才,但也带来明显的关键人物依赖:创始人的技术信誉与公司的市场定位高度绑定。公司尚未公开披露董事会组成或独立董事信息。[CO004, CO005, CO006, CO007, CO026, CO027]

管理层与创始人表
姓名职务教育 / 竞赛背景过往经历关键人物依赖
Scott WuCEO 兼联合创始人3 次 IOI 金牌;2011 年 MATHCOUNTS 全国冠军;Harvard(经济学);Codeforces 传奇宗师Lunchclub 联合创始人关键 — 品牌和技术可信度与产品绑定
Steven HaoCTO 兼联合创始人IOI 金牌曾任职:Scale AI、DeepMind、Waymo、Nuro、Modal、Cursor关键 — 核心模型和系统架构
Walden YanCPO 兼联合创始人IOI 金牌AI 产品领导经验高 — 产品愿景和路线图

董事会构成、独立董事以及更广的管理层团队(VP 级及以下)未公开披露。公开确认的只有三位联合创始人。

[CO004, CO005, CO006, CO007, CO026, CO027]

1.3 融资历史与投资人

Cognition AI 的融资速度在近年风险投资史上名列前茅。March 2024 的首轮约 $21M 种子 / Series A 融资,在尚无实质收入之前就给出约 $350M 估值,反映投资人押注创始团队的执行能力。仅数周后的 April 2024,Cognition 以 $2B 估值完成 $175M 融资,成立六个月内进入独角兽行列。Founders Fund——Peter Thiel 的风险投资机构——领投两轮,不仅提供资本,也向市场释放声誉信号。到 September 2025,随着 Windsurf 收购完成、ARR 快速增长,Cognition 以 $10.2B 投后估值融资 $400M,累计融资约 $696M。September 2025 轮的共同投资人包括 Lux Capital、8VC、Elad Gil(知名天使投资人和运营者)、Bain Capital Ventures、D1 Capital、Definition Capital 和 Swish Ventures。截至 May 2026,据报道公司正洽谈新一轮融资,潜在估值 $25B+,说明投资人对智能体 AI 编码基础设施仍有强烈兴趣。值得注意的是,公司披露从成立到 Q3 2025 的净现金消耗低于 $20M,说明收购前的高速增长以极高资本效率实现。[CO008, CO009, CO010, CO020, CO030, CO031]

Cognition AI 快照 KPI 表
指标数值 / 状态截至日期置信度缺口 / 限制
估值(上一轮)$10.2BSep 2025投后;投前未披露
累计融资~$696MSep 2025已知披露轮次合计
ARR(Devin,收购前)$73MJun 2025公司通过博客披露
ARR(Windsurf 后合并)~$155MJul 2025部分为公司声称;Windsurf 部分为估计
初始融资(种子轮 / Series A)$21M,估值 $350MMar 2024首轮披露的机构融资
Series A 延伸轮$175M,估值 $2BApr 2024公司公告 + 媒体报道
净现金消耗(成立至 Q3 2025)<$20MSep 2025公司声称;未经审计
员工数(估计)~50-2502025收购前约 49 人;Windsurf 整合情况不清楚
企业客户数(Windsurf)350+Jul 2025收购时仅 Windsurf 口径
价格 / ARR 倍数(Sep 2025)~65-140xSep 2025取决于采用哪一套 ARR 基数

收入和估值数字来自公司披露或分析师估计;公开渠道没有经审计财务报表。ARR 数字代表订阅和用量收入的年化口径。倍数区间反映 Windsurf 收购前后两套 ARR 基数。

[CO008, CO009, CO010, CO014, CO015, CO016]
利益相关方 / 投资者图谱
投资方轮次角色控制权 / 经济重要性尽调问题
Founders Fund种子轮 / Series A + $400M领投方最大单一机构持有人;可能拥有董事会席位核实董事会席位、反稀释和按比例认购权
Lux Capital$400M (Sep 2025)跟投方聚焦深科技;可能为少数股权核实持股比例
8VC$400M (Sep 2025)跟投方聚焦科技的成长基金核实持股比例
D1 Capital$400M (Sep 2025)跟投方后期成长股权核实持股比例和治理权
Bain Capital Ventures$400M (Sep 2025)跟投方有企业 SaaS 经验;少数股权核实持股比例
Elad Gil多轮个人投资者知名天使投资人;可能有顾问关系核实是顾问角色还是正式董事会角色
Definition Capital$400M (Sep 2025)跟投方聚焦科技核实出资规模
Swish Ventures$400M (Sep 2025)跟投方早期 / 成长期科技核实角色

持股比例未公开披露。轮次参与信息来自新闻稿和媒体报道。董事会权利、反稀释条款和清算优先权未知。

[CO008, CO009, CO010, CO020, CO036]
FO002: Cognition AI ARR 与估值快照 KPI

截至最新披露数据点的关键财务与运营指标。

ARR 由公司披露的 Devin ARR(2025 年 6 月 $73M)与 Windsurf 收购时 ARR(2025 年 7 月 $82M)合计。估值为 2025 年 9 月融资后的投后估值。累计烧钱来自公司口径,未经审计。

[CO037, CO042, CO015, CO016]

1.4 规模与里程碑

按任何 SaaS 基准看,Cognition AI 的收入曲线都异常陡峭。公司在 2024 年末达到约 $1M ARR(September 2024),到 June 2025 加速至 $73M ARR;九个月约增长 73x,动力来自企业采用 Devin 处理代码迁移、修 bug 和功能自动化任务。July 2025 收购 Windsurf 后,来自 350+ 企业账户的 $82M ARR 被并入,合并业务在数周内升至约 $155M ARR。Goldman Sachs(据报道正在让 12,000 名开发者试点 Devin)、Citi、Dell、Cisco、Ramp、Palantir、Nubank 和 Mercado Libre 等大型企业客户已被公开提及。Nubank 据称用 Devin 处理代码迁移工作流,实现 12x 效率提升。April 2026,Cognition 在 Singapore 开设 APAC 总部,释放国际扩张意图。同月,公司宣布与 Mercedes-Benz 建立合作。员工数未详细公开;收购前外部估计 Cognition 团队约 49 人,而 Windsurf 收购在裁员和买断离职之前可能带来 200 名员工。[CO011, CO012, CO013, CO014, CO015, CO021]

里程碑表
日期事件类型金额 / 估值 / 状态参与方影响
Nov 2023Scott Wu、Steven Hao 和 Walden Yan 创立公司创立N/A三位 IOI 金牌得主首家专为智能体 AI 编程打造的公司启动
Mar 2024$21M 种子轮 / Series A,估值 $350M融资$21M / $350MFounders Fund首款产品发布前获得机构背书
Mar 2024Devin 公开发布;公布 SWE-bench 13.86% 成绩产品N/ACognition AI「世界首个 AI 软件工程师」主张引发全球媒体关注
Apr 2024融资 $175M,估值 $2B融资$175M / $2BFounders Fund + 其他投资方成立约 5 个月即成为独角兽
Sep 2024$1M ARR 里程碑规模$1M ARRN/A自助式企业客户带来早期商业牵引
Jan 2025Devin 2.0 发布,支持多智能体并行产品N/ACognition AI公司声称较 v1.x 效率提升 83%;任务可并行执行
Jun 2025ARR 达到 $73M(仅 Devin)规模$73M ARRN/A9 个月 ARR 增长 73 倍;企业采用取得突破
Jul 2025收购 Windsurf(估计约 $250M)产品~$250M / +$82M ARR / 350+ 客户Cognition AI / Windsurf增加 AI 原生 IDE 产品和企业客户基础
Aug 2025裁撤 30 名 Windsurf 员工;向约 200 人提供买断方案反向9 个月薪资买断包Cognition AI CEO Scott Wu 发声文化整合争议;人才流失风险
Sep 2025融资 $400M,估值 $10.2B融资$400M / $10.2BFounders Fund、Lux、8VC、D1、Bain、Elad Gil + 其他投资方成为十角兽;18 个月估值扩大 5 倍
Apr 2026新加坡 APAC 总部启用规模N/ACognition AI首个海外办公室;切入 APAC 企业市场
Apr 2026宣布与 Mercedes-Benz 达成合作合作N/ACognition AI, Mercedes-Benz进入欧洲汽车企业市场的切入口

日期和金额来自公司公告和媒体报道。Windsurf 收购价格为业内估计;官方条款未披露。里程碑表只反映公开已知事件;内部转向和监管活动未知。

[CO001, CO008, CO009, CO011, CO014, CO017]
FO001: Cognition AI 公司里程碑时间线

2023 年 11 月至 2026 年 4 月的关键创立、融资、产品、规模化和负面事件。

[CO034, CO037, CO041, CO042, CO025, CO033]
FO003: Cognition AI 业务系统流

Cognition AI 的身份、产品、客户和资本如何相互连接。

[CO003, CO011, CO021]

1.5 治理与关键人物风险

Cognition AI 是一家私营 C-corporation,没有公开财务披露,没有公开确认的董事会组成,除股权结构表内相关方外,投资人与分析师也不知道是否存在独立董事。公司治理结构仍不透明:投资人权利、Founders Fund 及其他领投方获得的董事席位、二级市场交易条款均未披露。产品愿景和技术信誉高度集中在三位联合创始人身上,尤其是 Scott Wu;他的公众形象和编程传奇身份与 Devin 品牌密不可分,由此形成重大关键人物风险。2025 年,公司向 Windsurf 员工提出自愿买断,原因是 CEO 公开要求 80-hour、每周六天工作制;这种强制性文化引发人才留存疑问,尤其是被收购员工原本并未主动选择这种环境。收购后 Windsurf 工程师流失,可能削弱支撑约 $250M 交易合理性的 IDE 产品能力。批评者和独立评测者指出,Devin 在真实任务中的成功率显著低于公司筛选后的基准演示所暗示的水平,“完全自主软件工程师”的叙事可能高估当前能力,进而带来未来投资人和客户预期风险。[CO034, CO035, CO041, CO022, CO023]

1.6 关键证据

Chapter 02

02市场分析

2.1 市场定义与边界

Cognition AI 的 Devin 所在市场,处在两类重叠市场的交叉处:AI 驱动的开发者工具(copilot、代码补全、代码审查自动化)和自主软件工程智能体(端到端任务执行,需要规划、多步推理和自我纠错)。前者是成熟大市场,GitHub Copilot 以 $2B+ ARR 锚定;后者是新兴细分市场,Devin 占据先行者位置。测算市场规模时,相关总可用市场覆盖专业软件开发团队可能采购、用于提升工程速度的工具——从自动补全插件,到无需人工监督即可执行工单的完全自主智能体。 纳入的支出包括:企业和 SMB 对 AI 编码助手的订阅、面向工程团队销售的自主智能体平台,以及用于程序化代码生成的 API 消耗。不纳入的支出包括:并非面向代码场景销售的通用大语言模型 API 访问,面向业务用户的低代码 / 无代码应用构建器(买方和工作流不同),基础设施软件(计算、存储、网络),以及传统外包软件开发服务。现状替代品包括离岸人力外包、无 AI 的传统 IDE、基于底层 LLM API 自建的内部开发者工具,以及云厂商捆绑的 IDE 扩展,如 AWS CodeWhisperer 和 Azure GitHub Copilot bundles。 不同分析机构对市场边界仍有争议。Grand View Research 以限制性口径测算“生成式 AI 编码助手”,认为 2030 年仅 $92–98M;Mordor Intelligence 则把“AI 代码工具”宽口径定义到 2030 年 $24B。Cognition 的 Devin 位于高端智能体细分市场,Team 方案 $500/month,企业合同在每年 $500K+ 区间,远高于 $10–$19/seat 的 copilot 档位。因此,可服务市场由愿意为自主工单解决支付生产率溢价的大型企业构成,估计占全球 27–28.7 million 专业工程师中的 5–15%。 [CM001, CM003, CM007, CM021, CM023]

市场定义表
类别纳入支出排除支出买方 / 付款方对 Cognition 的意义
AI 编程助手(copilot 层)IDE 插件订阅、自动补全 API、代码审查机器人非代码用途销售的通用 LLM API 访问开发者 / 工程预算间接:确立 Devin 升级所参照的品类常态
自主软件工程智能体(agent 层)端到端工单解决平台、工程师型智能体订阅、企业智能体部署RPA 工具、测试自动化(非 LLM)工程 VP / CTO / IT 预算直接:Devin 的主要 TAM
低代码 / 无代码平台可视化应用构建器、公民开发者工具、表格自动化完全排除在 Cognition TAM 之外业务用户(非工程师)Devin 无法覆盖
传统外包开发服务人员扩充、咨询、离岸工程团队内部员工薪资采购 / 财务替代目标:Devin 争夺这部分预算
云 IDE 和基础设施捆绑包与云服务捆绑的 AWS CodeWhisperer、Azure Dev tools、GCP AI核心云计算、存储、网络工程 / 云买方(捆绑)与 Cognition 部分争夺企业工程钱包

分析机构对市场边界没有共识;本表反映 Cognition 可服务的范围,不代表行业一致定义。第 3 行(低代码)明确排除在 Cognition TAM 之外,尽管部分分析报告把它与 AI 代码工具混在一起。

[CM001, CM021, CM023]

2.2 市场规模测算 — TAM、SAM 与 SOM

为了约束机会空间并调和差异很大的分析师估计,本报告采用多组独立测算镜头。考虑到不同来源的方法差异显著,镜头法优于单一自上而下估计。 镜头 1 — AI 代码工具市场(Mordor Intelligence,2025):广义 AI 代码工具市场 2025 年价值 $7.37B,预计 2030 年达到 $23.97B,CAGR 为 26.6%。该口径覆盖从 IDE 插件到完整智能体的所有 AI 辅助编码工具。TAM 候选值:$7.4B(2025)、$24B(2030)。局限:口径过宽;包含低代码工具、文档 AI 和测试自动化,与 Devin 的买方画像并不直接可比。 镜头 2 — 开发者席位 × 支付意愿:全球有 27–28.7 million 专业开发者(Evans Data / Statista 2024)。GitHub Copilot 在 $10–$19/month 档位,2025 年从这部分人群中获得约 1.3 million 付费席位。面向企业买方、价格为每团队 $500–$5,000/month 的自主智能体工具,可能可触达约 10% 的企业工程团队——约 500K–1M 个开发者席位 × $2,000/year 平均合同价值 = 2026 年 $1B–$2B SAM;随着采用扩散,2030 年扩展到 $5B–$8B。 镜头 3 — 企业软件工程支出占比:2024 年全球软件支出约 $675–$700B(WIPO)。即便软件项目劳动力成本的 1% 转向 AI 工具,也意味着 $6.7B 市场;3% 转移则意味着 $20B。这个上限镜头与 Mordor 的 2030 年 $24B 数字相互印证。 镜头 4 — Gartner 采用曲线:Gartner 预测,到 2028 年 90% 企业工程师将使用 AI 代码助手,early 2024 为 14%。全球企业工程师总数估计约 15 million。按 90% 渗透率 × $500/year 平均价格计算,仅企业细分市场到 2028 年就达到 $6.75B。这些估计彼此冲突明显:Grand View Research 的 2030 年 $92–98M 狭义数字完全排除了代码相邻 AI 工具,而 ResearchAndMarkets 的 $97.9B 数字似乎使用了异常宽泛的口径,抬高了可触达市场。尽调应主要参考 $1–8B 的 SAM 区间。 [CM001, CM002, CM003, CM004, CM005, CM006]

TAM/SAM/SOM 或规模测算视角表
发布方年份地区数值CAGR方法置信度局限
Mordor Intelligence2025→2030全球$7.37B→$23.97B26.6%自下而上的供应商收入 + 采用率建模口径宽,包含低代码工具;高估可与 Devin 对比的市场
Grand View Research2024→2030全球$25.9M→$92.5M(狭义)24.8%仅狭义「生成式 AI 编程助手」排除更广的开发者 AI 工具,可能低估
ResearchAndMarkets / BusinessWire2025→2030全球→$97.9B24.8%覆盖包含开发基础设施在内的广义 AI 工具TAM 数字大到不可信;包含不可比细分市场
开发者席位视角(本分析)2025→2030全球$1–2B SAM→$5–8B~40%27M 开发者 × 企业渗透率 × ASPASP 假设(平均 $2K/yr)对企业 / SMB 组合敏感
Gartner 采用率视角(本分析)2028全球企业~$6.75B 企业细分市场N/A90% 企业渗透率 × 15M 工程师 × $500/yr假设 Gartner 采用曲线到 2028 年完全兑现
智能体 AI 市场(CMR / VC Cafe)2025→2034全球$4.35B→$103B>40%跨行业自主智能体支出包含非软件垂直行业;仅作方向性支撑

TAM 区间很宽,根源在口径定义不同。对 Cognition 规划最可用的是 SAM 视角(当前 $1–2B)。广义 AI 代码工具市场到 2030 年的低 / 基准 / 高情景为 $1B / $7B / $24B。

[CM001, CM002, CM003, CM004, CM005, CM024]
FM001: 市场规模测算视角

截至 2025 年,AI 自主软件工程市场的 TAM/SAM/SOM 分层,展示从广义 AI 工具到 Cognition 目标 SOM 的可触达市场漏斗

SAM 和 SOM 为分析师估计,基于开发者席位和企业支出两种视角;研究机构对实际市场边界存在争议。

[CM001, CM005, CM007, CM035]
FM002: 市场估计区间

2030 年 AI 代码工具市场的低 / 基准 / 高估计,展示分析师在范围定义上的分歧

所有数值均为 2030 年预测,单位为十亿美元。区间很宽,反映分析师来源的范围定义根本不同。ResearchAndMarkets 的 $97.9B 数据因口径过宽、可信度不足,作为离群值剔除。

[CM002, CM003, CM004, CM005, CM024]

2.3 买方、用户与付费方分层

Cognition AI 面向三类主要买方,它们的预算归属和采用触发点不同。第一类、也是核心细分市场,是企业软件团队:工程副总裁或 CTO 是买方,软件工程师是用户,IT 预算是付费方。采用触发点是可衡量的生产率缺口,通常通过把 AI 工具 ROI 与人力成本做基准对比来发现。Goldman Sachs(12K 开发者试点)、Citi、Cisco、Dell 和 Nubank 等参考客户展示了这个细分市场。企业合同每年规模从 $100K 到 $2M+ 不等,销售周期 3–9 个月,包含法务和安全审查。 第二类是高增长科技创业公司:CTO 或创始工程师同时充当买方和用户,创始人把种子资金分配为付费来源。采用触发点是速度:Devin 能让 10 人团队打出 50 人团队的执行力。这类买方通常从 $20/month 的 Core 方案开始,再升级到 Team。关键参考客户包括 Ramp 和 Nubank,后者报告 12× 效率提升。交易价值较低,每年 $500–$6,000,但量大,且随着公司扩张,转为企业合同的可能性高。 第三类是个人开发者和自由职业者,战略优先级较低。买方和用户是同一人;付费方是按小时计费的自雇开发者。采用触发点是竞争必要性,因为 AI 工具正成为专业开发者的基本门槛。$20/month 的 Core 方案面向该群体;该群体对流失高度敏感,主要承担产品反馈循环和获客漏斗功能,而不是核心收入引擎。金融服务企业是横向优先细分市场,原因是 Goldman Sachs 的 12K 开发者 Devin 试点和 Citi 据报道的部署;这些买方在企业签约前需要 SOC 2 Type II、数据驻留和监管合规承诺。 [CM010, CM011, CM018, CM019, CM022, CM025]

细分市场 / 买方图谱
细分市场买方用户付款方工作流预算负责人采用触发点
F500 企业工程团队工程 VP / CTO软件工程师、DevOps工程 / IT 预算工单解决、代码审查、新功能开发CTO 或工程 VP开发者人力成本压力;试点证明 AI ROI
中型 SaaS 公司CTO / 工程经理后端 + 前端开发者工程预算功能交付速度、技术债清偿CTO上市速度竞争压力;从融资到发布的周期
高增长初创公司(种子轮–Series B)创始工程师 / CTO开发者 + CTO创始人预算 / 种子资金全栈功能开发、MVP 构建创始人资本效率:替代 2–3 名开发者的人力成本
金融服务企业CIO / 平台工程负责人重视合规的工程师IT / 数字化转型预算内部工具、监管报告代码CIO监管技术现代化 + 开发者成本下降
个人开发者 / 自由职业者本人本人本人客户项目交付、副业项目个人竞争所需;同行在用 AI 工具

预算归属差异很大:企业交易走正式采购,安全审查持续 3–9 个月;初创公司交易由创始人在几天内批准。金融服务买方在企业签署前要求 SOC 2 Type II 和数据驻留承诺。

[CM018, CM019, CM025, CM026]
FM003: 买方 / 细分市场图

Cognition AI 关键市场细分中,从预算所有者到企业部署的买方、用户、付费方关系和价值流

[CM018, CM019, CM025, CM026, CM028]

2.4 增长驱动因素与采用约束

市场同时拥有强结构性顺风和具体约束,这些约束决定采用速度,也决定 Cognition 将机会货币化的节奏。增长驱动因素包括已记录的开发者生产率提升——GitHub Copilot 用户完成任务快 51–55%,46% 的代码由 AI 生成,使用 AI 工具的组织报告开发者生产率提升 3.2×——这些数据验证 ROI 逻辑,也降低采购审批门槛。向智能体 AI 的快速迁移尤其重要:企业对智能体 AI 系统的支出预计从 2024 年低于 $1B,激增到 2028 年 $51B+,CAGR 约 150%,使企业技术组织中的自主智能体预算正常化。 劳动力成本和人才稀缺构成强经济理由。美国软件工程师薪资中位数每年超过 $130K,意味着一个 $500/month 的工具,只要能自主解决 10–20% 的工程工单,ROI 就能用周而非季度衡量。Gartner Magic Quadrant 对 AI 代码助手的覆盖,为企业采购提供合法性;Gartner 预测 2028 年企业采用率达到 90%,形成自上而下的组织命令,支持类别级支出审批。 主要约束来自信任和治理:企业在自主部署前普遍要求基于角色的访问控制、审计轨迹、IP 归属清晰和安全审查,这会给企业销售周期增加 3–9 个月。幻觉和代码质量风险意味着自主智能体可能引入难以捕捉的 bug;买方坚持保留审查层,从而限制完全自动化范围。与遗留工具链集成——本地 Git 服务器、隔离 CI/CD 流水线、自研 IDE——会增加部署成本和复杂度。EU AI Act 以及美国关于 AI 安全的行政命令,要求关键工作流中的企业 AI 系统具备文档和可解释性。竞争碎片化同样拖慢采购决策:GitHub Copilot(42% 市场份额)、Cursor($9B 估值)、Windsurf(现属 Cognition)和 OpenAI Codex 都在竞争,买方面临选择过载。 [CM008, CM009, CM012, CM013, CM014, CM015]

增长驱动因素与约束表
因素方向时间影响尽调问题
开发者生产力 ROI 已有记录(使用 AI 快 51–55%)顺风当前(2024–2026)验证经济性;加速 AI 工具采购审批核实客户给出的 Cognition 专属 ROI 数据,不能只看 GitHub Copilot 替代指标
Gartner:到 2028 年 90% 企业采用 AI 代码工具顺风近期(2026–2028)品类合法性增强;在大型企业形成自上而下的采购要求确认企业客户是否在 2026 年加快承诺
智能体 AI 支出:到 2028 年 $1B→$51B(150% CAGR)顺风近期(2026–2028)预算大规模转向自主工具,TAM 快速扩大跟踪企业 CFO 是否把智能体预算与 copilot 预算分开审批
软件工程师人才短缺与薪资通胀顺风持续从经济账看,$500/month 的 Devin 相比 $130K+ 工程师招聘更有吸引力跟踪开发者薪资调查;评估 AI 工具是否放缓目标客群的招聘需求
幻觉 / 代码质量信任缺口逆风近期(2025–2027)拖慢自主部署;需要人工审查层,限制完整 TAM 转化询问客户:Devin 输出中有多少百分比无需人工审查就上线生产
企业安全与 IP 归属顾虑逆风当前且持续销售周期拉长 3–9 个月;法律审查前限制 Fortune 500 部署审查 Cognition 的 IP 条款、数据处理和 SOC 2 / ISO 27001 认证
EU AI Act 与美国 AI 治理监管逆风中期(2026–2028)增加合规文档负担;可能要求 Devin 目前缺少的可解释性功能审查 Cognition 的监管合规路线图;评估 EU 市场准备度
竞争碎片化(Copilot、Cursor、Windsurf、Codex)逆风当前买方选择过载拖慢采购;copilot 层存在商品化风险跟踪 Devin 的智能体差异化能否守住,或竞品是否缩小差距

时点反映每个因素预计开始显现或达到相关性峰值的时间。所有顺风都是结构性(长期趋势)而非周期性因素, 意味着有利动态在 3–5 年投资周期内更耐久。

[CM008, CM009, CM012, CM013, CM014, CM020]
FM004: 采用漏斗或价值链图

Cognition AI Devin 在企业环境中,从开发者认知到完整自主部署的采用旅程

漏斗阶段比例基于 Gartner 采用数据和报道的客户数量估计;中间阶段(评估、部门级)为分析师估计,未由任何研究机构直接发布。

[CM008, CM010, CM015, CM017, CM019]

2.5 关键证据

Chapter 03

03竞争格局

3.1 竞争格局概览

2026 年,AI 辅助软件工程市场分成三大竞争层。第一层是成熟工作流 copilot,包括 GitHub Copilot(Microsoft)和 Amazon Q Developer,背后都有云超大规模厂商分发支持。GitHub Copilot 到 early 2025 拥有估计 42% 的品类市场份额、20 million 用户和超过 $2 billion 年经常性收入,已经深嵌企业开发者工具链。Amazon Q Developer 提供类似的补全和聊天能力,并针对 AWS 工作负载优化,面向庞大的 AWS 企业客户存量,带有 SOC 2 Type II 合规和原生 IAM 集成。 第二层是 IDE 原生 AI 编码助手,领头者是 Cursor(Anysphere)。Cursor 在 mid-2024 推出首个付费档位后,到 February 2026 已达到 $2 billion ARR,当前估值 $29.3 billion。Cursor 的 VS Code fork 模型带来插件型竞争对手无法获得的深度编辑器集成;其多智能体并行执行架构(最多八个并发智能体)与 Devin 的自主编码用例越来越重叠。Windsurf(原 Codeium)在被 Cognition 收购前也属于这一层。 第三层是完全自主的 AI 软件工程师,Cognition 的 Devin 直接定位在这里。Claude Code(Anthropic)、OpenAI Codex(以网页智能体形式重新推出)和 SWE-agent(开源学术基准框架)都在不同程度上参与竞争。Claude Code 的终端原生路径和在 SWE-bench 评测上的更高代码质量构成质量威胁;OpenAI Codex 以智能体执行环境重新发布,直指 Devin 的核心定位。差异化前沿正从基准分数转向可衡量的企业工作流集成和可验证 ROI 指标;在这片地形上,Cognition 与 Goldman Sachs 等客户的部署证据目前提供优势。 [CP001, CP002, CP003, CP033]

竞品画像表
竞品公司成立时间层级ARR(估计)估值主要投资方
GitHub CopilotMicrosoft/GitHub2021超大云厂商副驾驶$2B+N/A(MSFT)Microsoft
CursorAnysphere2022IDE 原生助手$2B$29.3B(Nov 2025)投资方:Thrive、a16z、Accel、Nvidia、Google
Claude CodeAnthropic2023自主智能体N/A(捆绑)$60B+Google、Amazon、Salesforce
Amazon Q DeveloperAmazon Web Services2023超大云厂商副驾驶N/A(AWS 捆绑)N/A(Amazon)Amazon
OpenAI CodexOpenAI2025(重新发布)自主智能体N/A(捆绑)$300B+Microsoft、a16z、Thrive
ReplitReplit Inc.2016云端 IDE 平台$80M 估计$1.2BAndreessen Horowitz、Coatue
SWE-agentPrinceton NLP2024开源框架免费 / 开源N/A学术机构
Windsurf(收购前)Codeium(已收购)2021IDE 原生助手$82M(收购时)$250M 估计General Catalyst、Kleiner

ARR 和估值数字来自截至 2026 年初的二手来源估计;缺少一手来源时采用这些估计。

[CP001, CP002, CP003, CP005, CP006, CP016]
FP001: 竞争定位图
[CP025, CP033, CP037]

3.2 第一梯队:超大规模云厂商支持的竞争对手

Microsoft 收购 GitHub 后拥有 GitHub Copilot,Copilot 受益于嵌入全球最大开发者平台的无可匹敌分发。90% 的 Fortune 100 公司使用 Copilot,活跃开发者超过 20 million,这种规模形成强切换成本惯性。Copilot 定价从个人 $10/month,到 Business 档 $19/month、Enterprise 档 $39/month,是多数开发团队默认的低成本入口。平台已经从行内代码补全逐步扩展到聊天、PR 摘要和工作区级 Copilot Workspace;后者是多步规划和执行智能体,正在侵入 Devin 的自主领地。GitHub Copilot 到 early 2025 年经常性收入超过 $2 billion,是全球最大的纯 AI 编码收入线。 Amazon Q Developer 最直接回应企业细分市场的合规和安全需求,提供 Free 档和 $19/user/month 的 Pro 档。它的原生 AWS 服务集成、200,000-token 上下文窗口,以及 Agents for Amazon CodeWhisperer 能力组合,使其成为 AWS 标准化组织的优选。Q Developer 已获得 SOC 2 Type II 认证,并支持 VPC 隔离——多数创业公司竞争对手尚不具备这样的安全姿态。与 Cognition 相比,超大规模云厂商层的主要限制是缺少全周期自主任务完成能力:Copilot 和 Q Developer 都不能在每个阶段没有人工在环确认的情况下,自主完成规划、实现、测试和端到端部署。这一架构约束正是 Devin 设计要利用的可触达缺口,尽管两家超大规模云厂商都在积极把智能体能力延伸到这一前沿。 [CP004, CP009, CP010, CP011, CP019, CP022]

功能 / 能力矩阵
功能Devin(Cognition)GitHub CopilotCursorClaude CodeAmazon Q Dev
自主端到端执行是(全周期)部分(Workspace)部分(8 个智能体)部分(人工监督)部分(Agents 功能)
IDE 集成SaaS + Windsurf IDEVS Code、JetBrains、VimVS Code 分叉版终端 / CLIVS Code、JetBrains、CLI
上下文窗口未披露未披露未披露200K tokens200K tokens
多智能体并行是(Devin Teams)是(最多 8 个)
开源模型选项模型路由(多个)
SOC 2 / 企业合规部分企业版企业版通过 API 提供企业能力SOC 2 Type II
免费层免费(有限)免费(个人爱好)是(有限)免费(50 次请求 / 月)
VPC / 本地部署未见文档披露通过 API 网关是(VPC)

能力评级基于截至 2026 年 5 月公开记录的功能;不纳入路线图功能。

[CP008, CP010, CP012, CP022, CP025, CP030]
定价 / 打包对比
工具免费层入门付费版中档套餐企业版
Devin(Cognition)$20/mo(5 ACUs)$500/mo(250 ACUs,Team)定制
GitHub Copilot是(2k completions)$10/mo(个人版)$19/mo(Business)$39/mo(Enterprise)
Cursor是(个人爱好)$20/mo(Pro)$40/mo(Business)定制
Claude Code是(有限)$10/mo(Pro)$100/mo(Max)企业 API
Amazon Q Developer是(50 次请求 / 月)$19/mo(Pro)N/A定制 / Marketplace
Replit$25/mo(Core)$40/mo(Teams)定制

定价来自截至 2026 年 5 月的官方产品页面;可能频繁变化。

[CP009, CP015, CP019, CP023]
FP002: 功能广度 / 能力图
[CP008, CP010, CP022, CP029]

3.3 IDE 原生竞争对手 — Cursor 与 Windsurf

从纯 ARR 维度看,Cursor(Anysphere)是 Cognition 最强的商业竞争对手。它在公开发布后约 24 个月内从零扩张到 $2 billion ARR。November 2025 的 Series D 融资以 $29.3 billion 估值融资 $2.3 billion,是当时 AI 编码史上最大的一笔风险融资。投资人包括 Thrive Capital、Andreessen Horowitz、Accel、Nvidia 和 Google。Cursor 的架构依赖 VS Code fork,并配有深度 token 预算控制、模型无关路由(Sonnet、GPT-4o、Gemini)和最多可运行八条并行任务线程的多智能体后台执行功能,正在缩小与 Devin 完全自主主张之间的功能差距。Cursor 拥有超过 1 million 付费客户和 50,000 个企业团队,证明其在个人开发者和团队细分市场都具备产品-市场匹配。Cursor 在 ARR 指标上增长快于 Cognition,付费客户更多,因此按比例获得更多训练信号和定价权。 Windsurf(原 Codeium)在 July 2025 被 Cognition 收购前,凭借免费档、兼容 VS Code 的 AI IDE 和超过 800,000 用户,建立了强开发者心智。收购后,Windsurf 的企业账本(贡献 $82M ARR,350+ 企业客户)并入 Cognition 收入基础,把直接竞争对手转为分发渠道。整合尚未完成,在 Cognition 品牌下留住 Windsurf 企业客户,是关键执行风险。此次收购也表明 Cognition 愿意通过外延增长抵御 IDE 层侵蚀;这释放一个信号:来自 Cursor 的 IDE 层竞争压力,被视为分发策略的生存级威胁。 [CP005, CP006, CP007, CP008, CP018, CP020]

3.4 自主智能体竞争对手 — Claude Code、OpenAI Codex 与 SWE-agent

在自主智能体层,Claude Code(Anthropic)可以说是 Cognition 最直接的质量型竞争对手。Claude Code 是终端原生的智能体编码助手,构建在 Anthropic 的 Claude 3.7 Sonnet 和 Claude 4 模型之上;这些模型在 SWE-bench Verified 上长期位居前列,该独立基准衡量智能体解决真实 GitHub issue 的能力。Devin 在 March 2024 的首个 SWE-bench 分数 13.86% 当时是突破,但 Anthropic 后续模型改进把 Claude 系智能体推到 50% 以上,削弱了 Devin 的基准差异化。Claude Code 定价(Pro $10/month,Max $100/month)在入门价格上低于或接近 Devin Core 档($20/month,5 ACUs),尽管用于智能体规模自主使用时经济性不同。关键架构差距在于 Claude Code 仍需要人工操作者在终端监督,而 Devin 设计为在企业云环境中完全无人值守运行;对大规模自动化来说,这是有意义的运营差别。 OpenAI 在 2025 年以基于网页的智能体编码环境重新推出 Codex,并由 o3 模型家族支持,定位为直接的 Devin 竞争对手。Codex Workspace 允许用户分配高层级功能任务并收到完成后的 PR,复制了 Devin 的核心价值主张。考虑到 OpenAI 的 $300 billion-plus 估值、API 分发和 ChatGPT 开发者生态,如果 OpenAI 优先投入执行质量和激进定价,Codex 的竞争入场将对 Cognition 差异化构成根本威胁。SWE-agent(Princeton NLP Group)是开源研究框架,不是商业产品;它的重要性主要在于基准比较器,也在于它释放了招聘信号:世界级 ML 研究者正积极在商业体系之外研究自主代码执行框架,可能加速核心技术的开源商品化。 [CP012, CP013, CP014, CP015, CP016, CP017]

护城河耐久性 / 竞争风险登记表
护城河因素Cognition 强度耐久性(1-5)主要威胁威胁来源
自主执行深度高 —— 完整无人值守任务周期3前沿实验室快速追平能力Anthropic、OpenAI
企业集成中 —— Goldman、Citi、Nubank 证明案例4超大云厂商捆绑激励Microsoft、Amazon
Windsurf IDE 分发中 —— 350+ 家企业客户3Cursor IDE 市场主导地位Cursor(Anysphere)
创始人人才 + 网络高 —— 3 名 IOI 金牌得主、前 Scale/DeepMind4前沿实验室挖角人才OpenAI、Google、Anthropic
训练数据飞轮中 —— 推测性质,未验证2更大的模型实验室拥有更多训练数据OpenAI、Google
基准领先低 —— 发布数月内失守1所有主要竞品都超过 SWE-bench 分数所有同业
定价竞争力低 —— 单任务成本最高2免费 / 低价层带来的商品化压力Claude Code、Copilot

耐久性分数是分析师的定性判断;尺度为 1(最弱)到 5(最强)。

[CP026, CP027, CP028, CP032, CP034]

3.5 护城河评估与竞争动态

Cognition 可防守的竞争优势可拆成三类:执行深度、企业数据飞轮,以及借 Windsurf 获得的分发。执行深度指完成多小时、多文件智能体编码会话的能力,包括规划、沙盒执行、测试和创建 PR;在已验证的企业部署中,Cognition 仍领先多数竞争对手(Goldman Sachs 12K 开发者试点、Nubank 12x 效率声明)。每个完成任务产生的企业数据飞轮,会给模型微调带来训练信号,理论上随时间复利;但在 Cognition 的模型质量相对前沿实验室得到独立验证之前,这一优势仍属推测。Windsurf 整合增加了 IDE 层分发,这是 Devin 单独难以快速自然扩张的能力。 质疑护城河持久性的反论也很强。Cursor 在纯 ARR 指标上增长快于 Cognition,付费客户更多。前沿模型实验室(Anthropic、OpenAI)控制底层推理能力,可以在不依赖授权的情况下部署等价智能体流水线,因此拥有结构性成本优势。基准可信度问题仍存在:Cognition 的 March 2024 Devin 演示部分受到质疑,SWE-bench 性能的独立评测显示,其他工具在 Devin 发布后数月内就超过了它。纯智能体差异化的竞争窗口正在收窄,Cognition 的企业销售与市场落地执行和客户留存指标,才是可持续竞争位置的真正领先指标。相对于 Amazon Q Developer 这类企业级竞争对手(SOC 2 Type II、VPC 隔离),Cognition 的合规姿态也存在已记录缺口,可能拖慢安全敏感垂直行业的企业合同周期。 [CP026, CP028, CP031, CP035]

FP003: 护城河 / 就绪度 KPI
[CP026, CP028, CP035]
Chapter 04

04财务情况

4.1 收入模型与 ARR 曲线

Cognition AI 运营的是基于消耗的 SaaS 收入模型,锚定 Agent Compute Units(ACUs)——一种按自主任务消耗计算资源的自有计量单位。个人开发者按 $20/month 订阅(Core 方案:5 ACUs),团队按 $500/month 订阅(Team 方案:250 ACUs),企业通过定制合同购买,含批量折扣和额外 ACU 池。这个价格阶梯按任务复杂度和数量等比例货币化,因此结构上不同于 GitHub Copilot 或 Cursor 这类按席位工具。但该模型也让收入更不平滑:企业客户的 sprint 工作负载波动,会带来不均匀的月度消耗;公司需要细致的 cohort 级分析,才能区分耐久 ARR 与一次性扩张。 Cognition 的 ARR 曲线是企业 AI 史上最快的一批。公司运营前六个月几乎没有收入,September 2024 跨过 $1M ARR,到 June 2025 达到 $73M ARR——九个月增长 73x——随后在 July 2025 收购 Windsurf 后跃升至约 $155M 合并 ARR(Windsurf 贡献 $82M ARR)。Cognition AI 在 September 2024 发布的博客《Funding, Growth, and the Next Frontier of AI Coding Agents》确认了 $73M ARR 数字,并概述了公司的双产品策略。Windsurf 之后,合并实体拥有分叉收入基础:Devin 的企业自动化收入(高 ACV、低席位数)和 Windsurf 的 IDE 订阅收入(较低 ACV、较高席位数)。一边整合团队,一边拆分并增长两条收入线,是重大执行风险。 截至 April 2026,报道显示公司正讨论以潜在 $25B+ 估值进行额外融资,意味着投资人相信 ARR 将在 $155M 之上继续显著增长。不过,收购后的 ARR 披露数字应谨慎解读:Windsurf 的 $82M ARR 贡献是在收购时点测量,未必反映迁移后的留存率。

收入来源表
收入来源产品模式ACV 区间推出时间ARR 贡献估计
个人订阅Devin Core按使用量计费,$20/mo(5 ACUs)$240/yr2024~$5M 估计
团队订阅Devin Teams按使用量计费,$500/mo(250 ACUs)$6,000/yr2024~$25M 估计
企业合约Devin Enterprise定制,VPC 部署$50K–$1M+ / 年2024~$43M 估计
IDE 订阅(Windsurf)Windsurf Free/Pro按席位计费,$0–$15/mo~$120/yr 付费席位收购前(2024)~$30M 估计
IDE 企业版(Windsurf)Windsurf Enterprise定制企业合约$50K–$500K / 年收购前(2024)~$52M 估计
用量超额Devin 全部套餐按 ACU 计量,~$2.25/ACU不定2024未披露

ARR 贡献估计是分析师近似值;Cognition 不披露各收入来源拆分。

[CI001, CI002, CI003, CI008]
FI001: 收入模型桥
[CI003, CI004, CI007]

4.2 融资历史与资本结构

Cognition AI 在不到 18 个月内通过三轮主要融资筹集约 $696 million 风险资本,成为有记录以来融资最快的 AI 基础设施公司之一。March 2024 的种子 / Series A 从 Founders Fund 筹集约 $21 million,投前估值 $350 million;在尚无产品收入前,这已经释放出对创始团队的早期信心。April 2024,发布数周后,Cognition 又以 $2 billion 投后估值完成 $175 million 融资,仍由 Founders Fund 领投,并在成立六个月内跨过独角兽门槛。September 2025 的 Series B 以 $10.2 billion 投后估值融资 $400 million,共同投资人包括 Lux Capital、8VC、Elad Gil、Bain Capital Ventures、D1 Capital、Definition Capital 和 Swish Ventures。VentureBeat 和 TechCrunch 在 September 2025 确认了这一轮融资。 公司披露从成立到 Q3 2025 的净现金消耗低于 $20 million;考虑到融资规模,这个烧钱水平极低,说明大部分资本被保留,而不是像同阶段公司那样快速花在人员或基础设施上。July 2025 的 Windsurf 收购估计约 $250 million,由现有资本储备支付,没有新一轮新股融资;相对于员工数和交易规模,这体现了财务纪律。收购后员工数据报道在裁掉 30 人、部分 Windsurf 员工接受 9 个月买断方案前达到约 249 人。

定价 / 变现表
套餐产品价格ACU 额度目标客群超额费率
CoreDevin$20/month5 ACUs个人开发者~$2.25/ACU
TeamDevin$500/month250 ACUs工程团队~$2.25/ACU
EnterpriseDevin定制定制池企业组织协商确定
免费版Windsurf IDE$0/month有限 AI 请求爱好者 / 学生N/A
ProWindsurf IDE~$15/month标准 AI 点数个人开发者不定
EnterpriseWindsurf IDE定制定制企业组织协商确定

Windsurf 定价截至 2025 年 7 月收购时;与 Cognition 定价结构整合后可能变化。

[CI001, CI002, CI005]
FI002: 单位经济模型桥
[CI006, CI035, CI011]

4.3 单位经济与利润率结构

Cognition 不披露毛利率、获客成本(CAC)或 cohort 级留存数据,因此很难精确建模单位经济。但仍可做结构性推断。ACU 模型意味着收入成本一部分可变(每个任务的推理算力),一部分固定(模型托管、沙盒环境、CI/CD 基础设施)。前沿模型规模的企业 AI 推理(每个任务很可能调用 GPT-4 或 Claude API)按当前云 spot 价格可达到每小时智能体运行 $5–$50;企业 ACU 包约按 $2.25/ACU 定价,Team 方案每月含 250 ACUs,因此毛利率高度取决于任务长度和模型调用效率。 高价值企业合同(Goldman Sachs、Citi、Nubank)可能采用带批量折扣的定制价格;如果基础设施成本能在大任务量上摊薄,规模化后利润率可能改善。Nubank 案例研究称代码迁移工作流效率提升 12x,说明客户认为 Devin 相对节省的开发者工时价值很高,这是定价权的有利信号。不过,完全自主编码智能体的内在难点在于任务复杂度差异很大:简单 bug 修复执行成本低,复杂的多周功能构建却可能不成比例地消耗算力预算。除非 Cognition 在企业层面限制任务范围,否则规模化利润率管理需要成熟的算力预测能力。 公司收购前资本效率(收入规模化前,在融资 $350M+ 的情况下烧钱低于 $20M)与收购后的成本结构可能形成鲜明反差。把 250 名员工(Windsurf 人员)加到约 49 人团队之上,员工数增加四倍以上,短期内会显著压缩人均收入,除非整合协同迅速兑现。

单位经济模型表
指标估计依据置信度注意事项
毛利率(估计)50–70%根据算力成本与 ACU 定价推断未披露;随任务复杂度高度波动
每 ACU 推理成本(估计)$0.50–$1.50前沿模型推理市场费率取决于模型组合和任务长度
企业客户 ACV年合同 $50K–$1M+行业可比数据、客户案例研究区间很宽;没有一手披露
净留存率(NRR,估计)100–130%从 ARR 增长轨迹推断未公开披露 NRR 数据
CAC(估计)未披露N/Aunknown未发布营销支出或销售数据
员工人均 ARR(Windsurf 前)~$1.5M2025 年 6 月约 49 名员工,ARR 为 $73M员工数为近似估计
员工人均 ARR(Windsurf 后)~$623K2025 年 7 月约 249 名员工,ARR 为 $155M包含离职前收购来的所有 Windsurf 员工

所有单位经济数据均为分析师估计;不存在一手财务披露。所有数值仅作方向性参考。

[CI006, CI009, CI010, CI011]
FI003: 财务估计区间
[CI012, CI013, CI014, CI016]

4.4 公开财务数据缺口与披露限制

Cognition AI 是一家私营 Delaware C-corporation,没有公开申报义务,没有披露 GAAP 财务报表,也没有投资者日或业绩电话会。本章所有财务数据依赖:(1)公司撰写的博客文章;(2)引用匿名来源的第三方新闻报道;(3)投资人沟通和新闻稿摘要;(4)二级市场研究估计。这种披露姿态在 IPO 前 AI 公司中常见,但会给试图验证 ARR 曲线可持续性的投资人和客户带来重大风险。 最重要的未披露财务指标包括:按产品线划分的毛利率(Devin vs. Windsurf)、客户流失和净留存率(NRR)、CAC 和回本周期、企业交易规模与合同期限(ACV),以及 Windsurf 整合后的成本。没有 NRR 数据,就无法判断 $73M 到 $155M ARR 的跃升来自自然客户扩张,还是完全来自非有机收购贡献。其他公司的企业 AI 工具合同在采用强劲时通常显示 100–140% NRR;如果 Cognition 的 NRR 低于 100%,即使有 Windsurf 注入,仅 Devin 的底层 ARR 也可能停滞。Sacra、CB Insights 和其他二级研究平台提供了部分收入和员工数估计,但必须把这些数字视为近似值。 第二个担忧是 early 2026 据报道的 $25B+ 估值讨论,这意味着基于 $155M ARR 的前瞻收入倍数约 160x。极高倍数在早期高速增长 AI 公司中历史上并不少见,但只有在 ARR 继续每年增长 2x+ 时才可持续;这个门槛要求大量新增企业客户获取,并留住 Windsurf 基础。任何 ARR 增长放缓都会显著压缩估值倍数。

资金充足性表
轮次日期金额投后估值领投方已确认来源
种子轮 / Series AMar 2024$21M$350MFounders Fund多家新闻报道
Series A 延展轮Apr 2024$175M$2BFounders FundAxios、CNBC 等多方
Series BSep 2025$400M$10.2B投资方:Lux Capital、8VC、Elad Gil、BCV、D1TechCrunch、VentureBeat、CNBC
累计融资(截至 Series B)Sep 2025~$596M$10.2B多方由上方合计
收购 WindsurfJul 2025~$250M(估计)N/ACognition(收购方)多方报道估计
潜在新一轮融资(传闻)2026未披露$25B+(据报道)未披露媒体报道,未确认

Series B 总额不含 Windsurf 收购成本;合计投入资本(融资 + 收购)约为 $846M。Founders Fund IX Form D/A(SEC,Oct 2025)确认 LP 资本基础为 $972M。

[CI012, CI013, CI014, CI015, CI016, CI024]
FI004: 资本强度 / 现金流图
[CI015, CI019, CI020]

4.5 资本充足性与现金跑道

Cognition 已融资约 $696M,并披露截至 Q3 2025 烧钱低于 $20M,因此进入 2026 年时,即使没有新融资,也拥有充足财务跑道。Windsurf 收购估计消耗 $250M 储备,剩余约 $426M,再扣除持续运营费用。按保守的 200 名员工、每人每年 $400K 烧钱(混合全负担成本)计算,不含算力的年度运营成本可能达到 $80M 或更高。如果公司以中等毛利率(考虑推理成本,估计 50–70%)产生 $155M ARR,毛利可能足以支持持续运营,无需立即补充资本。不过,$25B+ 估值讨论说明公司寻求资本是为了加速,而不仅是维持运营。新资本的预期用途包括 Devin 下一代模型研发、扩建企业销售队伍、为大规模推理建设全球数据中心容量,以及为整合 AI 编码基础设施市场而进行潜在后续收购。 Founders Fund 的领投关系带来额外跑道可选性:Founders Fund 过去愿意为表现突出的被投公司领投后续轮,降低融资风险。D1 Capital、Bain Capital Ventures 等大型机构共同投资人的存在,也说明公司可触达跨界投资和后期资本。总体资本充足性评级为中高:按当前烧钱估计,公司至少有 24 个月跑道,但激进增长执行和 Windsurf 整合成本可能让烧钱速度快于现有估计。

公开财务缺口表
缺失指标重要性严重程度使用的代理指标
按产品线拆分的毛利率决定盈利能力和长期单位经济模型行业可比:估计 50–70%
净留存率(NRR)判断 ARR 是否具备内生韧性无可用数据;假设 100–130%
客户流失率验证企业合同粘性未披露;从案例研究推断
CAC 与回本周期决定规模化后的 GTM 效率无可用数据
GAAP 收入与 ARR递延收入可能不同于 ARR使用 ARR 作代理
Windsurf 收购后留存决定合并后 ARR 能否守住尚未披露

所列缺口是截至 2026 年 5 月研究日期最影响决策、但尚未披露的财务数据。

[CI017, CI018]
Chapter 05

05产品与技术

5.1 产品定义与客户工作流

Cognition AI 提供两个核心产品:Devin,一个云端托管的自主 AI 软件工程师;Windsurf,一个用于本地和云辅助开发的智能体 IDE。Devin 接收自然语言任务描述——可通过网页仪表盘、Slack、Linear 或 Jira 提交——然后自主执行完整端到端软件工程任务:规划路径、编写和修改代码、运行测试、按失败结果迭代,并打开带可审查 diff 和置信度指标的 PR。不同于代码补全工具(GitHub Copilot、Cursor),Devin 接管的是整个任务,而不是辅助人类写单行代码。 在典型企业工作流中,开发者或工程经理把 Jira 工单或 Slack 消息分配给 Devin。Devin 读取工单,使用 DeepWiki server 探索代码库上下文,搭建必要环境,编写实现,运行测试套件,并以 PR 形式回报。PR 包含 Devin 内部决策轨迹,供人工审查。Devin 2.0 版本新增置信度仪表,在提交前量化任务成功概率,帮助团队高效分流任务。Devin 可以在多个会话中并发处理任务,使团队能够委托整个 sprint 待办列表,而不是逐个工单分配。 Windsurf IDE 通过集成 AI Cascade agent,为 Devin 补充本地开发环境。开发者可在 Windsurf 本地启动编码会话,并在会话中途把复杂、算力密集型工作交给云端 Devin,状态在切换中保留。April 2026 公布的 “Devin in Windsurf” 功能正式定义了本地到云端的交接工作流。两个产品合在一起覆盖完整开发者生命周期:Windsurf 服务日常编码辅助,Devin 负责被委托的自主任务执行。

产品 / SKU 映射表
产品交付模式目标用户关键功能价格档位上线时间
Devin Free云端智能体个人开发者有限 ACU、网页仪表盘、基础集成免费Apr 2026
Devin Pro云端智能体个人开发者更多 ACU、API 访问、置信度计量器付费(基础价 $20/mo)Apr 2025
Devin Max云端智能体重度用户高 ACU 池、优先队列付费(定制)Apr 2026
Devin Teams云端智能体开发团队共享 ACU 池、团队管理、Slack/Jira 集成团队定价2025
Devin Enterprise云端或 VPC企业组织VPC 部署、定制训练的 Devin、专属支持定制合同2024
Windsurf IDE (Free/Pro)桌面 IDE个人开发者Cascade AI 智能体、编辑器内补全、Codemaps免费 / $15/mo2024
Windsurf Enterprise桌面 IDE + 云端企业组织SSO、集中计费、企业支持定制合同2024
Devin APIREST APIDevOps / CI以编程方式管理会话、CI/CD 集成包含在付费计划中2025

截至 2026 年 5 月研究日期的计划结构;2026 年 4 月计划刷新以 Free/Pro/Max/Teams/Enterprise 结构取代了旧 Core/Team 计划。

[CE001, CE008, CE009, CE025, CE026]
FE001: Devin 任务生命周期流
[CE003, CE010, CE017, CE021]

5.2 产品与技术地图

截至 2026 年 5 月,Cognition AI 的产品组合由四个清晰单元组成:(1) Devin Core/Pro/Max/Enterprise——面向个人、团队和企业客户的云端智能体套餐;(2) Windsurf IDE——带有 Cascade AI 智能体的智能体桌面 IDE,2025 年 7 月从 Codeium 收购而来;(3) Devin API——用于编程式创建会话和 CI/CD 流水线自动化的 RESTful API;(4) Windsurf Enterprise——IDE 的企业层,提供 SSO、集中计费和高级管理员控制。第五个元素是 Windsurf Codemaps:由 SWE-1.5 和 Claude Sonnet 4.5 驱动的 AI 标注结构化代码库地图,把导航锚定到精确行号和可视化节点图,帮助团队快速上手和调试。 2026 年 4 月的套餐调整取消了原有 Core 和 Team 计划,改成五层结构:Free(ACU 有限)、Pro、Max、Teams 和 Enterprise。这个改动指向先落地再扩张的策略:Free 层让个人开发者和编程训练营学生低门槛接触,Pro 和 Max 服务高强度用户,Teams/Enterprise 则承接组织级部署。SWE-Check 是 2026 年 4 月与 Applied Compute 合作推出的专用缺陷检测模型,通过强化学习训练,内部评测追平 Opus 4.6,同时运行速度约快 10x——定位是在 Devin 智能体循环里更省成本的质量闸门。 Cognition 的自研模型 SWE-1.5 同时支撑 SWE-Check 和 Windsurf Codemaps,代表公司在 Devin 运行时模型组合之外第一次发布自有模型。所有产品都以云优先方式交付;需要本地部署或隔离云基础设施的企业客户可以选择 VPC 部署。2026 年发布的 Devin for Terminal 允许开发者从本地终端启动智能体会话,并在任务超出本地资源时升级到云端,把 CLI 和云部署两种模式接了起来。

集成生态表
类别平台集成类型能力状态
源代码管理GitHub原生创建 PR、访问 repo、跟踪 issue正式可用(GA)
源代码管理GitLab原生分支管理、创建 MR正式可用(GA)
源代码管理Bitbucket原生创建 PR、贡献代码正式可用(GA)
项目管理Jira原生工单分配、状态同步正式可用(GA)
项目管理Linear原生跟踪 issue、委派 sprint正式可用(GA)
沟通Slack原生启动会话、进度更新正式可用(GA)
沟通Microsoft Teams原生启动会话、通知正式可用(GA)
监控Sentry / Datadog / PagerDutyMCP告警触发会话、日志上下文MCP 市场
数据库PostgreSQL / MySQL / MongoDBMCP数据任务读写访问MCP 市场
文档Notion / ConfluenceMCP为任务检索上下文MCP 市场

MCP Marketplace 集成由第三方配置;可用性取决于客户配置。原生集成由 Cognition 维护。

[CE005, CE006, CE007, CE018, CE030]
FE002: SWE-bench 表现柱状图
[CE002, CE016, CE020, CE023]

5.3 技术架构与智能体运行时

Devin 的核心技术架构围绕一个智能体运行时展开,负责在沙箱化云环境中管理长周期规划和多步执行。智能体接收自然语言任务说明后,会拆成有序子步骤:探索代码库、搭建环境、实现功能、执行测试、创建 PR。每一步里,智能体都会在隔离计算容器中调用专用工具——代码编辑器、Unix shell 和无头浏览器——防止横向访问其他会话或客户数据。这种沙箱设计把 Devin 的执行上下文与客户生产环境切开;代码变更只通过 PR 交付,合并前需要人工批准。 DeepWiki 服务器通过构建向量化项目图来理解大型代码库,图中表示文件、函数和模块之间的关系。相比只靠 token 上下文窗口,这种图表示让 Devin 更能驾驭百万行代码库。BlockDiff 快照会在任务执行过程中记录增量状态检查点;测试失败或方案被放弃时,可以快速回滚。DeepWiki 和 BlockDiff 是 Cognition 最有辨识度的技术能力之一,因为它们正面处理了基于 LLM 的编码智能体两类主要失败模式:大型仓库里的上下文丢失,以及不可恢复的错误级联。 2025–2026 年发布的自研 SWE-1.5 模型支撑 SWE-Check 和 Windsurf Codemaps。它专门针对软件工程任务通过强化学习训练,区别于多数竞品以通用 LLM 作为基础层的做法。智能体运行时还支持 Model Context Protocol(MCP),让 Devin 可以通过可扩展插件市场连接 Sentry、Datadog、PostgreSQL、MongoDB、Notion 等外部工具,以及数百个其他服务。ACU(Agent Compute Units)是按任务消耗计量的计算 token,价格约为每 ACU $2.25,各套餐包含的配额不同。

SWE-bench 性能对比表
智能体 / 模型基准版本得分(% 已解决)评估模式日期来源
Devin(Cognition AI)Full(25% 子集)13.86%无辅助Mar 2024Cognition AI 博客
此前 SOTA(最佳 LLM)Full1.96%无辅助Mar 2024SWE-bench 排行榜
最佳辅助 LLMFull4.80%辅助(提供文件)Mar 2024SWE-bench 排行榜
SWE-agent(开源)Lite12.47%无辅助Mar 2024SWE-bench / Princeton
Claude Code (Opus 4)Verified(500)72.5%无辅助2025Anthropic
mini-SWE-agentVerified(500)65%无辅助Jul 2025SWE-bench 更新
Devin PR 接受率趋势AIDev 数据集(7,156 个 PR)+0.77%/week真实世界 PR32 weeksArxiv 2026

不同基准版本(Full、Verified、Lite)不能直接比较。Cognition 的 13.86% 使用 Full 基准的 25% 随机子集。Claude Code 的 72.5% 使用 500 个实例的 Verified 子集。

[CE002, CE020, CE023, CE024, CE035]
FE003: 产品能力 KPI
[CE002, CE004, CE014, CE016, CE026]

5.4 部署、集成与可靠性

Devin 支持两种主要部署模式:云托管(默认,运行在 Cognition 管理的基础设施上)和 VPC 部署,后者面向需要数据驻留、网络隔离或监管合规的企业客户。VPC 模式允许 Devin 在客户自己的 AWS、GCP 或 Azure 环境中运行,避免把源代码传到 Cognition 服务器。企业层还提供定制训练的 Devin 实例,客户可以用自有代码库和内部规范微调智能体。 平台与 GitHub、GitLab、Bitbucket 原生集成,用于源码管理(创建 PR、代码审查自动化、分支管理);与 Jira 和 Linear 集成,用于工单分派和状态跟踪;与 Slack 和 Microsoft Teams 集成,用于从沟通渠道发起会话并汇报进度。REST API 支持 CI/CD 流水线里的编程式会话管理,工程组织可以在特定事件发生时自动触发 Devin 会话(例如测试套件失败、客户缺陷报告)。MCP Marketplace 把集成延伸到监控平台(Sentry、Datadog、PagerDuty)、数据库(PostgreSQL、MySQL、MongoDB)和文档工具(Notion、Confluence)。Windsurf IDE 在本地集成 Cascade AI 智能体,并通过「Devin in Windsurf」把任务交接到云端。 Cognition 公开材料没有通过正常运行时间 SLA 或状态页数据正式披露可靠性。单个会话的任务执行时间上限为 45 分钟,但更长项目可以拆成连续子会话。Devin 2.0 的信心仪表会在执行前估计任务成功概率,让团队在消耗 ACU 前过滤低概率任务。这些质量控制能减少浪费,也说明 Devin 在复杂任务上的可靠性仍是概率性的,而不是有保证的。

技术架构组件表
组件功能输入输出自研?
智能体运行时任务拆解、多步规划自然语言任务说明有序子任务计划
沙箱环境每个会话独立计算容器智能体命令代码编辑、shell 输出、PR diff是(云托管)
DeepWiki 服务器向量化代码库图谱,用于上下文仓库文件图索引代码映射
BlockDiff 快照用于回滚的增量状态检查点每一步的会话状态检查点索引、回滚目标
SWE-1.5 模型Bug 检测(SWE-Check)+ Codemaps代码片段、仓库图谱Bug 可能性评分、标注地图
MCP 层外部工具连接符合 MCP 的工具规范智能体上下文中的实时工具数据开放协议
Devin API以编程方式管理会话带任务说明的 HTTP 请求会话 ID、状态、PR URL是(REST)

自研组件由 Cognition 开发,未开源;MCP 是 Anthropic 开发并被 Cognition 采用的开放协议。

[CE003, CE013, CE021, CE032, CE033]
FE004: 技术架构分层流
[CE003, CE008, CE013, CE032, CE033]

5.5 技术差异化与竞争护城河

Cognition AI 最核心的技术差异化,是端到端自主接管任务:Devin 不是给人类推荐代码再等待接受或拒绝,而是独立规划、执行并交付完整任务产物。这让 Devin 与代码补全工具(GitHub Copilot、Cursor、Claude Code)形成对位,而不是这些工具上的边际改进。2024 年 3 月,Devin 在 SWE-bench Full 上拿到 13.86%,刷新无辅助自主智能体的 SOTA,比此前最佳 1.96% 高 7x,也比此前有辅助 SOTA 4.80% 高 3x。Devin 的 PR 接受率在 32 周内呈现每周 +0.77% 的稳定正向趋势;在一项 7,156 个 PR 的比较研究中,它是唯一展现持续改善的智能体(Arxiv,2026)。 自研 SWE-1.5 模型和 SWE-Check 工具代表 Cognition 在模型训练上的早期纵向整合。多数竞品 AI 编码智能体(GitHub Copilot、Cursor、收购前的 Windsurf)搭建在第三方基础模型(GPT-4、Claude、Gemini)之上,基础模型提供商因此对工具厂商拥有结构性议价权。Cognition 专门为软件工程任务训练 SWE-1.5,正在建立一个模型层,使最专门化的用例不再依赖 OpenAI 或 Anthropic 的 API 定价。收购 Windsurf 带来了重要 IP:Cascade 智能体工作流引擎、Windsurf IDE 用户基础,以及 350+ 企业账户;这让产品套件加速成型,不必经历多年 IDE 开发。 数据飞轮优势存在,但没有公开量化。每个 Devin 任务都会产生执行轨迹,可用于微调未来模型版本。企业客户在自有代码库上使用定制训练的 Devin 实例时,Cognition 会积累任务特定的性能数据;没有同等企业关系的竞争对手很难复制。DeepWiki 代码库图表示和 BlockDiff 快照是自有工程创新,在大型代码库场景下提供功能优势;纯注意力 LLM 在这些场景中更容易失去连贯性。

信任、安全与合规表
控制项类型状态详情已披露?
会话沙箱隔离架构已确认每个会话都在临时容器中运行是(官方文档)
PR 把关的代码变更流程已确认所有代码变更合并前都需要人工批准
VPC 部署架构可用(Enterprise)客户控制的云边界
执行追踪日志审计已确认shell 命令、文件编辑、API 调用的完整日志
SOC 2 Type II认证未披露截至 2026 年 5 月无公开证明
ISO 27001认证未披露无公开证明
提示词注入防御(已修补)安全2024 年 12 月已修补直播中发现严重漏洞;Cognition 已修补反向
数据留存政策隐私未披露代码提交到云会话;留存不清楚

未披露 SOC 2 不等于该认证不存在;它可能只在 NDA 下提供给企业客户。研究日期,Cognition 的 trust.devin.ai 页面信息有限。

[CE015, CE028, CE029, CE031]

5.6 信任、安全、安保与合规

Devin 的主要安全架构依赖沙箱化计算隔离:每个会话都运行在临时容器中,除非通过集成凭证明确授权,否则无法访问其他会话或客户生产环境。代码变更以 PR 形式呈现并需要人工批准;Devin 执行轨迹会记录会话中每一条 shell 命令、每一次文件编辑和每一次 API 调用,形成可审计日志。VPC 部署把源代码留在客户自己的云边界内,又增加了一层隔离。 2024 年末,一次直播演示暴露了重大安全事件:Devin 的系统提示处理存在严重漏洞,实际上允许提示注入攻击操纵 Devin 行为。Cognition 承认问题并迅速修补,但事件凸显了部署自主智能体的内在风险——这类智能体可以根据文本输入执行 shell 命令。Hacker News 评论把这次失败称为「考虑到严重性,显得很业余」,并指出对具备 shell 执行能力的智能体来说,提示注入防御本应是第一优先级安全控制。 截至 2026 年 5 月,Cognition 尚未公开披露 SOC 2 Type II、ISO 27001 或同等级企业安全认证。trust.devin.ai 子域名存在,但研究时返回的公开信息有限。依赖 VPC 部署和定制训练模型的企业客户可能受单独安全协议约束,但这些协议无法公开审计。Devin 会话中执行代码的数据隐私与留存政策,在 Cognition 面向公众的文档中没有完整披露;对金融服务和受监管行业客户来说,这是实质性合规缺口。Mercedes-Benz 和 Goldman Sachs 的合作关系显示,合规障碍可能已通过定制协议解决;但缺少公开披露,外部无法独立评估。

Chapter 06

06客户情况

6.1 客户基础分层

Cognition AI 面向技术领先企业、中端市场公司和开发者优先组织中的专业工程团队。买方画像通常是希望在不按比例增加人手的情况下放大开发吞吐量的工程负责人或 CTO;使用者通常是单个软件工程师或工程经理,通过 Slack、Jira 或 Devin 网页界面直接分派任务。付款方则是工程预算负责人,可能是部门预算所有者,也可能是大型企业里的集中 IT 采购部门。 已确认生产部署的垂直行业包括金融服务和金融科技(Nubank——拉美最大数字银行)、汽车 / 制造(Mercedes-Benz)、IT 服务与外包(Cognizant)、企业技术,以及美国联邦政府。COBOL 现代化博客文章确认,财富 500 强中仍运行大量 COBOL 工作负载的行业已经部署,覆盖金融服务、保险和公用事业。2026 年,Cognition 推出 Government 垂直线,瞄准美国联邦机构的遗留软件现代化。日本和新加坡企业市场在 2026 年 4 月开放,说明东南亚金融服务、制造和技术公司可能是下一批目标细分市场。 地理分布以美国优先(截至 2026 年初,多数收入和具名客户来自美国),欧洲(伦敦办公室 2026 年 1 月开设)和 APAC(日本、新加坡办公室 2026 年 4 月开设)是增长中心。渠道分布包括直销企业销售、Windsurf IDE 的免费转付费漏斗(收购时日活跃用户 250K+),以及 Cognizant 转售合作,把 Devin 和 Windsurf 部署到 Cognizant 自有工程团队及其全球客户基础。Windsurf 集成在自上而下的企业采购之外,又创造了自下而上、开发者主导的采用路径。

客户分层表
客群购买方类型用户类型用例规模 / 价值证据等级
拉美金融科技企业CTO / 工程负责人软件工程师 / 工程经理老旧 ETL 迁移、代码现代化100M 用户、6M+ LoC 代码库已披露名称(Nubank),已量化
全球汽车 / 制造CTO / 工程副总裁软件工程团队老旧系统现代化、云原生开发、物流全球企业已披露名称(Mercedes-Benz),已公告
IT 服务 / 外包IT 高管 / 渠道软件工程团队 + 客户工程团队将 Devin 部署到自有工程团队和客户工程团队300+ 家客户组织已披露名称(Cognizant),合作伙伴
Fortune 500 COBOL 遗留系统工程副总裁 / CIO遗留系统软件工程师COBOL 现代化,迁移到现代技术栈运行数十年的 COBOL 系统未披露名称,仅博客披露
美国联邦政府机构 IT 高管机构软件工程师关键基础设施现代化联邦机构规模垂直业务已上线(2026 年 2 月),未披露机构名称
亚太企业CTO / 工程负责人软件工程团队东南亚、日本的软件生产大型企业2026 年 4 月开放市场,未披露客户名称
个人开发者(SMB / 独立开发者)本人本人编码自动化、个人项目低价值、高数量自助式 Free/Pro 档位

证据等级反映公开披露深度。Mercedes-Benz、Cognizant 和 Fortune 500 COBOL 案例仍停留在公告阶段, 尚未发布成果指标。

[CU001, CU002, CU003, CU004, CU005]
FU001: ARR 增长轨迹柱状图
[CU006, CU007, CU023]

6.2 采用轨迹与增长指标

Cognition AI 最突出的采用信号,是年经常性收入(ARR)的增长轨迹:从正式可用时(2024 年 12 月)约 $1M ARR,到 2025 年 4 月约 $73M ARR——四个月内大约增长 73x。这个增速与 Sacra 对 2024 年 10 月约 $15M ARR 的估计、以及 Growjo 平台当前 $73M 的估计相吻合,但两组数字都是第三方估计,不是审计后或公司披露的财务数据。2025 年 4 月的定价重置(Core 计划从 $500/月降到 $20/月基础价)在维持企业收入规模的同时,加速了个人开发者采用。 内部使用信号很有信息量。到 2026 年 2 月,Cognition 自己的工程团队每周合并 659 个 Devin PR,是 2025 年最佳周 154 个的四倍。这个「自用」信号重要:打造 Devin 的团队也是最重度的用户,形成高质量反馈循环,很可能加快迭代速度。PR 合并率从 Devin 发布时的 34% 提高到 2025 年 4 月的 67%——可视为输出质量改善的代理指标——但该指标为自报,分母方法未说明。 2026 年上半年在伦敦、东京、新加坡扩张,说明管理层相信收入足以支撑多办公室开销。Cognizant 合作尤其重要,因为它是渠道放大器:Cognizant 全球客户基础覆盖金融服务、制造和医疗等 300+ 组织,即便部分渗透,也可能在不同比例增加 Cognition 销售人手的情况下,显著拉动 Devin 的企业账户数。不过,关于活跃付费账户数量、平均合同价值或 ARR 估计之外的客户增长率,公开数据不存在。

客户增长和采用路径表
指标数值日期来源置信度含义
GA 发布时 ARR~$1M2024 年 12 月第三方估算(Sacra/imseankim)仅早期采用者;当时定价为 $500/month
定价重置后 ARR~$73M2025 年 4 月第三方估算(Growjo / imseankim)低至中降价至 $20/month 后约 4 个月增长 70x+
Growjo 口径 ARR(2025)$73M(估计)2025 年 9 月Growjo 收入估算第三方模型;非公司披露
发布时 PR 合并率34%2024 年 3 月公司声称(imseankim/VentureBeat)初期每 3 个 PR 有 2 个被拒;初始质量低
2025 年 4 月 PR 合并率67%2025 年 4 月公司声称(imseankim)约 1 年质量翻倍;仍有 33% 被拒
Cognition 自用 Devin 每周 PR 数6592026 年 2 月公司博客(官方)高强度内部试用;较 2025 年最佳周增长 4x
收购时 Windsurf DAU250,000+2025 年 7 月公司声称(Cognition 博客)大规模 Windsurf 用户基础,可作为收购后的转化漏斗
员工数增长~102% YoY2025Growjo 估算员工数快速增长,与收入加速相符

ARR 数字来自第三方估算,并非公司披露。PR 合并率由公司声称。缺失分母:总账户数和 ACV 拆分未知。

[CU006, CU007, CU008, CU009, CU010]
FU002: 关键客户与采用 KPI
[CU006, CU007, CU008, CU009, CU010, CU012]

6.3 具名客户证明与引用质量

质量最高的具名客户证明是 Nubank 案例研究,也是唯一公开详细且量化的生产部署。Nubank 是拉美最大数字银行(约 1 亿客户),使用 Devin 将一个 600 万行 ETL 单体迁移到子模块;该任务原本估计需要 1,000+ 名工程师投入 18 个月。借助 Devin,每个业务单元(Data、Collections、Risk)的迁移以周为单位推进,相比全人工基线,实现了按节省工时计算的 12x 工程效率提升和 20x 成本节省。关键在于,案例研究记录了针对客户特定迁移模式的微调,带来 4x 速度提升(单任务从 40 分钟降到 10 分钟)和内部基准上 2x 准确率提升。 Mercedes-Benz 在一篇博客文章(2026 年 4 月 27 日)中被列为客户,称其在「全球工程组织内」部署 Devin 和 Windsurf,用于遗留系统现代化、云原生开发和物流。公开信息没有披露结果指标;该公告仍处于合作 / 宣布阶段,而非事后案例研究。Cognizant(2026 年 1 月 28 日宣布)正在把 Devin 和 Windsurf 部署到「其工程组织和全球客户」中——这更像渠道合作,而不是具名终端客户,也没有生产结果数据。 其他未具名客户证据包括:财富 500 强 COBOL 现代化部署(2026 年 4 月 8 日博客文章,无具名客户)、美国政府(Cognition for Government 2026 年 2 月 25 日推出,无具名机构)、日本企业(2026 年 4 月 9 日推出,无具体客户)。证据深度和新鲜度差异极大:Nubank 是带量化结果的生产案例;其他仍在公告阶段。这造成证据质量集中风险——如果 Nubank 不能代表典型部署,客户证明基础就很浅。

具名客户验证表
客户客群用例生产环境 / 试点成果限制
Nubank(拉美新银行)金融科技ETL 单体迁移(6M LoC、100K 个数据类)生产环境效率提升 12x、成本节省 20x;数周而非数月需要微调;Cognition 发布的案例研究
Mercedes-Benz汽车老旧系统现代化、云原生开发、物流已公告(2026 年 4 月)未披露仅公告;无成果指标
CognizantIT 服务 / 渠道在自有组织和客户群中部署 Devin + Windsurf合作(2026 年 1 月)未披露渠道合作;无终端客户或成果数据
Fortune 500 COBOL(未披露名称)多个遗留系统垂直领域COBOL 现代化,迁移到现代技术栈生产环境(暗示)未量化;博客称部署仍在推进未披露名称;仅博客披露
Cognition AI 自身AI SaaS / 内部Devin 构建 Devin;每周合并 659 个 PR生产环境(内部)每周 Devin PR 数增长 4x(2025→2026)自我引用;无独立验证
美国政府(机构未披露)联邦政府关键基础设施软件现代化垂直业务已上线未披露未披露机构名称;FedRAMP 状态未知

只有 Nubank 在发布的案例研究中给出了量化生产成果。其他具名关系要么停留在公告阶段,要么是渠道合作, 没有披露终端客户。

[CU003, CU004, CU005, CU011, CU012]
FU003: 获客与扩张流
[CU017, CU018, CU019, CU021]

6.4 留存、满意度与耐久性

Cognition AI 公开可得的留存和满意度指标有限。公司没有披露净留存率(NRR)、总留存率(GRR)、客户流失率或队列分析数据。最接近满意度的公开指标是 PR 合并率(截至 2025 年 4 月为 67%,发布时为 34%),它衡量 Devin 打开的 PR 被人工审查者接受的比例——这是输出质量的合理代理,但不是标准客户满意度指标。67% 的 PR 接受率意味着 Devin 三分之一的工作产物会被拒绝或需要大量返工;对高量自主任务或许可以接受,但相较人类工程师仍是明显质量缺口。 独立来源的开发者情绪参差不齐。Hacker News 社区指出,早期 Devin 演示看起来打磨得很好,但复杂任务的真实表现落后于最初 13.86% SWE-bench 公告设定的预期。imseankim.com 的独立评测者在等待六周观察产品生产表现后指出,相比 Cognition 官方性能数字,「独立测试者讲述的是更复杂的故事」,并提示 ACU 成本在实践中累积很快:中等使用量就很容易让名义 $20/月 Core 计划下的月账单超过 $100–200。2024 年 12 月披露的安全漏洞也短暂损害了注重安全的企业客户信任。 耐久性信号方面,Nubank 部署描述了持续且扩大的使用(数据、催收和风险业务单元逐步采用 Devin)。Cognition 内部使用 Devin(2026 年 2 月每周合并 659 个 PR)说明产品驱动留存较强。Cognizant 和 Mercedes-Benz 合作一旦嵌入企业工作流,会形成结构性切换成本。不过,具名企业客户的合同期限、续约率和部署后满意度并不公开。

留存、重复使用和满意度表
指标数值客群置信度尽调请求
NRR(净收入留存率)未披露所有细分客群Unknown向 Cognition 索取:按队列和 ACV 区间拆分的 NRR
GRR(总收入留存率)未披露所有细分客群Unknown向 Cognition 索取:年度合同续约率
流失率未披露所有细分客群Unknown索取:按计划档位拆分的月度和年度订阅用户流失率
PR 合并率67%(2025 年 4 月)所有 Devin 用户中(公司声称)用独立 PR 审计或客户访谈验证
开发者满意度(HN)评价不一——有人热情,有人失望开发者社区委托对企业用户做独立满意度调查
Nubank 扩张深度3+ 个业务单元使用 Devin金融科技企业高(案例研究)索取:Nubank 合作时长、当前 ACU 支出
Cognition 内部 Devin 使用每周 659 个 PR(2026 年 2 月)AI SaaS(内部)高(博客)与 Git 统计交叉核对;确认 PR 类型分布
平均任务成功率~67% PR 接受率(代理指标)通用按任务类型、语言和复杂度区间拆分

留存指标未公开披露。PR 合并率是唯一可得的公开质量代理指标,而且由公司报告,未说明分母方法。 所有留存数字都需要直接向公司尽调索取。

[CU013, CU014, CU015, CU016]
FU004: 客户分层金字塔
[CU001, CU002, CU004, CU005, CU022]

6.5 扩张与集中风险

Cognition AI 的先落地再扩张策略很清楚:Nubank 部署从一个业务单元的 ETL 迁移开始,随后扩展到 Data、Collections 和 Risk 单元。Cognizant 合作也暗示类似结构——Cognizant 先内部部署,再向全球客户基础增购。Free 和 Pro 套餐面向开发者主导采用:个人开发者以低成本试用,推动内部采用,并制造自下而上的企业采购压力。Windsurf IDE 的 250,000+ 日活跃用户提供了庞大自然安装基础,可转化为付费 Devin 使用。 集中风险显著,且没有公开披露。在 ARR 约 $73M、已确认具名账户很少的情况下,少数大型企业合同(Nubank 量级、Mercedes-Benz 量级)很可能构成收入主导份额。如果 Cognizant 渠道合作表现良好,它本身也可能通过单一转售商带来不成比例的收入集中,而不是多元化直客关系。COBOL 现代化垂直线瞄准高价值项目(多年遗留系统),但这类项目在结构上通常是一次性工作,除非重新谈成长周期支持合同;初始迁移结束后,收入可能面临断崖风险。 受监管行业采购摩擦很高。金融服务、医疗和政府客户要求安全认证(SOC 2、FedRAMP)和数据驻留控制,而 Cognition 未公开披露已经满足这些要求。VPC 部署缓解了数据驻留担忧,但缺少已发布的合规证明,会限制其在监管最严格行业垂直中的可服务市场。2026 年 2 月推出 Government 垂直线,说明 Cognition 正在主动处理 FedRAMP 等效认证要求,但时间表和当前状态并不公开。这个合规缺口是 Cognition 面向目标大型企业客户的主要企业采购障碍。

扩张与集中度风险表
驱动因素 / 风险类型描述影响尽调路径
先落地再扩张(用例深度)扩张驱动Nubank 从 1 个业务单元起步,扩至 3+;Cognizant 先内部部署,再销售给客户绘制每个账户的扩张时间线和 ACU 增长
Windsurf 向 Devin 转化扩张驱动250K+ Windsurf 日活用户可作为云端智能体增购转化漏斗追踪 Windsurf Free 到付费 Devin 会话的转化率
COBOL 现代化(项目制)集中度风险高价值一次性迁移;完成后缺少结构性复购需求评估 COBOL 账户是否签约持续支持,还是迁移后离开
头部客户集中度(未知)集中度风险ARR 约 $73M 且具名账户很少时,一个 Nubank 量级账户 ≈ 重大占比索取前 10 大客户集中度;询问最大客户收入占比
Cognizant 渠道依赖渠道风险单一转售商可能主导新企业账户获取理解收入分成条款;评估锁定效应与直销竞争动态
合规缺口(SOC 2、FedRAMP)采购摩擦受监管行业没有认证就无法采购确认 SOC 2 Type II 和 FedRAMP In Process 状态;了解时间表
开发者疲劳 / ACU 成本敏感流失风险Pro/Core 套餐实际费用高(中等使用量 $100-200+/month);开发者可能降级索取 Pro 到 Free 的降级率和 ACU 使用分布

头部客户集中度是关键尽调缺口;没有公开数据。COBOL 项目制特征会制造收入断崖风险,需要追问。

[CU017, CU018, CU019, CU020]
Chapter 07

07风险

7.1 监管与法律风险格局

Cognition 处在快速演变的监管环境中。三大框架会在近期形成实质义务。EU AI Act(Regulation 2024/1689)对 GPAI 模型提供商的要求在 2026 年 8 月生效,要求透明度文档、版权合规政策和 EU AI Office 注册。由 ICO 管理的 UK GDPR 适用于英国 / 欧盟居民的所有数据处理,并要求对高风险 AI 处理开展数据保护影响评估;Devin 的自主 PR 合并和代码部署能力很可能触发 Article 35。2025 年 1 月的白宫行政令 撤销了此前拜登时期的 AI 安全义务,使美国联邦环境走向放松监管,但后续机构层面规则(尤其是政府 AI 采购)仍在生效。California SB 1047 原本会施加前沿模型问责要求,已在 2024 年 9 月被否决;后续立法(AB 2013、SB 1235)在 2025–2026 会期仍处于活跃状态。另一起 Doe v. GitHub、Microsoft 和 OpenAI 版权诉讼已在加州北区联邦地区法院合并审理,主张 AI 代码模型用开源仓库训练构成版权侵权;Cognition 未披露其模型训练数据来源,形成类似的潜在 IP 暴露。Cognition 服务条款限制使用输出训练竞争模型,但上游模型训练数据的版权状态无法独立验证。相较 Cognition 的规模及其横跨欧盟、日本和新加坡的企业客户基础,合规监控、跨司法辖区 DPA 覆盖和 GPAI 透明度文档都明显投入不足。若没有专职监管事务职能,到 2026 年 Q3,Cognition 可能面临执法行动、GPAI 未注册处罚,或欧盟市场准入受限。当前国际扩张阶段,主动监管项目应被视为业务关键投资,而不是可有可无的法务开销。

监管 / 法律风险登记表
规则 / 案件司法辖区状态可能性严重性缓释措施剩余风险敞口尽调路径
欧盟《AI Act》GPAI 条款(Reg. 2024/1689)欧盟2026 年 8 月生效发布 GPAI 透明度文件;向 EU AI Office 注册验证 GPAI 注册;取得 AI 透明度证明
GDPR / UK GDPR(AI 数据处理)欧盟 / 英国生效中DPA 协议;不使用 Enterprise 数据训练(公司声称)审计 DPA 覆盖范围;验证智能体处理是否符合 ICO DPIA 要求
Doe v. GitHub / Microsoft / OpenAI(AI 版权)美国加州北区联邦法院合并诉讼进行中ToS 限制将输出用于竞争模型训练;训练数据来源未披露获取关于 Devin 模型训练数据来源和 OSS 许可证敞口的法律意见
CA SB 1047 后续法案(AB 2013)美国加州2025–2026 年监测中当前无约束性义务;SB 1047 于 2024 年 9 月被否决按季度监测;若 AB 2013 通过委员会则重新评估
美国联邦 AI 行政令 — Jan 2025(放松监管)美国已生效受益于放松监管立场;关注各机构 AI 采购规则跟踪 OSTP AI Action Plan(预计 Jul 2025);监测政府部署的 FedRAMP 要求

可能性和严重性评级是定性尽调评估。缓释后的剩余敞口反映分析师判断,并非 Cognition 披露。

[CR001, CR002, CR003, CR004, CR005]
FR001: 风险热力图 — 影响 vs. 发生概率
[CR006, CR007, CR011, CR018, CR019, CR024]

7.2 安全与技术风险

Cognition 的核心产品风险在于 Devin 以较高系统权限运行——自主编写、执行、测试并部署代码——这相较被动代码助手大幅扩大了攻击面。2024 年 12 月,一次直播演示公开暴露了提示注入漏洞:攻击者可以在仓库 README 文件中嵌入恶意指令,使 Devin 泄露密钥、发起未授权 API 调用,或在客户代码库中植入后门。Cognition 承认并修补了问题。不过,事件表明,提示注入防御——OWASP Top 10 for LLM Applications 中排在首位的 LLM01——在 GA 发布前没有充分加固。Cognition 安全页面确认其拥有 SOC 2 Type II 认证(2024 年 3 月审计)、传输中和静态数据加密,以及全员 MFA。Trust Center 要求先签署 NDA,才披露渗透测试报告、审计范围和第三方评估结果——限制了企业采购团队独立评估控制有效性。关键未解缺口包括:没有公开漏洞赏金计划;没有正式 Agent Security Framework;没有公开渗透测试范围或红队披露;也没有发布 SLA 或历史正常运行时间数据。Devin 2.0(2025 年 4 月)把能力扩展到直接合并 PR,并在客户基础设施上调度智能体,显著扩大了未来任何提示注入漏洞的爆炸半径。OWASP GenAI Security Project 将智能体自主性识别为超越经典 LLM 漏洞的新兴独立风险类别。随着企业采用增长到数百家公司中的 VPC 部署,未被发现的 Devin 漏洞可能同时污染多个企业代码库——这是单个客户安全团队无法独立缓解的系统性供应链尾部风险。

运营与安全风险登记表
失效模式可能性严重性缓释成熟度剩余敞口未解决缺口
提示词注入 / 智能体操纵(OWASP LLM01)严重部分 — Dec 2024 已修补;未公开漏洞赏金计划或智能体安全框架未公布漏洞赏金计划;没有独立红队披露;没有 Devin 2.0 之后的渗透测试摘要
自主合并 PR 导致供应链代码污染严重部分 — 已有 SOC 2 Type II;建议代码审查,但未强制执行未公布渗透测试范围;信任中心细节受 NDA 限制;Devin 2.0 扩大影响半径
SOC 2 Type II 范围缺口 — 报告无法公开获取部分 — Mar 2024 取得 SOC 2 Type II;报告需签 NDA 才能获取签署 NDA 并审阅 SOC 2 报告范围、控制描述和例外项
AI 幻觉 / 在生产环境生成不安全代码部分 — 建议代码审查和分支保护;默认不强制没有独立研究比较 Devin 生成代码与人工编写代码的漏洞率
平台宕机 / ACU 配额耗尽部分 — 托管在 AWS;未公布 SLA 或历史正常运行时间索取 SLA 和历史正常运行时间数据;获取企业故障赔付条款

失效模式按严重性排序。缓释成熟度反映公开可得证据;未披露的内部控制可能更成熟。

[CR006, CR007, CR008, CR009, CR010]
FR002: 风险传导图
[CR006, CR007, CR011, CR018, CR019, CR020]

7.3 竞争与市场风险

Cognition 最尖锐的竞争风险,是曾经定义其领先地位的智能体编码基准正在快速商品化。Devin 在 2024 年 3 月发布时,SWE-bench Full 得分 13.86% 曾具有突破性。到 2025 年中,Claude Code Opus 4 在 SWE-bench Verified 上达到 72.5%——约 15 个月里提升五倍。OpenAI Codex 和 GPT-5 也达到可比的 Verified 分数。这样的改善速度威胁 Devin 的技术差异化:其多步规划、执行和测试循环可能在 12–18 个月内被基础模型提供商复制。GitHub Copilot(250M+ 安装量、Microsoft 分发)、Cursor($500M+ ARR、$9B 估值、$40/月定价)和 Claude Code(Anthropic)都在争夺与 Devin($500/月 Team 计划)相同的企业开发者生产力预算。FTC 已经担心,基础 AI 输入要素的集中控制——包括 Devin 依赖的模型——可能让这些提供商扭曲下游 AI 应用市场竞争,这对 Cognition 是结构性风险。Cognition 收购 Windsurf(2025 年 7 月)带来 350+ 企业账户和 250K+ DAU,但也带来集成风险:两个独立的 AI 智能体代码库、模型流水线、计费系统和销售动作必须同时合并,还要维持 ARR 增长。Devin 发布时,开发者社区围绕基准可靠性、封闭演示和误导性优势主张产生怀疑,形成不利声誉动态;这增加了技术成熟买家的采用摩擦,并且随着企业评估更严格,仍是逆风。定价压缩已经出现:Devin 2.0 将有效价格下调约 3×;未来竞争压力可能要求进一步降价,压缩每会话收入。

合作伙伴与依赖风险登记表
依赖项对手方角色集中度失效情景严重性剩余敞口
基础 LLM 提供商(主要)Anthropic / OpenAI所有 Devin 会话的核心模型智能极高提供商将 API 价格提高 3–5×,或限制智能体用例条款严重高 — 未披露多模型备用方案或自研模型时间表
云基础设施Amazon Web ServicesVPC 计算;企业部署AWS 调整 AI 智能体条款;价格大幅上涨中 — 未公开多云或本地部署备用路径
开发者工作流集成GitHub(Microsoft)PR 审查;代码 diff;CI 集成;Copilot 分发Microsoft 限制 AI 竞争者访问 GitHub API,或调整 Copilot 集成政策中 — API 层集成;没有独家安排
AI IDE 平台(Windsurf,Jul 2025 收购)内部(Cognition)企业 IDE 分发;350+ 企业账户;250K+ DAU整合失败;收购后人才流失;产品蚕食 Devin 会话中 — 整合可控;收购时保留管理层

列出的所有对手方都存在一定竞争利益冲突。依赖严重性按失效或重大涨价对收入的影响评级。

[CR011, CR012, CR013, CR014]
FR003: 关键合作伙伴与监管依赖图
[CR011, CR012, CR013, CR014, CR001, CR002]

7.4 财务与运营风险

截至 2026 年中,Cognition 已披露融资约 $1.575B。公司有 222 名员工(按 Growjo,年化员工增长 102%),产品又高度消耗算力,估计月烧钱为 $5–15M;保守情景下现金跑道约 18–36 个月。从 GA(2024 年 12 月)约 $1M ARR 增长到约 $73M ARR(2025 年 4 月估计)非常出色,但也引入客户集中风险:少数具名企业账户支撑 $73M ARR,说明可能存在几个超大合同,续约时会带来实质流失风险。Nubank 是唯一披露且量化的生产案例研究。如果 Nubank 或同等规模客户因安全事件、预算重分配或转向竞品而流失,ARR 下滑可能削弱未来以 $10.2B 假定估值融资的能力。ACU 定价模型(超额使用每 ACU $2.25)会带来不可预测的使用量收入,也会让低估消耗的客户因账单冲击而流失。LLM 推理成本带来的毛利压缩——以商业 Anthropic/OpenAI 价格调用第三方 API——会限制毛利率扩张,直到 Cognition 构建自有模型或拿到规模定价。公司没有公开披露盈利路径、EBITDA 或毛利率目标。Windsurf 集成增加了运营复杂度:同时合并计费系统、企业合同和产品团队,会提高执行风险并可能拖慢战略路线图。2025 年 1 月 Series A 的 $2B 估值与 Windsurf 后融资所暗示的 $10.2B Growjo 估值估计,意味着任何 ARR 增长轨迹的实质失误、流失事件或下一轮融资下调,都可能打击团队动机并触发关键员工离职。公司拒绝披露毛利率、NRR 或队列留存指标,使财务风险进一步升高。

人员与执行风险登记表
角色 / 职能依赖或缺口可能性严重性缓释措施尽调路径
CEO / 联合创始人(Scott Wu)产品愿景单点依赖;投资人信任背书;技术领导力;竞赛程序员身份严重联合创始人团队;董事会监督;未公开接班计划索取治理结构;确认董事会独立性;审阅联合创始人归属悬崖期安排
AI 研究 / ML 工程人才Anthropic、OpenAI、Google DeepMind 和 Microsoft 争抢同一批竞赛编程与 ML 人才,竞争极端激烈有竞争力的薪酬(估计 $500K+ TC);Cognition 在 AI 社区品牌强索取按职能拆分的员工数;审阅高级 ML 工程师任期和股权刷新政策
国际法律 / 监管运营向欧盟、日本、新加坡和美国政府客户扩张,但未披露专门法律基础设施;尚未取得 FedRAMP刚起步;可能按司法辖区临时聘用外部律师核实欧盟 DPO 任命;索取日本 PIPL 和新加坡 PDPA 合规路线图;确认 FedRAMP 时间表

风险评级基于公开披露的组织结构。内部接班计划、人才梯队厚度和留才项目未公开披露。

[CR015, CR016, CR017]

7.5 执行、人员与合作伙伴风险

Cognition 三位联合创始人——Scott Wu(CEO)、Steven Hao 和 Walden Yan——都是竞技编程冠军,技术资历极强,但在早期创业之外没有构建企业软件公司的历史业绩。关键人风险集中在 Scott Wu 身上,他驱动技术愿景、投资人关系和产品战略。Anthropic、OpenAI、Google DeepMind 和 Microsoft 争夺 AI 人才异常激烈,且都在以超过 $500K 总薪酬的待遇运营前沿 AI 项目,直接竞争 Cognition 依赖的竞技编程和 ML 工程人才池,给其 AI 人才管线造成压力。Devin 平台高度依赖第三方伙伴:AWS(VPC 算力)、GitHub(PR 工作流集成)以及 Slack/Teams(企业沟通)。这些平台中任何一家修改 API 访问、定价或集成政策,都可能破坏 Devin 的企业价值主张。Anthropic 和 OpenAI 既是供应商又是竞争者:它们供应 Devin 编排的底层模型,同时分别通过 Claude Code 和 OpenAI Codex 竞争,形成利益冲突激励。2026 年 Q1–Q2 宣布向日本、新加坡和欧盟扩张,引入了司法辖区特定的合规义务,需要在收入兑现前招聘法律和运营人员。美国政府客户通常要求 FedRAMP 或等效授权,流程一般需要 12–24 个月;已宣布政府垂直线与实际 FedRAMP 授权之间的时间差,形成声誉和合同风险窗口。国际法律基础设施——欧盟 DPO 任命、日本 PIPL 合规、新加坡 PDPA——正在成为缺口;这家公司此前更优先产品速度,而不是监管脚手架。

缓释措施与止损标准表
风险可监测触发因素止损标准阈值行动含义
提示词注入 / 供应链泄露披露 CVE;客户代码库污染;第二起安全事件任何已确认的生产客户代码库供应链污染,或 12 个月内发生第二起公开安全事件立即复盘投资逻辑;评估流失影响;判断修复可信度;考虑退出持仓
竞争对手将智能体编程商品化每月 SWE-bench 排行榜;Cursor、GitHub Copilot、Claude Code 价格公告竞争对手 SWE-bench Verified ≥80% 且价格 ≤Devin 的 50%,或 Devin ARR 同比增长 <50%减仓;重新评估基准分数之外的护城河逻辑;评估 Windsurf 带来的差异化
客户集中 / Nubank 流失续约时披露 NRR;最大客户 ARR 占比;管线多元化证据单一客户占 ARR >30%,或任一滚动季度 NRR <90%立即审查 ARR 质量;建模流失情景;索取管线多元化证据
LLM 提供商价格冲击或 API 限制Anthropic / OpenAI 价格公告;面向智能体用例的 API 条款更新LLM API 成本上涨 >3×,但 Devin 未同步涨价;或 API 条款限制自主合并 PR测算模型毛利压缩敞口;评估自研模型时间表和成本转嫁可行性

止损标准是投资逻辑破坏触发点。阈值反映当前共识;每个监测周期重新评估。

[CR018, CR019, CR020, CR021]
Chapter 08

08估值

8.1 投资逻辑与反向逻辑

Cognition AI 的乐观逻辑建立在三条相互强化的主张上。第一,自主软件工程是真正的市场品类迁移,而不是增量生产力提升。全球软件开发劳动力的总可用市场超过 $650B;即便自主智能体只替代 10% 的开发者时间,也意味着 $65B+ 的收入机会。Cognition 的 Devin 是品类先锋,在多数买方理解这一品类之前,就把自主编码智能体变成领先企业的采购行项目。先发品牌认知、灯塔客户证明(Nubank)和 Windsurf IDE 网络(250K+ DAU)共同形成平台护城河,让其定价权不完全依赖纯基准表现。第二,ARR 从 $1M(GA,2024 年 12 月)到估计 $73M(2025 年 4 月)用时九个月,意味着增长率与历史 SaaS 基准结构性脱钩。按这个增速,公司有可能在 2025 年 4 月数据点后的 18 个月内达到 $500M ARR,进入历史上增长最快的企业 SaaS 公司行列,并在前瞻 ARR 倍数压缩的情况下支撑 $10B+ 估值。第三,联合创始人深厚的技术背景和 Founders Fund 的高信念投资,传递出创始团队能够维持产品速度、延续差异化的信号。 反向逻辑同样有力。基准商品化是生死风险:Claude Code Opus 4 在 2025 年的 SWE-bench Verified 得分 72.5%,对比 Devin 发布时的 13.86% Full 得分,说明性能差距正在快速缩小,可能到 2026 年末消除 Cognition 的技术差异化。$10.2B 估值约为估计历史 ARR 的 140×,定价已经假设近乎完美执行,而除 Growjo 估计外没有公开佐证。客户集中风险无法量化:$73M ARR 来自未披露数量的企业账户,可能只意味着三到五个合同,其中单个客户流失就会触发实质 ARR 下滑。2024 年 12 月直播中发现的提示注入安全事件,在关键早期采用拐点损害了 Cognition 在技术成熟买家中的可信度。对 Anthropic 和 OpenAI 的 LLM 提供商依赖也引入结构性利益冲突,因为两者都直接与 Devin 竞争;这种冲突可能表现为 API 定价突然上调或功能限制,抬高 Devin 的边际交付成本。没有 NRR、毛利率或队列留存数据,外部无法独立评估报告 ARR 的财务质量。

建议摘要表
维度评估证据基础置信度含义
建议观察名单 — 入场估值低于 $5B 时有条件投资ARR 增长异常强劲;估值倍数高达 140× ARR;安全和集中度风险未解决跟踪 NRR、毛利率和整合进展;必须严守入场纪律
风险评级有安全事件历史;基准商品化;依赖 LLM;客户集中度未知若执行失速,以 $10.2B 入场有较高资本减值风险
估值立场投机性溢价 — 当前 $10.2B、约 140× ARR,是 Cursor 倍数的 7×倍数对比:Cursor $9.9B / $500M ARR = 20× ARR;Cognition $10.2B / $73M ARR = 140× ARR基准情景下公允价值区间为 $3–7B;高于 $7B 入场需要乐观情景证据点
乐观情景目标$18–25B 退出(持有 3 年,$700M+ ARR,25–35× 前瞻 ARR)增长率可持续;Windsurf 整合成功;NRR ≥120%低-中以 $10.2B 入场可获 1.5–2.5× 回报;只有近乎完美执行,才足以支撑风投 IRR
悲观情景目标$800M–$1.8B 退出(第二起安全事件,或重大流失,或竞争对手降价 >3×)安全阴影;基准商品化;LLM 定价风险以 $10.2B 入场可能损失 90%+ 资本;没有更高置信度的 NRR 数据,不接受这种不对称

该建议面向正在评估以当前估值买入 Cognition 老股或新股的机构投资者。这不是投资建议。

[CV001, CV002, CV003]
最终尽调请求表
尽调请求优先级关键原因可接受阈值
按队列划分的净收入留存率(季度,滚动 4Q)关键ARR 质量;判断增长是扩张驱动,还是被流失掩盖NRR ≥110% 证实健康;<90% 则破坏投资逻辑
按产品线划分的毛利率(Devin ACU、Windsurf、Enterprise)关键LLM 成本结构;盈利路径;毛利压缩风险目标毛利率 50%+;<30% 会让规模化可持续性存疑
前五大客户 ARR 及其占总 ARR 比例关键集中度风险;单一客户流失事件的影响前五大客户 <40% 总 ARR;任一单一客户 >15% ARR 即为高集中
Devin 2.0 后渗透测试报告或红队摘要(December 2024 补丁后)安全可信度;智能体能力扩展后需要重新审计签署 NDA 后开放 SOC 2 Type II + 至少 1 次 Devin 2.0 后独立渗透测试
Windsurf 整合后的客户留存率(收购后 6 个月)整合执行证明;IDE 客户流失风险收购后 6 个月,Windsurf 企业账户留存率 ≥85%
LLM API 合同条款(Anthropic、OpenAI):批量价格、智能体用例条款、控制权变更供应商利益冲突;毛利率底线判断价格锁定至 2027 年,或多模型分散策略已落地
欧盟 GPAI 合规计划与监管事务人员配置2026 年 8 月合规截止;若不合规,欧盟收入面临风险已明确欧盟监管事务负责人;GPAI 透明度文件时间表 <6 个月
美国政府垂直市场 FedRAMP 授权状态与时间表政府 ARR 管线;政府垂直市场上线可信度(2026 年 2 月)FedRAMP 授权时间表清晰;具名机构确认过渡期用例

优先级:关键 = 投资决策前必备;高 = 投资条款清单前必备;中 = 交割前必备。

[CV023, CV024, CV025, CV026, CV027, CV028]
FV001: 投资建议决策逻辑
[CV001, CV002, CV018, CV019]

8.2 估值背景与融资历史

Cognition 已完成三轮披露融资,合计约 $1.575B。$21M 种子轮在 2023 年确立了创始团队。$175M Series A(2025 年 1 月至 4 月,Founders Fund IX 领投)把最初机构轮投后估值定在 $2B。SEC EDGAR 记录确认 Founders Fund IX(CIK 0001971631)是向 Cognition 出资的大型机构载体;Cognition Capital SPV I(CIK 0002072175,Form D 于 2025-06-11 提交)显示,Windsurf 收购交割同期或之后不久设立了后续结构化载体。$400M 融资(2025 年 9 月)将公司投后估值定在 $10.2B,较 Series A 在约九个月内上升 5.1×——这是 AI 软件史上最陡峭的估值跃升轨迹之一。总融资约 $1.575B 暗示在 $10.2B 投后基础上稀释约 15%+,但具体优先权堆叠和清算瀑布未披露。以 $10.2B 估值计算,主要收入倍数约为 2025 年 4 月约 $73M ARR 估计的 140×,高于所有已披露 AI 软件同业。Cursor/Anysphere 于 2025 年 9 月以约 $500M ARR 完成 $9.9B Series D,对应 20× ARR 倍数。Cognition 的溢价部分由更高增长率和品类先锋定位支撑,但对以 $10B+ 入场的投资人来说,执行风险已经被计入价格。Windsurf 收购(2025 年 7 月)以 Cognition 股权和现金完成,收购价未披露;Cognition Capital SPV I 指向一个结构化投资人层面载体,可能用于出资收购对价或为合并实体经营计划兜底。Winbuzzer 在 2025 年 8 月报道的「文化重置」——Cognition 向 Windsurf 员工提供九个月买断,并据称要求每周工作 80 小时——引入了头条估值没有体现的人才整合风险。任何以 $10.2B 入场的新机构投资人,在低于 $10.2B 的退出价值下回本前,都要面对种子轮、Series A 和 $400M 轮形成的实质优先权悬挂。

投资逻辑 / 反向逻辑表
投资逻辑支柱支持证据反向证据确信度
自主工程成为品类开创事件ARR 在 9 个月内从 $1M 增至 $73M;Nubank 效率提升 12×;Cognition 内部使用 Devin(659 PRs/week)基准商品化在加速;同行可能不需要 Cognition 工具,也能复制编排层
Windsurf IDE + Devin 云智能体整合形成网络护城河收购时带入 Windsurf 250K+ DAU 和 350+ 企业账户(Jul 2025)存在整合执行风险;据报道 Aug 2025 进行了文化重置;两个代码库需同时合并低-中
Founders Fund IX 背书是质量信号Founders Fund IX(CIK 0001971631)领投 Series A;有 Peter Thiel 相关投资记录Founders Fund 组合集中;过去也有投资逻辑失误;不保证结果
降价作为放量策略Devin 2.0 将有效价格降至原来的 1/3;每个 ACU 任务量增加 83%;暗示成本曲线在下行毛利率未知;3× 降价要么说明成本结构改善,要么是竞争回应;两者都会压低总收入天花板

确信度评级是分析师基于公开证据的判断。内部财务指标(NRR、毛利率)不可得,无法修正这些评估。

[CV004, CV005, CV006, CV007]
FV002: 估值对 ARR 与倍数的敏感性
[CV008, CV009, CV010, CV011]

8.3 乐观、基准与悲观情景分析

乐观情景假设 Devin 的 ARR 到 2027 年保持每季度 50% 的持续增长率,2027 年底达到约 $700M–$800M ARR。在这个规模上,随着模型成本优化和批量合同改善毛利率,Cognition 可能获得 $15–25B 的 IPO 或并购估值,相比 $10.2B 入场价回报 1.5–2.5×。乐观情景需要满足:2024 年 12 月之后没有重大安全事件;Windsurf 集成成功并补上可信 IDE 渠道;NRR 披露 ≥120%;竞争模型未能在 Devin 的企业工作流集成深度上实现商品化。基准情景假设 ARR 中速增长(从 2025 年冲刺期逐季衰减到 30–40%),到 2027 年底达到 $200–350M ARR,毛利率改善到 50–60%。按标准化 25–35× 前瞻 ARR 倍数(与高增长 AI SaaS 一致),对应 $5–12B 估值——与 $10.2B 入场价持平或接近,意味着该入场价投资人的回报几乎可以忽略。基准情景需要没有重大流失事件、Windsurf 集成合理推进,以及企业持续投资 AI 基础设施。悲观情景假设 18 个月内出现重大负面事件:第二次安全事件、Nubank 流失,或 Claude Code / OpenAI Codex 商品化引发 Devin 价格战。在悲观情景下,ARR 平台化在 $80–120M,NRR 披露低于 90%,估值压缩到 10–15× ARR,对应 $800M–$1.8B 退出价值——以 $10.2B 入场的投资人亏损 90%+。考虑到已知风险因素集中,尤其是安全阴影和结构性基准商品化趋势,悲观触发概率估计为 20–30%。回报预期不对称:乐观上行(3 年 1.5–2.5×)不足以补偿 $10.2B 入场估值下 30%+ 严重资本损失概率。这是核心投资逻辑挑战:除非投资人对 NRR 和毛利率拥有市场没有的高信念、差异化信息,否则当前入场价的风险调整后回报不具吸引力。

乐观 / 基准 / 悲观情景表
情景ARR(2027 年底估计)ARR 倍数隐含估值概率信号关键假设
乐观$700–800M25–35× 前瞻 ARR$18–28B20–25%(需要近乎完美执行)无重大安全事件;NRR ≥120%;Windsurf 整合成功;LLM 定价稳定
基准$200–350M20–25× 前瞻 ARR$4–9B45–55%(中等增长执行)中等流失;Windsurf 部分整合;季度 ARR 增长率衰减 30–40%
悲观$80–120M8–12× 滚动 ARR$640M–$1.4B25–30%(重大负面事件)第二起安全事件,或重大客户流失,或竞争对手价格下探 >3×,或监管执法

情景区间是分析师基于可比公司基准和已披露增长指标作出的估计。概率是定性信号,不是量化模型。

[CV008, CV009, CV010]
FV003: 各情景估值与回报区间
[CV008, CV009, CV010, CV003]

8.4 可比公司估值分析

Cognition 的可比公司组包括 AI 原生开发者生产力工具(Cursor、Replit)、拥有编码产品的基础模型提供商(Anthropic/Claude Code、OpenAI/Codex),以及 ARR 运行率相近的更广义企业 AI SaaS 公司。Cursor/Anysphere 是最接近的可比对象:两者都是面向企业开发者的 AI 编码工具,都是 VC 支持,并且都在 2025 年下半年达到最新估值。Cursor 以约 $500M ARR 完成 $9.9B Series D(Bloomberg,2025 年 11 月),对应 18–20× ARR 倍数。Cognition 以约 $73M ARR 对应 $10.2B 估值,意味着约 140× ARR 倍数,按 ARR 口径相对 Cursor 有 7–8× 溢价。Cognition 的溢价部分由更高增长率(Devin 从发布起增长更快)支撑,但随着 Cursor 的 ARR($500M)已经证明更高收入质量和更宽客户基础,这种溢价在结构上不可持续。Harvey AI(法律 AI,$3B 估值,约 $100M ARR,约 30× ARR)是垂直 AI SaaS 可比对象,高质量 NRR 支撑其溢价倍数。Glean(企业 AI 搜索,$4.6B,约 $100M ARR,约 46× ARR)说明企业 AI SaaS 在 $100M ARR 时可以维持 40–50× ARR 倍数,但 Glean 的 NRR 据报道高于 150%。Replit(2023 年 $1.16B 估值,带 AI 转型的开发平台)代表品类风险:一个先锋转型故事,后来把市场份额输给更聚焦的 AI 原生竞争者。2025 年,接近盈利或已盈利 AI SaaS 公司的隐含 IPO 入场倍数约为 15–25× 前瞻 ARR;Cognition 若以 $10.2B 和约 $73M ARR 为基准,需要在标准倍数下达到 $400–680M ARR,IPO 才能守住 $10.2B 入场估值。没有公开市场可比对象:GitHub Copilot 嵌在 Microsoft Azure 收入中,Amazon Q 嵌在 AWS 中,Claude Code 嵌在 Anthropic API 收入中。私募市场相对公开可比的溢价历史上为 20–40%;这意味着在基准情景假设下,Cognition 的公允价值区间为 $3–7B,低于上一轮 $10.2B,说明当前轮次价格较内在价值带有 1.5–3× 投机溢价。合适的投资人回应是保持入场纪律——等待 ARR 在 $300M+ 得到确认,或估值修正到 $4–6B。

可比估值表
公司估值ARR / 收入ARR 倍数阶段与 Cognition 的可比性
Cursor (Anysphere)$9.9B(Nov 2025 Series D)~$500M ARR(2025 估计)~20× ARR后期私有公司,IPO 前最接近可比公司:AI 编程工具,企业客户,聚焦开发者生产力;ARR 倍数低 7×
Glean$4.6B(2024)~$100M ARR(估计)~46× ARR后期私有公司企业 AI 搜索;高 NRR(≥150%)支撑溢价;Cognition NRR 未知
Harvey AI$3B(2025)~$100M ARR(估计)~30× ARR成长期垂直 AI SaaS;聚焦法律;企业 NRR 强;TAM 小于 Cognition
Cohere$5.5B(2024)~$100M ARR(估计)~55× ARR后期私有公司基础模型提供商;基础设施层;受基准商品化影响较小
Replit$1.16B(2023)低于 $50M ARR(估计)~25× ARR成长期品类风险先例:曾向 AI 原生新进入者丢失份额的开发者工具;关注是否重演
GitHub CopilotN/A(嵌入 Microsoft)~$200M+ ARR(估计)N/A上市公司(Microsoft)最强分发护城河(250M+ GitHub 用户);价格 $10–$39/user/month;将 Devin 的 $500/month 价位商品化
Amazon Q DeveloperN/A(嵌入 AWS)~$50M ARR(估计)N/A上市公司(Amazon)AWS 捆绑定价;与 Devin 的企业云部署竞争相关

所有非公开市场估值均基于已披露或据报道的融资轮。私营公司的收入倍数为分析师估计;实际指标未经公司确认。

[CV011, CV012, CV013, CV014, CV015, CV016]
FV004: 投资质量 KPI:Cognition 与阈值对比
[CV001, CV030, CV031, CV032, CV033]

8.5 退出准备度与最终尽调清单

以当前 ARR 轨迹和公开市场要求的披露缺口看,Cognition 距离 IPO 准备就绪仍有 2–4 年。若要以 $8B+ 市值 IPO,公司需要约 $400–500M ARR、正向单位经济(40%+ 毛利率)、披露 NRR 高于 120%,以及清晰的 SOC 2 Type II + GDPR 合规基础设施。M&A 退出路径存在:Microsoft(GitHub Copilot/Azure 定位)、Alphabet(Google Cloud 开发者工具)、Salesforce(Einstein 平台内的开发者生产力)或 Snowflake/Databricks(数据工程自动化)都是合理收购方。$10.2B 估值让战略收购路径变难:Microsoft 此前 2018 年以 $7.5B 收购 GitHub 时,GitHub 已有规模;若以 $10.2B 收购 Cognition,将跻身 Microsoft 最大 AI 收购之一,需要董事会级别信念。最可能的退出路径仍是后期私募股权 / pre-IPO 跨轮资金参与;如果 ARR 轨迹持续,IPO 窗口在 2028–2030 年。投资前最终尽调优先事项:(1) 按队列披露的 NRR——高于 110% 是绿灯,低于 90% 会打破投资逻辑;(2) 规模化毛利率——目标 50%+ 毛利率;(3) Devin-2.0 后审计的渗透测试报告;(4) Windsurf 集成路线图,以及被收购客户基础的留存数据;(5) LLM 提供商合同条款,包括智能体用例定价底线;(6) GPAI 合规计划和欧盟监管时间表;(7) 前五大客户 ARR 集中度(前五大超过 50% 是实质集中风险);(8) 按职能划分的员工数,用于评估烧钱速度相对 ARR 覆盖率。打破投资逻辑的条件包括:12 个月内第二次披露安全事件、NRR 披露低于 90%、ARR 增速降至 <50% YoY,或确认重大企业客户流失。

投资逻辑破坏与止损触发表
触发事件阈值监测指标行动含义
第二起安全事件或供应链泄露Dec 2024 补丁后 18 个月内,任何已确认的第二起 Devin 安全事件CVE 披露;企业安全公告;HN / 安全社区立即复盘投资逻辑;评估修复可信度;可能退出
重大客户流失最大客户 ARR 同比下降 >25%,或丢失任何 ARR >$10M 的单一客户融资时 ARR 披露;新闻公告;客户访谈减仓;用流失情景重建 ARR 轨迹模型
能力相当时,竞争对手价格比 Devin 等价方案低 3× 以上竞争对手 SWE-bench Verified ≥80%,且价格 ≤$160/month(Devin 等价)SWE-bench 排行榜;每月查看竞争对手定价页审查结构性差异化;加快 Windsurf 护城河评估
披露 NRR 低于 90%任何已披露或估计的队列测算中,NRR <90%下一轮融资财务披露;资料室重建 ARR 增长模型;规模化后的高流失会破坏 $10B+ 入场逻辑
因 GPAI 不合规遭遇欧盟监管执法EU AI Office 对 Cognition 启动正式调查或发出合规命令EU AI Office 公告;DPC 新闻稿;Reuters/FT 监管报道评估欧盟收入风险;建模执法成本;可能触发更大范围的企业采购暂停

止损标准不是法律阈值,而是供监测委员会使用的投资风险管理指标。

[CV018, CV019, CV020, CV021, CV022]

免责声明

本报告是基于公开证据的尽调快照,不构成投资建议。重要的财务、法律、技术和合同事实仍未公开;作出任何投资决定前,应向管理层及一手文件直接核验。

证据索引

结论
编号陈述可信度来源
CO001 Cognition AI was founded in November 2023 in San Francisco, California. SO002, SO005, SO010
CO002 Cognition AI is headquartered in San Francisco, California. SO001, SO005
CO003 Cognition AI's flagship product, Devin, is an autonomous AI software engineer capable of interpreting tickets, planning, writing code, debugging, testing, and deploying software with minimal human oversight. SO002, SO025
CO004 Scott Wu is the co-founder and CEO of Cognition AI. SO001, SO005, SO010, SO011
CO005 Steven Hao is the co-founder and CTO of Cognition AI. SO001, SO005, SO010
CO006 Walden Yan is the co-founder and CPO (Chief Product Officer) of Cognition AI. SO001, SO005, SO010
CO007 Scott Wu is a three-time IOI (International Olympiad in Informatics) gold medalist. SO011, SO010
CO008 Cognition AI raised $175M at a $2B valuation in April 2024, led by Founders Fund. SO006, SO007, SO008
CO009 Cognition AI raised $400M at a post-money valuation of $10.2B in September 2025, led by Founders Fund. SO006, SO007, SO008, SO009
CO010 Cognition AI's total capital raised is approximately $696M as of September 2025. SO008, SO014
CO011 Cognition AI acquired Windsurf, an AI-native IDE company, in July 2025. SO004, SO009, SO015
CO012 Windsurf had approximately $82M ARR at the time of its acquisition by Cognition AI. SO003, SO008, SO009
CO013 Windsurf had more than 350 enterprise customers at the time of its acquisition by Cognition AI. SO004, SO009
CO014 Cognition AI's ARR grew from $1M in September 2024 to $73M in June 2025, representing approximately 73x growth in nine months. SO008, SO024
CO015 Following the Windsurf acquisition, Cognition AI's combined ARR reached approximately $155M in July 2025. SO003, SO008
CO016 Cognition AI's net cash burn was under $20M from founding through Q3 2025, indicating high capital efficiency. SO003, SO014
CO017 Devin was publicly announced in March 2024. SO002, SO005
CO018 Cognition AI initially explored cryptocurrency before pivoting to building autonomous AI coding agents. SO005, SO010
CO019 The founding team of approximately 10 people collectively held 10 IOI gold medals at company launch. SO002, SO011
CO020 Founders Fund led both the initial seed/Series A and the $400M Series B round for Cognition AI. SO006, SO007, SO008
CO021 Goldman Sachs is among Cognition AI's enterprise customers and is reportedly running a large-scale Devin pilot with approximately 12,000 developers. SO012, SO003
CO022 Cognition AI laid off 30 former Windsurf employees following the July 2025 acquisition. SO013, SO016
CO023 Cognition AI offered nine-month salary buyout packages to approximately 200 remaining Windsurf employees who did not want to commit to the company's 80-hour, six-day workweek culture. SO013, SO016, SO017
CO024 Cognition AI's enterprise ARR grew more than 30% in seven weeks following the Windsurf acquisition. SO003
CO025 Cognition AI opened its Singapore APAC headquarters in April 2026. SO001, SO003
CO026 Scott Wu studied economics at Harvard University. SO011, SO010
CO027 Scott Wu previously co-founded Lunchclub, an AI-powered professional networking platform. SO011, SO010
CO028 Devin's Core plan is priced at $20/month for individual developers with pay-as-you-go compute at approximately $2.25 per ACU. SO026, SO025
CO029 Devin's Team plan costs $500/month and includes 250 Agent Compute Units plus full API access. SO025, SO026
CO030 Cognition AI's $175M Series A was conducted at a $2B post-money valuation in April 2024. SO006, SO008
CO031 Cognition AI raised approximately $21M in March 2024 at a valuation of approximately $350M in its initial institutional round. SO008, SO021
CO032 Scott Wu holds the 'Legendary Grandmaster' designation on the Codeforces competitive programming platform. SO011, SO010
CO033 Mercedes-Benz announced a partnership with Cognition AI in April 2026. SO001, SO025
CO034 Independent product reviewers have found that Devin's real-world task success rates require significant human oversight on complex tasks, suggesting performance falls below the 'fully autonomous' marketing framing. SO018, SO019
CO035 CEO Scott Wu has stated publicly that Cognition AI's culture requires six-day workweeks of more than 80 hours, framing it as a company-wide commitment to the mission. SO013, SO016, SO017
CO036 Investors in the September 2025 $400M round included Lux Capital, 8VC, Elad Gil, Definition Capital, Swish Ventures, Bain Capital Ventures, and D1 Capital. SO006, SO009, SO014
CO037 As of April 2026, Cognition AI is reportedly in discussions for a new financing round that could value the company at over $25 billion. SO008
CO038 Walden Yan is an IOI gold medalist who leads product management and the product roadmap at Cognition AI. SO010, SO005
CO039 Steven Hao previously held roles at Scale AI, Cursor, Modal, DeepMind, Waymo, and Nuro before co-founding Cognition AI. SO010, SO011
CO040 Cognition AI had approximately 49 employees before the Windsurf acquisition according to revenue-per-employee analyses. SO008
CO041 Windsurf had been in acquisition discussions with OpenAI for approximately $3B and was the subject of a Google talent licensing deal worth approximately $2.4B before Cognition acquired the remaining team and IP. SO015, SO009
CO042 Cognition AI's price-to-ARR multiple at the September 2025 valuation was approximately 65x (using $155M combined ARR) to 140x (using $73M pre-acquisition Devin ARR). SO023, SO008
CO043 Devin operates within a sandboxed environment that includes a shell, code editor, and browser, using long-context model inference to sequence planning, implementation, and iterative testing steps autonomously. SO002, SO025
CM001 The broad AI code tools market was valued at $7.37 billion in 2025, according to Mordor Intelligence, and is forecast to reach $23.97 billion by 2030 at a 26.6% CAGR. SM001
CM002 Grand View Research forecasts the narrow generative AI coding assistants market at $92.5 million by 2030 at a 24.8% CAGR—a significantly smaller scope than Mordor's $24B figure. SM002
CM003 ResearchAndMarkets published a $97.9 billion by 2030 forecast for the generative AI coding assistants space, though this figure likely uses an atypically broad scope including non-comparable segments. SM003
CM004 No single analyst consensus exists for the AI coding tools TAM: estimates range from $92M (Grand View, narrow) to $97.9B (ResearchAndMarkets, broad) for 2030, with the most operationally relevant SAM estimate at $1–8B. SM001, SM002, SM003
CM005 The estimated serviceable addressable market for premium agentic developer tools (autonomous tier) is $1–2B in 2025–2026, growing to $5–8B by 2030 based on developer-seat and enterprise-spend lenses. SM006, SM007, SM008
CM006 Global software spending reached approximately $675–$700 billion in 2024, according to WIPO's Global Innovation Index and multiple market research sources. SM008, SM007
CM007 The number of professional software developers worldwide in 2024 is estimated at 27–28.7 million by Evans Data, Statista, and JetBrains Research. SM006, SM007, SM020
CM008 Gartner forecasts that 90% of enterprise software engineers will use AI code assistants by 2028, up from less than 14% in early 2024. SM004, SM005, SM009
CM009 GitHub Copilot holds approximately 42% market share among paid AI coding tools in 2025, with $2 billion in ARR and over 20 million all-time users. SM017, SM018, SM019
CM010 GitHub Copilot is used by over 50,000 organizations including 90% of Fortune 100 companies in 2025, with enterprise adoption up 75% quarter-over-quarter. SM017, SM019
CM011 Developers using AI coding assistants complete code tasks 51–55% faster, with AI-generated code constituting 46% of all code written by Copilot users on GitHub. SM017, SM018
CM012 Enterprise spend on agentic AI systems is projected to surge from less than $1 billion in 2024 to over $51 billion by 2028 at approximately 150% CAGR. SM023, SM024
CM013 Enterprise buyers universally require sandbox environments, audit logs, and human-in-the-loop review gates before deploying autonomous coding agents in production workflows. SM013, SM014, SM015
CM014 The biggest barriers to enterprise adoption of agentic AI include trust deficits from hallucination risk, legacy system integration complexity, and unclear ROI measurement frameworks. SM013, SM014, SM015, SM016
CM015 65% of enterprises were regularly using generative AI as of 2025, up from 33% in 2023, indicating rapid mainstream adoption of AI tools in enterprise workflows. SM012, SM023
CM016 Only a minority of AI use cases in enterprises have reached full production in 2025; primary barriers are legacy systems, data silos, compliance concerns, and talent shortages. SM024, SM013
CM017 72% of enterprises plan to deploy AI copilot and agent technologies by 2026, according to aggregated industry surveys. SM023
CM018 The primary enterprise buyer for autonomous software engineering agents is VP Engineering or CTO, with budget sourced from developer productivity or digital transformation allocations. SM012, SM024
CM019 Goldman Sachs (12,000-developer pilot), Citi, Cisco, Dell, Nubank, and Ramp are among Cognition AI's enterprise reference customers, representing the financial services and high-growth tech segments. SM011, SM023
CM020 The EU AI Act and evolving US AI governance regulations impose documentation and explainability requirements that will affect enterprise deployment of autonomous coding agents beginning 2026–2027. SM013, SM015
CM021 Status-quo substitutes for autonomous AI coding agents include offshore staff augmentation, traditional IDEs, internal developer tooling built on base LLM APIs, and cloud-vendor bundled IDE extensions. SM001, SM011
CM022 Switching costs for enterprise AI coding tools include integration migration effort, process overhaul, vendor lock-in risks, and cultural reorientation, making incumbent vendor stickiness a meaningful competitive moat. SM015, SM016
CM023 North America leads the AI coding tools market with the largest share; cloud deployment dominates though enterprise demand for on-premises solutions is growing, particularly among regulated industries. SM001, SM002
CM024 The autonomous agents market (cross-sector) is forecast to grow from $4.35B in 2025 to $103B by 2034 at over 40% CAGR, providing a directionally supportive upper-bound context for the AI software engineering agent TAM. SM015, SM025
CM025 Gartner's Magic Quadrant for AI Code Assistants provides enterprise procurement legitimacy for the category, with GitHub Copilot named a Leader in both 2024 and 2025 editions. SM004, SM010
CM026 87% of developers are using AI tools daily as of 2025, indicating that AI-assisted development has reached mainstream adoption among professional software engineers. SM011, SM023
CM027 Developers using AI tools report 3.2× productivity improvements on average, with 67% using AI tools at least five days per week according to 2025 surveys. SM011, SM017
CM028 Financial services enterprises present a high-value segment for Cognition AI given Goldman Sachs' 12,000-developer Devin pilot; these buyers require SOC 2 Type II, data residency, and regulatory compliance before enterprise sign-off. SM013, SM024
CM029 The global software development market is valued at $675–$730B in 2024; a 1% shift of project labor costs to AI tools implies a $6.7B+ market, corroborating the Mordor $24B broad TAM estimate for 2030. SM008, SM021
CM030 AI-assisted pull request cycle times have been reduced by up to 75% (from 9.6 days to 2.4 days) for organizations using GitHub Copilot, per reported enterprise metrics. SM017, SM018
CM031 Unpredictable agent behavior including hallucinations, lack of transparency, and security concerns are the primary trust barriers slowing enterprise deployment of autonomous AI coding agents. SM014, SM016
CM032 77% of companies are currently using or actively testing AI tools, with the global AI market estimated at approximately $298 billion in 2025. SM023
CM033 Agentic AI adoption in software engineering is constrained not just by technology but by organizational readiness: workforce resistance, lack of AI talent, and unclear business cases are the dominant enterprise blockers as of 2025. SM013, SM015, SM024
CM034 Competitive fragmentation in AI coding tools—GitHub Copilot (42% market share), Cursor ($9B valuation), Windsurf (acquired by Cognition), and OpenAI Codex—can produce buyer choice overload that stalls enterprise procurement decisions. SM011, SM017
CM035 Cognition AI's $155M ARR as of July 2025 represents approximately 2% penetration of the estimated $7.37B AI code tools TAM, suggesting material white space remains if the market definition is correct. SM001
CP001 GitHub Copilot holds approximately 42% of the AI coding assistant market share as of 2025. SP005, SP017
CP002 GitHub Copilot had over 20 million active users as of early 2025. SP005, SP012
CP003 GitHub Copilot exceeded $2 billion in annual recurring revenue by early 2025. SP002, SP004
CP004 GitHub Copilot is used by 90% of Fortune 100 companies. SP012, SP021
CP005 Cursor (Anysphere) reached $2 billion ARR by February 2026. SP001, SP003, SP004
CP006 Cursor raised a $2.3 billion Series D in November 2025 at a $29.3 billion valuation. SP001, SP003
CP007 Cursor has over 1 million paying customers and 50,000 enterprise teams. SP003, SP006
CP008 Cursor's multi-agent background agents feature runs up to eight parallel task threads simultaneously. SP005, SP016
CP009 Amazon Q Developer offers a Free tier and a Pro tier priced at $19 per user per month. SP013, SP018
CP010 Amazon Q Developer achieved SOC 2 Type II certification with VPC isolation support. SP007, SP013
CP011 Amazon Q Developer has a 200,000 token context window. SP007, SP018
CP012 Claude Code operates as a terminal-native agentic coding assistant without requiring a custom IDE. SP008, SP009
CP013 Devin's initial SWE-bench score in March 2024 was 13.86%, a breakthrough at the time of announcement. SP010, SP015, SP025
CP014 Subsequent model improvements from Anthropic pushed Claude-based agents above 50% on SWE-bench Verified. SP010, SP015
CP015 Claude Code is priced at $10 per month on Pro and $100 per month on Max tiers. SP027, SP008
CP016 OpenAI relaunched Codex as a web-based agentic coding environment backed by the o3 model family in 2025. SP014, SP011
CP017 SWE-agent is an open-source research framework developed at Princeton NLP Group, not a commercial product. SP025, SP010
CP018 Cursor investors include Thrive Capital, Andreessen Horowitz, Accel, Nvidia, and Google. SP001, SP003
CP019 GitHub Copilot individual pricing is $10 per month; Business tier is $19 per month; Enterprise tier is $39 per month. SP021, SP012
CP020 Windsurf (formerly Codeium) was acquired by Cognition in July 2025, adding approximately $82M ARR and 350-plus enterprise customers. SP024, SP006
CP021 Cognition's acquisition of Windsurf converted a direct IDE-tier competitor into a distribution channel for Devin. SP024, SP019
CP022 GitHub Copilot expanded from inline code completions to Copilot Workspace, a multi-step planning and execution agent. SP012, SP005
CP023 Cursor pricing is $20 per month for Pro tier and $40 per month for Business tier as of 2025. SP023, SP005
CP024 Replit offers an AI-native development platform with cloud execution environments targeting student and hobbyist developers. SP022
CP025 The core architectural gap between Devin and Claude Code is that Claude Code still requires a human operator at the terminal, while Devin runs fully unattended. SP009, SP008
CP026 Cognition's defensible competitive advantages include execution depth, enterprise data flywheel from completed tasks, and Windsurf distribution. SP019, SP011
CP027 Benchmark credibility concerns about Devin emerged after independent evaluations showed other tools surpassing it within months of the March 2024 announcement. SP010, SP015, SP025
CP028 The competitive window for pure agentic differentiation is narrowing as frontier model labs deploy equivalent agentic pipelines. SP011, SP019
CP029 Cursor is growing faster than Cognition on ARR metrics and has more paying customers, providing proportionally more training signal. SP001, SP003, SP004
CP030 Cursor uses a VS Code fork architecture that provides deep editor integration unavailable to plugin-based competitors. SP005, SP016, SP023
CP031 Amazon Q Developer's native AWS service integrations position it as preferred for organizations standardized on AWS infrastructure. SP007, SP013
CP032 GitHub Copilot's enterprise distribution advantages create high switching-cost inertia across Fortune 100 organizations. SP012, SP021
CP033 The AI coding tools market in 2026 is stratified into hyperscaler-backed co-pilots, IDE-native assistants, and fully autonomous agent tiers. SP011, SP017, SP019
CP034 Frontier model labs such as Anthropic and OpenAI control the underlying reasoning capabilities and can deploy equivalent agentic pipelines without licensing dependencies. SP008, SP014
CP035 Cognition's enterprise go-to-market execution and customer retention metrics are the true leading indicators of sustainable competitive position, more so than benchmark scores. SP009, SP011
CP036 Cursor's growth from zero to $2B ARR in approximately 24 months represents the fastest known ARR scale-up in AI coding tools. SP001, SP003
CP037 Both GitHub Copilot and Amazon Q Developer lack end-to-end autonomous task completion without human-in-the-loop confirmation at each stage. SP007, SP012
CI001 Cognition AI's Core plan is priced at $20 per month for 5 ACUs; Team plan at $500 per month for 250 ACUs. SI004, SI022
CI002 Enterprise ACU overages are priced at approximately $2.25 per ACU beyond the plan inclusion. SI004, SI001
CI003 Cognition AI reached approximately $1M ARR in September 2024. SI001, SI005
CI004 Cognition AI grew from $1M to $73M ARR between September 2024 and June 2025—a 73x increase in nine months. SI001, SI003, SI005
CI005 Windsurf (formerly Codeium) was priced at $0 (Free) and approximately $15 per month (Pro) before the Cognition acquisition. SI006, SI007
CI006 Cognition AI's gross margin is estimated at 50–70%, based on inference compute costs versus ACU pricing—no primary disclosure exists. SI011, SI005
CI007 The Windsurf acquisition added approximately $82M ARR and 350-plus enterprise customers to Cognition's combined revenue base. SI006, SI010, SI012
CI008 Cognition AI's combined ARR reached approximately $155M post-Windsurf acquisition in July 2025. SI009, SI010, SI013
CI009 ARR per employee at Cognition was approximately $1.5M prior to the Windsurf acquisition (49 employees, $73M ARR in June 2025). SI003, SI011
CI010 ARR per employee dropped to approximately $623K post-Windsurf, reflecting the quadrupling of headcount to ~249 employees. SI019, SI003, SI011
CI011 Net revenue retention rate for Cognition AI is not publicly disclosed; analyst estimates range from 100–130% based on ARR growth trajectory. SI005, SI011
CI012 Cognition AI's seed/Series A in March 2024 raised approximately $21M at a $350M valuation, led by Founders Fund. SI002, SI023, SI020
CI013 Cognition AI's April 2024 funding round raised $175M at a $2B post-money valuation, led by Founders Fund. SI002, SI014, SI020
CI014 Cognition AI's September 2025 Series B raised $400M at a $10.2B post-money valuation with Lux Capital, 8VC, Elad Gil, Bain Capital Ventures, and D1 Capital as co-investors. SI008, SI009, SI010
CI015 Cognition AI disclosed a net cash burn of under $20M from founding through Q3 2025—extraordinary capital efficiency for a company at this ARR growth rate. SI001, SI009
CI016 Reports as of April 2026 indicate Cognition AI is in discussions for new financing at a $25B+ valuation. SI011, SI012
CI017 Cognition AI does not disclose gross margin, customer churn, NRR, CAC, or GAAP financial statements. SI005, SI007, SI016
CI018 All ARR figures for Cognition AI rely on company blog posts, third-party news citing unnamed sources, or secondary market research—no audited financial data is available. SI003, SI005, SI016
CI019 The Windsurf acquisition was estimated at approximately $250M, funded from Cognition's existing capital reserves without a new primary fundraise. SI006, SI010
CI020 Founders Fund led both the March 2024 seed/Series A and the April 2024 Series A extension, providing deep capital commitment and reputational signal. SI002, SI008, SI020
CI021 A $25B+ valuation on $155M ARR implies a forward revenue multiple of approximately 160x, elevated even by hypergrowth AI company standards. SI011, SI012
CI022 Cognition AI's total capital raised through Series B is approximately $596M excluding the Windsurf acquisition; total capital deployed is approximately $846M when including the acquisition. SI008, SI009, SI010
CI023 Post-Windsurf headcount reportedly reached approximately 249 employees before layoffs of 30 and additional voluntary departures from buyout offers. SI019, SI023
CI024 SEC Form D/A for Founders Fund IX, LP (CIK 0001971631, filed October 2025) discloses total capital raised of approximately $972M across the LP and principals fund, with Peter Thiel as managing member. SI026
CI025 Founders Fund IX Principals Fund (CIK 0002090410) is a co-issuer with Founders Fund IX, LP per the amended SEC Form D/A, indicating a typical GP co-investment vehicle structure. SI026
CI026 Cognition AI opened its Singapore APAC headquarters in April 2026, hiring Richard Spence as VP and General Manager APAC, targeting Southeast Asia, Australia, India, and South Korea. SI027, SI018
CI027 Mercedes-Benz announced deployment of Devin and Windsurf across its global engineering organization in April 2026, focusing on legacy modernization, cloud-native development, and logistics applications. SI028, SI018
CI028 Cognition AI introduced a revised pricing structure in April 2026, replacing the Core and Team plans with Free, Pro, Max, Teams, and Enterprise tiers, signaling a land-and-expand monetization strategy. SI018, SI004
CI029 Cognition AI's annual Team plan pricing of $6,000 per developer is approximately 15x more expensive than GitHub Copilot Enterprise ($390/yr) and 25x more than Cursor Pro ($240/yr), justified by task-completion vs. code-suggestion value. SI004, SI011
CI030 Total capital raised ($596M through Series B) relative to ARR at time of last raise ($73M Devin ARR at Series B in Sep 2025) implies a capital-to-ARR ratio of approximately 8x—high but declining as ARR compounds. SI008, SI009, SI003
CI031 The 9-month severance buyout offered to approximately 200 Windsurf employees represents an estimated one-time integration cost of $15–30M, calculated on average startup engineer salaries of $100–150K. SI019
CI032 Cognition AI's April 2026 APAC office opening and Mercedes-Benz partnership signal material increases to operating expenditure in H1 2026 through regional headcount and GTM investment. SI027, SI028
CI033 Enterprise AI SaaS tools with variable compute cost structures typically achieve gross margins of 55–75% at scale, per industry benchmarks; Cognition's agentic model adds inference cost variance absent from seat-based tools. SI011, SI015
CI034 Cognition AI has not filed Form D or equivalent exempt offering notice with the SEC as of May 2026; the company's capital raises are documented only through investor disclosures and third-party press reports. SI016, SI023, SI026
CI035 If the Windsurf $82M ARR churns by 30% post-acquisition migration, combined Cognition ARR falls to approximately $128M—a scenario not publicly addressed by company management. SI006, SI003, SI011
CE001 Devin is a cloud-hosted autonomous AI software engineer that accepts natural-language task specifications and executes complete software engineering tasks—planning, coding, testing, and deploying—without human intervention during execution. SE002, SE004
CE002 In the original SWE-bench evaluation (March 2024), Devin resolved 13.86% of issues unassisted—far exceeding the prior state-of-the-art of 1.96% for unassisted agents and 4.80% for assisted LLMs. SE003, SE006, SE020
CE003 Devin's agent runtime executes tasks within a sandboxed cloud environment equipped with a code editor, Unix shell, and headless browser—isolated from other sessions and the customer's production environment. SE002, SE001
CE004 Each Devin session is bounded by a maximum runtime of 45 minutes; longer tasks must be broken into sequential sub-sessions. SE003, SE001
CE005 Devin integrates natively with GitHub, GitLab, and Bitbucket for source control, enabling automated PR creation, branch management, and code contributions. SE007, SE001
CE006 Devin integrates natively with Jira and Linear for project management ticket assignment and sprint delegation. SE007, SE001
CE007 Devin integrates natively with Slack and Microsoft Teams for session initiation via chat messages and progress reporting. SE007, SE001
CE008 Agent Compute Units (ACUs) are the proprietary metered compute token consumed per Devin task, priced at approximately $2.25 per ACU with plan-included allocations varying by tier. SE004, SE008
CE009 Devin Enterprise tier offers VPC deployment allowing the agent to operate within customer-controlled cloud infrastructure, and custom-trained Devin instances tuned to proprietary codebases. SE001, SE004
CE010 Devin accepts task inputs from multiple channels: web dashboard, Slack messages, Jira tickets, and Linear issues—translating natural-language instructions into executable multi-step plans. SE001, SE002
CE011 Nubank achieved 12x engineering efficiency improvement and 20x cost savings using Devin to migrate a 6-million-line ETL monolith; Data, Collections, and Risk business units completed migrations in weeks instead of months. SE016, SE014
CE012 SWE-Check, released April 2026 in collaboration with Applied Compute, is a reinforcement-learning-trained bug detection model that matches Anthropic Opus 4.6 performance while running approximately 10x faster. SE024, SE002
CE013 The DeepWiki server builds vectorized project graphs of entire codebases, enabling Devin to navigate and reason about multi-million-line repositories beyond standard LLM context window limits. SE001, SE002
CE014 The Windsurf acquisition (July 2025) added 350+ enterprise customers, 250,000+ daily active users, and the Windsurf IDE product including its Cascade agentic workflow engine. SE015, SE022
CE015 In late 2024, a live-streamed demonstration publicly exposed a major security vulnerability in Devin's system prompt handling, potentially enabling prompt-injection attacks. Cognition acknowledged and patched the issue. SE017, SE009
CE016 Devin 2.0 (April 2025) achieved 83% more tasks completed per ACU compared to Devin v1.x, representing a significant efficiency improvement for existing subscribers. SE008, SE019
CE017 Devin 2.0 introduced a confidence meter that quantifies the probability of task success before and during execution, allowing teams to filter low-probability tasks and prioritize high-value delegations. SE008, SE001
CE018 Model Context Protocol (MCP) integration allows Devin to connect to hundreds of external tools including monitoring platforms, databases, and documentation systems through the MCP Marketplace. SE007, SE001
CE019 The Devin REST API enables programmatic session creation, status retrieval, and CI/CD pipeline integration for DevOps automation workflows. SE007, SE001
CE020 A 32-week empirical study of 7,156 real-world pull requests showed Devin's PR acceptance rate increased by +0.77% per week—the only agent in the study with a consistent positive trend. SE011, SE010
CE021 BlockDiff snapshotting records incremental state checkpoints during Devin's execution, enabling rapid rollback to a known-good state when tests fail or an approach proves unworkable. SE001, SE002
CE022 Devin supports parallel task execution across multiple concurrent sessions, allowing enterprise teams to delegate entire sprint backlogs simultaneously rather than one task at a time. SE001, SE004
CE023 Claude Code (Opus 4) achieved 72.5% on SWE-bench Verified (500 instances) in 2025, far exceeding Devin's original 13.86% benchmark score—though the two benchmarks use different task sets. SE010, SE012
CE024 mini-SWE-agent achieved 65% on SWE-bench Verified in July 2025, matching most commercial agents at a fraction of the computational cost—a signal that benchmark performance is rapidly commoditizing. SE006, SE020
CE025 Devin for Terminal (2026) allows developers to start an agent session locally from the terminal and escalate to cloud execution when the task outgrows local compute—preserving session state across the transition. SE024, SE001
CE026 The CognitionAI/devin-swebench-results GitHub repository has 124 stars and 20 forks as of research date, indicating modest developer interest in the benchmark methodology transparency. SE005
CE027 Cognition AI's 2026 product roadmap includes Devin-in-Windsurf integration, SWE-1.5 model improvements, SWE-Check bug detection, APAC regional expansion (Singapore), and Mercedes-Benz enterprise deployment. SE024, SE015
CE028 Cognition AI has not publicly disclosed SOC 2 Type II, ISO 27001, or equivalent compliance certifications as of May 2026, representing a gap for regulated-industry enterprise procurement. SE001, SE004, SE021
CE029 All Devin sessions run in isolated ephemeral compute containers, and code changes are surfaced only as pull requests requiring human approval before merging—preventing unauthorized production deployments. SE001, SE004
CE030 The MCP Marketplace integrates Devin with monitoring tools (Sentry, Datadog, PagerDuty), databases (PostgreSQL, MySQL, MongoDB), and documentation (Notion, Confluence), enabling context-rich agent execution. SE007, SE001
CE031 Enterprise VPC deployment allows Devin to execute tasks within the customer's own AWS, GCP, or Azure environment, eliminating source-code transmission to Cognition's external servers. SE001, SE004
CE032 Windsurf Codemaps, powered by SWE-1.5 and Claude Sonnet 4.5, provides AI-annotated structured maps of codebases with 'trace guide' expansions, enabling navigation to exact file and line references. SE024, SE013
CE033 The Devin REST API uses standardized HTTP endpoints for session creation, status polling, result retrieval, and webhook-based notifications, enabling integration into CI/CD pipelines. SE007, SE001
CE034 GitHub Copilot was ranked a Leader in Gartner's 2025 Magic Quadrant for AI Code Assistants, representing the current benchmark for enterprise adoption against which Devin competes. SE010, SE012
CE035 A 2026 empirical study of AI coding PRs across 7,156 tasks found documentation tasks achieved 82.1% acceptance versus 66.1% for new feature development—a 16-percentage-point gap exceeding inter-agent variance. SE011, SE018
CU001 Cognition AI targets engineering teams at technology-forward enterprises, midmarket companies, and developer-first organizations, with buyer profiles including CTOs and VPs of Engineering seeking developer throughput multiplication. SU001, SU021
CU002 Confirmed enterprise customer verticals include fintech/financial services (Nubank), automotive/manufacturing (Mercedes-Benz), IT services/outsourcing (Cognizant), US Federal Government (Feb 2026), and COBOL-heavy Fortune 500 legacy sectors. SU002, SU003, SU004, SU006
CU003 Mercedes-Benz was announced in April 2026 as a production customer deploying Devin and Windsurf across its global engineering organization for legacy modernization, cloud-native development, and logistics. SU003
CU004 Cognizant partnered with Cognition in January 2026 to deploy Devin and Windsurf across its own engineering organization and its global client base—a channel partnership that could expose Cognition to Cognizant's 300+ clients. SU004, SU014
CU005 Cognition launched a Government vertical (Cognition for Government) in February 2026 to target US federal agencies for critical infrastructure software modernization; no named agencies have been publicly disclosed. SU006
CU006 Cognition AI's estimated ARR grew from approximately $1M at general availability (December 2024) to approximately $73M by April 2025—a roughly 70x increase in approximately four months, coinciding with the price cut from $500/month to $20/month. SU007, SU008
CU007 Devin's PR merge rate (fraction of Devin-opened PRs accepted by human reviewers) improved from 34% at launch (March 2024) to 67% by April 2025—a doubling of accepted output quality in approximately 12 months. SU008, SU009
CU008 By February 2026, Cognition's own engineering team was merging 659 Devin-authored pull requests per week—a 4x increase from their best week of 154 PRs in 2025. SU017, SU025
CU009 Independent reviewers note that real-world ACU consumption on the $20/month Core plan can easily result in monthly bills of $100-200 for moderate usage, contradicting the headline price reduction's accessibility promise. SU008, SU009
CU010 Windsurf IDE had 250,000+ daily active users and 350+ enterprise customers at the time of acquisition (July 2025), providing Cognition with a large top-of-funnel developer install base for Devin conversion. SU010, SU014
CU011 Nubank used a fine-tuned custom Devin instance for ETL migration; fine-tuning doubled task completion scores and reduced per-task time from 40 minutes to 10 minutes (4x speed improvement). SU002
CU012 Cognition opened offices in London (January 2026), Tokyo (April 2026), and Singapore (April 2026) within a four-month period, signaling management's confidence in revenue sufficiency for global expansion. SU005, SU006
CU013 Net Revenue Retention (NRR), Gross Revenue Retention (GRR), and customer churn rate have not been publicly disclosed by Cognition AI; retention durability cannot be independently assessed. SU007, SU008
CU014 Hacker News developer community expressed skepticism about Devin's real-world performance versus demo quality; early reviews noted that complex or novel tasks frequently required significant rework, exceeding expectations set by the SWE-bench launch announcement. SU011, SU023
CU015 The security vulnerability disclosed in December 2024—a live-streamed prompt-injection attack on Devin—temporarily damaged trust among security-conscious enterprise buyers, requiring Cognition to accelerate security remediation communication. SU018, SU011
CU016 Nubank's use of Devin expanded from the initial Data unit pilot to Data, Collections, and Risk business units, confirming land-and-expand progression within a single named enterprise customer. SU002
CU017 Cognition AI's land-and-expand strategy relies on starting with a discrete project (migration, modernization) within one business unit, demonstrating ROI, then expanding to adjacent teams—as demonstrated by the Nubank case. SU002, SU004
CU018 Windsurf's 250,000+ daily active users serve as a top-of-funnel for Devin cloud agent adoption: Windsurf Free users experiencing AI-assisted local development are natural conversion targets for cloud-based autonomous task delegation. SU010, SU014
CU019 Devin's COBOL modernization use case is inherently project-based—once a legacy migration is complete, there is no structural recurring need unless customers contract for ongoing support or continued modernization work. SU004, SU016
CU020 Top-customer concentration is a material but undisclosed diligence gap: with ~$73M ARR and a small set of confirmed named accounts, a single Nubank-scale contract could represent more than 10% of ARR. SU007, SU008
CU021 Cognizant's reseller partnership creates a single-channel concentration risk: if Cognizant accounts for a material share of Cognition's enterprise pipeline, the loss of that relationship would disproportionately impact new account growth. SU004, SU022
CU022 The self-serve developer tier (Free and Pro plans) provides a low-ACV but high-volume customer base that can generate bottom-up enterprise procurement pressure, analogous to Slack's or Atlassian's PLG motion. SU008, SU009, SU015
CU023 Growjo estimates Cognition AI's total funding at $1.4B and current valuation at $10.2B as of September 2025, consistent with reports of a $400M fundraise following the Windsurf acquisition. SU007, SU010
CU024 Cognition AI had approximately 222 employees as of 2025 and grew headcount by approximately 102% year-over-year—a pace consistent with rapid revenue scaling but also requiring significant talent investment. SU007, SU020
CU025 The absence of SOC 2 Type II and FedRAMP compliance attestations in public documentation limits Cognition's addressable market among regulated enterprises (financial services, healthcare, US government) that require these certifications for procurement. SU016, SU021
CU026 Devin's PR merge rate of 67% implies a 33% rejection or rework rate on AI-generated pull requests, which may be acceptable for high-volume routine tasks but represents a significant quality gap versus a senior human engineer. SU008, SU019
CU027 Fortune 500 COBOL modernization deployments are confirmed via Cognition's April 2026 blog post, but no named customers, outcome metrics, or contract structures are disclosed—limiting diligence quality. SU004
CU028 Geographic expansion from US-only to Europe and APAC in Q1-Q2 2026 within four months demonstrates aggressive international GTM investment, creating both market opportunity and potential cash burn risk from multi-office overhead. SU005, SU006
CU029 The imseankim.com review noted ARR growth from $1M to $73M in nine months while also identifying that 'independent testers tell a more complicated story' compared to official benchmarks—suggesting a gap between marketing and user-level satisfaction at the margin. SU008
CU030 Devin supports concurrent multi-session task execution, allowing enterprise teams to delegate entire sprint backlogs simultaneously, which is cited as a key adoption driver for teams managing high-volume routine work. SU021, SU024
CU031 The Devin PR acceptance rate empirical study (Arxiv 2026) showed a +0.77%/week positive trend—the only agent with sustained improvement—suggesting iterative model quality gains that support long-term customer retention. SU019
CU032 Cognition AI's Government vertical launch (February 2026) positions Devin as a tool for US federal legacy software modernization, a multi-billion dollar addressable budget segment, but no government contract or agency name has been publicly disclosed. SU006
CU033 Nubank's ETL migration case study is the only published, quantified production customer proof; Mercedes-Benz, Cognizant, and Fortune 500 COBOL deployments lack independent third-party verification or outcome data. SU002, SU003, SU004
CU034 Devin Autofix Review Comments (January 2026) closes a key customer workflow loop—Devin automatically addresses pull request review feedback—substantially improving the PR review cycle productivity for enterprise teams. SU029, SU017
CU035 The Cognizant channel partnership bypasses individual enterprise sales cycles by embedding Devin into Cognizant's managed services offering—a go-to-market shortcut that scales account coverage rapidly but creates single-reseller revenue concentration risk. SU004, SU021
CR001 The EU AI Act (Regulation 2024/1689) imposes GPAI transparency and documentation obligations on providers of general-purpose AI models effective August 2026; Cognition has not publicly disclosed its GPAI registration status, model transparency documentation, or training data copyright compliance policy. SR009, SR007
CR002 Under GDPR and UK GDPR, organizations must conduct Data Protection Impact Assessments for high-risk AI processing; Devin's autonomous code writing, PR merging, and production deployment capabilities likely qualify as high-risk processing under GDPR Article 35, requiring DPIAs from EU/UK enterprise customers. SR010, SR009
CR003 California SB 1047, which would have imposed safety requirements on frontier AI developers, was vetoed by Governor Newsom in September 2024; however, successor legislation including AB 2013 remained active in the 2025–2026 California legislative session, creating ongoing monitoring obligation. SR006, SR008
CR004 The Doe v. GitHub, Microsoft, and OpenAI class action alleges that training AI coding models on public GitHub repositories without license compliance constitutes copyright infringement; Cognition has not disclosed its model training data provenance, creating analogous latent IP exposure under the same copyright theory. SR011, SR007
CR005 The January 2025 White House Executive Order revoked prior Biden-era AI safety mandates, creating a deregulatory US federal environment for AI developers; however, agency-level rules for government AI procurement—relevant to Cognition's announced Government vertical—remain under development. SR008, SR006
CR006 In December 2024, Cognition publicly disclosed a prompt-injection vulnerability in Devin discovered during a live-streamed demonstration; the vulnerability could enable an adversary to embed malicious instructions in repository content, causing Devin to exfiltrate credentials or insert backdoors in customer codebases. SR001, SR013
CR007 The OWASP Top 10 for LLM Applications lists Prompt Injection (LLM01) as the highest-severity risk for LLM-based applications; for autonomous agents with code execution, PR merge, and terminal access permissions, a successful prompt-injection exploit has a blast radius materially larger than for passive coding assistants. SR003, SR004
CR008 Cognition obtained SOC 2 Type II certification in March 2024, prior to the December 2024 prompt-injection incident; the SOC 2 report is not publicly accessible without executing an NDA through the Cognition Trust Center, limiting independent assessment of control effectiveness by enterprise procurement teams. SR001, SR002
CR009 Devin 2.0 (April 2025) introduced the ability to directly merge pull requests and schedule autonomous agent runs on customer infrastructure, materially expanding the attack surface and potential blast radius of any future prompt-injection or supply-chain compromise beyond what existed at the December 2024 incident. SR016, SR003
CR010 Cognition's security documentation acknowledges that Devin code output may contain hallucinations, bugs, or insecure code, and recommends code reviews and branch protections as mitigations; this acknowledged limitation creates ongoing liability exposure for enterprise customers deploying Devin in production code pipelines without mandatory human review gates. SR001, SR004
CR011 Cognition is structurally dependent on Anthropic and OpenAI as foundation model API providers; these same companies compete with Devin through Claude Code and OpenAI Codex, creating a supply chain where the primary suppliers have economic incentives to disadvantage a dependent customer. SR023, SR015
CR012 AWS provides the cloud infrastructure for Devin's VPC deployment option; Devin's enterprise architecture is AWS-native with no disclosed multi-cloud or on-premises fallback, creating material infrastructure concentration risk in a single cloud vendor. SR001, SR016
CR013 GitHub (Microsoft) provides the primary code repository integration for Devin's PR workflow; Microsoft also distributes GitHub Copilot as a direct Devin competitor, meaning Cognition depends on a competitor-owned platform for core product functionality. SR015, SR007
CR014 Cognition's acquisition of Windsurf (July 2025) added 350+ enterprise accounts and 250K+ daily active users but introduced integration execution risk: two AI agent codebases, model pipelines, enterprise sales motions, and billing systems must be combined simultaneously while maintaining ARR growth momentum. SR018, SR030
CR015 Cognition's three co-founders have exceptional technical credentials as competitive programmers but no prior track record building enterprise software companies beyond early-stage startups; execution risk is elevated as the company scales from startup to enterprise sales motion across multiple geographies simultaneously. SR025, SR017
CR016 AI engineering talent competition is severe; Anthropic, OpenAI, Google DeepMind, and Microsoft actively recruit competitive programming and ML talent at packages estimated above $500K total compensation, directly competing with Cognition's talent pool and creating ongoing retention risk. SR023, SR015
CR017 Cognition's international expansion to Japan, Singapore, EU, and US Government sectors requires jurisdiction-specific compliance infrastructure—EU DPO appointment, Japan PIPL compliance, FedRAMP authorization—with no publicly disclosed plans or timelines, creating a gap between announced customer segments and compliance readiness. SR009, SR010
CR018 At Devin's March 2024 launch its SWE-bench Full score of 13.86% was the then-best result for autonomous coding agents; by mid-2025 Claude Code Opus 4 achieved 72.5% on SWE-bench Verified—a competitive improvement rate that threatens to eliminate Devin's technical differentiation within 12–18 months. SR015, SR023
CR019 The FTC stated that concentrated control of foundational AI inputs—including foundation models and cloud compute—could allow providers to distort competition in downstream AI application markets; this is a direct structural risk for Cognition given its dependence on Anthropic/OpenAI (models) and AWS (compute), both of which are also competitive threats. SR007, SR011
CR020 ARR growth from ~$1M at GA (December 2024) to ~$73M estimate (April 2025) across a small number of named enterprise customers implies high per-customer ARR concentration; individual churn events at top customers could cause material ARR declines capable of impairing the $10.2B valuation estimate. SR019, SR020
CR021 Devin's ACU pricing model ($2.25 per ACU overage) creates unpredictable usage-based cost exposure for enterprise customers; 'bill shock' from unexpected ACU consumption is a cited churn risk factor in independent community analysis of early Devin adopters. SR020, SR028
CR022 Cognition has raised approximately $1.575B in disclosed funding through mid-2026; with 222 employees, 102% annual headcount growth, and compute-intensive AI infrastructure, estimated monthly burn is $5–15M, implying 18–36 months of runway under conservative scenarios assuming $73M ARR and aggressive growth continuation. SR024, SR021
CR023 LLM inference costs for Devin sessions—paid at commercial API rates to Anthropic and OpenAI—compress gross margins structurally; no path to proprietary model development or renegotiated volume pricing is publicly disclosed, leaving margin compression risk unaddressed. SR023, SR016
CR024 GitHub Copilot's distribution through Microsoft enterprise sales channels—250M+ installs, Microsoft 365 bundling potential, Azure integration—represents a structural distribution advantage that benchmark performance alone cannot overcome for Cognition. SR015, SR027
CR025 Cursor's approximately $500M ARR at $9B valuation with $40/month pricing represents a 12.5× price advantage over Devin's Team plan at $500/month; this price sensitivity gap creates pricing risk as competitors improve and enterprise buyer ROI scrutiny increases. SR027, SR019
CR026 Nubank is the only named customer with published and quantified production deployment outcomes; $73M ARR from an undisclosed number of additional enterprise accounts implies high revenue concentration risk and limits independent verification of Cognition's enterprise product-market fit beyond one case study. SR026, SR020
CR027 Developer community analysis of the December 2024 prompt-injection incident described Cognition's prior credibility as damaged, noting: Devin's claimed product superiority was inconsistent with observed PR quality; the security gap was described as 'amateurish' for a product in development for over a year. SR013, SR014
CR028 The OWASP GenAI Security Project identifies agent autonomy as an emerging, distinct risk category beyond classical LLM vulnerabilities; agents capable of modifying code, merging PRs, and executing shell commands face compound risks where a single exploit can traverse multiple security boundaries simultaneously. SR004, SR003
CR029 CVE-2024-5185 (EmbedAI CSRF data-poisoning) demonstrates that AI application platforms face real-world vulnerability classes—CSRF, data poisoning—beyond prompt injection; the NVD tracks similar classes across AI applications, indicating a broad and evolving vulnerability landscape for agentic coding platforms like Devin. SR005, SR003
CR030 EU AI Act GPAI provisions require frontier model providers to publish summary information about training data, establish copyright compliance policies, and publish technical documentation; Cognition has not publicly disclosed model training data provenance, open-source license compliance scope, or copyright policy. SR009, SR001
CR031 The ICO's AI and data protection risk toolkit requires organizations to conduct DPIAs for high-risk AI processing; Devin's autonomous PR merges and production deployments on customer infrastructure in EU member states likely qualify under GDPR Article 35, creating DPIA obligations for UK/EU customers. SR010, SR009
CR032 Doe v. GitHub and OpenAI consolidated copyright litigation (In Re: OpenAI Copyright Infringement, S.D.N.Y. 2025) involves claims that AI model training on copyrighted code and text violates copyright law; an adverse ruling could require Cognition to disclose or remediate its training data, imposing significant operational and legal costs. SR011, SR004
CR033 Cognition's January 2025 Series A was led by Founders Fund at a $2B post-money valuation; a security incident, major regulatory enforcement action, or sustained competitor improvement could trigger down-round financing pressure at the next capital raise given the high valuation multiple relative to current ARR. SR017, SR024
CR034 Developer community skepticism at Devin's launch—documented on HN and via independent benchmarking—highlighted concerns about benchmark reliability, limited demo scope, and performance gaps for non-standard tasks; this adverse reputation dynamic increases adoption friction among technically sophisticated enterprise evaluators. SR014, SR013
CR035 Cognition's headcount grew 102% year-over-year to approximately 222 employees (Growjo, April 2025); this rapid hiring pace, without disclosed revenue-per-employee targets or profitability milestones, increases burn rate risk if ARR growth decelerates below the ~70×/year pace observed in the early 2025 growth sprint. SR021, SR020
CR036 arXiv research on AI-assisted PR acceptance rates over 32 weeks shows a measurable +0.77%/week improvement trend, validating that autonomous coding agent productivity is improving; however, the long-run acceptance ceiling and transition from supervised to fully autonomous PR merges at enterprise scale remains empirically unproven. SR022, SR015
CR037 Under GDPR Article 35, automated decision-making with significant effects on individuals requires a DPIA; Devin's autonomous code deployments affecting production systems could qualify in EU deployments, creating DPIA obligations that enterprise customers must fulfill before deploying Devin at scale. SR010, SR002
CR038 Cognition's pricing dropped by approximately 3× with Devin 2.0 (April 2025), from higher effective ACU pricing to lower tiers; while improving adoption, this pricing compression requires proportional volume growth to sustain ARR trajectory and reduces the pricing power available to offset future LLM API cost increases. SR028, SR019
CR039 The White House AI EO of January 2025 creates a deregulatory federal environment that reduces US-federal compliance burden for Cognition in the near term; but agency rules for government AI procurement—relevant to Cognition's announced Government customer vertical—may introduce FedRAMP and NIST AI RMF requirements on a 12–24 month timeline. SR008, SR006
CR040 Independent developer analysis consistently identifies Devin's ACU cost model opacity and unpredictable session-level billing as adoption barriers; enterprise customers in community analysis report difficulty estimating total cost of ownership, creating adoption friction beyond the pilot phase. SR013, SR020
CR041 SWE-bench Verified (500 human-filtered instances) is considered more reliable than SWE-bench Full (2294 instances); Claude Code Opus 4's 72.5% Verified score versus Devin's 13.86% Full score reflects both a benchmark type difference and a fundamental performance gap that, unless Cognition closes it, will increasingly drive enterprise procurement decisions toward Anthropic's product. SR015, SR027
CR042 Cognition's terms of service prohibit using Devin outputs to train competing models, but do not publicly disclose the provenance of Devin's own training data; this asymmetry creates potential legal exposure under the same copyright theories being litigated against GitHub and OpenAI in the Doe v. GitHub class action. SR001, SR011
CV001 Cognition AI's $10.2B September 2025 post-money valuation divided by an estimated $73M trailing ARR (April 2025) implies approximately a 140× ARR multiple—a 7× premium to the closest comparable Cursor/Anysphere, which closed a $9.9B Series D at approximately $500M ARR, implying ~20× ARR. SV011, SV012, SV010
CV002 The recommendation for Cognition AI at $10.2B valuation is watchlist / conditional invest: the bull thesis (70×/year ARR growth, Windsurf integration moat, enterprise lighthouse customers) is structurally compelling but insufficient to justify a risk-adjusted positive return at 140× ARR without NRR, gross margin, and top-customer concentration data. SV012, SV013, SV017
CV003 Entry discipline recommendation: investors should target a Cognition AI entry valuation below $5B (implying ≤70× current estimated ARR), at which the bull case returns are 3.5–5× in 3 years and the bear case returns imply losses limited to 50–80% of investment, reflecting a more acceptable risk-reward asymmetry. SV012, SV013
CV004 Cognition AI raised $175M at a $2B valuation in its Series A round led by Founders Fund IX (CIK 0001971631, EDGAR); this represents a 5.1× step-up to the September 2025 $10.2B round in approximately nine months—one of the fastest large-round step-up trajectories in AI software startup history. SV002, SV011
CV005 Cognition Capital SPV I (CIK 0002072175, Form D filed June 2025) was structured contemporaneously with the Windsurf acquisition; the SPV structure suggests one or more institutional investors required a structured co-investment vehicle, which is a common signal of concentration risk in a single institution funding the Windsurf consideration. SV001, SV003
CV006 A 'culture reset' at Cognition post-Windsurf acquisition—reported as Cognition offering nine-month buyouts to Windsurf staff and requiring 80-hour workweeks—introduces talent attrition risk that is not reflected in the $10.2B headline valuation and could impair the 250K+ DAU Windsurf user base that underpins the IDE moat thesis. SV021, SV014
CV007 Founders Fund IX's portfolio page confirms Cognition AI as a portfolio company; Founders Fund IX filed an amended Form D (D/A) in October 2025, indicating continued capital activity in the fund consistent with ongoing Cognition investment or follow-on management activity. SV003, SV007
CV008 Under the bull scenario, Cognition AI reaches $700–800M ARR by end-2027 (implying ~45% quarterly growth sustained for 10 quarters from the April 2025 base), at which a 25–35× forward ARR exit multiple would imply an exit valuation of $18–28B, representing a 1.5–2.5× return from the $10.2B September 2025 round. SV012, SV013
CV009 Under the base scenario, ARR growth decelerates from 9-month 70×/year pace to 30–40% quarterly growth, reaching $200–350M ARR by end-2027; at 20–25× forward ARR multiple, the implied exit valuation is $4–9B—flat to slight loss relative to the $10.2B entry, implying negative risk-adjusted returns for investors entering at $10.2B. SV012, SV013
CV010 Under the bear scenario, a material adverse event (second security incident, major customer churn, or competitor price undercut >3×) causes ARR to plateau at $80–120M with NRR below 90%; at 8–12× trailing ARR, the implied exit valuation is $640M–$1.4B, representing a 90%+ capital loss for investors at $10.2B entry. SV017, SV019
CV011 Cursor/Anysphere's Series D at $9.9B on approximately $500M ARR implies a 20× ARR multiple; Cognition's $10.2B valuation at approximately $73M ARR implies a 140× multiple, representing a 7× premium to Cursor on a comparable product category—enterprise AI coding tools—that is only sustainable if Cognition grows 7× faster to ARR parity with Cursor within 12–18 months. SV010, SV012
CV012 Harvey AI (legal AI) at approximately $3B valuation and ~$100M ARR implies a 30× ARR multiple; high NRR (reportedly above 150% for legal enterprise accounts) justifies this premium. Cognition at 140× ARR requires NRR disclosure at a comparable level to avoid multiple compression, but NRR is currently undisclosed. SV018, SV013
CV013 Glean (enterprise AI search) at $4.6B valuation and approximately $100M ARR implies a 46× ARR multiple; its NRR above 150% and lack of benchmark commoditization risk distinguish it from Cognition. Cognition's 140× ARR multiple exceeds even high-NRR AI SaaS peers, suggesting the market is pricing in either a unique category premium or future ARR disclosure that narrows the apparent gap. SV018, SV004
CV014 Replit (developer platform, $1.16B valuation, 2023) represents the category risk for Cognition: a developer productivity pioneer that lost relative market share to more focused AI-native competitors (Cursor, Devin) during the 2024–2025 agentic coding transition. The Replit trajectory is a base-case downside reference for Cognition's own category risk. SV018, SV017
CV015 GitHub Copilot, embedded in Microsoft's GitHub Enterprise product at $19–$39 per seat per month, represents a structural pricing threat: Microsoft can subsidize Copilot through Azure and M365 bundling, enabling sustained price pressure at a price point approximately 15–25× below Devin's $500/month Team plan while offering comparable pass-level coding assistance. SV009, SV017
CV016 OpenAI Codex positions directly as an autonomous coding agent—similar to Devin—in Q2 2025, representing a new competitive threat from a foundation model provider that both supplies Devin's underlying intelligence and competes for the same enterprise developer productivity budget. SV009, SV020
CV017 Amazon Q Developer (AWS-bundled AI coding assistant) competes with Devin for enterprise developer productivity budget within AWS cloud deployments; its bundling within AWS Enterprise Support creates lock-in dynamics that disadvantage Devin for customers already committed to the AWS ecosystem. SV009, SV018
CV018 A second security incident within 18 months of the December 2024 prompt-injection disclosure is the primary thesis-break trigger; the probability is moderate given Devin 2.0's expanded attack surface (autonomous PR merges, scheduled agents) and the absence of a published post-Devin-2.0 penetration test or bug-bounty program. SV019, SV020
CV019 Customer concentration risk—undisclosed top-customer ARR share—is the primary financial thesis-break variable. If Nubank accounts for greater than 30% of the estimated $73M ARR, a Nubank renewal risk is a material single-event that could reduce ARR by $20M+ without any other change in business trajectory. SV027, SV012
CV020 Benchmark commoditization risk is the primary competitive thesis-break: Claude Code Opus 4's 72.5% SWE-bench Verified score (2025) versus Devin's 13.86% Full score (2024 launch) demonstrates a 5× improvement rate in 15 months that, if sustained, eliminates Devin's technical differentiation by late 2026. SV017, SV020
CV021 NRR below 90% would be a thesis-break for any investment at $10.2B; NRR below 90% implies net negative ARR cohort contribution, meaning the existing customer base is contracting, and sustaining headline ARR growth requires ever-larger new logo acquisition to compensate—a structurally unsustainable growth model at enterprise SaaS scale. SV013, SV012
CV022 EU regulatory enforcement action under the EU AI Act GPAI provisions (effective August 2026) could restrict Cognition's EU market access; Cognition expanded to Europe (London office, January 2026) without disclosed GPAI compliance infrastructure, creating a potential Q3 2026 enforcement risk. SV022, SV023
CV023 Net Revenue Retention (NRR) by cohort is the single most critical undisclosed financial metric for valuation purposes; it determines whether the $73M ARR estimate reflects durable enterprise contracts or early-adopter pilots with high churn, a distinction worth $2–5B in implied valuation. SV012, SV013
CV024 Gross margin is the second critical undisclosed metric; Devin's per-session LLM API costs (Anthropic/OpenAI commercial rates) are substantial for compute-intensive multi-hour sessions, and gross margin below 30% would signal an unsustainable cost structure that requires either proprietary model development or volume pricing renegotiation before profitability. SV020, SV028
CV025 Post-Devin-2.0 penetration testing and red-team disclosure is required for enterprise procurement security sign-offs; the December 2024 incident occurred before Devin 2.0's expanded agentic capabilities (PR merges, scheduling), meaning the existing SOC 2 Type II audit (March 2024) does not cover the current expanded attack surface. SV019, SV028
CV026 Windsurf customer retention at 6 months post-acquisition (January 2026) is a critical proof point for the IDE moat thesis; if the 350 enterprise accounts acquired with Windsurf show greater than 15% churn, the Windsurf contribution to ARR is overstated and the platform differentiation thesis weakens. SV021, SV014
CV027 LLM API contract terms with Anthropic and OpenAI—specifically volume pricing floors, agentic use-case restrictions, and change-of-control provisions—are undisclosed but material; a scenario where Anthropic raises API pricing 3× as Claude Code gains enterprise share would compress Devin's gross margin materially and impair the $10.2B valuation. SV020, SV009
CV028 FedRAMP authorization status for the US Government vertical (launched February 2026) is a credibility question for government procurement; without FedRAMP In Process or Authorized status, government agency sales are limited to informal pilots, capping US Government revenue contribution at a level well below the implied market opportunity. SV022, SV006
CV029 The Founders Fund IX Form D/A amendment (October 2025, EDGAR) confirms continued fund activity consistent with Cognition follow-on capital management or new investment in the post-$400M round period; combined with the Cognition Capital SPV I (June 2025), the structured capital base is complex and the preference stack is undisclosed. SV003, SV001
CV030 The $73M ARR estimate at April 2025, growing from $1M at GA (December 2024) in nine months, represents a 73× top-line growth rate—exceptional even by AI-native startup standards—but is based on third-party estimates (Growjo, Sacra) that have not been confirmed by Cognition in any official disclosure. SV012, SV013
CV031 Devin's pricing dropped approximately 3× with Devin 2.0 (April 2025), with the Team plan remaining at $500/month but providing 83% more tasks per ACU; this effective price reduction improves adoption but compresses the revenue-per-session metric, requiring proportionally greater volume to maintain ARR trajectory. SV028, SV012
CV032 Total disclosed funding across all rounds is approximately $1.575B; at a $10.2B post-money valuation, early rounds are already at 60–140× capital returned on paper, creating exit pressure for early investors that may accelerate secondary activity and reduce inside-round price discovery for new institutional investors. SV011, SV023
CV033 Mercedes-Benz announced a partnership with Cognition AI in April 2026, adding a second named enterprise customer alongside Nubank; while terms and ARR contribution are undisclosed, the automotive/manufacturing vertical validates enterprise cross-sector demand and reduces the market risk of over-dependence on financial services verticals. SV008, SV022
CV034 a16z's 100 Gen AI Apps report identifies coding tools and developer productivity as the highest-revenue AI application category, with enterprise developer tools generating $5B+ in collective ARR across the cohort as of their most recent analysis; this market context supports the bull-case premise that Cognition is entering the most valuable AI application segment. SV004, SV018
CV035 Replit's valuation of $1.16B in 2023 as a developer platform that failed to capture the AI coding wave represents the primary bear-case precedent for Cognition; developer tool platforms with strong user bases but insufficient enterprise NRR can see rapid valuation compression when more focused AI-native competitors commoditize their core value proposition. SV018, SV017
CV036 Cognition AI's geographic expansion to Europe (London, January 2026), Japan, and Singapore (April 2026) adds TAM and demonstrates customer pull outside the US but introduces compliance infrastructure gaps—EU GPAI registration, UK GDPR DPO appointment, Japan PIPL—that create near-term regulatory and reputational risk before those markets generate material ARR. SV022, SV025
CV037 The Cognition Capital SPV I Form D (CIK 0002072175, filed 2025-06-11, Claymont, DE) is consistent with a structured co-investment vehicle used to fund the Windsurf acquisition consideration or provide a liquidity backstop to Windsurf investors, adding a layer of financial structure complexity to the Cognition cap table beyond a standard venture equity round. SV001, SV003
CV038 The Dealroom.co listing for Cognition AI reflects the company's early financing stages at a $8M valuation figure—consistent with seed-round entry—while TechCrunch and CNBC confirm the September 2025 round at $10.2B; this 1,275× valuation step-up across approximately 24 months is unprecedented in enterprise SaaS history and indicates market expectation of near-monopoly market capture. SV005, SV029
CV039 For an IPO at $8B+ market capitalization on conventional public market multiples (15–25× forward ARR), Cognition would need to demonstrate approximately $320–535M ARR in the 12 months preceding the IPO; at base-case growth rates, this milestone is not reached before 2028–2029, suggesting the IPO window is approximately 2–4 years from May 2026. SV011, SV024
CV040 Strategic acquirer scenarios exist: Microsoft (GitHub distribution), Alphabet (Google Cloud developer tools), Salesforce (Einstein developer productivity), and Databricks/Snowflake (data engineering automation) are all logical fits; however, the $10.2B entry price for Cognition makes a strategic acquisition at premium unlikely unless ARR scales to $400M+ first, as acquirers would need to pay a 20–30% acquisition premium above the last private valuation. SV011, SV023
CV041 Cursor/Anysphere's November 2025 Series D at $29.3B (per Business Wire) represents an even more recent and higher precedent valuation for AI coding tools; if confirmed, this would shift the peer set valuation anchor upward and could partially justify Cognition's 140× ARR multiple relative to Cursor's even higher implied multiple. SV010, SV012
CV042 Cognition AI's Mercedes-Benz partnership announcement (April 2026) and Cognizant channel partnership (January 2026) diversify the named customer base beyond Nubank, reducing—but not eliminating—customer concentration risk; the absence of disclosed outcome metrics for these relationships limits the financial materiality assessment. SV008, SV022
来源
编号出版方标题引文
SO001 Cognition AI Cognition AI — Official Homepage
SO002 Cognition AI Introducing Devin, the first AI software engineer Devin is a tireless, skilled teammate, equally ready to build alongside you or independently complete tasks for you to review.
SO003 Cognition AI Funding, growth, and the next frontier of AI coding agents Enterprise ARR grew over 30% in just seven weeks since we acquired Windsurf.
SO004 Cognition AI Cognition's acquisition of Windsurf
SO005 Wikipedia Cognition AI — Wikipedia
SO006 TechCrunch Cognition AI defies turbulence with a $400M raise at $10.2B valuation The round values Cognition at $10.2 billion post-money.
SO007 CNBC Cognition valued at $10.2 billion two months after Windsurf purchase
SO008 Sacra Cognition revenue, valuation & funding Cognition ARR: $1M (Sep 2024) → $73M (Jun 2025) → $155M (Jul 2025 post-Windsurf)
SO009 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise, showing strong enterprise momentum
SO010 FavTutor Meet The Team Behind Devin AI, Its Founders & Investors
SO011 Analytics India Magazine Meet the Creator of Devin: A Child Prodigy Who is Making Coding Obsolete
SO012 Observer Cognition, Maker of Goldman's First 'A.I. Employee'
SO013 Economic Times Work-life imbalance: After buying remnants of Windsurf, Cognition lays off 30, tells rest to work long hours Cognition laid off 30 employees and offered buyouts to approximately 200 remaining Windsurf staff with a requirement to commit to 80-hour six-day workweeks.
SO014 The AI Insider Cognition AI Closes $400M in Funding to Reach $10.2B Valuation Amid Rapid Growth
SO015 AI Invest (ainvest.com) Cognition AI's Windsurf Acquisition: A Masterstroke in the AI Talent & IP War
SO016 WinBuzzer Cognition AI's Culture Reset: Offers Nine-Month Buyouts to Windsurf Staff, Demands 80-Hour Weeks
SO017 SFGate SF tech CEO offers buyouts to let workers flee 'extreme' work culture
SO018 Eesel AI A deep dive into Cognition AI reviews: Hype vs. Reality Independent testing shows real-world task success rates well below the 'fully autonomous' marketing framing.
SO019 Sean Kim (imseankim.com) Devin AI Review: From $500 to $20 — 6 Weeks With Cognition's AI Software Engineer
SO020 Tech Funding News Cognition AI scores $400M at $10.2B valuation as demand spikes for coding agents
SO021 Stepmark AI Company Spotlight: Cognition AI – Devin, your Autonomous Software Engineer
SO022 Tech Funding News Cognition raises $500M at nearly $10B valuation following Windsurf acquisition
SO023 AI Invest (ainvest.com) AI-Driven Software Development and Startup Valuation: The Cognition AI Case Study
SO024 ARR Club Cognition ARR, Revenue Growth & Milestones
SO025 Devin.ai Devin — AI Software Engineer
SO026 Devin.ai Devin Pricing
SM001 Mordor Intelligence AI Code Tools Market Size, Share & 2030 Trends Report
SM002 Grand View Research Generative AI Coding Assistants Market Size Report, 2030
SM003 BusinessWire / ResearchAndMarkets Generative AI Coding Assistants Strategic Research Report 2025 — Market to Reach $97.9 Billion by 2030
SM004 Gartner Gartner Identifies the Top Strategic Trends in Software Engineering for 2025 and Beyond
SM005 Futuretechmag Gartner Highlights Strategic Software Engineering Trends Beyond 2025
SM006 Evans Data Corporation Worldwide Developer Population Grows to 27 Million
SM007 Statista Global developer population 2024
SM008 WIPO (World Intellectual Property Organization) Global Software Spending Surges to Close to USD 700 Billion in 2024
SM009 Signisys AI Code Assistant: 90% by 2028 — Gartner forecast summary
SM010 GitHub (official) Gartner positions GitHub as a Leader in the 2025 Magic Quadrant for AI Code Assistants GitHub Copilot surpassed 20 million users and is used by over 50,000 organizations.
SM011 Kingy.ai AI Coding Agents in 2025: The Ultimate Battle for Developer Supremacy
SM012 OpenAI (official) The State of Enterprise AI — 2025 Report
SM013 Deloitte AI Trends 2025: Adoption Barriers and Updated Predictions
SM014 Forbes The Biggest Barriers Blocking Agentic AI Adoption
SM015 California Management Review (UC Berkeley) Adoption of AI and Agentic Systems: Value, Challenges, and Pathways
SM016 VC Cafe Which Barriers Still Block Agentic AI Adoption?
SM017 Second Talent GitHub Copilot Statistics and Adoption Trends 2025
SM018 AI Business VC GitHub Copilot Crosses $2B ARR — 46% of Code Is Now AI-Generated
SM019 Endroid GitHub Copilot Surpasses 20 Million Users, Eyes Enterprise AI Coding Dominance
SM020 JetBrains Research Global Developer Population 2024
SM021 Grand View Research Software Market Size, Share and Trends — Industry Report, 2030
SM022 ResearchAndMarkets AI Coding Assistant Tools Global Market Insights 2025
SM023 GeniusAI Tech AI Statistics 2025: Market Growth, Usage, Trends and Adoption
SM024 ISG (Information Services Group) State of Enterprise AI Adoption Report 2025
SM025 MarketsAndMarkets AI Assistant Market Report 2025–2030, by Application, Geo, Tech
SP001 BusinessWire Cursor Secures $2.3 Billion Series D Financing at $29.3 Billion Valuation
SP002 devclass.com GitHub Copilot tops $2B ARR, confirms enterprise dev tools push
SP003 CNBC Cursor AI startup funding round valuation November 2025
SP004 The Next Web Cursor Anysphere $2B ARR funding $50B valuation talks
SP005 BetterStack GitHub Copilot vs Cursor vs Windsurf: AI Coding Assistant Comparison
SP006 Digital Applied Cursor AI $29B Valuation Agent Revolution
SP007 Amazon Web Services Amazon Q Developer — AI Coding Assistant for AWS
SP008 Anthropic Claude Code — Agentic coding in your terminal
SP009 Lowcode.agency Claude Code vs Devin: Autonomous Coding Agent Comparison
SP010 ArXiv SWE-bench and AI software engineering benchmark analysis 2026
SP011 Benched.ai Top Coding Agents 2025 — Feature and performance guide
SP012 GitHub GitHub Copilot — Features and Plans
SP013 Amazon Web Services Amazon Q Developer Pricing
SP014 GitHub / OpenAI OpenAI Codex GitHub Repository
SP015 SWE-bench (Princeton NLP) SWE-bench: Evaluating LLM Agents on Real-World Software Engineering Tasks
SP016 Educative.io Cursor vs Windsurf vs Copilot: AI IDE Comparison 2025
SP017 WeCompareAI Best AI Coding Tools 2025: GitHub Copilot vs Cursor vs Windsurf
SP018 CreateAIAgent.net Amazon Q Developer vs GitHub Copilot Workspace Comparison
SP019 Kingy.ai AI Coding Agents in 2025: The Ultimate Battle for Developer Supremacy
SP020 aloa.co GitHub Copilot vs Cursor vs Windsurf: Complete Comparison
SP021 GitHub Docs Subscription plans for GitHub Copilot
SP022 Replit Replit Pricing — AI development platform
SP023 Cursor Cursor Pricing
SP024 Windsurf Windsurf IDE Editor — formerly Codeium
SP025 GitHub (Princeton NLP / SWE-bench) SWE-bench: Software Engineering Benchmark GitHub Repository
SP026 Sean Kim Tech Blog Cursor vs Windsurf vs GitHub Copilot: AI IDE Battle October 2025
SP027 Anthropic Anthropic Claude Pricing — Pro and Max plans
SI001 Cognition AI (Official) Funding, Growth, and the Next Frontier of AI Coding Agents
SI002 Axios Cognition raises $175M at $2B valuation for AI software engineer Devin
SI003 Sacra Research Cognition AI revenue and growth analysis
SI004 Devin.ai (Official) Devin AI Pricing — Plans and ACU details
SI005 ARR.club Cognition AI ARR and revenue tracking
SI006 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise
SI007 Stepmark AI Company Spotlight: Cognition AI — Devin Autonomous Software Engineer
SI008 Axios Cognition AI raises $400 million at $10 billion valuation for Devin
SI009 TechCrunch Cognition AI defies turbulence with a $400M raise at $10.2B valuation
SI010 CNBC Cognition valued at $10.2 billion two months after Windsurf acquisition
SI011 AInvest AI-Driven Software Development Startup Valuation: Cognition AI Case Study
SI012 TechFundingNews Cognition AI scores $400M at $10.2B valuation as demand spikes for coding agents
SI013 TechFundingNews Cognition raises $500M at nearly $10B valuation following Windsurf acquisition
SI014 Observer Cognition AI startup valuation $10B
SI015 TheAIInsider Cognition AI closes $400M in funding to reach $10.2B valuation amid rapid growth
SI016 CBInsights Cognition Labs company profile and financials
SI017 GrowJo Cognition AI company revenue and funding estimates
SI018 Cognition AI (Official) Cognition AI Blog
SI019 Winbuzzer Cognition AI culture reset: offers nine-month buyouts to Windsurf staff
SI020 PitchBook Cognition AI Series A valuation and Devin funding profile
SI021 Growjo Cognition AI funding and employee count 2025
SI022 Devin.ai (Official) Devin AI — Official product and pricing page
SI023 Wikipedia Cognition AI — Wikipedia article
SI024 AIForDevelopers AI Startup Cognition Labs raises $175M valued at $2 billion
SI025 Towards AI (Medium) Cognition AI's Devin: Is it worth the hype?
SI026 SEC EDGAR Form D/A: Founders Fund IX, LP — Amended Notice of Exempt Offering of Securities (CIK 0001971631)
SI027 Cognition AI (Official) Devin heads east: Cognition opens its Singapore APAC headquarters
SI028 Cognition AI (Official) Engineering in the fast lane: Mercedes-Benz partners with Cognition
SE001 Devin (Official Documentation) Devin Documentation — Getting Started and Integrations
SE002 Cognition AI (Official) Introducing Devin, the first AI software engineer
SE003 Cognition AI (Official) SWE-bench Technical Report
SE004 Devin (Official) Devin — The AI Software Engineer
SE005 Cognition AI (GitHub) CognitionAI/devin-swebench-results — SWE-bench evaluation results and methodology
SE006 SWE-bench (Princeton / CMU) SWE-bench Leaderboard and Benchmark Documentation
SE007 Devin (Official Documentation) Devin Documentation — Integrations
SE008 imseankim.com (Developer Blog) Devin 2.0 AI Software Engineer Review: Cognition Pricing and Benchmark
SE009 Hacker News (Y Combinator) Hacker News: Introducing Devin (community discussion thread)
SE010 Benched.ai Top AI Coding Agents 2025 — Benchmarks and Capabilities Snapshot
SE011 arxiv.org / MSR 2026 Empirical Study Comparing AI Coding Agents: Temporal Trends in PR Acceptance
SE012 Artificial Analysis AI Coding Agent Comparison — Artificial Analysis
SE013 Codeium / Windsurf (now Cognition) Windsurf — The AI Code Editor
SE014 Devin (Official) Devin Customer Case Studies
SE015 Cognition AI (Official) Cognition AI acquires Windsurf — blog announcement
SE016 Devin (Official) — Customer Proof How Nubank refactors millions of lines of code with Devin
SE017 Hacker News (Y Combinator) Hacker News: Streamer discovers major vulnerability in Cognition's Devin live on air
SE018 Towards AI (Medium) What happened when Devin AI took on 2,294 GitHub bugs — the 13.86% that changed everything
SE019 TechCrunch Devin AI software engineer is now just $20 a month, down from $500
SE020 SWE-bench (GitHub) SWE-bench GitHub Repository — Benchmark and Leaderboard
SE021 Kingy.ai (Technical Blog) AI Coding Agents in 2025: The Ultimate Battle for Developer Supremacy
SE022 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise showing startup resolve
SE023 UC Berkeley Haas / CMR Adoption of AI and Agentic Systems: Value, Challenges, and Pathways
SE024 Cognition AI (Official) Cognition AI Blog — Product Updates and Technical Posts
SE025 Artificial Analysis AI Agent Benchmark Performance — Coding Agents Evaluation
SU001 Devin (Official) Devin AI — Customer Case Studies
SU002 Devin (Official) — Customer Case Study How Nubank refactors millions of lines of code with Devin
SU003 Cognition AI (Official Blog) Engineering in the fast lane: Mercedes-Benz partners with Cognition
SU004 Cognition AI (Official Blog) Cognizant partners with Cognition; COBOL modernization; Government vertical
SU005 Cognition AI (Official) Devin heads east: Cognition opens its Singapore APAC headquarters
SU006 Cognition AI (Official Blog) Launching in Japan; Cognition for Government; London office
SU007 Growjo Cognition AI: Revenue, Competitors, Alternatives
SU008 imseankim.com (Developer Blog) Devin 2.0 AI Software Engineer Review: Cognition Pricing and Benchmark
SU009 TechCrunch Devin AI software engineer is now just $20 a month, down from $500
SU010 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise
SU011 Hacker News (Y Combinator) Hacker News: Introducing Devin — Developer community discussion
SU012 Artificial Analysis AI Coding Agent Comparison — Artificial Analysis
SU013 Benched.ai Top AI Coding Agents 2025 — Benchmarks and Capabilities Snapshot
SU014 Cognition AI (Official Blog) Devin in Windsurf and Singapore APAC expansion
SU015 Kingy.ai (Technical Blog) AI Coding Agents in 2025: The Ultimate Battle for Developer Supremacy
SU016 UC Berkeley Haas / CMR Adoption of AI and Agentic Systems: Value, Challenges, and Pathways
SU017 Cognition AI (Official Blog) How Cognition Uses Devin to Build Devin
SU018 Hacker News (Y Combinator) Streamer discovers major vulnerability in Cognition's Devin live on air
SU019 arxiv.org / MSR 2026 Empirical Study: PR Acceptance Rates for AI Coding Agents
SU020 Growjo (Employee/Revenue Intelligence) Cognition AI headcount, revenue growth, funding data
SU021 Cognition AI (Official) Devin AI — Product Page
SU022 VentureBeat Cognition AI $175M Series A at $2B valuation to build AI software engineers
SU023 Towards AI (Medium) What happened when Devin AI took on 2,294 GitHub bugs
SU024 Cognition AI (Official) Devin Documentation — Integrations, Enterprise Deployment
SU025 Cognition AI (Official Blog) Introducing Devin 2.2 — Most important update since launch
SU026 Stack Overflow (Developer Survey 2024) 2024 Stack Overflow Developer Survey — AI Tools Adoption
SU027 GitHub Blog How GitHub Copilot is getting better at understanding your code
SU028 State of AI Report State of AI Report — Enterprise AI Adoption Trends
SU029 Cognition AI (Official Blog) Closing the Agent Loop: Devin Autofixes Review Comments
SU030 Stanford HAI Stanford AI Index 2025 — Enterprise AI Adoption and Coding Agents
SR001 Cognition AI (Official Docs) Security at Cognition — SOC 2 Type II, Data Privacy, Intellectual Property
SR002 Cognition AI (Trust Center) Cognition Trust Center — Security Documentation Portal (NDA-gated)
SR003 OWASP Foundation OWASP Top 10 for Large Language Model Applications
SR004 OWASP Foundation (GenAI Security Project) OWASP GenAI Security Project — LLM Top 10 Risks
SR005 NIST National Vulnerability Database CVE-2024-5185 — EmbedAI CSRF Data Poisoning Vulnerability
SR006 California Legislative Information SB 1047 — Safe and Secure Innovation for Frontier Artificial Intelligence Models
SR007 Federal Trade Commission (US) Generative AI Raises Competition Concerns
SR008 The White House Executive Order — Removing Barriers to American Leadership in Artificial Intelligence
SR009 European Parliament EU AI Act — First Regulation on Artificial Intelligence
SR010 UK Information Commissioner's Office (ICO) AI and Data Protection Risk Toolkit
SR011 CourtListener (RECAP) Doe v. GitHub Inc. — AI Copyright Infringement Docket
SR012 Wikipedia (Wikimedia Foundation) Prompt Injection — Cybersecurity Attack Vector
SR013 Hacker News (Y Combinator) HN: Streamer discovers major vulnerability in Devin live on air (Dec 2024)
SR014 Hacker News (Y Combinator) HN: Cognition AI — Devin Launch Skepticism and Scope Analysis (Mar 2024)
SR015 SWE-bench (Princeton NLP / CMU) SWE-bench — Live Benchmark Leaderboard for AI Software Engineering Agents
SR016 Cognition AI (Official Blog) Introducing Devin 2.0 — 83% More Tasks per ACU
SR017 TechCrunch Cognition raises $175M at $2B valuation to build AI software engineers
SR018 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise
SR019 Sacra (Research) Devin ARR and Revenue Metrics Tracker — Cognition
SR020 imseankim.com (Independent Analysis) Devin ARR $1M to $73M in 9 Months — PR Merge Rate and Cost Analysis
SR021 Growjo (Business Analytics) Cognition AI — Headcount, Revenue, and Growth Analytics
SR022 arXiv (Cornell) AI-Assisted Pull Requests: Longitudinal Acceptance Rate Trends in Open-Source
SR023 Anthropic (Official) Claude Code — Agentic Coding Assistant Product Overview
SR024 PitchBook Data Cognition AI — Funding Rounds and Investor Profile
SR025 Cognition AI (Official Blog) Introducing Devin — The AI Software Engineer (March 2024)
SR026 Devin.ai (Official) Nubank Case Study — 12x Efficiency, 20x Cost Savings
SR027 benched.ai (Independent Benchmarking) Top Coding Agents 2025 — Devin vs. Competitors
SR028 TechCrunch Devin 2.0 — Price Drops 3× as Cognition Expands Agentic Capabilities
SR029 CNBC (Technology) Cognition AI Raises $175M Valuing AI Coding Startup at $2 Billion
SR030 Axios (Technology) Cognition AI — Windsurf Acquisition and $400M Raise
SV001 US Securities and Exchange Commission (EDGAR) Cognition Capital SPV I — Form D, CIK 0002072175 (June 2025)
SV002 US Securities and Exchange Commission (EDGAR) Founders Fund IX, LP — Form D Original Filing, CIK 0001971631 (April 2023)
SV003 US Securities and Exchange Commission (EDGAR) EDGAR Full-Text Search — Founders Fund IX Form D Filings
SV004 a16z (Andreessen Horowitz) 100 Gen AI Apps — 4th Edition: Top Consumer and Enterprise AI Applications
SV005 Dealroom.co Cognition AI — Funding History and Investor Profile
SV006 The New Stack Cognition Launches Devin AI Software Engineer for Enterprise
SV007 Founders Fund (Official) Founders Fund Portfolio — Official Website
SV008 Business Wire Cognition AI and Mercedes-Benz Partner to Accelerate Software Development
SV009 OpenAI (Official) OpenAI Codex — Agentic Coding Assistant Product Overview
SV010 Business Wire Cursor Secures $2.3 Billion Series D Financing at $29.3 Billion Valuation
SV011 TechCrunch Cognition AI defies turbulence with $400M raise at $10.2B valuation
SV012 Growjo (Business Analytics) Cognition AI — Revenue and Valuation Analytics
SV013 Sacra (Research) Cognition AI Revenue and Metrics Tracker
SV014 VentureBeat Cognition follows Windsurf acquisition with $400M fundraise
SV015 Cognition AI (Official Blog) Cognition Funding, Growth, and the Next Frontier of AI Coding Agents
SV016 imseankim.com (Independent Analysis) Cognition Devin Product-Led Growth and ARR Analysis
SV017 SWE-bench (Princeton NLP / CMU) SWE-bench Live Leaderboard — Autonomous Software Engineering
SV018 benched.ai (Independent Benchmarking) Top AI Coding Agents 2025 — Competitive Analysis
SV019 Hacker News (Y Combinator) Cognition AI Security Incident Live-Stream Discussion
SV020 Anthropic (Official) Claude Code — Agentic Coding Assistant Overview
SV021 Winbuzzer (Tech News) Cognition AI's Culture Reset Offers Nine-Month Buyouts to Windsurf Staff
SV022 Cognition AI (Official Blog) Cognition Expands to Europe — London Office Opening
SV023 Axios (Technology) Cognition AI raises $400M at $10.2B valuation
SV024 The AI Insider Cognition AI Closes $400M to Reach $10.2B Valuation Amid Rapid Growth
SV025 TechFundingNews Cognition AI Scores $400M at $10.2B Valuation as Demand Spikes for Coding Agents
SV026 Observer (Business) Cognition AI Startup Valuation $10B
SV027 Devin.ai (Official) Nubank Case Study — 12× Efficiency Gain, 20× Cost Savings
SV028 Cognition AI (Official Blog) Introducing Devin 2.0 — 83% More Tasks Per ACU
SV029 CNBC (Technology) Cognition valued at $10.2 billion two months after Windsurf acquisition
SV030 Ainvest.com (AI Investment Research) AI-Driven Software Development Startup Valuation: Cognition AI Case Study