Applied Compute
初创公司尽调报告——企业 AI 智能体、Specific Intelligence、$160M Series A、$1.3B 估值
观察:创始团队履历过硬,也已有具名企业客户牵引,足以进入跟踪池;但单位经济性未披露、客户高度集中,离买入结论还需要更多尽调。
封面要素
公司概况
Applied Compute 是一家位于 San Francisco 的企业 AI 公司,2025 年由 Yash Patil、Rhythm Garg 和 Linden Li 创立——三人都是 Stanford 校友,均从 OpenAI 的技术岗位离职创业。公司打造「Specific Intelligence」智能体:用客户自己的数据和工作流训练,部署在客户环境中,并通过强化学习持续改进。2026 年 4 月,公司完成 Kleiner Perkins 领投的 Series A 后,投后估值达到 $1.3B;已披露部署客户包括 DoorDash、Cognition 和 Mercor。收入、ARR 和毛利率仍未披露。
- 成立时间
- 2025-01-01
- 创始人
- Yash Patil, Rhythm Garg, Linden Li
- 创立地点
- San Francisco, California, USA
- 总部
- San Francisco, California, USA
- 产品
- 「Specific Intelligence」——专有 AI 智能体,用客户专有数据训练,并部署在客户自己的基础设施内。智能体覆盖企业工作流自动化、代码生成和数据运营,并利用生产使用信号上的强化学习持续改进。
- 客户
- 企业科技公司和高速运营团队;已确认部署在 DoorDash(配送运营)、Cognition(AI 软件工程)和 Mercor(人才市场)。目标市场是任何需要专用、可持续改进 AI 智能体,而不是通用模型的企业。
- 商业模式
- 未披露;预期模式是企业 SaaS 授权,或与算力和定制服务挂钩的按智能体订阅。公司未公开披露收入或定价。
- 阶段
- Series A (2026)
- 融资情况
- 累计融资 $160M,其中包括 2026 年 4 月完成、由 Kleiner Perkins 领投的 Series A,投后估值 $1.3B。此前种子轮来自 Mayfield 和未披露的天使投资人。
执行摘要
主要优势
- 三位前 OpenAI 创始人直接参与过生产级 AI agent 建设,履历可比 2025–2026 年企业 AI 队列中的头部公司。
- DoorDash、Cognition、Mercor 等具名企业部署,在 Series A 完成前就给出了早期商业验证。
- Kleiner Perkins 领投、以 $1.3B 估值融资 $160M,显示一线投资人已有信心,也把现金跑道推到 2027–2028 年。
- Specific Intelligence 的差异化在于用客户数据训练、可本地部署的 agent,直接回应企业对数据主权和准确性的真实要求。
主要风险
- OpenAI、Anthropic 和超大规模云厂商(AWS Bedrock、Azure AI)都在向企业 agent 定制收敛,差异化窗口被快速压缩。
- 目前只确认 3 家客户,收入、ARR 和留存均未披露;客户集中风险很高。
- 创始人仍是早期职业阶段研究者,未披露过往 CEO/CRO 经验;把企业 go-to-market 放大,和做 AI 研究不是同一套能力。
- 在收入未披露的情况下给到 $1.3B 估值,隐含很高的前瞻倍数,需要持续超高速增长才能消化。
未决问题
- 需要披露收入运行率、ARR 和毛利率,才能按基本面承保。
- 需要厘清 DoorDash、Cognition 和 Mercor 的客户合同条款,包括 ARR、NRR 和续约选择权。
- 需要说明未来 12 个月内,面对 OpenAI 企业微调和 Microsoft/Copilot Studio 的竞争护城河在哪里。
- 需要披露三位联合创始人之外的团队规模,以及关键工程 / 销售招聘。
- 需要厘清 Series A 后的股权结构表和优先权安排。
目录
01公司概览
1.1 身份、总部与商业模式
Applied Compute, Inc. 是一家企业 AI 公司,打造其所谓的「Specific Intelligence」:专有智能体用客户自己的数据和工作流训练,部署到客户自身环境的生产系统中,并从真实使用轨迹中通过强化学习持续改进。多方来源共同证实,公司由三名前 OpenAI 研究员在 2025 年创立;独立报道(TechStartups、Grokipedia)进一步将时间缩小到 2025 年 5 月。California Secretary of State 的备案记录和 Comcast NBCUniversal LIFT Labs 的投资组合页面都显示,公司总部位于 San Francisco,地址为 251 Rhode Island Street #207;监管备案还确认,该实体最初在 Delaware 注册成立,并于 2025 年 10 月 10 日以外州股份公司身份在 California 登记,文件编号 B20250336266。商业上,Applied Compute 自己的产品材料把平台描述为 Train、Serve、Improve 三段式:在客户数据上对会用工具的智能体做后训练;在模型训练时使用的同一框架中提供低延迟生产推理;再从生产反馈中持续做在线强化学习。公司把自己的技术栈定位为「模型灵活」:客户以后可以换用更新的前沿基础模型,不必重建框架、数据管道或部署栈;公司还称,其工程师直接嵌入客户团队,而不是把开发外包。本轮研究没有把这些公司自述主张独立复核到客户内部工程流程层面。[CO003, CO004, CO006, CO007, CO008, CO009]
Applied Compute 的身份、产品平台、客户、资本和关键人依赖如何彼此连接。
[CO006, CO007, CO009, CO012, CO030, CO041]1.2 创始人、领导层与治理
Applied Compute 由 Yash Patil、Rhythm Garg 和 Linden Li 共同创立,三人都是 Stanford University 校友,并从 OpenAI 的技术岗位离职创业。Patil 是 OpenAI 智能体式 Codex 软件工程项目的关键成员,现任 Applied Compute CEO;Garg 是 OpenAI o1 的核心贡献者,o1 是首个经强化学习训练的推理模型;Li 则负责强化学习训练所需的 ML 系统和基础设施。不同来源无法完全对齐高管头衔:Applied Compute 的 California Secretary of State 备案将 Rhythm Garg 列为 Chief Financial Officer 和 Secretary,并把 Yash Patil 列为唯一的 Chief Executive Officer 和注册代理人;但 Comcast NBCUniversal LIFT Labs 的投资组合页面则把 Garg 的头衔列为 Chief Technology Officer,把 Linden Li 列为 Chief Architect。这不是编辑取舍,而是所审阅来源中真实存在、尚未调和的冲突,本身就是一个值得标记的治理披露信号:未找到独立且有日期的高管声明文件来解决这个差异。公司自述称,团队中三分之二曾是创业公司创始人,包括前顶尖 AI 研究员和数学奥林匹克获奖者;Lux Capital 的投资组合资料还称,团队校友包括 OpenAI 强化学习基础设施老兵,以及来自 Scale AI、Together、Two Sigma 和 Watershed 的人员。所审阅来源中,三位联合创始人之外没有出现任何具名高管(例如独立总法律顾问或销售负责人);公司成立约一年时仍如此,也是关键人依赖信号。[CO012, CO013, CO014, CO015, CO016, CO017]
| 人员 | 角色 | 背景 | 创始人-市场匹配 / 职能覆盖 | 关键人依赖 |
|---|---|---|---|---|
| Yash Patil | CEO 兼联合创始人;据加州州务卿文件,为注册代理人 | OpenAI Codex 智能体式软件工程项目核心成员;Stanford University 校友 | 直接参与构建智能体编码系统的深度经验,直接支撑 Applied Compute 的智能体训练产品和公开叙事 | 高 — 唯一点名 CEO、注册代理人,并在已审阅所有融资公告中持续担任公开发言人 |
| Rhythm Garg | 联合创始人;Comcast NBCUniversal LIFT Labs 称其为首席技术官,但公司加州州务卿文件列为首席财务官兼秘书(尚未调和) | OpenAI o1 核心贡献者;o1 是首个经强化学习训练的推理模型;Stanford University 校友 | RL 推理背景与 Applied Compute 的强化学习训练栈高度贴合 | 高 — 官员头衔冲突本身说明可独立验证的治理披露有限 |
| Linden Li | 联合创始人;首席架构师 | 曾在 OpenAI 从事强化学习训练的 ML 系统和基础设施;Stanford University 校友 | 基础设施背景直接支撑 Applied Compute 的训练与服务平台 | 高 — 已审阅来源中唯一被点名的技术基础设施负责人 |
仅限于在 Applied Compute 自有材料、投资人的投资组合页面和加州州务卿文件中识别出的三位具名创始高管;在本章审阅的 28 个来源中,未找到其他非创始人高管(例如独立总法律顾问或销售负责人)。
[CO012, CO013, CO014, CO015, CO016, CO017]1.3 融资历史、估值与投资人基础
Applied Compute 已披露的融资经历了四个公开节点。Upstarts Media 最先以未公告独家消息报道,公司已融资 $20 million,估值 $100 million,由 Benchmark 合伙人 Victor Lazarte 领投,Sequoia、Conviction、Hanabi Capital、Definition 和独立投资人 Zach Frankel 也参与;另一篇二手文章(StartupsUnion)却把同一轮日期写成「June 2024」,比其他来源证实的 2025 年创立时间早整整一年,这是一个尚未调和且很可能错误的日期,本文按原样记录而不悄悄修正。随后,The Information 约在 2025 年 9 月报道 Applied Compute 正洽谈按约 $500 million 估值融资;公司又在 2025 年 10 月 29 日退出隐身状态,公开宣布来自 Benchmark、Sequoia、Lux Capital、Hanabi、Neo、Definition、Elad Gil、Victor Lazarte 和 Omri Casspi 的 $80 million 轮融资;独立媒体将融资后估值放在约 $700 million,但公司当时没有披露这个数字。到 2026 年 1 月,The Information 又报道新一轮洽谈估值为 $1.3 billion,是三个月前 $500 million 数字的两倍多;公司在 2026 年 4 月 8 日确认完成 Kleiner Perkins 领投的 $80 million Series B(Elad Gil、Lux、Greenoaks、Neo 和 Hanabi 继续参与),使已披露总融资达到 $160 million。Applied Compute 自己的公告给出该数字,TechCrunch 也援引 PitchBook 数据独立印证。Latham & Watkins LLP 担任 Applied Compute 2026 年 4 月融资的外部法律顾问。合起来看,公司估值不到一年约涨 13x;这种抬升速度本身就是尽调警示,独立于公司底层基本面。[CO018, CO019, CO020, CO021, CO022, CO023]
| 利益相关方 | 角色 | 控制或经济重要性 | 尽调问题 |
|---|---|---|---|
| Kleiner Perkins | 领投方,2026 年 4 月 Series B($80M) | 迄今披露的最大单笔支票,投后估值 $1.3 billion | 确认 Series B 的董事席位 / 观察员权利及任何保护性条款 |
| Benchmark(Victor Lazarte) | 领投方,2025 年种子轮($20M) | 最早披露的机构领投方;Lazarte 也被点名参与 2025 年 10 月轮次 | 确认多轮稀释后的当前股权结构中持股比例 |
| Sequoia Capital | 种子轮和 2025 年 10 月轮次参与方 | 至少两轮已披露融资的重复投资人 | 确认累计总投资额及任何董事 / 观察员权利 |
| Lux Capital | 2025 年 10 月和 2026 年 4 月轮次参与方 | 重复投资人,并发布关于公司的公开投资组合评论 | 确认经济权益及 Lux 是否持有董事席位 |
| Elad Gil | 天使 / 个人投资人,2025 年 10 月和 2026 年 4 月轮次 | 两轮公开融资均重复参与 | 确认持股规模及任何顾问角色 |
| Latham & Watkins LLP | Applied Compute 外部法律顾问,2026 年 4 月轮次 | 顾问关系;据已审阅来源,并非股权持有人 | 确认持续法律服务范围,以及该律所是否为更早轮次提供建议 |
| Greenoaks / Neo / Hanabi Capital | 2026 年 4 月 Series B 参与方 | 在 2026 年 4 月轮次中与 Kleiner Perkins 同投的少数股参与方 | 确认相对出资规模及是否参与更早轮次 |
根据点名各方角色的融资公告和法律顾问来源整理;“控制或经济重要性”仅反映来源所述内容(领投 / 参与状态、重复参与),不代表已验证股权结构百分比,后者未披露。
[CO018, CO021, CO024, CO026, CO027, CO011]1.4 规模、已披露客户与封面指标缺口
Applied Compute 在自己的案例研究和融资公告中点名了三家企业客户:Cognition(Devin AI 软件工程师和 Windsurf IDE 的开发商)、DoorDash 和 Mercor。与 Cognition 的合作中,Applied Compute 共同开发了「SWE-check」,这是一个实时漏洞检测模型,原生嵌入 Windsurf;Cognition 自己的博客称,该模型大约比其替换的前沿模型(Opus 4.6)快 10x,同时在同分布评测中把准确率差距收窄到接近前沿模型水平。与 DoorDash 的合作中,Applied Compute 构建了一个校准过的自动评分器,并训练了一个强化学习模型,用来纠正商户入驻时 AI 生成菜单中的错误;独立 LLMOps 案例库(ZenML)证实,该方案让低质量菜单相对减少 30%,并已覆盖美国全部菜单流量。不过,同一篇独立文章也提醒,其叙述「由 Applied Compute 提供,因而自然会正面呈现其工具」,这种宣传偏差警示需要延续。与 Mercor 的合作中,Applied Compute 定制训练的「Applied Compute: Small」模型在 Mercor 的 APEX-Agents 排行榜公司法类别中排名第 1、总榜第 4;这是截至 2026 年 2 月案例研究的结果,领先 Opus 4.5 和 GPT-5.2,但依据的是公司托管材料,未独立复核。Applied Compute 自己的融资文章提到正在与「F500 中的企业」合作,暗示三家之外还有客户,但所审阅来源没有披露任何其他名称。公司未披露员工数;RocketReach 估计截至 2026 年中为 21-29 人。所审阅来源没有披露收入、ARR,或三家点名部署之外的客户数量。[CO030, CO031, CO032, CO033, CO034, CO035]
| 指标 | 数值或状态 | 日期 | 置信度 | 缺口 |
|---|---|---|---|---|
| 估值 | $1.3 billion | 2026-04-08 | 高 | |
| 累计披露融资 | $160 million | 2026-04-08 | 高 | |
| 种子轮估值 | ~$100 million | 2025(年中,据 Upstarts Media) | 中 | 一级文件未确认确切交割日期;二级来源将同一轮融资定在 2024 年中 |
| 过渡期融资谈判估值(The Information) | ~$500 million | 2025-09(约) | 中 | 由二级引用报道;公司未确认 |
| 过渡期已披露融资估值 | ~$700 million | 2025-10-29 | 中 | 公司在 2025 年 10 月公告中未披露估值;该数字为第三方估计 |
| 已点名企业客户 | 3(DoorDash、Cognition、Mercor) | 2025-10-29 | 高 | 公司提到更广的 “F500” 工作,但未点名其他客户 |
| 员工人数 | 21-29(第三方估计) | 2026(年中) | 低 | Applied Compute 未在任何已审阅来源中直接披露 |
| 总部 | San Francisco, CA 总部(251 Rhode Island St #207) | 2025-10-10 | 高 | |
| 注册 | 特拉华州实体;加州外州注册于 2025-10-10 提交(doc. B20250336266) | 2025-10-10 | 高 | |
| 收入 / ARR | 低 | Applied Compute 或任何已审阅来源均未披露 | ||
| 成立日期 | 2025(据报为 2025 年 5 月) | 2025 | 中 | 确切月份未经公司一手来源确认 |
数值尽可能来自 Applied Compute 自身披露(accessDate 2026-07-05);只有二级、第三方估计或相互冲突来源时标为中 / 低置信度。null 表示本轮抓取的任何来源均未找到该指标。
[CO003, CO004, CO024, CO025, CO018, CO020]截至 2026 年 4 月 B 轮和 2026 年 7 月独立确认时的核心规模指标。
员工数来自第三方估计(RocketReach),并非公司披露;估值跃升由约 $100M 种子轮估值与 2026 年 4 月 $1.3B 估值对比得出,两者来自不同来源,而非同一连续备案。
[CO024, CO025, CO030, CO040, CO003, CO029]1.5 里程碑时间线与负面信号
Applied Compute 的时间线从 2025 年离开 OpenAI 并创立公司开始,随后是 2025 年中种子轮、2025 年 9 月融资洽谈泄露、2025 年 10 月 10 日 California 公司备案、2025 年 10 月 29 日退出隐身状态的融资、2026 年 1 月新一轮独角兽级融资洽谈报道、2026 年 2 月 Mercor 基准结果、2026 年 4 月 8 日 Kleiner Perkins 领投的 $1.3 billion Series B、2026 年 5 月 Cognition 产品案例研究,以及 2026 年 7 月 TechCrunch 对独角兽状态的独立确认。除了尚未调和的高管头衔冲突(California 备案中 Rhythm Garg 为 CFO/Secretary,而 Comcast LIFT Labs 中为 CTO)和种子轮日期冲突(StartupsUnion 的「June 2024」与其他来源证实的 2025 年创立相冲突)之外,本轮所审阅来源没有找到任何点名 Applied Compute 的诉讼、监管执法、裁员或安全事件披露;鉴于公司非常年轻且为私营公司,本轮法律数据库覆盖有限,这里记录为「缺乏证据」发现,而不是确认其记录清白。来自一手报道的最清晰负面信号不是事件型,而是结构型:不到一年估值约涨 13x;在 $1.3 billion 估值下完全没有披露收入或 ARR;治理披露(高管头衔、员工数)彼此冲突,或完全依赖第三方估算而非公司声明。[CO001, CO018, CO021, CO023, CO024, CO025]
| 日期 | 事件 | 类型 | 金额 / 估值 / 状态 | 参与方 | 含义 |
|---|---|---|---|---|---|
| 2025(据报为 5 月) | Yash Patil、Rhythm Garg 和 Linden Li 离开 OpenAI,创办 Applied Compute | 成立 | Yash Patil、Rhythm Garg、Linden Li 三位创始人 | 确立创始团队和本报告通篇采用的 2025 年起点 | |
| 2025(年中,当时未公告) | Applied Compute 完成 $20 million 种子轮融资 | 融资 | $20M 融资,~$100M 估值 | Benchmark(Victor Lazarte)、Sequoia、Conviction、Hanabi Capital、Definition、Zach Frankel 等投资方 | 第一笔机构资本;为后续估值跃升设定基线 |
| 2025-09(约) | The Information 报道 Applied Compute 正洽谈以 ~$500M 估值融资 | 融资 | ~$500M 估值(报道,未获公司确认) | n/a | 公司公开亮相前,估值快速上行的第一信号 |
| 2025-10-10 | Applied Compute, Inc. 在加州州务卿处注册为外州股份公司 | 治理 | 备案(document B20250336266) | Applied Compute, Inc. | 建立可被监管机构看到的公司记录,用于交叉核对高管头衔和总部地址 |
| 2025-10-29 | Applied Compute 走出隐身模式,并宣布 $80 million 融资 | 融资 | $80M 融资;~$700M 估值(据独立媒体) | Benchmark、Sequoia、Lux Capital、Hanabi、Neo、Definition、Elad Gil、Victor Lazarte、Omri Casspi 等投资方 | 首次公开融资公告;同时披露 DoorDash、Cognition、Mercor 为客户 |
| 2025-10-29 | DoorDash、Cognition 和 Mercor 被点名为早期客户 | 合作 | 3 个具名客户 | DoorDash、Cognition、Mercor | 确立 Applied Compute 的初始商业牵引力和参考客户 |
| 2026-01-20 | The Information 报道新一轮融资谈判估值为 $1.3 billion | 融资 | ~$1.3B 估值(报道) | Kleiner Perkins(据报道为潜在领投) | 创立不到一年便出现独角兽级融资信号 |
| 2026-02-24 | Mercor 案例研究发布;Applied Compute:小模型在 APEX-Agents 榜单的公司法项目排名 #1 | 产品 | #1 公司法排名,总榜第 4 | Applied Compute、Mercor | 展示相对于前沿实验室自有模型的竞争性模型表现 |
| 2026-04-08 | Applied Compute 宣布由 Kleiner Perkins 领投、投后估值 $1.3 billion 的 $80 million Series B | 融资 | $80M 融资;$1.3B 估值;累计融资 $160M | Kleiner Perkins、Elad Gil、Lux、Greenoaks、Neo、Hanabi 等投资方 | 独角兽里程碑;约一年内估值较 2025 年种子轮跃升约 13 倍 |
| 2026-05-11 | Applied Compute 和 Cognition 发布 SWE-check 案例研究及产品博客文章 | 产品 | bug 检测速度比 Opus 4.6 快 10 倍 | Applied Compute、Cognition | 展示联合训练专精模型的生产部署 |
| 2026-07-05 | TechCrunch 独角兽追踪器将 Applied Compute 列为 2026 年新晋独角兽,估值 $1.3 billion | 规模 | $1.3B 估值;$160M 融资(据 PitchBook) | n/a | 对独角兽状态的独立、同期确认 |
时间线根据 Applied Compute 自有新闻稿 / 博客材料、融资法律和媒体报道、加州州务卿文件,以及本轮抓取的独立融资和独角兽追踪报道重建;内部创立前活动和任何未披露不利事件必然未纳入。
[CO001, CO002, CO018, CO020, CO021, CO023]Applied Compute 自 2025 年离开 OpenAI 并创立,到 2026 年 7 月获得独立独角兽确认的时间线。
创立月份、种子轮时间和 2025 年 9 月融资谈判日期均为近似值,是根据二手报道拼合而来,而非来自单一带日期的一手来源;另一家二手来源将种子轮日期列为早一年(“June 2024”),本报告记录这一差异,但未采用。
[CO001, CO018, CO020, CO021, CO023, CO024]1.6 图表
02市场分析
2.1 市场边界:什么算企业 AI 智能体
Applied Compute 竞争的不是宽泛的「所有 AI」类别,而是企业 AI 智能体和面向内部业务工作流的智能体软件。公司自己的材料把机会表述为弥合前沿模型原始能力与单家公司特定工作流中实际效用之间的「overhang」;做法依赖两类前线部署角色——Forward Deployed Engineers(FDE),负责搭建评测框架和生产环境;Applied Research Engineers(ARE),负责训练和调优底层模型。因此,Applied Compute 落在三个相互重叠的支出类别里:企业智能体平台与编排、强化学习后训练和评测框架基础设施、前线部署式定制劳动力。同样的前线部署模式由 Palantir 在十多年前开创,现在已产品化为 Palantir Foundry 内的 AI 智能体,并正被大规模复制:AWS 在 2026 年承诺投入 $1 billion,组建专门的、智能体优先的 Forward Deployed Engineering 单元,服务 NFL、NBA、Cox Automotive 和 Southwest Airlines 等具名客户。本边界排除两类相邻替代品:没有评测框架或 RL 定制层、按 token 计费的无差异通用模型 API 消耗;以及早于智能体 AI 出现的既有工作流 / RPA / CRM 软件。企业很少只选一条路:65% 的受访技术领导者称采用「自建 + 采购」混合架构,只有约 10% 仅依赖供应商。这说明内部自建不是边缘情况,而是 Applied Compute 必须逐单替代的现实选择。[CM001, CM002, CM003, CM004, CM005, CM006]
| 细分 / 类别 | 纳入支出 | 排除支出 | 买方 / 付款方 | 与 Applied Compute 的相关性 |
|---|---|---|---|---|
| 企业智能体平台与编排 | 智能体构建平台、编排 / 运行时、智能体原生应用套件(例如 Agentforce 类产品) | 无工作流写入权限的消费者聊天机器人和个人效率副驾驶工具 | 业务线负责人 + IT | Applied Compute 智能体部署进入的核心应用层 |
| RL / 后训练基础设施与评测框架 | 强化学习微调、奖励模型 / 评分器构建、任务专属评测框架 | 通用基础模型预训练算力和前沿实验室 R&D 本身 | 数据科学 / ML 平台团队,通常与供应商共用 | Applied Compute 声称的核心方法(RL 加定制评测框架) |
| 前线部署定制与嵌入式工程服务 | 按项目或包月顾问费计费的现场工程人力(FDEs/AREs、AWS FDE、Palantir FDE) | 没有 ML / 智能体专业化的通用系统集成商人力派遣 | 业务线预算加采购 / 专业服务 | Applied Compute 将其 FDE/ARE 结构营销为核心差异化 |
| 应用层垂直智能体(法律、编码、市场运营、招聘) | 嵌入单一工作流的领域专属智能体(Harvey、Cognition、DoorDash、Mercor) | 没有工作流集成的横向、无差异聊天助手 | 领域职能负责人(总法律顾问、工程负责人、运营负责人、产品负责人) | Applied Compute 四个已披露具名案例研究所在位置 |
| 排除:通用模型 API 消费 | 不计入核心细分 | 按 token 计费、没有智能体评测框架或 RL 定制层的原始 LLM API 使用 | IT / 财务,作为转嫁型基础设施成本 | 相邻替代供应商可在其上构建,但不属于核心智能体供应商收入 |
| 排除 / 相邻:既有工作流与 RPA/CRM 软件 | 不计入核心细分 | 早于智能体式 AI 的规则型自动化和案件管理软件许可费 | 业务线软件预算 | 现状替代品和渠道竞争者;在竞争者章节覆盖 |
| 相邻:AI 算力 / 数据中心基础设施 | 不计入核心细分 | 支撑所有 AI 工作负载的 GPU 资本开支、电力和机房托管支出 | 基础设施 / 云预算 | 整个品类的资本强度约束,而非 Applied Compute 特定收入 |
边界受证据约束,来自 Applied Compute 自身定位和第三方规模测算方法;“纳入 / 排除”反映被引用分析机构对自身估算的口径,而不是经审计的市场研究分类。
[CM001, CM002, CM003, CM005, CM007, CM009]2.2 市场规模:多重视角,但没有单一可靠 TAM
这个市场没有一个宽口径 TAM 数字可靠,因此规模判断必须叠加几个受证据约束的视角。MarketsandMarkets、Grand View Research、Precedence Research 和 Fortune Business Insights 四家独立市场研究机构,都把 2025-2026 年全球独立「AI 智能体」软件市场规模放在大约 $7.3-8.5 billion 区间,但对轨迹判断分歧很大:MarketsandMarkets 预计 2030 年达到 $52.62 billion(46.3% CAGR),Grand View 预计 2033 年达到 $182.9 billion(49.6% CAGR),Precedence 预计 2035 年约 $294.66 billion(43.57% CAGR)。再往上一层,是一个更大、口径不同的估算:Gartner 把 2026 年嵌入所有企业软件中的智能体 AI 能力支出定为 $201.9 billion——不只计算独立智能体供应商——约为独立口径数字的 25 倍,因为它把现有应用里的内嵌助手和智能体功能也算进去。这个差异是计量边界问题,不是事实冲突;Gartner 自己对 2026 年全球 AI 总支出的估计又进一步放大了边界复杂性:总额 $2.52 trillion(同比增长 44%,约 $1.37 trillion 为基础设施、$452.5 billion 为软件、$588.6 billion 为服务),其中「agentic AI」子类别预计以 119% CAGR 从约 $15 billion 增长到 2029 年接近 $753 billion。Applied Compute 自己可触达的切片——企业智能体平台 + RL/后训练基础设施 + 前线部署定制——位于独立到嵌入口径之间,但没有独立第三方估算把它单独拆出来;这个缺口应保留为明确尽调问题,而不是用编造数字填补。[CM011, CM012, CM013, CM014, CM015, CM016]
| 发布方 | 年份 | 地域 | 数值 | 复合年增长率(CAGR) | 方法 | 置信度 | 局限 |
|---|---|---|---|---|---|---|---|
| MarketsandMarkets | 2025 -> 2030 | 全球 | $7.84B -> $52.62B | 46.3% | 围绕独立 AI 智能体软件细分市场,结合二手研究与专家访谈 | 中 | 一手研究由供应商提供;方法未经独立审计 |
| Grand View Research | 2025 -> 2033 | 全球 | $7.6B -> $10.9B (2026) -> $182.9B | 49.6% (2026-2033) | 在智能体系统与技术细分市场中,自下而上与自上而下交叉校验 | 中 | 八年预测期会放大终值不确定性 |
| Precedence Research | 2025 -> 2035 | 全球 | $7.92B -> $11.55B (2026) -> ~$294.66B | 43.57% | 细分市场交叉校验方法,类似同类市场研究机构 | 中 | 十年周期;独立市场同行中给出的期末市场规模最大 |
| Deloitte TMT Predictions(汇编) | 2026 -> 2030 | 全球 | $8.5B -> $35B-$45B | 未披露 | Deloitte 的科技 / 媒体 / 电信预测团队,通过独立预测追踪器引用 | 中 | 通过二级汇编引用,而非 Deloitte 原始报告 |
| Fortune Business Insights(汇编) | 2025 -> 2034 | 全球 | $7.29B -> $139.19B | 40.5% | 通过独立预测追踪器引用的第三方估算汇编 | 低 | 未独立获取原始报告;完全依赖二级聚合方 |
| Gartner(独立智能体 AI 支出,汇编) | 2025 -> 2029 | 全球 | $15B -> $753B | 119% | Gartner 在其整体 AI 市场模型中单列的智能体 AI 支出子类 | 中 | 通过二级汇编引用,并非 Gartner 原始研究简报 |
| Gartner(广义嵌入智能体 AI 的企业软件) | 2026 | 全球 | $201.9B | 点估计 | 统计所有企业软件中嵌入的智能体能力,而非独立智能体厂商 | 中 | 为独立估算的 ~25x;市场边界不同,不能直接比较 |
所有数值都是第三方分析师 / 市场研究机构估算,且口径不同(独立智能体软件 vs. 更广义企业软件中嵌入的智能体能力);各行不能相加,也不应跨发布方求和。
[CM011, CM012, CM013, CM014, CM015, CM016]从全球 AI 总支出逐层收窄到 Applied Compute 有证据约束的可服务层。
这是分层视角,不是严格的 TAM-SAM-SOM 漏斗,因为底层来源的市场边界不同,也没有任何来源单独隔离 Applied Compute 实际销售所处的精确层级。
[CM018, CM016, CM012, CM014, CM020]2026 年独立 AI 智能体软件市场规模的低 / 基准 / 高分析师估算,统一采用十亿美元口径。
2030 年行的中点取 Deloitte 公布的 $35-45B 区间算术中心,并与 MarketsandMarkets 的点估计并列展示;两行都采用同一单位(独立市场规模,十亿美元),因此可彼此比较,但不能与 Gartner 口径不同的 $201.9B 数字直接比较。
[CM014, CM012, CM013, CM011, CM020]2.3 买方、用户与付款方分层
Applied Compute 已披露的四个企业部署横跨四个不同垂直领域,每个项目都由领域职能负责人推动,而不是同一个横向买方画像:法律(Harvey,双方共同后训练的模型在 Harvey 的 1,250 任务 Legal Agent Benchmark 上超过 Opus 4.8 Max 和 GPT-5.5 xhigh)、软件工程(Cognition,SWE-check 漏洞检测智能体运行速度约为其替换前沿模型的 10x)、市场运营(DoorDash,强化学习训练的评分器纠正 AI 生成商户菜单),以及劳动力市场评测(Mercor,按公司托管、未独立复核的案例研究,Applied Compute 自己的小模型在 Mercor 的 APEX-Agents 排行榜公司法类别第 1、总榜第 4)。这类部署的预算所有权正在从中央 IT 转移:一项 2026 年对 266 名 Fortune 50-Global 2000 技术领导者的调查发现,业务线负责人已成为最大的 AI 工具购买群体,占 46%,首次追平或超过 CIO(38%)和 CTO(38%)。典型采用路径仍然要先通过采购和治理关口才会承诺预算——84% 的企业把安全 / 合规签核视为不可谈判条件,70% 希望先有自助沙盒试用——这有利于 Applied Compute 这类能把工程师直接嵌入工作流负责人身边的供应商,而不是只向中央 IT 销售标准化产品。[CM036, CM037, CM038, CM039, CM040, CM041]
| 细分市场 / 垂直领域 | 买方 | 用户 | 付款方 | 工作流 | 预算负责人 | 采用触发因素 |
|---|---|---|---|---|---|---|
| 法律服务 | 总法律顾问 / 法务运营负责人(Harvey 企业客户) | 律师和法律助理 | 律所或法务部门预算 | 合同审查、法律研究和起草智能体,并用评分基准打分 | 业务组 / 业务线预算 | 基准测试验证了相对前沿模型的准确率(Harvey 的 Legal Agent Benchmark 结果) |
| 软件工程 / 开发者工具 | AI 原生工具厂商的工程负责人(Cognition) | 在 IDE 内工作的开发者 | 工程 / 产品预算 | 原生嵌入 IDE 的实时代码审查和 bug 检测 | 产品工程预算 | 需要前沿级推理,同时满足实时延迟和成本 |
| 市场平台运营(外卖) | 商户运营 / ML 平台负责人(DoorDash) | 入驻平台的商户 | 中央运营 / ML 预算 | 商户入驻时的菜单准确性评分与纠错 | 中央运营预算 | 在不按比例加人的情况下提升入驻准确率 |
| 招聘 / 劳动力市场评估 | 劳动力市场平台的产品与评测负责人(Mercor) | 人才评估员和平台匹配引擎 | 产品 / ML 预算 | 特定领域智能体基准测试和排行榜排名(APEX-Agents) | 产品工程预算 | 需要在任务特定评测排行榜上拉开差距 |
| 广义 F500 企业(通用) | CIO/CTO 与业务线负责人混合 | IT、运营、财务、客户体验等一线员工 | 业务线预算影响力上升(46%),CIO/CTO 各为 38% | 通过前置部署项目交付跨职能工作流自动化 | 业务线预算占比相对中央 IT 上升 | 前置部署工程进入主流(AWS 的 $1B 团队、Palantir 的 AI FDE),降低集成摩擦 |
前四行是 Applied Compute 自己披露的具名部署;第五行来自 2026 年 Mayfield CXO 调查的全市场综合样本,并非 Applied Compute 专属数据点。
[CM037, CM040, CM041, CM042, CM043, CM044]企业智能体工作流中的预算所有者、买方、用户和采购关卡关系。
这些关系是具名案例研究与 2026 年 Mayfield CXO 调查的定性综合,不是单一来源图。
[CM037, CM039, CM044, CM038]2.4 增长驱动因素与采用约束
采用率正在多个方向同时加速。Gartner 预计,到 2026 年底,最多 40% 的企业应用会搭载任务特定智能体(2025 年不到 5%);McKinsey 发现,88% 的组织现在报告至少在一个业务职能中经常使用 AI;一项 2026 年 CXO 调查则显示,42% 的企业已经在生产环境运行智能体 AI(生产 + 试点合计 72%)。AWS 和 Palantir 的前线部署工程模式走向主流,本身也是驱动因素:它验证了 Applied Compute 的 FDE/ARE 结构已经成为行业模式,但也表明超大规模云厂商可能会直接争夺同一批客户。约束同样具体。Deloitte 称,只有五分之一公司具备成熟的智能体治理模型;Gartner 预计,到 2027 年,40% 的企业会在生产中暴露治理缺口后,将自主智能体降级或下线。数据就绪度连续第五年位居受访 CXO 提到的首要阻碍(58%);投资回报门槛也很苛刻:Gartner 预计,到 2027 年,超过 40% 的智能体 AI 项目会被取消;MIT 相关 NANDA 研究发现,95% 的企业生成式 / 智能体 AI 试点无法展现可衡量的 P&L 影响,而与供应商合作的部署成功率约为纯内部自建的两倍。所有这些之下,还叠加了任何重 RL 供应商都会面对的资本强度约束:IEA 预计,到 2030 年,数据中心用电量将几乎翻倍至约 945 TWh;Goldman Sachs 预计,到 2030 年,全球数据中心电力需求将比 2023 年上升 165%。这种成本和供给背景会影响 Applied Compute 训练和评测工作所依赖算力的价格与可得性。[CM021, CM022, CM023, CM024, CM025, CM026]
| 驱动 / 约束 | 方向 | 时间 | 影响 | 尽调问题 |
|---|---|---|---|---|
| 生产阶段采用加速 | 驱动 | 当前(2026) | Mayfield 显示 42% 企业已在生产环境运行智能体;McKinsey 显示经常使用 AI 的比例升至 88% | 核实 Applied Compute 自身生产环境客户数(相对于试点客户),尤其是三个完全具名部署之外的客户 |
| 预算向业务线买方重新分配 | 驱动 | 当前至 2026 年 | 业务线负责人(46%)的预算影响力已持平或超过 CIO(38%),企业销售摩擦降低 | 确认 Applied Compute 合同由谁签署:CIO、业务线,还是两者共同签署 |
| 前置部署工程模式进入主流 | 驱动兼约束 | 2026 | AWS 投入 $1B,Palantir 嵌入 FDE 智能体工具,验证 FDE/ARE 模式,但也引入超大规模云厂商竞争 | 评估 AWS 或 Palantir FDE 团队是否直接争夺 Applied Compute 的目标客户 |
| 治理与监督缺口 | 约束 | 持续 | Deloitte 显示只有 1/5 公司拥有成熟智能体治理模型;Gartner 预计到 2027 年,40% 企业会降级或停用智能体 | 尽调 Applied Compute 自身治理与护栏工具的成熟度 |
| 数据就绪瓶颈 | 约束 | 持续 | Mayfield 调查中 58% CXO 将数据就绪度和质量列为采用智能体 AI 的最大阻碍 | 评估 Applied Compute 前置部署人力有多少被数据 / 集成工作吃掉,而不是用于模型训练 |
| 试点转生产失败与 ROI 证明压力 | 约束 | 当前至 2027 年 | MIT NANDA:95% 企业 GenAI / 智能体试点未能体现 P&L 影响;Gartner:到 2027 年,40%+ 智能体项目会被取消 | 要求 Applied Compute 提供客户级 ROI 证据,不能只依赖三份完整记录的案例研究 |
| 算力 / 电力资本强度 | 约束 | 2025 - 2030 | IEA 预计到 2030 年数据中心用电需求将接近翻倍;Goldman 预计全球数据中心电力需求到 2030 年增长 165% | 厘清 Applied Compute 的算力采购模式(自有或云租 GPU)及其对算力成本通胀的暴露 |
方向标签反映截至本报告生成日引用证据的总体平衡;若干项目(例如前置部署模式)既推动采用,也会通过竞争加剧构成约束。
[CM021, CM025, CM028, CM029, CM038, CM045]Gartner 将企业智能体 AI 演进分为五个阶段:从嵌入式助手到全面智能体化的“新常态”。
数值取自 Gartner 针对阶段 2、3、4、5 报告的具体百分比;阶段 1 的数值是示意性占位(“大多数企业应用”),因为 Gartner 没有公布该阶段的精确百分比。
[CM021, CM022, CM023, CM024]2.5 图表
03竞争对手
3.1 格局:直接同行、平台既有玩家与自建替代
企业提到「AI 智能体」时常放在同一组里的大多数公司,并不是 Applied Compute 面对的同一购买决策。Applied Compute 自己的公开材料把公司定位为强化学习后训练和评测基础设施伙伴:它为 Harvey、Cognition 和 DoorDash 等具名客户训练专有、任务调优模型,再由这些客户以自己的品牌推出相应智能体。这样一来,实际竞争集异常分层。直接垂直同行——Harvey(法律)、Glean(企业搜索 / 智能体)、Sierra 和 Decagon(客服)、Cognition/Devin(软件工程)——各自在特定领域拥有面向终端用户的品牌智能体产品。平台既有玩家——OpenAI 的 Frontier/AgentKit 技术栈、Salesforce 的 Agentforce、Microsoft 的 Copilot Studio——则把智能体构建工具打包进企业已经付费的开发者平台、CRM 席位或 Microsoft 365 订阅。一个公开创业公司聚合器还列出几十家规模更小、阶段更早的进入者(LangChain、Agent Bricks、Manus AI 等)围绕同一广义空间展开竞争,尽管它们都没有本章所分析直接同行的融资规模或具名客户证据。 本章最重要的细节是竞合:Harvey 同时是 Applied Compute 已披露的合作客户,用于其法律智能体的强化学习后训练;又是独立运营的垂直既有玩家,直接向律所销售自己的 Assistant/Vault/Knowledge 套件。同样的前线部署基础设施与品牌产品之间的区分,也出现在 Cognition 和 DoorDash 身上。与此同时,Sierra 管理层称其构建在「一组」第三方基础模型之上,并叠加自有微调专有层——这是一条内部自建替代路径;如果扩散,即便企业没有打造竞争性品牌产品,也会减少对 Applied Compute 这种外部 RL 后训练专家的需求。[CP001, CP002, CP003, CP004, CP008, CP010]
Applied Compute 在品牌化终端产品所有权上偏低,在专业化深度上跨垂直;直接垂直同行则在产品所有权和自身领域深度上都更高。
坐标位置是基于各实体披露产品范围和案例研究做出的证据型序位判断,不是标准化第三方指数;应按方向性理解,而非精确坐标。
[CP002, CP006, CP008, CP010, CP014, CP018]3.2 竞争对手画像:规模、融资、目标客户与战略方向
同一 12 个月窗口内,每个具名直接同行都披露了比 Applied Compute 更大、更新的融资轮。Harvey 在 2026 年 3 月按 $11B 估值融资 $200M,由 GIC 和 Sequoia 联合领投,并称其律所和法律部门客户运行了超过 25,000 个定制智能体。Sierra 在 2026 年 5 月按 $15.8B 估值融资 $950M,是本组最大一轮;公司在八个季度内 ARR 突破 $150M,并拥有超过 40% 的 Fortune 50 作为客户;创始人 Bret Taylor 称 Sierra 的收入比第二大竞争对手「大出数倍」。Decagon 在 2026 年 1 月完成 $250M Series D,估值增至 $4.5B,是六个月前 $131M Series C($1.5B 估值)的三倍,并列出 Avis、Hertz、Block、Affirm、Duolingo 和 Oura 等具名客户。Devin 编码智能体的开发商 Cognition 于 2025 年 9 月按 $10.2B 估值融资 $400M,披露 ARR 在一年内从约 $1M 升至 $73M;同一季度还伴随裁员报道和高强度工作时长预期。Glean 完成 $150M Series F,估值 $7.2B,并将自己定位为横向构建 / 部署 / 编排层,而不是单一垂直楔子。 三家平台既有玩家竞争的是分发,而不是垂直深度。OpenAI 目前报告企业收入占总收入超过 40%,并在 AgentKit 开发者工具之外推进跨系统「Frontier」编排层,背后还有与 McKinsey、BCG、Accenture 和 Capgemini 的「Frontier Alliances」合作。Salesforce 的 Agentforce 在 2026 财年 Q4 达到约 $800M ARR(同比增长 169%),已完成超过 29,000 笔交易,不过超过 60% 的预订来自现有客户扩张,而不是净新客户。Microsoft 的 Copilot Studio 已被超过 230,000 个组织采用来构建定制智能体,并作为 Microsoft 365 和 Teams 订阅中的低代码附加项分发。独立买方指南指出,即便在两个最接近的客服同行中,Decagon 更偏向客户自有工程集成,Sierra 更偏向完全托管、前线部署的交付模式——差异在交付机制,而不是底层模型技术。[CP007, CP006, CP012, CP011, CP016, CP017]
| 竞争对手 | 类别 | 规模 / 融资 | 目标细分市场 | 差异化 | 局限 |
|---|---|---|---|---|---|
| Applied Compute | RL 后训练 / 智能体基础设施(标的公司) | 据其 2026 年 4 月融资披露,估值 ~$1.3B(见公司概况);远小于下列同行 | 需要为自有智能体做定制 RL 后训练的企业(Harvey、Cognition、DoorDash、Mercor) | 通过前置部署的 RL / 评测基础设施训练客户自有模型,而不是交付一个带品牌的终端用户智能体 | 未披露公开定价,具名客户基数小,且部分具名客户本身资金充裕,未来可能内建 |
| Harvey | 直接垂直同行(法律) | 2026 年 3 月以 $11B 估值融资 $200M | 律所和企业法务部门 | 统一的 Assistant / Vault / Knowledge 套件;披露客户运行的自定义智能体超过 25,000 个 | 未公开标价;同时披露为 Applied Compute 的 RL 后训练客户,使纯竞争对手定位变得模糊 |
| Glean | 直接横向同行(企业搜索 / 智能体) | Series F 轮融资 $150M,估值 $7.2B | 大型企业内跨职能知识工作者 | 定位为在公司既有知识和应用栈之上构建、部署、编排智能体的一层 | 横向覆盖更广,可能意味着在任一单一领域的垂直深度不及 Harvey、Sierra 或 Decagon |
| Sierra | 直接垂直同行(客户服务) | 2026 年 5 月以 $15.8B 估值融资 $950M | 大型消费品牌和 Fortune 50 企业 | 托管式前置部署交付;8 个季度 ARR 超过 $150M;Fortune 50 中超过 40% 是客户 | 托管服务模式把迭代控制权交给 Sierra 自有团队,而不是买方工程师 |
| Decagon | 直接垂直同行(客户服务) | 2026 年 1 月 Series D 融资 $250M,估值 $4.5B | 面向消费者的旅游、金融科技、医疗和零售品牌 | "Agent Operating Procedures" 让 CX 团队用自然语言写逻辑,再编译成受治理的代码 | 独立买方指南认为,相比 Sierra,它需要更多客户侧工程集成 |
| Cognition / Devin | 直接垂直同行(软件工程) | 2025 年 9 月以 $10.2B 估值融资 $400M;Devin ARR ~$73M(2025 年 6 月) | 需要自主编码智能体吞吐量的工程团队 | 在所审阅同行中自主性最高的编码智能体;沙盒化云开发环境 | 2026 年独立测试发现,只有边界清晰的 bug 修复成功率 ~78%,且安全审查缺口未解决;2025 年据报道出现裁员和高压文化 |
| OpenAI(Frontier / AgentKit) | 平台在位者 | 企业业务占总收入 >40%,目标到 2026 年底与消费者业务持平 | 已基于 GPT 模型构建的企业;Oracle、State Farm、Uber 是具名 Frontier 客户 | 由单一供应商打包模型、AgentKit 工具,以及跨系统 Frontier 编排层 | 企业智能体编排比 Sierra / Decagon / Harvey 的垂直产品更晚,纵深上的验证也更少 |
| Salesforce(Agentforce) | 平台在位者 | Agentforce ARR ~$800M(同比 +169%);FY26 Q4 与 Data 360 合计 $2.9B | 销售、服务和行业云中的现有 Salesforce CRM 客户 | 绑定进 150,000+ 客户安装基数的续约;已成交 29,000+ 笔 Agentforce 交易 | 60%+ 预订来自现有客户扩张,而非从专业厂商手中赢得的净新增客户 |
| Microsoft(Copilot Studio) | 平台在位者 | 2026 年初,230,000+ 组织在构建智能体;15M+ 付费 M365 Copilot 席位 | Microsoft 365 和 Teams 企业客户 | 低代码智能体构建器直接嵌入企业已经续费的订阅 | 采用数据把智能体构建器使用量与更广义 Copilot 席位数混在一起,难以拆出净增智能体牵引力 |
| 内部构建 / 通用模型栈 | 替代方案 / 现状 | 不适用 -- 成本是企业自己的工程预算 | 拥有内部 ML / 工程团队的企业(例如 Sierra 自身的 "constellation of models" 做法) | 完全掌控模型选择、微调和路线图,不依赖供应商 | 需要企业内部具备大多数企业缺乏的 RL / ML 专长,而这正是 Applied Compute 与垂直同行共同销售时瞄准的缺口 |
融资和估值数字均取各厂商截至本报告生成日的最新公开披露;Applied Compute 估值从公司概况重述用于比较,未在本节重新推导。
[CP005, CP007, CP009, CP012, CP016, CP018]3.3 能力、定价、GTM / 分发与信任姿态对比
包括 Applied Compute 在内,本章画像中的公司没有一家公布核心企业智能体产品的自助标价。Harvey、Glean 和 Sierra 都销售经济条款未披露的定制企业合同;独立 2026 年买方指南估计,Decagon 和 Sierra 合同通常从每年约 $95,000-$150,000 起,加入前线部署集成后可达到 $200,000-$350,000,不过这些是第三方估算,不是供应商公布的数字。Cognition 的 Devin 是所审阅产品中唯一披露定价机制的产品——基于使用量的 activity credit units——评论者称,这造成约 $500 / 月的成本下限,并让长时间任务的总成本不可预测。OpenAI 对旗舰 API 定价最透明,但与其他公司一样,大企业合同条款仍为定制且未披露。Salesforce 和 Microsoft 都把智能体定价折入既有 CRM 和 Microsoft 365 订阅包中,这让买方在评估独立专营供应商与既有玩家附加产品时,更难做同口径价格比较。 能力上,功能 / 能力矩阵显示出一致模式:垂直同行(Harvey、Sierra、Decagon、Cognition)披露了很强的领域专精,但底层模型 / RL 训练深度大多未披露;OpenAI 则披露了很强的模型深度,但在垂直领域上保持横向。分发和打包能力集中在三家平台既有玩家手中,它们可以把智能体工具附着到企业本来就会预算的续约中——Salesforce 自己披露 Agentforce 超过 60% 的预订来自扩张而非新 logo,就是这一动态的直接证据。就信任和监管姿态而言,所审阅来源没有披露覆盖全部九个实体且详细可比的合规 / 认证矩阵;本章把它视为开放证据缺口,而不是根据不完整公开材料推断排名。[CP036, CP035, CP034, CP031, CP021, CP027]
| 能力标准 | Applied Compute | Harvey | Sierra | Decagon | Cognition/Devin | OpenAI | Salesforce Agentforce | Microsoft Copilot Studio |
|---|---|---|---|---|---|---|---|---|
| 交付带品牌的终端用户智能体产品 | 否 | 是 | 是 | 是 | 是 | 部分(Frontier / AgentKit 是构建层) | 是 | 是 |
| RL / 后训练定制深度 | 强(核心产品) | 未知(从 Applied Compute 采购) | 未知(据 CNBC,拥有自有微调层) | unknown | unknown | 强(前沿实验室) | unknown | unknown |
| 垂直领域专精 | 跨垂直基础设施 | 强(法律) | 强(客户服务) | 强(客户服务) | 强(软件工程) | 无(横向) | 中(CRM 相邻工作流) | 无(横向) |
| 是否披露公开标价 | 否 | 否 | 否 | 否 | 部分(评测者报道 Devin 基于 ACU 的定价) | 部分(API 公开标价;企业条款定制) | 否(与 CRM 授权打包) | 否(与 M365 授权打包) |
| 披露的具名企业客户数 | 4 个具名案例研究 | 25,000+ 自定义智能体(汇总数,不是具名客户数) | Fortune 50 中 40%+(汇总数) | 2025 年新增 100+ 企业客户(汇总数) | unknown | 600,000+ 客户账户(ChatGPT Enterprise / Business) | 29,000+ 笔 Agentforce 交易 | 230,000+ 组织使用 Copilot Studio |
| 前置部署 / 托管工程交付模式 | 是(据公司概况,设有 FDE / ARE 岗位) | unknown | 是 | 是 | 否(自助式云智能体) | 部分(与咨询公司合作的 Frontier Alliances) | 部分(专业服务生态) | 部分(伙伴生态) |
标为「未知」的单元格表示所审阅来源未找到公开披露证据,并非假定不存在;汇总客户数不能直接跨厂商比较,因为各家披露单位不同(智能体、账户、交易或组织)。
[CP006, CP013, CP015, CP018, CP021, CP025]| 厂商 | 定价模式 | 价格 / 档位信号 | 包含能力 | 折扣或未知项 | 含义 |
|---|---|---|---|---|---|
| Applied Compute | 定制项目(隐含) | 未披露公开价目表 | RL 后训练、评测 / 评分基础设施、测试框架工程 | 完整定价结构未知;仅披露协作式案例研究 | 买方不直接接触销售,就无法把 Applied Compute 成本与同行对标 |
| Harvey | 定制企业合同 | 未披露公开价目表 | 覆盖法律工作流的 Assistant、Vault、Knowledge 模块 | 合同经济性未披露;只披露智能体运行汇总数 | 相对公开标价替代方案,买方只能在信息不透明下谈判 |
| Glean | 定制企业合同 | 未披露公开价目表 | 搜索、知识锚定,以及智能体构建 / 编排工具 | 所审阅来源未披露定价 | 不透明模式与其他垂直 / 横向同行相同 |
| Sierra | 托管式年度合同 | 独立买方指南估计为每年 $150K-$350K,含前置部署集成 | Agent Studio、Journeys 构建器、语音 / 品牌克隆、托管迭代 | Sierra 自身未发布标价;数字来自第三方估计 | 高价托管服务定价反映 Sierra 的前置部署交付模式,而不只是软件授权 |
| Decagon | 年度合同 | 独立买方指南估计每年 $95K-$200K+,取决于集成范围 | Agent Operating Procedures 平台、分流 / CSAT 工具 | 未发布标价;数字来自第三方估计 | 第三方指南中的入门估计低于 Sierra,符合更偏自助 / 工程团队主导的模式 |
| Cognition / Devin | 按使用量计费(活动信用单位) | 评测者称,有意义使用至少要每月 $500 | 自主编码智能体,沙盒化云端开发环境 | ACU 计价让长时间运行或复杂任务的总成本难以预测 | 按量计费相比固定按席位模式更容易带来预算不确定性 |
| OpenAI | API 用量 + 按席位企业层级 | GPT-5.5 API 公开标价;ChatGPT Business/Enterprise 席位另行计价 | 前沿编排、AgentKit 工具、ChatGPT Enterprise 席位 | 大型企业实际成交价 / 定制价未完全公开 | API 经济性透明,但大型定制企业交易不透明,和较小专业厂商相似 |
| Salesforce (Agentforce) | 作为 CRM 许可的捆绑加购项 | 已审阅来源未把 Data 360/CRM 捆绑价格单独列示 | Agentforce 360、Data 360、Sales/Service/Analytics 智能体 | 附加价格可能随既有 CRM 合同而变;未公开披露 | 捆绑定价让买家难以把它与独立智能体厂商做同口径价格比较 |
| Microsoft (Copilot Studio) | 与 Microsoft 365 Copilot 许可捆绑 | Copilot 按席位价格已公开;Copilot Studio 用量 / 消耗计价另计 | 低代码智能体构建器、连接器、治理工具 | 总混合成本取决于席位数加消耗量,不是一个公开单价 | 和 Salesforce 一样,捆绑让价格不透明,也更利好已有席位基础的既有厂商 |
Sierra/Decagon 的美元区间来自第三方买方指南估算,并非厂商公开标价;应把它看成价格层级的方向性信号,而不是精确合同金额。
[CP036, CP035, CP034, CP031, CP019, CP027]垂直同行(Harvey、Sierra、Decagon、Cognition)在领域专业化和 RL/模型深度信号上领先;超大规模云既有厂商(OpenAI、Salesforce、Microsoft)在分发 / 捆绑能力和披露规模上领先。
强弱序位标签是基于功能 / 能力矩阵表标准的证据综合,并围绕五个总结维度重新切分,而非沿用表中更细的逐项标准。
[CP002, CP006, CP011, CP015, CP031, CP021]3.4 切换成本、锁定、多供应商并用与分发能力
这个竞争格局中的切换成本来自三种不同机制,而不是单一来源。第一,前线部署或托管工程交付——Sierra、Decagon,以及(按公司概览)Applied Compute 自己的 FDE/ARE 模式都采用——会把供应商员工嵌入客户运营工作流,实际切换成本因此高于简单软件切换。第二,数据重力:Harvey 的 Vault 和 Knowledge 模块会在 Harvey 平台内存储并索引律所自己的文档和先例;切换意味着重新索引一家律所的机构知识,而不仅是换一个模型端点。第三,打包和分发能力:Salesforce 和 Microsoft 可以把智能体工具附着到客户每年本来就会签的 CRM 和 Microsoft 365 续约上,这种销售动作在结构上不同于每个独立专营供应商——包括 Applied Compute——必须赢下的净新采购决策,也更低摩擦。 模型层的多供应商并用,和应用层并不一样。Sierra 管理层称,其自有微调层之下运行着「一组」OpenAI 和 Anthropic 模型;这意味着 Sierra 的企业买方即便承诺使用 Sierra 的应用层,也没有被锁定在单一基础模型供应商上。正是同一个多供应商选项,让内部自建对部分企业成为可信替代:原则上,拥有内部 ML 人才的公司可以直接基于 OpenAI 或 Anthropic API,复刻任何垂直同行技术栈的轻量版本。OpenAI 自己的 Frontier 话术正是围绕这一张力展开,明确把能够「穿行于一家公司的系统和数据之间」的跨系统智能体,定位为比「嵌入单一产品内」的智能体更持久;Sierra 的 Bret Taylor 则把 Sierra 的 $950M 融资描述为一项防御性动作,要「积极投资」并守住对大量资金充足对手的领先——这等于隐含承认,即便是被报道的品类领导者,也把多供应商并用和被替代视为现实风险。[CP037, CP038, CP039, CP040, CP041, CP042]
3.5 护城河耐久性、商品化风险与竞争对手负面证据
Applied Compute 的核心护城河主张——深度强化学习后训练和评测基础设施能力——同时面临两个方向的直接内生化威胁。OpenAI 等超大规模云 / 模型厂商正在推动跨系统 Frontier 编排和自己的后训练研究预算;Sierra 等资金雄厚的垂直既有玩家也公开称,会构建自有微调专有层,而不是从外部购买这项能力。更复杂的是,Applied Compute 自己的证明点——Harvey、Cognition 和 DoorDash——本身正在向数十亿美元估值扩张,或已经跨过该门槛;这正好让这些被用来证明产品市场匹配的客户,在初始能力转移完成后,拥有将 RL 后训练内生化的资产负债表能力。资本差距很刺眼:Sierra、Harvey、Decagon 和 Cognition 在同一 12 个月内各自完成 $200M-$950M 单轮融资,估值是 Applied Compute 自己披露轮的 3x-12x,让它们每一家都有更多资本去争夺算力和研究人才。 独立证据也给整个品类,而不只是 Applied Compute,增加了更多谨慎信号。Gartner 预测,到 2027 年底,超过 40% 的智能体 AI 项目会因成本、价值或风险控制问题被取消——这是需求侧风险,会压缩本章所有供应商的可触达管道。2026 年对 Cognition Devin 的独立测试发现,即便是所审阅项目中资金最充足、自治程度最高的产品,也存在真实可靠性缺口:在边界清晰的 bug 修复上成功率约 78%,安全漏洞盲点未解决,并且记录到接近 $500 / 月的成本下限;Cognition 自身在 2025 年 8 月裁员约 30 人,并向另外 200 人提供买断,同时伴随高强度工作时长预期报道,尽管其估值同季度翻倍以上。Sierra 创始人也公开预测,约两年内市场会修正并出现「淘汰效应」。合起来看,Applied Compute 的基础设施层位置避开了部分正面产品竞争,但避不开资本强度、内生化和需求耐久性风险;这些风险已经出现在其直接客户和同行自己的披露中。仍有一个真正开放的证据缺口:所审阅来源没有说明 DoorDash 同时出现在 Applied Compute 案例研究和 OpenAI 企业客户名单中,代表的是重叠工作负载,还是完全不同的工作负载。[CP043, CP044, CP045, CP030, CP031, CP046]
| 护城河主张 | 威胁 | 严重性 | 缓解措施 / 尽调问题 |
|---|---|---|---|
| Applied Compute 的 RL 后训练和评测基础设施专长构成可防守的技术护城河 | 超大云厂商(OpenAI Frontier)和资金充足的垂直既有厂商(Harvey、Sierra)正在自建或收购相近的内部后训练能力 | 高 | 跟踪 Harvey、Sierra 或 Cognition 是否披露内部 RL/后训练招聘,或随着时间推移减少对外部 RL 厂商的依赖 |
| Harvey、Cognition、DoorDash 等具名标杆客户验证了 Applied Compute 的路径 | 这些客户本身估值正在迈向或已超过数十亿美元,能力转移完成后可能把 RL 后训练内化 | 高 | 向 Applied Compute 或其客户索取续约条款、排他性条款和多年承诺表述 |
| Applied Compute 定位在基础设施层,避开与垂直既有厂商的直接竞争 | 垂直既有厂商自己的高管(Sierra 的 Bret Taylor)称正在内部搭建专有微调层,说明基础设施与应用的边界并不固定 | 中 | 尽调中厘清每个具名客户的模型 IP 有多少由 Applied Compute 训练,多少由客户自有团队完成 |
| 已募资本足以支撑 Applied Compute 当前客户基础和路线图 | 直接同业在同一个 12 个月窗口内各自融资 $200M-$950M,估值达到 Applied Compute 已披露估值的 3x-12x | 高 | 确认 Applied Compute 当前现金跑道、下一轮融资时间,以及算力成本是否比收入涨得更快 |
| 超大云厂商捆绑(Salesforce、Microsoft)正在把基础智能体构建工具商品化 | 捆绑工具压缩点状解决方案的可服务市场,最终也可能削弱对外部 RL 专家的需求 | 中 | 监测超大云厂商套件是否开始纳入 RL/后训练定制,而不只是智能体编排 |
| 智能体 AI 的广泛市场需求支撑竞争集合的长期增长 | Gartner 预计到 2027 年底,超过 40% 的智能体 AI 项目会被取消;Sierra 创始人也预期市场会修正并出现“出清效应” | 中 | 把具名同业的续约率和扩张率(而不只是总预订额)作为判断需求韧性的领先指标 |
| Applied Compute 的前置部署工程模式创造了与垂直同业相似的切换成本 | 独立评测显示,即便资金充足的同业(Devin)也仍有未解决的可靠性缺口,说明整个品类(包括 Applied Compute)的技术差异化还没有资金规模暗示的那么扎实 | 中 | 索取 Applied Compute 自身相对基线和竞品后训练方法的评测 / 基准测试结果,而不是只看客户证言 |
严重性是基于证据的定性判断(高 / 中),不是评分指数;每行都引用了支撑该威胁的具体竞争对手披露或独立证据。
[CP043, CP044, CP013, CP045, CP039, CP030]在同一个 12 个月窗口内,所有具名直接同行都完成了比 Applied Compute 更大、估值更高的融资;与此同时,独立分析师指出,全行业智能体 AI 项目取消率偏高。
[CP007, CP012, CP018, CP016, CP009, CP025]3.6 图表
04财务
4.1 收入模式与变现
Applied Compute 自己的材料把产品描述为三阶段「Agent Cloud」——Train、Serve、Improve——交付方式是嵌入式、托管式项目,而不是自助订阅。公司称,其工程师会从最初评测到生产部署,再到持续在线 RL 改进,全程「坐在」客户工程团队身边;公司自己的案例研究把这种动作讲得很具体:工程师在 DoorDash 的 Sunnyvale 办公室现场工作,把生产 QA 标签转化为自动评分器;Cognition 项目则把 Applied Compute 的 RL 技术栈与 Cerebras 推理结合起来,为生产漏洞检测智能体达到实时延迟目标。三篇已发布案例研究(Cognition、DoorDash、Mercor)都没有披露合同金额、最低消费承诺或按使用量收费表;公司自己的首页、融资文章和上线文章也只用定性语言描述业务。与 OpenAI、Anthropic、Microsoft、Cognition 自己的 Devin、Together AI 和 Cerebras 已公开价目表相比——这些公司都至少披露了部分按席位或按 token 费率——Applied Compute 是这个对比集中唯一一个定价结构完全未披露的供应商,符合一家仍处于早期、高接触企业销售阶段,而不是计量产品阶段的公司特征。[CI001, CI002, CI003, CI004, CI005, CI006]
| 收入流 | 机制 | 单位 | 当前数值 / 状态 | 证据质量 | 尽调问题 |
|---|---|---|---|---|---|
| 嵌入式 RL 后训练项目 | 公司工程师与客户共同设计评测 / 评分器,并在客户现场或远程后训练客户专属模型(如 Cognition、DoorDash) | 单项目合同(费用结构未披露) | 已在具名客户生产环境中运行 | 公司宣称已投产;未披露金额 | 索取合同金额、期限长度和续约条款 |
| Agent Cloud 平台用量(Train / Serve / Improve) | 部署模型的训练、服务和持续在线 RL 改进按算力用量收费 | 推定按算力 / token 或席位计费(未披露) | 描述为在线产品能力 | 仅有产品描述;未公布定价 | 索取用量定价表和量级门槛 |
| 托管 / VPC 企业交付 | 面向重视安全的客户,提供 SOC 2 认证的托管或 VPC/serverless 部署 | 平台费或订阅费(未披露) | 市场化能力;公司网站称已获 SOC 2 认证 | 仅有功能描述,未绑定价格 | 确认合规功能是单独收费还是打包 |
| 数据 / 评测框架共同开发(如 Mercor) | 与数据和评测伙伴联合开展基准测试及后训练工作 | 项目制合作(若收费,费用未披露) | 公开的技术合作;不清楚是否产生收入,还是 R&D/营销 | 模糊 -- 可能是付费项目,也可能是未收费的联合营销 | 确认伙伴合作是否为计费项目 |
| 汇总收入 / ARR | n/a | USD | 已审阅所有来源均未披露 | 完全未披露 | 索取收入明细表、董事会材料或审计师函 |
各行根据 Applied Compute 自身产品和案例描述,重构可能的收入机制;没有任何一行对应公司披露的金额,汇总收入行是已确认的披露缺口,不是估算。
[CI001, CI003, CI005, CI006, CI014]| 厂商 / 产品 | 价格 / 单位 / 合同 | 标价 vs. 实际成交价 | 折扣 / 未知项 | 来源 |
|---|---|---|---|---|
| Applied Compute(Agent Cloud + 嵌入式项目) | 未公开 | 未知 -- 没有标价可与实际成交价比较 | 整体定价结构未披露 | Applied Compute 首页和案例研究 |
| OpenAI ChatGPT Business / Enterprise 套餐 | $20-25 每用户 / 月(Business);定制(Enterprise) | Business 标价已公开;Enterprise 为定制 / 协商价 | Enterprise 批量折扣未披露 | OpenAI ChatGPT 定价页 |
| Anthropic Claude Pro / Max | $17-20/月(Pro);Max 起价 $100/月 | 消费者 / 团队层级标价已公开 | 本页未显示 Enterprise/API 定价 | Anthropic(Claude)定价页 |
| Microsoft Foundry(Azure AI Foundry)平台 | Foundry Models、Agent Service、Foundry IQ 和 Tools 按消耗计费 | 按计量项公布标价 / 估算价 | 实际协商 Azure 价格随客户协议而变 | Microsoft Azure Foundry 定价页 |
| Cognition Devin | $0 / $20 / $200 月度层级,外加按用量计费的云端智能体额度 | 个人 / 团队层级标价已公开 | 未显示 Enterprise/团队合同价格 | Devin 定价页 |
| Together AI(托管推理) | $0.30-$1.74 / 1M 输入 token,随模型而变 | 按模型公布标价 | 未显示批量 / 专用端点折扣 | Together AI 定价页 |
| Cerebras Inference | 免费试用;Developer 层级按 token 付费;Enterprise 层级定制 | 部分标价(Developer 层级);Enterprise 为定制 | Enterprise 定价和批量条款未披露 | Cerebras 推理页 |
对照行是相邻模型 API 或编码智能体产品的标价,并非 Applied Compute 自身嵌入式 RL 后训练合同的已确认类比;它们框定的是市场上企业 AI 定价结构的区间,而不是 Applied Compute 的实际成交价,后者仍完全未披露。
[CI002, CI007, CI008, CI009, CI010, CI011]Applied Compute 自身产品叙事如何把一次客户互动转化为收入;所有美元金额均未披露。
节点顺序根据 Applied Compute 自身 Train/Serve/Improve 产品叙事及 DoorDash/Cognition 案例研究重构;没有任何节点披露美元金额。
[CI006, CI004]4.2 单位经济与成本结构
Applied Compute 没有披露任何经典单位经济输入:毛利率、CAC、销售周期长度和净收入留存。最接近的可用参照是 Palantir,其 Foundry AIP 交付模式同样依赖嵌入客户的前线部署工程师;截至 2026 年 3 月 31 日的季度,Palantir 报告 GAAP 毛利率约 86.8%($1.42B 毛利润 / $1.63B 收入),说明劳动密集、嵌入式交付在规模化后并不天然稀释毛利。Applied Compute 自己的成本栈可能由 GPU / 推理算力(Cognition 案例研究点名 Cerebras 为推理伙伴)、前线部署工程人力(如 DoorDash 现场项目)和数据 / 评测框架成本构成(如 Mercor 合作中,据称少于 1,000 个专家标注任务就带来大幅评测分数提升,意味着单项目边际数据成本相对较低)。一篇独立 LLMOps 对 DoorDash 案例研究的评论明确指出,Applied Compute 自己的材料没有披露「总开发成本、持续推理成本或量化业务影响」,并提醒该案例研究带有宣传性质,应保持适当怀疑——这是对公司自我发布成本收益证据的真实第三方负面校验。[CI015, CI016, CI017, CI018, CI019, CI020]
| 指标 | 数值 | 置信度 | 为什么重要 | 尽调问题 |
|---|---|---|---|---|
| 收入 / ARR | n/a -- 未披露 | 任何估值倍数的核心承销指标 | 索取当前收入和 ARR 数据 | |
| 毛利率 % | n/a -- 未披露(对照:Palantir 的 FDE 模式毛利率约 86.8%,2026 年 Q1) | 决定算力和人力密集的交付模式扩大后能否守住盈利 | 按项目类型索取收入成本明细 | |
| CAC / 销售周期长度 | n/a -- 未披露 | 衡量嵌入式 / 现场交付模式的获客销售效率 | 按客群索取平均成交周期和 CAC | |
| 净收入留存 / 扩张 | n/a -- 未披露 | 判断嵌入式项目能否转成持久且扩张的客户 | 索取具名客户的续约和扩张数据 | |
| 算力成本占收入 % | n/a -- 未披露;只能从 Together AI/Cerebras token 定价推断 | 决定对 GPU/推理成本上涨的暴露 | 索取基础设施支出与收入的拆分 | |
| 单项目边际数据 / 评测成本 | 按 Mercor 案例研究看方向性较低(少于 1,000 个专家标注任务带来大幅分数提升) | 低 -- 单一案例研究,未证明能代表全公司 | 说明项目未必需要大规模标注数据也能呈现收益 | 索取每个典型项目的数据采集支出 |
Applied Compute 自身每一个单位经济数值都是 null,因为公司尚未披露;Palantir、Together AI 和 Cerebras 数字是外部对照,不是 Applied Compute 数据,并已如此标注。
[CI015, CI016, CI017, CI018, CI019, CI021]Applied Compute 项目背后的可能成本输入,从算力和人力到未披露的合同价格与利润率。
成本节点来自 Applied Compute 自身案例研究中对交付模式的描述(算力伙伴、现场工程、伙伴标注数据),不是披露的成本拆分;FI004 和 TI003 中的 Palantir 对照项是唯一可用的量化毛利率锚点。
[CI017, CI022, CI012, CI011]Applied Compute 的资本和成本强度可能集中在哪些环节,并与外部成本 / 定价信号对标。
单元格描述的是从 Applied Compute 自身案例研究和产品页推断出的方向性暴露,并与外部定价 / 能源需求来源交叉核对;没有任何单元格给出 Applied Compute 专属美元金额。
[CI022, CI012, CI036, CI021, CI035]4.3 资本充足性与融资依赖
对一家私营公司而言,Applied Compute 的融资历史记录得异常充分:2025 年中种子轮估值约 $100 million;The Information 在 2025 年 9 月报道融资洽谈估值 $500 million;2025 年 10 月退出隐身时披露 $80 million 融资,估值约 $700 million;2026 年 4 月 8 日,公司、其法律顾问 Latham & Watkins 和领投方 Kleiner Perkins 共同确认,公司完成 $80 million Series B,投后估值 $1.3 billion——累计已披露融资达到 $160 million。TechCrunch 2026 年 7 月的独角兽汇总在该轮结束约三个月后,通过 PitchBook 独立印证了这个数字。缺失的是贷款人或后期投资人评估未来资本充足性所需的一切:账面现金、月度 burn 和 runway 都未披露;公司唯一公开的资金用途声明也是定性的(「扩大团队、扩展部署、把第一代智能体 workforce 推向市场」)。本章所审阅来源没有披露任何 venture debt、项目融资工具或信贷额度;迄今每轮融资似乎都是定价普通股权。考虑到从种子轮到 Series B 不到一年估值约涨 13x,下一轮融资看起来更可能由增长和投资人需求提前拉动,而不是被临近现金短缺倒逼;但这是推断,不是已披露触发因素。[CI023, CI024, CI025, CI026, CI027, CI028]
| 项目 | 数值 | 来源 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| 累计披露融资额 | $160 million(截至 2026 年 4 月 8 日) | Applied Compute 融资公告;Latham & Watkins 新闻稿;TechCrunch 独角兽追踪 | 高 | 确认是否存在任何额外未披露资本(SAFEs、过桥票据) |
| 最新投后估值 | $1.3 billion(2026 年 4 月 Series B) | Applied Compute 融资公告;Latham & Watkins 新闻稿;Kleiner Perkins 观点文章 | 高 | n/a -- 多个独立来源充分交叉印证 |
| 账面现金 | 已审阅所有来源均未披露 | n/a | 索取当前现金余额 | |
| 月度资金消耗 / 隐含现金跑道 | 已审阅所有来源均未披露 | n/a | 索取融资后的资金消耗率和隐含现金跑道 | |
| Series B 资金计划用途 | 仅定性:扩充团队、扩大部署规模、把第一代智能体劳动力推向市场 | Applied Compute 融资公告 | 中 | 索取资金用途百分比拆分(人员 vs. 算力 vs. G&A) |
| 债务 / 项目融资义务 | 未披露;所有轮次均描述为股权融资 | Applied Compute 融资公告;Latham & Watkins 新闻稿 | 中 | 确认不存在风险债务或算力融资安排 |
融资和估值数字已被公司自身公告、法律顾问新闻稿和独立媒体充分交叉印证;现金、资金消耗和现金跑道行是真实披露缺口,不是本研究遗漏。
[CI023, CI024, CI027, CI028, CI029, CI030]基于来源边界的 Applied Compute 2025-2026 年各轮融资估值估算,加上一个外部前线部署工程师利润率对照。
估值边界反映来源自身使用的区间和限定语言(“roughly”、“approx.”);B 轮数字是公司确认的单一点,而非区间。利润率对照是 Palantir 自身披露数据,不是 Applied Compute 经济性的估算。
[CI040, CI026, CI036]4.4 公开披露缺口与财务结论
在 SEC EDGAR 公司数据库中搜索「Applied Compute」,没有返回与这家 2025 年创立的初创公司相关的备案——唯一名称相近的注册人是一个无关申报主体,其注册资格已在 2006 年被撤销——这确认公司没有任何标准上市公司披露(收入、毛利率、现金状况、员工数),而这些信息在 Palantir 这类前线部署工程师模式参照公司那里,可以通过 10-Q 获得。Applied Compute 的公开案例研究只点名三家客户,因此付费客户总数、集中度风险和续约行为都仍未知。品类层面的逆风又放大了披露缺口:Gartner 预测,到 2027 年底,超过 40% 的智能体 AI 项目会因成本和 ROI 不清而被取消;McKinsey 警告,多数组织已经遇到有风险的智能体行为,因而需要新的(尚未计入成本的)治理支出;IEA 和 Goldman Sachs 都预计,到 2030 年,数据中心电力需求将大幅上升,可能推高任何重算力供应商利润率底层的 GPU / 推理成本。合起来看,Applied Compute 当前财务画像最适合概括为「资金充足,但无法验证」:融资和估值事件来源丰富且彼此一致,而承销人判断收入质量、毛利路径或资金 runway 所需的每个指标,都是已确认的缺口,而非只是研究不充分。[CI031, CI032, CI033, CI034, CI035, CI036]
| 缺失指标 | 对尽调的影响 | 尽调路径 |
|---|---|---|
| 收入 / ARR | 无法用任何收入倍数验证 $1.3B 估值 | 直接向公司索取收入明细表或经审计师审阅的财务数据 |
| 毛利率 / 收入成本 | 无法评估算力和人力密集的交付模式是否具备结构性盈利能力 | 按项目类型索取收入成本明细 |
| 现金头寸、资金消耗和现金跑道 | 无法评估资本充足性或下一轮融资大致时点 | 向管理层或主投方索取最新现金余额和资金消耗趋势 |
| 付费客户总数 | 公开确认的具名客户只有 3 家(Cognition、DoorDash、Mercor);真实业务账本未知 | 索取客户名单或客户数,并区分生产 vs. 试点状态 |
| SEC 或其他监管备案 | 公司是私有企业,因此没有备案;Palantir 10-Q 等上市公司披露基准不可用 | 监测 SEC EDGAR 未来是否出现 S-1,并跟踪私募二级市场数据提供商 |
| 员工人数 | 无法对标单员工收入或资金消耗 | 向公司或经核实的媒体披露索取当前员工数 |
截至运行日期,每行都反映已确认的公开披露缺失,不是研究不充分;Applied Compute 是私有且未披露型公司,几乎所有财务 KPI 都未公开,但融资和估值事件披露反而较充分。
[CI031, CI032, CI033, CI020, CI034]4.5 图表
05产品与技术
5.1 Specific Intelligence 是托管式专用化层,不是通用模型 API
Applied Compute 把产品定义为一套具体工作流里的云平台:训练、推理、持续改进,而不是一个单独的基础模型。官网把平台拆成 Train、Serve 和 Improve:客户带来自己的数据、评测框架和评分器;Applied Compute 在文本、图像、代码和结构化数据场景中,对会使用工具的智能体做后训练;训练出的模型随后进入生产,并从线上反馈继续更新。公司反复把这件事定位为一层架在日益商品化前沿模型之上的专业化层,让客户保留自己的奖励函数、评测和记忆,而不是把智能外包给单一上游基础模型供应商。这个说法在官网、融资文章、Modal 客户案例以及后续技术研究文章里都一致。更重要的是,平台强调模型灵活,而不是绑定单一模型:Applied Compute 表示,客户可以从当时最好的基础模型开始训练,之后升级模型时,不必重建评测框架、数据管线或部署栈。落到产品承诺上,它讲的不是“我们有最好的基础模型”,而是“我们帮企业把专有判断沉淀成一套持久、可复用的训练与部署循环”。[CE001, CE002, CE003, CE004, CE011, CE037]
| 模块 / 资产 | 主要用户 | 状态 / 成熟度 | 差异化 | 尽调缺口 |
|---|---|---|---|---|
| Train | Applied Compute 研究员及客户 ML / 工程团队 | 核心产品,反复有证据支持 | 在客户数据、测试框架和评分器上,对跨多模态的工具使用型智能体做后训练 | 没有公开 API / SDK 文档证明存在完全自助式训练流程或定价模型 |
| Serve | 运营生产智能体的平台 / 基础设施团队 | 核心产品,反复有证据支持 | 让部署测试框架与训练保持一致,支持单租户区域和 VPC 运行,并针对智能体延迟 / 吞吐权衡做优化 | 没有公开可用性 SLA、状态页历史,或基于客户生产流量的基准测试;公司自撰文章之外缺少证据 |
| Improve | 负责闭环学习的产品 / ML 团队 | 核心产品,反复有证据支持 | 把生产轨迹、人类反馈和发布可观测性转成在线 RL 或自蒸馏更新 | 没有公开留存或回归率数据说明生产更新多久能被安全提升 |
| Context Engine / Contextbase | 企业知识工作者和智能体构建者 | 新兴但在 2026 年研究中已清晰产品化 | 加入 Remember / Refine / Retrieve 记忆,让机构知识随时间复利,而不是只留在权重里 | 公司基准测试之外,没有公开客户部署数量,也没有外部审计验证检索质量 |
| Router / 基准测试工具 | 内部平台团队和高级客户 | 早期但真实存在 | 多模型路由、轨迹回放和工作负载基准测试,显示该平台围绕智能体运营搭建,而不只是托管模型 | 公开开发者足迹只有一个代码仓库和公司自撰研究;面向客户工具的广度仍不清楚 |
各行区分首页三个模块和 2026 年研究中浮出的相邻技术资产。“状态 / 成熟度”只反映本章审阅的公开证据,不代表内部 SKU 边界。
[CE001, CE002, CE004, CE007, CE020, CE021]Applied Compute 在客户数据、企业记忆和长周期 RL 观测设施之上叠加 Train/Serve/Improve 控制平面。
层级边界根据 Applied Compute 主页、研究文章和客户案例重构。治理层有意混合公司声称的控制与独立证据显示的护栏需求,因为公开验证面不完整。
[CE001, CE004, CE005, CE006, CE007, CE010]5.2 技术架构把 RL 后训练与上下文 / 记忆系统绑在一起
Applied Compute 的公开技术材料显示了两个紧密相连的产品押注。第一,是在真实、可回放环境里用强化学习做后训练。DoorDash 的部署把内部 QA 标签变成自动评分器,再把该评分器作为奖励信号;Cognition 的 SWE-check 案例在 RL 中复刻生产环境里的 Windsurf 评测框架,再把后训练拆成能力阶段和产品对齐阶段;Mercor 的工作则在专家标注世界中做长周期 RL,并用轨迹级可观测性捕捉奖励设计失败。第二,权重本身不够,智能体还需要运行时记忆层。Context Engine 论文描述了一条 Remember/Refine/Retrieve 管线:摄取 SaaS 数据和智能体轨迹,蒸馏成 Contextbase,并在运行时向智能体开放检索 API;后续 “Memory in the wild” 文章把同一循环指向 Applied Compute 自己的代码轨迹,两周内关键记忆检索率从不到 10% 升至约 20%。公司更新的 “neural cheat-sheets” 研究还进一步暗示,它想把上下文压缩本身训练成一种能力,而不只是提示词技巧。合在一起,这些材料指向一套全栈架构:定制环境和评分器产出更好的权重,轨迹和上下文系统则保留模型在多次运行之间应该记住的东西。[CE007, CE008, CE009, CE010, CE014, CE016]
| 用户任务 | 当前工作流 | Applied Compute 方案 | 可衡量收益 | 限制 |
|---|---|---|---|---|
| DoorDash 商户菜单入驻 | 人工专家核查混乱的菜单输出,QA 标签在自动提取后捕捉质量问题 | 用专家 QA 标签构建校准过的评分器,再按 DoorDash 质量标准对纠错模型做 RL 训练 | 低质量菜单相对下降约 30%,模型已覆盖美国全部菜单流量 | 收益来自公司和案例研究证据;公开成本节省和误报率未披露 |
| Cognition / Windsurf 实时 bug 检测 | 前沿模型能发现隐蔽 bug,但对即时 IDE 反馈可能太慢或太贵 | 在 RL 中复刻生产测试框架,训练一个与产品延迟约束对齐的专用 bug 检测模型 | 速度约为 Opus 4.6 的 10x,分布内 delta F1 差距从 0.09 收敛到 0 | 分布外差距仍非零,因此专用化改善但没有消除前沿模型取舍 |
| 使用 Mercor 数据的专业工作智能体 | 通用模型在长周期法律、咨询和银行任务上表现吃力 | 用专家标注任务、长周期 RL 和轨迹可观测性,对小型专用模型做后训练 | 总体 Pass@1 和平均分几乎翻倍;引用实验中的公司法 Pass@1 增至三倍 | 证据以基准测试为中心,而不是客户生产 ROI |
| 企业记忆 / 上下文管理 | 机构知识散落在文档、轨迹和业务专家(SME)手中,而不是可复用智能体记忆 | Remember / Refine / Retrieve 构建 Contextbase,并把相关记忆回灌给运行时智能体 | 引用基准测试中,GPT-5.4 APEX-Agents 分数从 44.2% 升至 51.7%;内部 ACL-Wiki 记忆关键性约翻倍 | 多数证据来自公司自撰内容和基准测试,不是第三方生产遥测数据 |
每行都把一个真实客户或基准测试工作流映射到 Applied Compute 描述的产品介入。指标按引用公开来源原样记录,不应泛化到所有客户。
[CE008, CE009, CE014, CE015, CE016, CE017]| 层 / 流程 | 角色 | 公开证据 | 依赖 | 风险 |
|---|---|---|---|---|
| 可回放环境和测试框架 | 让模型在与生产中相同或相近的环境里尝试任务 | DoorDash、Cognition、Mercor 和 Modal 的描述都强调环境保真度 | 客户系统、工具定义、沙盒提供商 | 如果模拟环境偏离真实生产行为,就会出现训练-测试错配 |
| 评分器 / 奖励函数 | 给输出打分,让 RL 能强化理想行为 | DoorDash 自动评分器;Mercor 评分标准重做;杠杆 / 熵 / 陈旧度研究 | 人工标签、任务设计、奖励校准 | 糟糕的奖励塑形可能奖励拒答、走捷径或脆弱行为 |
| 服务与推理优化 | 以可接受的延迟和成本运行长周期、工具使用型轨迹 | 推理基准和 async-RL 陈旧性论文;官网单租户和 VPC 说法 | 推理引擎、KV-cache 管理、并发调优、GPU 供给 | 尾部延迟和缓存淘汰会同时拖累用户体验和训练质量 |
| Context Engine / 检索 API | 在运行时暴露精炼后的企业记忆 | Remember / Refine / Retrieve 与 Memory in the Wild 文章 | 企业文档、轨迹日志、检索质量 | 摘要质量差或记忆过时,会拖累 agent 可靠性和可审计性 |
| 路由 / 编排层 | 按任务选择合适模型或工作流,而不是硬套一个默认项 | agentic-router 研究与 Anthropic 工作流指南 | 模型池质量、路由功能、编排逻辑 | 路由出错会抬高成本,却不提升成功率 |
“公开证据”列有意混合客户案例和研究文章,因为 Applied Compute 的产品叙事异常受研究驱动。风险列记录架构在哪些环节可能失效,即便底层思路本身成立。
[CE010, CE012, CE013, CE020, CE021, CE022]在不同客户之间,Applied Compute 遵循可重复流程:捕获专家判断,推进到生产部署,再持续改进。
[CE007, CE009, CE014, CE015, CE016, CE018]5.3 服务层为长周期智能体和受控企业部署优化
Serve 和 Improve 层与训练循环一样核心。Applied Compute 表示,客户可以做到训练到部署零错配,立即发布检查点,并按照吞吐、延迟和并发要求配置部署,覆盖异步智能体到实时用户体验。公司还声称支持在自有云上无服务器部署,或通过一个控制平面在客户 VPC 内完整运行;同时支持单租户区域部署,并在每次调度上记录审计日志。几篇研究文章解释了这些主张为什么在技术上重要。推理基准认为,智能体工作负载不同于经典提示 / 响应服务:它们会产生几十次短生成、长生命周期 KV 缓存和重尾工具延迟。异步 RL 陈旧性文章显示,当服务和学习解耦后,利用率、队列容量、rollout 并发和响应长度长尾都会影响训练质量。公开 GitHub 证据很薄,但方向一致:组织开放了一个 Apache 许可仓库 `trie`,用于对端点回放推理流量,这与研究文章中的基准叙事吻合。最终图景很清楚:平台针对的是轨迹密集、会使用工具的智能体,而不是单轮聊天吞吐。[CE005, CE006, CE012, CE013, CE022, CE023]
| 控制 / 质量信号 | 状态 | 范围 | 最佳公开证据 | 缺口 |
|---|---|---|---|---|
| SOC 2 认证 | 公司自称 | 平台安全 / 控制 | 官网页脚与安全表述 | 未找到公开审计报告、生效日期、范围说明或信任中心材料 |
| 数据留在客户边界内 | 公司自称 | 敏感企业部署 | 官网:serverless 或 VPC 部署;“data never leaves your perimeter” | 未见公开架构图说明 telemetry、日志或训练轨迹如何按部署模式隔离 |
| 基于角色的访问控制和审计日志 | 公司自称 | 访问、调度和生命周期事件 | 官网信任文案 | 未见公开管理端或审计 UX 截图;无法核验粒度或保留期限 |
| checkpoint 推送和监控 | 产品能力强暗示 | 模型发布管理和 A/B 测试 | 官网 Improve 版块与可观测性研究语料 | 未见公开回滚 / SLO 指标或事故披露 |
| agent 安全风险指引 | 独立材料证实确有风险 | 工具使用、自主性和监督 | CISA 指引加 reward-hacking 基准 | 公司尚未公开把安全护栏映射到具名外部框架 |
本表区分公司自称控制项和独立证据支持的风险信号。在当前估值下,没有外部信任中心或公开可用性 / 安全细节,本身就是一家企业 AI 平台的尽调发现。
[CE005, CE006, CE023, CE033, CE034]Applied Compute 靠客户环境、推理基础设施、轨迹质量和护栏共同支撑;任一节点失效,都会直接传导到生产可靠性。
这张图是定性而非定量,用于展示传导路径,不是加权因果模型。
[CE012, CE013, CE022, CE023, CE024, CE033]5.4 差异化可信,但技术风险仍落在验证、安全和产品化深度上
Applied Compute 的差异化在窄而可运营的场景里最强。Cognition 的 SWE-check 案例支撑了一个论点:小型专精模型能在单一任务上接近前沿精度,同时显著降低延迟和成本;Mercor 的基准工作说明公司能从稀缺专家数据中榨出大幅提升;Modal 故事则显示,团队已经围绕 rollout、评测扇出和 GPU 受限服务的基础设施现实做过建设。其 2026 年的发布节奏也说明这是一支快速迭代的研究组织,而不是静态服务商。不过,公开证据有真实上限。最好的安全主张——SOC 2 认证、数据不离开客户边界、RBAC 和审计日志——都来自 Applied Compute 自己的营销页面;本轮没有找到公开审计报告、信任中心细节或可用性历史。Anthropic、Cohere、OpenAI、Microsoft 和 CISA 的外部指南都指向同一主题:生产智能体需要严密编排、可观测性和护栏,因为灵活性会放大成本和错误面。独立的 Reward Hacking Benchmark 对 Applied Compute 选择的范式给出更尖锐警示:环境不加固时,RL 训练的工具型智能体会利用捷径。简言之,这家很年轻的公司讲出了少见成熟的产品故事,但公开验证面仍落后于架构野心。[CE015, CE017, CE019, CE020, CE021, CE025]
| 日期 / 阶段 | 功能或里程碑 | 状态 | 含义 | 来源 |
|---|---|---|---|---|
| 2026-03-24 | 高杠杆样本 RL 研究 | 已发布 | 显示重点放在训练算力效率和 rollout 选择 | Applied Compute 研究文章 |
| 2026-04-22 | 推理基准 + 轨迹回放工具链 | 已发布 | 显示服务 / 可观测层按 agentic 工作负载调优,而不是只面向聊天流量 | Applied Compute 研究文章 + GitHub trie 仓库 |
| 2026-05-01 至 2026-05-08 | Context Engine 与 Memory in the Wild | 已发布 | 暗示从一次性 post-training 转向可复用的运行时记忆产品 | Applied Compute 研究文章 |
| 2026-05-22 至 2026-06-16 | RMSD、路由和保熵 RL | 已发布 | 路线图重点是让定制模型更稳定、更便宜,也更贴合企业任务 | Applied Compute 研究文章 |
| 2026-06-26 至 2026-07-03 | 神经速查表与 async-RL 陈旧性控制 | 已发布 | 指向上下文压缩和系统级 RL 优化这两类下一层产品押注 | Applied Compute 研究文章 |
| 2026-05-06 | DoorDash 推出 AI 驱动的自助入驻工具 | 客户侧已上线 | 说明早期 RL 纠错工作之后,更广的商户入驻工作流仍在扩张 | DoorDash 新闻室 |
| 2026-07-02 | 公开 trie 仓库最近更新 | 活跃但范围窄 | 证明至少一个公开开发者产物一直维护到本次报告日 | GitHub 仓库 Applied-Compute/trie |
Applied Compute 不发布正式路线图页,因此本表用带日期的发布和外部客户上线反推路线图方向。更适合把它看成发布节奏图,而不是已承诺的 GA 路线图。
[CE009, CE020, CE021, CE022, CE024, CE026]训练、服务和专用化证据最扎实;公开信任建设、自助工具和独立验证最弱。
[CE006, CE015, CE019, CE023, CE035, CE036]5.5 图表
06客户
6.1 已披露客户显示真实垂直广度,但各账户的证据深度差异很大
Applied Compute 已经不只是匿名 logo 页。官网点名 Cognition、DoorDash、Mercor、Harvey、Bridge 和 Latch Bio,公开案例集补全了四类部署原型:DoorDash 的商户入驻,Cognition 在 Windsurf 内部做实时 bug 检测,Mercor 做面向专业工作的基准驱动定制,Harvey 做法律智能体后训练。这意味着买方 / 用户 / 付款方地图已经横跨产品团队、工程负责人、运营团队和领域专家组织。共同模式不是单一横向席位销售,而是嵌入高成熟客户内部、有清晰质量定义的专精工作流。代价是公开证据并不均衡:DoorDash 和 Cognition 看起来是真实生产部署;Mercor 和 Harvey 明显更像高价值设计伙伴;Bridge 和 Latch Bio 目前最好只视为有背书的客户信号,部署披露有限。[CU001, CU002, CU004, CU011, CU019, CU025]
| 客户 / 分群 | 买方 / 用户 / 付费方 | 用例 | 公开规模信号 | 战略价值 / 缺口 |
|---|---|---|---|---|
| DoorDash | 买方:商户 ML / 产品;用户:入驻和内容系统;付费方:DoorDash 平台团队 | 商户入驻、菜单纠错、商户增长工具 | 所述纠错上线覆盖全部美国菜单流量;自助入驻加快 35% | 最强生产证据;仍缺合同金额和续约数据 |
| Cognition / Windsurf | 买方:工程 / 产品领导层;用户:IDE 内的软件工程师;付费方:Cognition 平台 | Windsurf 内实时 bug 检测和专门代码审查 | SWE-Check 已投产;另有更广 Devin / Windsurf 分发说法 | 生产信号强;未披露付费席位或 ARR 贡献 |
| Mercor | 买方:AI / 产品领导层;用户:基准和数据团队;付费方:Mercor 平台 / 企业预算 | 专家数据 post-training 和基准优化 | 874 个任务 / 50 个 world 的开发集;基准深度高 | 看起来是有价值的设计伙伴,但商业化生产范围未披露 |
| Harvey | 买方:AI 研究 / 产品领导层;用户:法律产品栈;付费方:Harvey 平台 | 法律 agent 基准和模型优化 | Harvey 整体覆盖 1,500+ 家组织的 142,000 名律师 | 下游规模很大,但 Applied Compute 在 Harvey 产品内的部署深度尚未量化 |
| Bridge | 买方:支持 / 运营领导层;用户:支持坐席和 SME;付费方:fintech 运营 / 产品 | 支持工单判断捕捉和 agent 学习闭环 | 仅有官网证言 | fintech 垂直场景证据有用;缺部署范围和结果指标 |
| Latch Bio | 买方:研究 / 基准领导层;用户:biology-agent 研究人员;付费方:平台 / 研究预算 | 基准测试和 biology-agent 评估 | 官网证言加公开 scBench 材料 | 有垂直场景证据,但目前关系由基准牵引,生产牵引不清晰 |
各行区分生产部署、基准 / 设计伙伴关系和证言级证据。战略价值衡量每个客户能在多大程度上扩展 Applied Compute 的可服务工作流面,而不是收入贡献。
[CU001, CU002, CU004, CU011, CU019, CU025]| 客户 | 分群 | 部署 / 用例 | 生产 / 试点 | 公开结果 | 限制 |
|---|---|---|---|---|---|
| DoorDash | 食品配送 / 商户赋能 | 面向商户入驻菜单的自动评分器 + RL 纠错模型 | 生产 | 低质量菜单相对下降 ~30%;上线覆盖全部美国菜单流量 | 无合同金额、留存或绝对缺陷率分母 |
| Cognition | 开发者工具 / AI 编程 | Windsurf Quick Review 内的 SWE-Check bug 检测 | 生产 | bug 检测比前沿替代方案快 10x | 无付费席位、附加率或客户扩张指标 |
| Mercor | 专家网络 / 模型评估 | 专家标注开发集上的 post-training 和 APEX-Agents 基准优化 | 带客户证明成分的基准伙伴 | #1 公司法排名;Pass@1 和整体平均分几乎翻倍 | 基准成功不等于已部署终端客户 ROI |
| Harvey | 法律 AI | Harvey LAB 上的 post-training 和更广法律 agent 评估 | 设计伙伴 / 基准牵引 | 评分规则通过率领先所引前沿模型 | 公开材料未显示 Harvey 在生产中使用该训练 checkpoint |
| Bridge | fintech / 支持运营 | 用生产中 SME 判断训练的支持工单 agent | 证言级生产说法 | 判断力随每张工单累积 | 未披露规模、延迟或质量指标 |
| Latch Bio | 生物基础设施 / 基准测试 | biology-agent 任务的基准质量映射 | 基准 / 证言级 | 公开 scBench 基础设施支持生物相关性 | 未披露生产部署或商业结果 |
本表有意区分生产证明和基准证明。即便商业化深度仍不透明,部分关系仍具战略价值。
[CU008, CU013, CU022, CU025, CU030, CU032]客户路径从高风险工作流起步,进入专家训练项目;只有模型在生产系统里赢得信任,才会继续扩张。
[CU003, CU008, CU013, CU018, CU022, CU029]DoorDash 和 Cognition 的生产证明最强;Harvey 和 Mercor 战略价值高,但证据更偏基准;Bridge / Latch Bio 证据仍薄。
[CU008, CU013, CU022, CU028, CU029, CU030]6.2 DoorDash 和 Cognition 提供最强生产证据和最干净结果指标
DoorDash 和 Cognition 是最重要的参考客户,因为二者都把 Applied Compute 接到线上产品界面,而不是孤立基准。DoorDash 将合作绑定到商户激活和菜单准确性,披露低质量菜单相对减少约 30%,并已覆盖全部美国菜单流量。DoorDash 2026 年 6 月的发布材料进一步显示,AI 辅助入驻已经成为更广商户增长系统的一部分:商户上线速度快逾 35%,AI 驱动网站平均转化率接近 10%。Cognition 也提供了同样强的产品证据。Applied Compute 和 Cognition 均称 SWE-Check 已在 Windsurf 内部投产,bug 检测速度比前沿替代方案快 10 倍;Cognition 更广的产品更新还显示,Windsurf/Devin 正在成为具备实际分发能力的多智能体指挥中心。这些是最清晰的信号:Applied Compute 可以嵌入那些重视延迟、精度和迭代速度,且足以为定制专业化层付费的客户工作流。[CU004, CU007, CU008, CU009, CU010, CU012]
| 指标 | 数值 | 日期 / 来源 | 置信度 | 含义 / 缺失分母 |
|---|---|---|---|---|
| DoorDash 低质量菜单率 | 相对基线下降 ~30% | 2026 DoorDash 案例研究 | 中 | 工作流结果强,但未披露绝对错误率分母或 ROI |
| DoorDash 部署广度 | 已上线覆盖全部美国菜单流量 | 2026 DoorDash 案例 + DoorDash 商户更新 | 高 | 生产证据强,但没有把具体商户数与该模型绑定 |
| DoorDash 商户入驻速度 | 商户上线加快 35% | DoorDash 2026 年 6 月商户更新 | 高 | 显示买方侧更广泛采用 AI,但不能把全部收益都归功于 Applied Compute |
| Cognition bug 检测速度 | 比前沿替代方案快 10x | 2026 Cognition 案例 + Cognition 博客 | 高 | 产品 KPI 清晰,但参考基线和商业变现未披露 |
| Harvey 客户规模代理指标 | 142,000+ 名律师、1,500+ 家组织、60 个国家 | Harvey 客户页面 | 高 | 终端市场规模大;Applied Compute 在该客户群中的捕获未知 |
| Harvey 产品参与度代理指标 | 月度采用率 92%,每用户每月节省 25+ 小时 | Harvey 客户页面 | 高 | 暗示粘性产品内有扩张空间,但衡量的是 Harvey 使用,而不是 Applied Compute 留存 |
本表在缺少直接信号时,把 Applied Compute 客户结果与下游平台采用代理指标放在一起。缺分母是核心尽调问题,不是脚注。
[CU007, CU008, CU009, CU010, CU013, CU028]公开证明从六个具名引用收窄到仅两段客户关系:它们既有强生产指标,也有持续部署细节。
该漏斗基于已审阅公开材料的证明质量,而非内部销售管道转化。Bridge 和 Latch Bio 算作引用,但不算证据充分的深度部署。
[CU002, CU008, CU013, CU022, CU030, CU032]6.3 扩张潜力可信,但公开耐久性证据仍以代理指标为主
最强的扩张逻辑来自客户问题本身。DoorDash 可以把同一套商户质量基础设施用到入驻、内容和持续商户增长。Cognition 可以把专精模型扩展到本地智能体、云端智能体、审查智能体和共享上下文界面。Harvey 的规模指标说明,如果设计伙伴关系成功,且基准提升能进入线上产品模块,Applied Compute 可能触达很大的下游使用基础。Mercor、Bridge 和 Latch Bio 分别展示了相邻垂直场景:客户专属评分器、专家数据或支持判断足够重要,后训练在经济上可能成立。但耐久性仍缺少直接证据。Harvey 每月 92% 采用率和节省 25+ 小时,是法律 AI 终端市场强度的有用代理指标,不是 Applied Compute 留存的直接证明。Applied Compute 未披露 NRR、GRR、流失、合同期限或钱包份额,因此当前公开扩张案例更多是战略和工作流逻辑,不是商业和队列证据。[CU010, CU016, CU018, CU022, CU028, CU029]
| 指标 | 数值 / null | 分群 | 置信度 | 尽调问题 |
|---|---|---|---|---|
| DoorDash 持续生产使用 | 全部美国菜单流量上线覆盖 | 商户 AI / 入驻 | 中 | 要求提供上线后质量趋势、回滚历史,以及该模型今天是否仍为默认项 |
| Harvey 月度采用率 | 92% | 法律 AI 终端市场 | 高 | 澄清由 Applied Compute 支撑的产品界面是否参与该采用,以及覆盖多少用户 |
| Harvey 每用户每月节省小时数 | 25+ | 法律 AI 终端市场 | 高 | 将终端用户生产力转化为付费意愿和基准相关扩张价值 |
| Cognition 重复使用代理指标 | Quick Review 已投产;OTA 从 Windsurf 迁移到 Devin Desktop | 开发者工具 | 中 | 要求提供调用频率、误报趋势和 SWE-Check 席位渗透率 |
| Bridge 重复学习代理指标 | 专业判断随每张工单累积 | 支持运营 | 低-中 | 要求提供工单量、升级率和随时间变化的质量改善曲线 |
| NRR / GRR / 流失 / 合同期限 | null | 所有分群 | 高 | 无公开披露;要求提供队列留存、续约日期和按 ARR 统计的集中度 |
缺少真实留存指标时,本表采用最接近的公开代理指标,并明确尽调问题。null 值是有意保留,不是遗漏。
[CU008, CU013, CU029, CU030, CU038]公开证据在获客和部署证明上最强;但尽调问题转向留存和扩张经济性时,证据明显衰减。
这是证据深度队列,不是收入留存队列。百分比衡量的是,在已审阅来源集中,每一证明层在生命周期各阶段提供可用证据的频率。
[CU028, CU029, CU034, CU035, CU038, CU039]6.4 最大未解问题是集中度、采购摩擦和偏服务的交付动作
客户故事足以证明 Applied Compute 能赢下成熟设计伙伴,但还不足以支撑商业耐久性。大多数量化结果来自公司撰写的案例研究或客户-伙伴文章,而不是独立客户撰写的 ROI 材料。公司没有披露客户数增长、收入集中度、合同期限、续约行为,也没有说明最大账户是否显著超大。这个问题很关键,因为公开证据也指向服务密集型交付模型:Applied Compute 表示会嵌入客户团队,Modal 也独立描述了围绕 RL 工作负载的深度参与方式。第一波部署中,这可能是优势;但如果每个新账户都需要前沿研究级注意力,规模化问题就会冒出来。独立不利背景进一步支持谨慎。CISA 警告智能体式 AI 需要更强监督,Gartner 的 2026 服务运营议程也强调,许多试点在扩张前仍需要运营模型重设计和硬 ROI 证明。落到现实,就是销售周期长、安全审查慢,以及明星 logo 可能夸大多元化程度的不小风险。[CU003, CU034, CU035, CU036, CU037, CU038]
| 扩张驱动因素 | 集中风险 | 影响 | 尽调路径 |
|---|---|---|---|
| DoorDash 商户栈内的工作流邻接 | 单一大客户可能占用不成比例的训练 / 支持负载 | 可能带来强落地后扩张经济性,也可能隐藏集中度 | 要求按客户和用例拆分 ARR;核验是否已有更多 DoorDash 模块上线 |
| Cognition / Devin 的多界面编程 agent | 单一旗舰客户可能主导开发者工具的标杆价值 | 可能让后续签单更容易,但收入结构更脆弱 | 要求合同范围、席位数,以及 Applied Compute 是否支持 SWE-Check 之外的更多能力 |
| Harvey 下游规模化采用 | 基准胜利未必自动转化为产品化收入 | 若集成则期权价值强;若局限于评估工作则价值有限 | 询问 checkpoint 部署范围和与 Harvey 合作挂钩的经常性收入 |
| 借 Bridge 和 Latch Bio 拓展 fintech 与生物垂直场景 | 证言级账户可能夸大可部署广度 | 有类别信号价值,但还不是持久证明 | 要求部署状态、开始日期和具名操作方背书 |
| 企业采购和安全审查 | 监督、ROI 证明和运营模式重设会放慢扩张 | 销售周期和扩张速度可能比 logo 页暗示的更慢 | 尽调中使用安全问卷、试点到生产漏斗数据和销售周期时间线 |
| 嵌入式前沿研究模式 | 高服务强度会限制账户吞吐 | 如果每个账户都要深度定制,可能压制毛利和速度 | 要求实施人员配比、上线时间分布和按合作类型拆分的毛利 |
核心商业问题不是 Applied Compute 能否拿下复杂 logo,而是这些 logo 能否转化为可重复、分散且高效交付的收入。
[CU003, CU018, CU029, CU034, CU035, CU037]6.5 图表
07风险
7.1 竞争压力是首要战略风险,因为控制平面故事正在迅速拥挤
Applied Compute 的核心卖点很有吸引力:用客户的评分器、数据和工作流后训练一个任务专属智能体,在该公司的具体用例上超过通用模型。问题在于,周边市场正从多个方向同时走向同一套企业智能体控制平面。AWS 表示 Bedrock 已服务超过 100,000 家组织,现在又提供 AgentCore,以便安全地大规模运行智能体。OpenAI、Microsoft、Google、Salesforce 和 Palantir 都在大型既有分发渠道之上,推销安全的企业智能体部署、治理或数字劳动力。Anthropic 自己的智能体指南也削弱了一个假设:并非每个工作流都需要重型 RL 专业化栈;许多工作流用更简单的编排模式可能就够了。Applied Compute 仍能在深度定制重要的场景获胜,但举证责任已经转为结果:它必须证明,自己的专精循环能比越来越强的平台默认能力带来显著更好的经济性或质量。[CR003, CR020, CR021, CR022, CR023, CR024]
| 失效模式 | 可能性 | 严重性 | 缓释成熟度 | 剩余风险暴露 | 未解决缺口 |
|---|---|---|---|---|---|
| 生产环境中的奖励黑客或不安全智能体行为 | 中 | 高 | 中 | 高 | 需要公开证据说明红队测试、回滚节奏和事故后控制 |
| 数据泄漏或工具访问权限过宽 | 中高 | 高 | 中 | 高 | 已有公开信任声明,但外部证明仍薄弱 |
| 实时工作流中的延迟 / 可靠性不达标 | 中 | 高 | 中 | 中高 | 服务层未公开正常运行时间或 SLA 披露 |
| 定制客户环境中的实施差异 | 高 | 中高 | 中低 | 高 | 未公开上线周期分布或人员配比证据 |
| 基准测试成功未能转化为持久产品使用 | 中 | 高 | 中低 | 高 | 需要旗舰案例之外的生产采用和留存证据 |
运营风险不在于某一次已知宕机,而在于定制化模型专精能否在不同客户环境中安全、反复跑通。
[CR005, CR010, CR011, CR012, CR031, CR035]纳入当前公开缓释后,竞争、治理负担和交付模式扩张是剩余风险最高的三项。
评分是对保留公开证据的定性综合,不是模型输出。它们用于强制排序剩余风险,而不是数学模拟概率。
[CR001, CR010, CR018, CR030, CR037, CR038]7.2 安全与合规负担真实存在,可能在引发头条事故前就拖慢企业采用
最清晰的非竞争风险不是今天已有已知执法行动,而是智能体式 AI 周围治理负担的广度。CISA 警告,组织需要更强监督来管理智能体式 AI 服务。NIST 的 AI RMF 把尽调面讲得很明确:安全、有韧性、负责、透明、增强隐私、管理偏见的系统才是目标标准。EU AI Act 又给 AI 风险加上正式法律框架;这很重要,因为 Applied Compute 正卖向高风险企业工作流,审计、日志、可解释性和人类监督设计都会变成采购问题。McKinsey 的调查显示 80% 组织已经见过有风险的智能体行为,也强化了这一点:买家不只是在购买准确率,而是在承保运营信任。Applied Compute 的公开信任主张方向正确,但缺少丰富的公开信任中心包,意味着每个受监管或安全敏感部署仍可能变成一次定制尽调。[CR010, CR011, CR012, CR013, CR014, CR015]
| 风险 / 框架 | 司法辖区 | 状态 | 可能性 | 严重性 | 缓释措施 | 剩余敞口 | 尽调路径 |
|---|---|---|---|---|---|---|---|
| 高风险 agent 工作流的 EU AI Act 义务 | EU / 跨境企业账户 | 框架已生效;义务取决于用例 | 中 | 高 | 风险分层、日志、文档和人类监督设计 | 在精确用例映射形成文档前,敞口高 | 将客户工作流映射到 AI Act 类别;要求内部合规矩阵 |
| agentic AI 的安全和隐私治理 | 美国及全球企业买家 | 以指引为主,不是审批制 | 高 | 高 | 与 CISA / NIST 对齐的控制、VPC 部署、可审计性说法 | 公开证明包薄,因此敞口高 | 要求信任中心材料、DPA 条款和红队 / 事故流程文档 |
| 私营公司披露不透明 | 美国投资人尽调 | 未见公开 SEC 运营申报 | 高 | 中 | 投资人尽调、董事会报告、NDA 下的经审计财务 | 审阅私有材料前为中-高 | 要求经审计报表、董事会材料摘录和控制叙事 |
| IP / 模型输出 / 训练数据争议 | 多司法辖区 | 留存来源中未发现公开争议 | 中 | 中高 | 客户自有数据 / 模型定位与法律审查 | 中,因为前沿 AI 的法律标准仍未稳定 | 审阅 MSA、赔偿条款、模型提供商下传条款和数据权利表述 |
| 受监管客户带来的合同合规负担 | 企业采购 | 法律、金融科技和公共部门类账户中可能较高 | 中高 | 高 | 内嵌合规支持和部署灵活性 | 中高,因为服务负载较重 | 抽样查看安全问卷、供应商风险请求和采购周期 |
各行按对投资者的实际严重性排序,而不是按法律新颖性。当前没有执法行动,并不降低企业 AI 部署的合规负担。
[CR010, CR014, CR015, CR035, CR036]| 依赖项 | 对手方 | 作用 | 集中度 | 失效场景 | 严重性 | 缓释措施 | 剩余风险暴露 |
|---|---|---|---|---|---|---|---|
| 基础模型 | OpenAI / Anthropic / 其他实验室 | 底层模型能力和路线图 | 高 | 基座模型供应商追平质量差距或改变经济性 | 高 | 兼容多模型的架构和客户专属训练数据 | 高 |
| 云 / 执行栈 | AWS / Modal / 基础设施合作伙伴 | 训练、沙箱和服务支持 | 中 | 定价、可用性或功能变化挤压毛利或拖慢上线 | 中高 | 多供应商评估和平台抽象层 | 中高 |
| 旗舰客户背书 | DoorDash / Cognition / Mercor / Harvey | 商业证明和未来管线可信度 | 高 | 主要背书客户暂停、流失或缩小范围 | 高 | 分散行业客户名单,增加更多独立客户背书 | 高 |
| 监管机构和标准制定者 | CISA / NIST / EU 主管机构 | 安全与合规预期 | 中 | 控制负担上升快于产品化 | 高 | 将合规控制写入产品和文档 | 中高 |
| 电力 / 算力供应链 | GPU / 数据中心生态 | 后训练和推理的成本与容量 | 中 | 算力成本飙升或可用性收紧 | 中高 | 优化工作负载,优先服务高价值用例 | 中高 |
依赖项按它们影响增长、毛利或客户信任的直接程度排序。现阶段,客户背书集中度几乎和基础设施集中度同样重要。
[CR003, CR008, CR018, CR019, CR037, CR038]多数风险沿同一条链传导:控制负担或竞争压力拖慢部署,削弱留存证明,压缩利润率,并动摇估值支撑。
该 DAG 是方向性、因果性的,不是定量模型。它用来说明,看似分散的风险为什么会互相强化。
[CR016, CR017, CR019, CR030, CR032, CR041]7.3 交付模型证明产品价值,但带来人员、伙伴和集中度风险
Applied Compute 的客户胜利说明公司为什么存在:DoorDash、Cognition、Mercor 和 Harvey 都需要客户专属的评测框架、评分器或基准环境。这种深度让产品短期内难以复制,但也意味着部署可能过久依赖稀缺研究人才。公开记录反复描述一种嵌入式、前线部署动作:研究人员与客户团队合作,DoorDash 需要现场投入,Cognition 使用 Windsurf 的副本,Modal 描述了深度集成模型。除此之外,公司依赖一组外部供应商和参考客户。它不拥有基础模型层,依赖云和执行伙伴,公开销售故事也由少数成熟旗舰账户支撑。如果其中任何一个账户或伙伴变弱,公司可能同时失去收入和社会证明。核心问题是,Applied Compute 能否在增长迫使其超出精品前线部署团队规模之前,把足够多交付动作标准化。[CR001, CR002, CR004, CR005, CR006, CR007]
| 角色 / 职能 | 依赖或缺口 | 可能性 | 严重性 | 缓释措施 | 尽调路径 |
|---|---|---|---|---|---|
| 创始人 / 技术领导层 | 公开战略和可信度集中在三位创始人身上 | 中 | 高 | 建立产品、平台和 GTM 的二线负责人 | 要求提供组织架构图、留任计划和授权责任划分 |
| 前置部署研究员 | 部署看起来高度依赖人才且定制化 | 高 | 高 | 标准化实施手册,降低单次上线的研究负载 | 要求提供实施人员配比和上线周期分布 |
| 安全 / 合规运营 | 公开信任证据弱于客户复杂度所要求的水平 | 中 | 高 | 投入正式信任中心资产和控制文档 | 要求提供审计范围、事故流程和客户安全材料包 |
| 销售 / 客户成功可复制性 | 案例研究很强,但覆盖面窄 | 中高 | 中高 | 扩大背书客户基础,发布更多标准化证明 | 要求提供从试点到生产再到扩张的转化漏斗 |
| 前沿 AI 人才招聘市场 | 这类研究员竞争仍激烈 | 高 | 中高 | 用资本在创始人光环之外招聘并留住人才 | 审阅招聘计划、薪酬理念和流失历史 |
执行风险离不开人才风险,因为当前产品看起来高度依赖高技能定制和紧密客户接触。
[CR001, CR002, CR004, CR030, CR031, CR040]Applied Compute 同时依赖基础模型厂商、基础设施伙伴、旗舰标杆客户和新兴监管框架。
该图隔离出最可能影响收入质量和商业速度的外部依赖,不试图覆盖每个供应商或监管方。
[CR003, CR008, CR029, CR036, CR037, CR039]7.4 资本强度今天尚可控,但估值预期几乎不给粗糙执行留空间
2026 年 4 月融资买来了时间,也固化了预期。投后估值 $1.3 billion、披露融资 $160 million 之后,Applied Compute 已不再按早期研究型精品团队来评判。公司必须先证明,重定制交付模型能复合成耐久收入,再面对算力、能源和安全需求超过现有经营杠杆的风险。Gartner 警告,到 2027 年底,超过 40% 智能体式 AI 项目可能被取消;这应该被直接视为市场风险信号:即使技术很强,成本、风险控制或运营模型变化跟不上,也可能商业失败。IEA 和 Goldman 的能源与数据中心需求数据又加了一层压力,因为算力密集型后训练在结构上昂贵。承销含义很清楚:观察公开基准胜利是否继续转化为可重复生产部署、更短实施周期和更清晰留存证明。如果不能,估值瓦解会快过技术叙事。[CR009, CR016, CR017, CR018, CR019, CR032]
| 风险 | 可监测触发因素 | 阈值 / 事件 | 行动含义 |
|---|---|---|---|
| 竞争护城河压缩 | 旗舰胜利不再相较平台默认方案体现清晰结果差异 | 连续两次重大上线都未体现实质性的质量 / 成本优势 | 重新评估差异化和估值溢价 |
| 安全 / 治理负担 | 客户安全审查扩张快于产品化控制 | 多个战略交易因信任 / 合规要求停滞 | 先推进控制产品化,再加速增长 |
| 服务密集型扩张失败 | 实施周期或人员配比没有改善 | 新上线仍需要创始人级别或研究密集支持 | 将业务视为低倍数的混合服务模式 |
| 客户背书集中度 | 某个旗舰账户显著缩小范围或流失 | 客户名单扩大前,顶级背书客户流失或降级 | 假设管线转化减弱、集中度风险上升 |
| 算力 / 资本挤压 | 训练或推理成本趋势超过商业化证明 | 毛利率轨迹或烧钱速度上升,但收入可见度未增强 | 收紧投资立场,要求更清晰的单位经济 |
| 披露不透明持续过久 | 尽调中没有经审计财务或留存材料包 | 即便在 NDA 下也无法核实烧钱速度、现金可维持时间或客户集中度 | 视为高确信度投资判断的未解决阻碍 |
每行定义一条可监测的投资假设破裂路径,而不是模糊担忧。最重要的区别在于,哪些风险能靠产品化消化,哪些风险暴露业务模式结构性更难扩张。
[CR016, CR017, CR030, CR032, CR041, CR042]7.5 图表
08估值
8.1 融资事实成立,但公开记录更支持“跟踪 / 继续研究”,而不是明确买入
Applied Compute 显然已跨过独角兽门槛:公司自己宣布完成 $80 million 融资,投后估值 $1.3 billion,TechCrunch 也独立印证了相同估值和累计融资额。这验证了市场标记,但不能证明 $1.3 billion 已是新投资人的公平入场价。公开证据支持高产品野心、可信旗舰客户,以及对一家如此年轻公司而言异常强的案例结果。但收入、利润率、留存和客户集中度均未披露,证据还不足以支撑有信心的价格判断。因此,估值敏感性才是主导框架。在这个价格上,即使用慷慨的 ARR 倍数假设,也需要公开记录尚未展示的有意义规模。仅凭公开证据,最诚实的建议是密切跟踪公司;只有私人尽调揭示更强经济性,或价格纪律改善时才行动。公开市场投资人通常会要求补齐这些缺失背景,才会把融资头条当作耐久估值锚。[CV001, CV002, CV003, CV013, CV025, CV026]
| 建议 | 信心 | 风险评级 | 估值立场 | 决策含义 |
|---|---|---|---|---|
| 跟踪 / 继续研究 | 中 | 高 | 估值偏高 / 尚未定论 | 没有私下核实收入、毛利、留存和轮次条款前,不要按当前估值标记投资 |
该建议对价格敏感,不是对公司质量的判决。Applied Compute 仍可能用业绩兑现当前估值,但公开证据集还未跨过这道门槛。
[CV033, CV036, CV037]建议从真实融资事件出发,经过客户证明和证据不透明,最终落到跟踪 / 继续研究结论。
[CV001, CV003, CV033, CV036]当前 $1.3B 标记需要多少 ARR,取决于投资人认为合理的倍数区间,差异很大。
数值用估值除以假设 ARR 倍数反推。它们是敏感性锚点,不是 Applied Compute 披露的 ARR 区间。
[CV001, CV025, CV026, CV027, CV028, CV029]8.2 正面论点是客户质量证据强;反面论点是溢价价格下证据不透明
看多逻辑很容易理解。Applied Compute 在 DoorDash 和 Cognition 有可信客户证明,在 Mercor 和 Harvey 有基准可信度,创始人履历强,投资人叙事也契合企业 AI 价值可能沉淀的位置。如果这些早期部署是可重复专业化平台的前沿,当前价格最终可能显得温和。看空逻辑更迫近。几乎所有硬商业问题在公开材料里仍无答案:ARR、现金消耗、毛利率、NRR、合同规模和轮次结构都缺失。ZenML 对 DoorDash 案例经济性的质疑抓住了核心问题:技术证明质量不等于承销证明质量。公开证据目前支持一家高质量公司,但还没有充分支持 $1.3 billion 公允价值结论。换句话说,市场标记方向可能没错,但证据基础仍不足以支撑精确判断。[CV004, CV005, CV006, CV008, CV015, CV016]
| 论点 | 改变判断所需证据 |
|---|---|
| 一家 2025 年成立的 AI 公司能做到这样的旗舰部署,说明产品-客户契合度异常强 | 证明这些部署可复制,并与有意义的 ARR 绑定,而不只是耀眼的技术胜利 |
| 顶级投资人和法律确认让融资事件可信 | 提供真实轮次条款、稀释和优先权细节,用来判断估值本身,而不是只看融资新闻的光环 |
| Harvey 式私募可比公司证明,市场会为企业 AI 支付极端倍数 | 展示相近的收入规模或采用深度;否则 Harvey 只是方向性可比,不是同类可比 |
| 竞争对手透明定价显示市场机会大且活跃 | 证明 Applied Compute 能在这些定价基准之上捕获价值,而不是被基准框住 |
| 公开披露缺口是最清晰的空头理由 | 公开或在 NDA 下提供当前 ARR、毛利率、NRR、集中度和烧钱数据 |
反假设不是 Applied Compute 缺乏质量,而是估值纪律需要经济性证明,公开文件目前还给不出来。
[CV008, CV015, CV016, CV030, CV032, CV035]用紧凑 KPI 展示判断当前估值标记是否站得住的关键指标,以及公开材料中最薄弱的部分。
[CV001, CV003, CV013, CV033, CV036]8.3 可比背景说明高溢价 AI 估值有可能,但 Applied Compute 尚未拿出最佳可比公司背后的公开披露
当前文件中最有用的私有可比公司是 Harvey。CNBC 报道称,Harvey 截至 2026 年 1 月估值 $11 billion、ARR 约 $190 million,这意味着私有市场愿意给拥有可见规模、强采用率和顶级投资人支持的企业 AI 公司极高倍数。这有助于证明,市场可以高价定价高溢价应用层 AI。但它不能证明 Applied Compute 今天也应得到同样待遇。Harvey 的估值旁边有披露 ARR 和可见下游用户规模,Applied Compute 只有少数近期公开部署,且没有披露经济性。Palantir、Salesforce、Microsoft 和 Alphabet 等上市公司锚点仍有用,但主要是披露和规模可比,而不是干净倍数可比。由此得到的情景框架是不对称的:如果私人尽调显示 ARR 和可重复性远强于公开记录暗示,上行存在;但除非这些私人指标异常强,基准情形仍低于当前标记。这也是为什么这里更需要情景纪律,而不是对类别的名义热情。[CV010, CV011, CV017, CV018, CV030, CV034]
| 情景 | 假设 | 估值 / 回报逻辑 | 关键风险 | 概率信号 |
|---|---|---|---|---|
| 乐观 | Applied Compute 已经具备高两位数 ARR 等效可见度,旗舰胜利沉淀成可复制部署引擎,高溢价私募 AI 倍数窗口仍然打开 | 若私下尽调显示约 $100M ARR 潜力,且高溢价倍数维持,价值可超过当前 $1.3B 估值标记 | 当前 ARR 有证据缺口;竞争和实施扩张仍是主要风险 | 中低 |
| 基准 | 公司质量真实,但经济性仍不成熟,且部分偏服务密集 | 仅凭公开证据,除非私下 ARR 和留存数据远强于已披露信息,否则价值应低于当前估值标记 | 毛利率、集中度和 NRR 不透明,使折价继续存在 | 中高 |
| 悲观 | 旗舰胜利停留在定制项目,买方因治理或 ROI 犹豫,平台透明定价挤压支付意愿 | 如果经济性更像高端服务层而非软件平台,公允价值可能显著低于独角兽估值 | 案例成功无法转化为持久 ARR 或毛利结构 | 中 |
情景有意保持方向性,因为公开文件缺少避免伪精确所需的输入。概率信号反映证据质量,也反映业务质量。
[CV031, CV032, CV038, CV039, CV040]| 可比对象 | 指标 | 倍数 / 估值 / 状态 | 参考价值 | 局限 |
|---|---|---|---|---|
| Harvey | 私募轮次 + 已披露 ARR | ~$11B 估值;~$190M ARR;据 CNBC 约 57.9x ARR | 企业 AI 应用可拿到极端高溢价倍数的最佳证据 | 已披露 ARR 和用户规模远大于 Applied Compute |
| Palantir | 公开申报的规模锚点 | 2026 Q1 申报文件显示季度毛利 $1.4B | 关键任务型企业 AI / 运营叙事的参照锚点 | 成熟度远高得多的上市公司,不是早期可比对象 |
| Salesforce / Agentforce | 公开套件厂商 + AI 分发 | 面向企业深度分发的公开 AI CRM / 智能体平台 | 可参照其分发和捆绑对退出倍数的压力 | CRM 平台更宽泛,不是聚焦 AI 的专业公司 |
| Microsoft / Azure Foundry | 公有云 + 定价透明度 | 已发布 AI Foundry 和模型定价;上市公司披露 | 可参照成本透明度和平台竞争 | 不是初创公司的直接估值可比对象 |
| Alphabet / Google AI | 公开 AI 套件 / 投资者披露 | 大盘股披露和企业 AI 分发 | 可作为披露和竞争天花板 | 规模和业务组合与 Applied Compute 不具可比性 |
| NVIDIA | 公开 AI 基础设施规模 / 投资者披露 | 拥有详细投资者关系材料的大盘 AI 基础设施龙头 | 可作为基础设施级 AI 股权热度和披露质量的天花板 | 基础设施经济性和业务组合与 Applied Compute 差异很大 |
由于没有干净的公开类比对象,本表有意混合私募轮次、上市公司和平台分发型可比对象。参考价值来自每个可比对象能说明什么,而不是来自虚假的对称。
[CV010, CV017, CV018, CV019, CV030, CV034]在几个示例 ARR 水平下,Applied Compute 是否配得上混合、强软件或溢价增长倍数,会让公允价值出现巨大差异。
低档使用 5x ARR,中档使用 12x ARR,高档使用 18x ARR。这些是场景分析的示例估值区间,不是市场报价。
[CV025, CV026, CV027, CV028, CV038, CV039]8.4 决定性问题是私人尽调能否把叙事强度转化为经济证明
这笔投资首先不会败在产品质量上。它会失败于:在市场变得更竞争、更透明之前,明星案例没有及时转化为可重复经济性。因此,最终尽调极其关键。投资人需要当前 ARR、毛利率、NRR、客户集中度,以及 2026 年 4 月轮次的精确结构,才能承销这个价格。没有这些,治理不透明和商业化风险就应该打折。Gartner 的智能体式 AI 项目取消警告,以及竞争对手定价透明度丰富,也进一步提高了一厢情愿的成本。如果尽调能证明 Applied Compute 已接近高溢价私有 AI 倍数隐含的 ARR 和留存门槛,当前标记可以辩护。否则,公司仍可能很优秀,但入场价依然偏贵。这个区分就是建议的核心。公司需要把技术声望转化为透明、可重复、耐久的经济性。[CV014, CV019, CV020, CV021, CV022, CV023]
| 触发因素 | 阈值 | 对投资假设的传导 | 行动含义 |
|---|---|---|---|
| ARR 显著低于敏感性阈值 | 私下尽调显示 ARR 远低于即便 12x-18x 倍数逻辑也需要的水平 | 削弱当前估值标记,也削弱高溢价软件叙事 | 按当前价格放弃,或要求更低进入价格 |
| 毛利率或服务密集度过弱 | 交付模式更像研究密集型服务,而非可扩张软件 | 压缩倍数,削弱退出可比性 | 重新按混合服务 / 软件风险定价 |
| 客户集中度严重 | 少数旗舰账户主导收入或证明叙事 | 提高波动性和背书脆弱性 | 要求集中度折价或延后 |
| 治理 / 合规摩擦拖慢增长 | 安全和采购负担显著拖慢部署 | 把技术护城河变成商业瓶颈 | 继续推进前,要求看到控制产品化证据 |
| 竞争性定价或捆绑挤压支付意愿 | 客户可替换为更便宜或捆绑的平台产品 | 降低长期高溢价倍数支撑 | 收紧估值纪律,监控赢单 / 输单数据 |
这些是技术上令人印象深刻的公司在当前估值标记下仍可能变成糟糕投资的最短路径。
[CV014, CV023, CV029, CV032, CV040]| 主题 | 缺失证据 | 重要性 | 负责人 / 尽调路径 |
|---|---|---|---|
| 按客户 / 用例拆分的当前 ARR 和增长 | 未公开披露 ARR | 用来检验当前价格是否至少在合理区间 | NDA 下的财务 / CEO 材料 |
| 毛利率和算力 COGS | 未公开披露毛利率 | 用来区分可扩张软件和高端服务 | 财务 + 基础设施负责人 |
| NRR、流失率和客户集中度 | 未公开披露留存或集中度 | 用来判断明星客户能否转化为持久收入 | 收入运营 / 客户成功 |
| 轮次结构与优先权条款 | 仅有估值大数 | 需要用来评估 2026 年 4 月估值标记的经济质量 | 法务顾问 + 融资文件 |
| 实施人员配比和上线周期 | 没有公开的可复制性指标 | 需要用来正确定价服务强度风险 | 运营 + 部署负责人 |
| 对平台既有厂商的赢单 / 失单数据 | 没有公开的竞争转化数据 | 需要判断当前价格下差异化能否守住 | 销售负责人 / 产品营销 |
在这些问题得到回答前,估值讨论应保持情景化和纪律性,而不是按高确信度定价。
[CV013, CV029, CV033, CV036, CV040]8.5 图表
免责声明
本报告是自动化流程于 2026-07-05 生成的研究和尽调材料。所有图表均来自公开证据;Applied Compute 管理层未审计或核验其中任何内容。前瞻性陈述和场景分析仅作示例。
证据索引
| 编号 | 陈述 | 可信度 | 来源 |
|---|---|---|---|
| CO001 | Applied Compute was founded in 2025 by three former OpenAI researchers: Yash Patil, Rhythm Garg, and Linden Li. | 高 | SO007, SO017, SO019 |
| CO002 | Independent reporting (TechStartups, Grokipedia) dates Applied Compute's founding specifically to May 2025. | 中 | SO017, SO019 |
| CO003 | Applied Compute is headquartered in San Francisco, California, with a principal office at 251 Rhode Island Street #207, and was originally incorporated in Delaware. | 高 | SO025, SO027 |
| CO004 | Applied Compute, Inc. registered as a foreign stock corporation with the California Secretary of State on October 10, 2025, under document number B20250336266. | 中 | SO025 |
| CO005 | Yash Patil is listed as Applied Compute, Inc.'s registered agent at the company's principal San Francisco address, per the California Secretary of State filing. | 中 | SO025 |
| CO006 | Applied Compute's product is 'Specific Intelligence': proprietary AI agents trained on a customer's own data, deployed into production, and continuously improved via reinforcement learning within the customer's own environment. | 中 | SO001, SO002, SO003 |
| CO007 | Applied Compute organizes its platform around three functions: Train (post-train tool-using agents on customer data), Serve (production-grade low-latency inference), and Improve (continuous online reinforcement learning from production traffic). | 中 | SO001 |
| CO008 | Applied Compute markets its training stack as 'model-flexible,' letting customers train from a chosen frontier base model and later upgrade to newer base models without changing the harness, data pipeline, or deployment stack. | 中 | SO001 |
| CO009 | Applied Compute embeds its own engineers directly with customer engineering teams rather than outsourcing or delegating model development. | 中 | SO003, SO005 |
| CO010 | Per the company's own account, two-thirds of Applied Compute's team are former startup founders, including former top AI researchers and Math Olympiad winners. | 中 | SO003, SO005 |
| CO011 | Lux Capital's portfolio profile describes Applied Compute's team as including alumni with reinforcement-learning-infrastructure experience at OpenAI, data-foundation experience at Scale AI, and additional experience from Together, Two Sigma, and Watershed. | 中 | SO015 |
| CO012 | Yash Patil is Applied Compute's CEO and co-founder. | 高 | SO025, SO003 |
| CO013 | Yash Patil was a key member of OpenAI's agentic Codex software-engineering effort before co-founding Applied Compute. | 高 | SO003, SO007 |
| CO014 | Rhythm Garg, a co-founder of Applied Compute, was a core contributor to OpenAI's o1, the first reinforcement-learning-trained reasoning model. | 高 | SO003, SO007 |
| CO015 | Linden Li, a co-founder of Applied Compute, worked on ML systems and infrastructure for reinforcement-learning training at OpenAI. | 中 | SO003, SO005 |
| CO016 | Per Applied Compute, Inc.'s California Secretary of State filing, Rhythm Garg is listed as Chief Financial Officer and Secretary of the company, with Yash Patil as the sole listed Chief Executive Officer. | 中 | SO025 |
| CO017 | Comcast NBCUniversal LIFT Labs' investor portfolio page instead lists Rhythm Garg's title as Chief Technology Officer, conflicting with the officer title recorded in the California Secretary of State filing. | 中 | SO027 |
| CO018 | Applied Compute's first institutional financing was a $20 million round at a $100 million valuation led by Benchmark partner Victor Lazarte, with Sequoia, Conviction, Hanabi Capital, Definition, and solo investor Zach Frankel also participating; Upstarts Media reported the round as still unannounced at the time of its scoop. | 中 | SO018 |
| CO019 | A separate secondary write-up (StartupsUnion) describes the same ~$100 million-valuation seed round as having closed in 'June 2024,' a date that conflicts with the 2025 founding and seed-round timing corroborated by Upstarts Media, TechStartups, and Grokipedia. | 低 | SO016 |
| CO020 | The Information reported in approximately September 2025 that Applied Compute was in talks to raise new funding at roughly a $500 million valuation. | 中 | SO017, SO007 |
| CO021 | Applied Compute publicly announced on October 29, 2025 that it had raised $80 million from investors including Benchmark, Sequoia, Lux Capital, Hanabi, Neo, Definition, Elad Gil, Victor Lazarte, and Omri Casspi. | 高 | SO003, SO007 |
| CO022 | Independent reporting (StartupHub.ai, AIM Media House) placed Applied Compute's valuation at approximately $700 million following the October 2025 $80 million round; the company itself did not disclose a valuation figure in its own announcement. | 中 | SO005, SO006 |
| CO023 | The Information reported in January 2026 that Applied Compute was in early talks to raise new funding at a $1.3 billion valuation — more than double the ~$500 million figure reported roughly three months earlier — with the round potentially reaching $70 million and Kleiner Perkins positioned to lead. | 中 | SO017 |
| CO024 | On April 8, 2026, Applied Compute announced it had raised $80 million in new financing at a $1.3 billion post-money valuation led by Kleiner Perkins, with continued participation from Elad Gil, Lux, Greenoaks, Neo, and Hanabi. | 高 | SO002, SO004 |
| CO025 | The April 2026 round brought Applied Compute's total disclosed funding to $160 million, per the company's own announcement and independently corroborated by TechCrunch citing PitchBook data. | 高 | SO002, SO023 |
| CO026 | Kleiner Perkins publicly described its April 2026 investment as Applied Compute's 'Series B,' framing its partnership as being with founders 'Yash, Rhythm, Linden.' | 中 | SO014 |
| CO027 | Latham & Watkins LLP represented Applied Compute in its April 2026 fundraise with an Emerging Companies & Growth team led by Bay Area partner Seth Gottlieb, with associates Kristine LaVeau, Camille N'Diaye-Muller, and Kavitha Babu. | 高 | SO004, SO002 |
| CO028 | TechCrunch's July 2026 tracker of 2026 unicorns lists Applied Compute at a $1.3 billion valuation, founded in 2025, with Benchmark and Sequoia among its investors and $160 million raised to date per PitchBook. | 高 | SO023, SO024 |
| CO029 | Applied Compute's valuation rose from roughly $100 million (mid-2025 seed) to an estimated ~$500 million (September 2025), ~$700 million (October 2025), and a company-confirmed $1.3 billion (April 2026) — an approximate 13x increase in under a year, a pace that is itself a notable diligence flag regardless of underlying fundamentals. | 中 | SO018, SO017, SO007, SO002 |
| CO030 | Applied Compute's publicly named early enterprise customers are Cognition (maker of Devin and the Windsurf IDE), DoorDash, and Mercor. | 高 | SO003, SO007 |
| CO031 | Applied Compute's own fundraise announcement references having 'learned from working with enterprises across the F500,' implying customers beyond the three named case studies, but no additional enterprise customer names are disclosed in any source reviewed. | 低 | SO002 |
| CO032 | Applied Compute and Cognition jointly developed SWE-check, a real-time bug-detection model embedded in the Windsurf IDE; Cognition states SWE-check runs roughly 10x faster than the frontier model it replaced (Opus 4.6) while narrowing the accuracy gap to frontier performance on in-distribution evaluations. | 高 | SO011, SO026 |
| CO033 | DoorDash used Applied Compute to build a calibrated automated grader and a reinforcement-learning-trained model that corrects errors in AI-generated merchant menus during onboarding. | 中 | SO012, SO028 |
| CO034 | The DoorDash deployment was a joint effort involving DoorDash ML engineer George Ignatius, DoorDash Head of Merchant ML Ying Yang, and the Applied Compute team working onsite at DoorDash's Sunnyvale office. | 中 | SO012, SO028 |
| CO035 | DoorDash co-founder Andy Fang credited Applied Compute's approach with helping DoorDash 'scale internal expertise and raise the bar on menu accuracy on the platform.' | 中 | SO012 |
| CO036 | An independent LLMOps case-study database (ZenML) reports that the DoorDash/Applied Compute reinforcement-learning menu-correction model achieved a 30% relative reduction in low-quality menus and was rolled out to all U.S. menu traffic. | 中 | SO028 |
| CO037 | The same independent ZenML case-study writeup cautions that its account 'is presented by Applied Compute, which naturally positions their tooling favorably,' flagging promotional bias as a limitation when evaluating the vendor's own effectiveness claims. | 中 | SO028 |
| CO038 | Applied Compute's custom-trained model 'Applied Compute: Small,' built using Mercor's APEX-Agents expert-graded benchmark, ranked #1 in the corporate-law category and 4th overall on the APEX-Agents leaderboard as of the case study's February 2026 publication, ahead of Opus 4.5 and GPT-5.2. | 中 | SO013 |
| CO039 | Mercor co-founder and Co-CEO Brendan Foody stated that 'Applied Compute gave us that answer quickly and precisely with frontier training infrastructure' when assessing whether expert-generated data could produce expert-level AI. | 中 | SO013 |
| CO040 | Third-party people-data aggregator RocketReach estimates Applied Compute's headcount at between 21 and 29 employees as of mid-2026; Applied Compute has not disclosed an official headcount figure in any source reviewed in this pass. | 低 | SO022 |
| CO041 | No source reviewed in this pass discloses Applied Compute's revenue, annual recurring revenue, or a customer count beyond the three named case-study customers (DoorDash, Cognition, Mercor). | 低 | SO002, SO003 |
| CO042 | Grokipedia's page on Applied Compute states the company emerged from stealth in October 2025 having secured a total of $100 million in funding at that time from Benchmark, Sequoia Capital, and Lux Capital. | 中 | SO019 |
| CO043 | Applied Compute's three co-founders are Stanford University alumni who were recent graduates or technical staff at the time they left OpenAI to start the company. | 中 | SO017, SO016 |
| CM001 | Applied Compute states it works with large enterprises to build and deploy agents inside customers' own production environments, including sitting in customers' offices to translate research into deployments. | 中 | SM003 |
| CM002 | Applied Compute frames the "AI overhang" as the gap between a model's raw capability and its realized utility in a specific enterprise workflow, and positions forward deployment as the mechanism that closes that gap. | 中 | SM003 |
| CM003 | Applied Compute operates two forward-deployed roles -- Forward Deployed Engineers (FDEs), who build evaluation frameworks and production environments, and Applied Research Engineers (AREs), who train and tune models -- to move agents from research into production. | 中 | SM003 |
| CM004 | Applied Compute's own positioning frames its addressable opportunity as training proprietary, task-specific models and agent platforms that outperform general frontier systems on an individual enterprise's own workflows, rather than competing as a general-purpose model lab. | 中 | SM002 |
| CM005 | The forward-deployed engineering model -- embedding vendor engineers directly inside customer teams to build and ship AI systems -- was pioneered by Palantir more than a decade ago. | 中 | SM025 |
| CM006 | In 2026 Palantir's Foundry platform productized this pattern as "AI FDE": an agent that executes Foundry data-pipeline, ontology, and code operations from natural-language requests, scoped to the requesting user's existing permissions. | 中 | SM025 |
| CM007 | AWS announced in 2026 a $1 billion investment to build a dedicated, agentic-first Forward Deployed Engineering organization that embeds engineers and AI agents directly inside customer teams to compress AI deployment timelines from months to days. | 中 | SM024 |
| CM008 | Named AWS Forward Deployed Engineering customers publicly disclosed as of mid-2026 include the Allen Institute, Cox Automotive, the NBA, the NFL, Ricoh, and Southwest Airlines. | 中 | SM024 |
| CM009 | Enterprise buyers increasingly mix internally built agent tooling with vendor-purchased agent platforms rather than choosing one exclusively: 65% of surveyed enterprise technology leaders report a hybrid build-and-buy architecture, with only about 10% relying on vendors alone. | 中 | SM026 |
| CM010 | General-purpose model API consumption billed per token, without an accompanying evaluation harness, reinforcement-learning customization layer, or forward-deployed integration work, functions as a substitute input rather than as standalone enterprise-agent spend in the sizing methodologies reviewed. | 中 | SM027 |
| CM011 | MarketsandMarkets estimates the global AI agents market at $7.84 billion in 2025, growing to $52.62 billion by 2030 at a 46.3% CAGR. | 中 | SM018 |
| CM012 | Grand View Research estimates the global AI agents market at $7.6 billion in 2025, $10.9 billion in 2026, and $182.9 billion by 2033 (49.6% CAGR from 2026-2033), with North America holding 39.6% share in 2025. | 中 | SM019 |
| CM013 | Precedence Research estimates the global AI agents market at $7.92 billion in 2025, $11.55 billion in 2026, and approximately $294.66 billion by 2035 (43.57% CAGR), with enterprises accounting for 67.1% of 2025 end-use revenue. | 中 | SM020 |
| CM014 | An independent tracker of agentic-AI market forecasts reports that Deloitte's TMT Predictions size the standalone agentic AI software market at $8.5 billion in 2026, growing to $35-45 billion by 2030. | 中 | SM027 |
| CM015 | The same tracker reports Fortune Business Insights sizing the standalone AI agent market at $7.29 billion in 2025, reaching $139.19 billion by 2034 at a 40.5% CAGR. | 中 | SM027 |
| CM016 | Gartner's broader lens -- counting agentic-AI capability embedded across all enterprise software rather than standalone agent vendors -- sizes 2026 agentic-AI-related enterprise software spending at $201.9 billion, roughly 25 times the standalone agent-software estimates from other analyst firms for the same year. | 中 | SM027 |
| CM017 | Within Gartner's total AI spending model, the standalone "agentic AI" spending sub-category is projected to compound at a 119% CAGR, expanding from roughly $15 billion toward $753 billion by 2029. | 中 | SM027 |
| CM018 | Gartner forecasts worldwide AI spending will reach $2.52 trillion in 2026 (up 44% year-over-year), split roughly $1.37 trillion infrastructure (54%), $452.5 billion software, and $588.6 billion services. | 中 | SM027 |
| CM019 | Four independent market-research firms (MarketsandMarkets, Grand View Research, Precedence Research, and Fortune Business Insights) each size the standalone AI-agent software market within roughly a $7.3-8.5 billion band for 2025-2026, despite disagreeing sharply on 2030s-decade CAGR and terminal value. | 中 | SM018, SM019, SM020, SM027 |
| CM020 | Independent analyst coverage agrees that standalone enterprise AI-agent software spending remains in the single-digit-to-low-teens billions in 2026 (roughly $8.5B-$11.6B), far below Gartner's own broader estimate of $201.9 billion for agentic-AI-embedded enterprise software the same year, reflecting a market-boundary difference rather than a factual disagreement. | 高 | SM011, SM019 |
| CM021 | Gartner predicts up to 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. | 中 | SM011 |
| CM022 | Gartner's best-case scenario projects agentic AI could drive approximately 30% of enterprise application software revenue by 2035 (surpassing $450 billion), up from about 2% in 2025. | 中 | SM011 |
| CM023 | Gartner predicts that by 2027 one-third of agentic AI implementations will combine multiple agents with different skills to manage complex tasks, and that by 2028 a third of user experiences will shift from native applications to agentic front ends. | 中 | SM011 |
| CM024 | Gartner predicts at least 15% of day-to-day work decisions will be made autonomously via agentic AI by 2028 (up from 0% in 2024), and that 33% of enterprise software applications will include agentic AI by 2028 (up from under 1% in 2024). | 中 | SM012 |
| CM025 | Gartner predicts more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value, or inadequate risk controls. | 中 | SM012 |
| CM026 | A January 2025 Gartner poll of 3,412 webinar attendees found only 19% of organizations had made significant investments in agentic AI, 42% conservative investments, 8% none, and 31% were still taking a wait-and-see approach. | 中 | SM012 |
| CM027 | Gartner estimates only about 130 of the thousands of vendors marketing agentic AI products have substantial agentic capability, with many others engaged in "agent washing" -- rebranding existing chatbots, RPA, or assistants without material autonomy. | 中 | SM012 |
| CM028 | Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents due to governance gaps identified only after production incidents occur, because applying uniform governance across agents of different autonomy levels creates either over-restriction or under-restriction failure modes. | 中 | SM013 |
| CM029 | MIT-affiliated NANDA research found that 95% of enterprise generative/agentic AI pilots fail to deliver measurable business value or P&L impact, with failures concentrated in flawed enterprise integration and a lack of workflow-specific learning loops rather than model quality; the same research found engagements built with external vendors saw roughly 2x higher success rates than purely internal builds. | 高 | SM023, SM012 |
| CM030 | McKinsey's 2025 Global Survey on AI found 88% of organizations report regular AI use in at least one business function (up from 78% a year prior), yet just 39% report AI's impact is visible at the enterprise EBIT level, and nearly two-thirds of organizations remain in the experimentation or piloting phase rather than at scale. | 中 | SM014 |
| CM031 | McKinsey's 2025 survey found 62% of organizations are at least experimenting with AI agents specifically. | 中 | SM014 |
| CM032 | Deloitte's 2026 State of AI in the Enterprise report finds worker access to AI rose 50% in 2025 and that the number of companies with 40%+ of AI projects in production is set to double within six months, but only one in five companies has a mature governance model for autonomous AI agent oversight. | 中 | SM015 |
| CM033 | Deloitte's enterprise agentification guide frames "agentification" as a gradual, workflow-by-workflow transformation that requires weighing cost, workforce, and risk factors, rather than a single wholesale AI rollout. | 中 | SM016 |
| CM034 | Salesforce's Agentic Enterprise Index reports that agent creation among first-mover Agentforce customers grew 119% between January and June 2025, with average customer-service conversations led by an agent increasing 22-fold over the same period. | 中 | SM017 |
| CM035 | Salesforce reports employee interactions with AI agents grew at an average monthly rate of 65%, and that 94% of consumers chose to interact with an AI agent when given the option, in the first half of 2025. | 中 | SM017 |
| CM036 | Mayfield's 2026 CXO Network survey of 266 Fortune 50-Global 2000 technology leaders found 42% of organizations already have agentic AI in production and 72% combined production-plus-pilot deployment. | 中 | SM026 |
| CM037 | The same Mayfield survey found line-of-business leaders are now the largest AI-tool buying group at 46%, matching or surpassing CIOs (38%) and CTOs (38%) for the first time, marking a shift in enterprise procurement power. | 中 | SM026 |
| CM038 | The Mayfield survey found 58% of technology leaders cite data readiness and quality as the top blocker to agentic AI adoption for a fifth consecutive year, and that 84% require security/compliance sign-off as non-negotiable while 60% report only an early-stage or no formal AI-governance framework. | 中 | SM026 |
| CM039 | 70% of enterprises surveyed by Mayfield want to test agentic AI tools in a self-serve sandbox before committing budget, reshaping vendor go-to-market and procurement design. | 中 | SM026 |
| CM040 | Applied Compute and Harvey jointly post-trained a legal agent that outperformed every other available model, including Opus 4.8 Max and GPT-5.5 xhigh, on rubric pass rate across Harvey's open Legal Agent Benchmark (LAB), which spans more than 1,250 tasks across 24 legal practice areas and over 75,000 binary grading criteria. | 中 | SM008 |
| CM041 | Applied Compute's SWE-check bug-detection agent, co-developed with Cognition for the Windsurf IDE, runs roughly 10x faster than the frontier model it replaced (Opus 4.6) while meeting the quality bar needed for real-time, in-IDE use. | 中 | SM006 |
| CM042 | Applied Compute built a reinforcement-learning-trained model for DoorDash that corrected errors in AI-generated merchant menus during onboarding, delivering an accuracy improvement that DoorDash's co-founder credits with raising the platform's menu-accuracy bar. | 中 | SM005 |
| CM043 | Applied Compute's custom-trained "Applied Compute: Small" model ranked #1 in the corporate-law category and 4th overall on Mercor's APEX-Agents leaderboard as of a February 2026 case study, per company-hosted material not independently re-verified. | 中 | SM007 |
| CM044 | Applied Compute's four disclosed enterprise deployments span four distinct verticals -- legal (Harvey), software engineering (Cognition), marketplace operations (DoorDash), and labor-marketplace evaluation (Mercor) -- each sponsored by a domain-function leader rather than a single horizontal buyer persona. | 中 | SM005, SM006, SM007, SM008 |
| CM045 | The IEA estimates data centers consumed about 415 TWh of electricity in 2024 (roughly 1.5% of global electricity consumption), growing about 12% per year over the prior five years, and projects consumption will nearly double to about 945 TWh by 2030 in its base case, with AI-accelerated servers driving about 30% annual growth in their own electricity use. | 中 | SM021 |
| CM046 | Goldman Sachs Research estimates U.S. data-center construction spending has tripled over the last three years and forecasts global data-center power demand will rise 165% by 2030 versus 2023 levels, with data-center capacity growing to about 92 GW by 2027. | 中 | SM022 |
| CP001 | Applied Compute publicly frames its differentiation as building proprietary, task-tuned "specific intelligence" rather than shipping a single horizontal end-user agent product. | 中 | SP002 |
| CP002 | Applied Compute's published case studies show it acts as an RL post-training and evaluation-infrastructure partner that trains customer-owned models, rather than selling its own branded legal, support, or coding agent to end users. | 中 | SP003, SP004, SP005 |
| CP003 | Harvey is simultaneously a collaboration customer of Applied Compute's Agent Cloud platform for reinforcement-learning post-training of its legal agent and an independent vertical incumbent that sells its own branded Assistant, Vault, and Knowledge products directly to law firms. | 高 | SP003, SP008 |
| CP004 | Applied Compute's other publicly named engagements -- Cognition (coding-agent bug detection) and DoorDash (merchant-onboarding automation) -- show the same infrastructure-provider pattern of training models for customers who then ship the resulting agent themselves. | 中 | SP004, SP005 |
| CP005 | Applied Compute's own disclosed valuation is materially smaller than the direct application-layer peers profiled in this chapter, giving it far less balance-sheet capacity to out-bid them for compute and research talent. | 中 | SP006, SP007 |
| CP006 | Harvey sells a unified legal-agent platform -- Assistant, Vault, and Knowledge -- directly to law firms and corporate legal departments, and reports more than 25,000 custom agents run by its customers as of its March 2026 funding announcement. | 高 | SP010, SP009 |
| CP007 | Harvey raised $200 million at an $11 billion valuation in a round co-led by GIC and Sequoia in March 2026, roughly doubling a valuation reported only months earlier. | 中 | SP010 |
| CP008 | Glean positions its product as a horizontal enterprise "AI Agents" platform for building, deploying, and orchestrating agents across a company's existing knowledge and application stack, rather than one narrow vertical use case. | 中 | SP011 |
| CP009 | Glean raised a $150 million Series F at a $7.2 billion valuation to accelerate its enterprise AI-agent product expansion globally. | 中 | SP013 |
| CP010 | Sierra sells a managed customer-service "Agent OS" that combines a no-code Agent Studio and Journeys builder with a forward-deployed engineering team that configures and iterates on agents on the customer's behalf. | 中 | SP014 |
| CP011 | Sierra crossed $150 million in annual recurring revenue within eight quarters of launch and serves more than 40% of the Fortune 50, according to founder Bret Taylor and Sierra's own customer materials. | 高 | SP016, SP015 |
| CP012 | In May 2026 Sierra raised $950 million at a $15.8 billion post-money valuation led by Tiger Global and Google's GV, making it, by its own investors' account, multiples larger by revenue than the next-largest company in its category. | 中 | SP016 |
| CP013 | Sierra's leadership describes its technology stack as a "constellation of" third-party foundation models from OpenAI and Anthropic combined with Sierra's own fine-tuned proprietary layers, rather than a single foundation model or a third-party RL-training vendor. | 中 | SP016 |
| CP014 | Decagon sells a conversational "AI concierge" product built around natural-language "Agent Operating Procedures" that compile into governed code, targeting consumer-facing customer-experience teams. | 中 | SP017, SP019 |
| CP015 | Decagon's disclosed customer roster spans travel, financial services, health, and retail brands including Avis, Hertz, Block, Affirm, Duolingo, and Oura, per its own January 2026 funding announcement. | 中 | SP019 |
| CP016 | Decagon raised a $250 million Series D led by Coatue Management and Index Ventures in January 2026, tripling its valuation to $4.5 billion in the six months since its prior round. | 中 | SP019 |
| CP017 | Decagon's prior Series C, announced about one year after the company emerged from stealth, raised $131 million at a $1.5 billion valuation -- roughly a third of its subsequent Series D valuation. | 中 | SP018 |
| CP018 | Cognition, maker of the Devin coding agent, raised $400 million at a $10.2 billion valuation in September 2025, up from a $4 billion valuation earlier the same year, with Devin's disclosed ARR climbing to $73 million in June 2025 from about $1 million a year earlier. | 中 | SP022 |
| CP019 | OpenAI bundles its own AgentKit tooling -- Agent Builder, ChatKit, evaluation tools, and Guardrails -- directly into its developer platform so customers can build, evaluate, and deploy agents without buying a separate agent-orchestration vendor. | 中 | SP023 |
| CP020 | OpenAI's newer "Frontier" enterprise offering lets partners such as Oracle, State Farm, and Uber build and manage agents that move across a company's own systems and data rather than staying inside one point product, positioning it against narrower agent-orchestration vendors. | 中 | SP024 |
| CP021 | OpenAI reports enterprise revenue is now more than 40% of its total revenue and is targeting parity with consumer revenue by the end of 2026, with Codex weekly active users reaching 3 million and its APIs processing more than 15 billion tokens per minute. | 中 | SP024 |
| CP022 | OpenAI runs "Frontier Alliances" go-to-market partnerships with McKinsey, BCG, Accenture, and Capgemini plus infrastructure partnerships with AWS, Databricks, and Snowflake to distribute its agent platform into enterprise accounts. | 中 | SP024 |
| CP023 | OpenAI's own enterprise materials cite Cursor, DoorDash, Thermo Fisher, and LY Corporation among existing customers building multi-agent systems directly on OpenAI's models, indicating at least one Applied Compute case-study customer (DoorDash) is simultaneously building agent capability on a hyperscaler platform. | 中 | SP024, SP005 |
| CP024 | OpenAI's developer-facing "Building agents" track packages its own frameworks, evaluation harnesses, and deployment guidance into a single funnel for teams building agents on OpenAI models, reducing the need to adopt a third-party agent-orchestration or RL-training vendor for common patterns. | 中 | SP025 |
| CP025 | Salesforce's Agentforce reached roughly $800 million in ARR in fiscal Q4 2026 (up 169% year-over-year) and, combined with Data 360, exceeded $2.9 billion in ARR, with more than 29,000 Agentforce deals closed since launch. | 中 | SP027 |
| CP026 | More than 60% of Agentforce and Data 360 Q4 FY26 bookings came from expansion within Salesforce's existing customer base, indicating the product is largely up-selling into Salesforce's incumbent CRM install base rather than winning net-new logos away from standalone agent specialists. | 中 | SP027 |
| CP027 | Salesforce markets Agentforce as a single platform spanning sales, service, analytics, and a broader "Agentforce 360" surface bundled with its core CRM subscription, directly overlapping with narrower customer-service specialists such as Sierra and Decagon. | 中 | SP026 |
| CP028 | Microsoft Copilot Studio has been adopted by more than 230,000 organizations to build custom AI agents, and Microsoft 365 Copilot separately reports more than 15 million paid seats as of early 2026. | 中 | SP029 |
| CP029 | Microsoft markets Copilot Studio as a low-code agent-building tool embedded directly inside the Microsoft 365 and Teams subscriptions its enterprise customers already pay for, giving it default distribution that narrower agent-platform vendors must win deals against. | 中 | SP028 |
| CP030 | Gartner forecasts that more than 40% of agentic-AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, a demand-side risk shared by every vendor profiled in this chapter, including Applied Compute. | 中 | SP032 |
| CP031 | Independent 2026 testing of Cognition's Devin found a task-dependent success rate (about 78% on well-scoped bug fixes) alongside documented limitations on ambiguous requirements and security-vulnerability awareness, plus a roughly $500-per-month cost floor, showing that even a heavily funded vertical incumbent has unresolved reliability gaps that keep human review necessary. | 中 | SP033 |
| CP032 | A public startup aggregator lists dozens of additional agentic-AI companies -- including LangChain, Agent Bricks, Manus AI, CopilotKit, and Sarvam AI -- as adjacent or potential entrants into the same agent-tooling space Applied Compute occupies, though most are earlier-stage and less funded than the named direct peers. | 低 | SP030 |
| CP033 | Independent 2026 buyer guides find Decagon requires more customer-side engineering integration while Sierra offers a more fully managed, forward-deployed-engineering delivery model, showing the two customer-service vertical peers differentiate on control versus convenience rather than on core model technology. | 中 | SP031 |
| CP034 | Independent 2026 buyer guides report Decagon and Sierra enterprise contracts typically start around $95,000-$150,000 per year and can reach $200,000-$350,000 once forward-deployed integration work is included, with neither vendor publishing self-serve list pricing. | 中 | SP031 |
| CP035 | Harvey, like Sierra and Decagon, does not publish self-serve list pricing; its own materials describe only that agents are sold and deployed inside law-firm and legal-department accounts, leaving contract economics undisclosed publicly. | 中 | SP008, SP009 |
| CP036 | Applied Compute discloses no public list pricing for its Agent Cloud RL post-training platform; its public materials describe collaborative engagements (e.g., with Harvey) rather than a subscription price card. | 中 | SP003, SP002 |
| CP037 | Decagon and Sierra both rely on named forward-deployed or professional-services engineering teams to configure and iterate on customer agents post-sale, which raises switching costs by embedding vendor staff in the customer's operational workflow rather than shipping a pure self-serve product. | 中 | SP017, SP014, SP031 |
| CP038 | Harvey's Vault and Knowledge modules store and index a law firm's own documents and precedent inside Harvey's platform, creating a data-gravity switching cost that is separate from and additive to underlying model quality. | 中 | SP009 |
| CP039 | Salesforce and Microsoft distribute their agent-building tools as add-ons inside CRM and Microsoft 365 subscriptions enterprises already renew annually, giving both hyperscalers a default incumbency advantage over standalone agent vendors that must win a net-new procurement decision. | 中 | SP026, SP028 |
| CP040 | Sierra's Bret Taylor states the company is deliberately raising outsized capital specifically to "invest aggressively" and preserve its lead against a large number of well-funded rivals, implying multi-homing and competitive-displacement risk are active concerns even for the category's reported revenue leader. | 中 | SP016 |
| CP041 | OpenAI's own Frontier pitch explicitly targets multi-product, cross-system agent orchestration as its differentiator versus point-product agent vendors, arguing that agents that move across a company's systems and data are stickier than agents embedded within a single product or environment. | 中 | SP024 |
| CP042 | Because Sierra, Harvey, and other application-layer peers build on interchangeable third-party foundation models (OpenAI, Anthropic, and others) rather than a single proprietary base model, enterprise buyers retain some ability to multi-home the underlying model layer even after committing to a vertical agent vendor. | 低 | SP016 |
| CP043 | Hyperscaler labs (OpenAI Frontier) and richly funded vertical incumbents (Harvey, Sierra) are increasingly building or buying their own post-training and fine-tuning capability in-house, directly threatening Applied Compute's core RL-post-training-as-a-service wedge. | 中 | SP024, SP016, SP010 |
| CP044 | Because Applied Compute's public proof points are the same handful of customers -- Harvey, Cognition, and DoorDash -- who are themselves scaling toward or past multi-billion-dollar valuations, those customers gain the balance-sheet capacity to insource RL post-training once initial capability transfer is complete. | 中 | SP003, SP004, SP005, SP022 |
| CP045 | Direct application-layer peers have raised far larger, faster-growing rounds than Applied Compute in the same 12-month window -- Sierra ($950M/$15.8B in May 2026), Harvey ($200M/$11B in March 2026), Decagon ($250M/$4.5B in January 2026), and Cognition ($400M/$10.2B in September 2025) -- giving them materially more capital to fund compute, talent, and their own in-house post-training than Applied Compute's own disclosed round. | 中 | SP016, SP010, SP019, SP022, SP006 |
| CP046 | Cognition laid off roughly 30 staff and offered buyouts to about 200 remaining employees in August 2025 amid reports of demanding 80-hour, six-day work-week expectations, even as its valuation more than doubled that same quarter, indicating execution and culture risk inside at least one heavily funded direct peer. | 中 | SP022 |
| CP047 | Sierra founder Bret Taylor forecasts a market correction within roughly two years and a "culling effect" where capital dries up for all but category leaders, a caution that applies to the broader set of richly valued agent vendors profiled in this chapter. | 中 | SP016 |
| CP048 | DoorDash appears in Applied Compute's own case studies as a customer of its RL post-training platform and separately in OpenAI's enterprise materials as a customer building multi-agent systems directly on OpenAI's models, and no reviewed source clarifies whether these are the same, overlapping, or entirely separate workloads inside DoorDash. | 低 | |
| CP049 | Applied Compute is named among the roughly 90 new unicorns minted in 2026 tracked by TechCrunch, placing it in a much larger cohort of newly minted billion-dollar AI startups competing for the same pool of enterprise agent budget and technical talent. | 中 | SP034 |
| CI001 | Applied Compute markets a three-stage 'Agent Cloud' platform (Train, Serve, Improve) that it sells as an embedded, managed engagement rather than a self-serve product. | 中 | SI001 |
| CI002 | Applied Compute's homepage and its Cognition case study do not publish any list price, per-token rate, subscription tier, or contract-minimum figure for its Agent Cloud platform or embedded engagements. | 高 | SI001, SI004 |
| CI003 | Applied Compute describes its go-to-market motion as embedded and managed, with its own engineers co-designing evals and sitting with customer engineering teams from the first eval through deployment. | 中 | SI001 |
| CI004 | In its DoorDash case study, Applied Compute engineers worked onsite at DoorDash's Sunnyvale office to translate production QA labels into an automated grader, illustrating a forward-deployed delivery motion. | 中 | SI005 |
| CI005 | None of Applied Compute's three published case studies (Cognition, DoorDash, Mercor) discloses a contract value, minimum spend commitment, or usage-based fee structure. | 中 | SI004, SI005, SI006 |
| CI006 | Applied Compute's own materials describe a revenue mechanism running from a customer's proprietary data and workflows through an embedded training engagement, to a production deployment, to an online-RL 'Improve' loop that could support renewal or expansion revenue, but no dollar figure anchors any stage of that chain. | 中 | SI001 |
| CI007 | OpenAI's ChatGPT Business plan for teams is listed at $20 per user per month when billed annually (or $25 billed monthly), while Enterprise pricing is custom and not published. | 中 | SI018 |
| CI008 | Anthropic's Claude plans range from a free tier to a Pro plan at $17-20 per month and a Max plan starting at $100 per month, well below the scale of Applied Compute's enterprise engagements and still fully published. | 中 | SI019 |
| CI009 | Microsoft's Foundry (Azure AI Foundry) publishes a consumption-based pricing structure across Foundry Models, Agent Service, Foundry IQ, and Foundry Tools, in contrast to Applied Compute's fully undisclosed enterprise pricing. | 中 | SI017 |
| CI010 | Cognition's Devin -- itself an Applied Compute customer -- prices its individual and team plans on a $0/$20/$200-per-month tier structure plus usage-based cloud-agent credits, a partial public reference point Applied Compute does not provide for its own services. | 中 | SI022 |
| CI011 | Together AI publishes per-million-token list prices for hosted open models, for example roughly $0.30 input / $1.20 output per million tokens for MiniMax M3 and $1.74/$3.48 for DeepSeek V4 Pro, giving an external anchor for raw inference costs. | 中 | SI023 |
| CI012 | Cerebras markets its wafer-scale inference cloud as materially faster than GPU-based inference at a lower cost per token, and Applied Compute's own Cognition case study confirms it used Cerebras inference to hit real-time latency targets for its SWE-check model. | 中 | SI027, SI004 |
| CI013 | OpenAI is winding down its public reinforcement fine-tuning (RFT) platform to new users as of the run date, narrowing the self-serve build-your-own-RL alternative enterprises might use instead of contracting Applied Compute. | 中 | SI026 |
| CI014 | Applied Compute markets SOC 2 certification and VPC/serverless deployment choice as built-in platform features rather than separately priced add-ons, so its security posture is bundled into an otherwise undisclosed price. | 中 | SI001 |
| CI015 | Palantir Technologies, whose Foundry AIP delivery model relies on forward-deployed engineers in a way Applied Compute's own case studies resemble, reported a GAAP gross profit of $1,416.8 million on $1,632.6 million of revenue (approximately 86.8% gross margin) for the quarter ended March 31, 2026. | 中 | SI021 |
| CI016 | Applied Compute has not disclosed a gross margin, cost-of-revenue figure, or any per-engagement profitability metric for its own business anywhere in its homepage, fundraise post, or launch essay. | 高 | SI001, SI002, SI003 |
| CI017 | In its Mercor case study, Applied Compute reports that fewer than 1,000 expert-labeled data points from Mercor were sufficient to nearly double a post-trained model's Pass@1 score and triple its score on a corporate-law evaluation, suggesting the marginal training-data cost of an engagement can be comparatively small relative to the performance gain. | 中 | SI015 |
| CI018 | Applied Compute has not disclosed its customer acquisition cost, average sales-cycle length, or any payback-period metric. | 高 | SI001, SI002 |
| CI019 | Applied Compute has not disclosed a net revenue retention rate, renewal rate, or expansion-revenue figure for any of its three publicly named customers. | 中 | SI004, SI005, SI006 |
| CI020 | An independent LLMOps case-study review of the DoorDash engagement notes that Applied Compute's own published material does not discuss the total cost of development, ongoing inference costs, or quantified business impact, and cautions that the case study is promotional material from Applied Compute that lacks the detailed methodology and statistical analysis expected of a rigorous technical publication. | 中 | SI016 |
| CI021 | Because Applied Compute's delivery model embeds engineers onsite with customers, as in the DoorDash engagement, a customer mix skewed toward smaller or shorter-duration accounts would plausibly compress margins in the same structural way that has historically challenged other forward-deployed-engineer businesses, even though Palantir itself now reports strong margins at scale. | 低 | SI021, SI005 |
| CI022 | Applied Compute's cost-to-serve stack plausibly includes GPU/inference compute (via partners such as Cerebras, as used in the Cognition engagement), forward-deployed engineering labor (as in the DoorDash onsite engagement), and licensed or partner-sourced training data/eval harnesses (as with Mercor), none of which the company breaks out publicly. | 中 | SI004, SI005, SI015 |
| CI023 | Applied Compute has raised a cumulative $160 million in disclosed external financing as of its April 8, 2026 Series B announcement. | 高 | SI002, SI007, SI011 |
| CI024 | Applied Compute's April 2026 Series B priced the company at a $1.3 billion post-money valuation, led by Kleiner Perkins with participation from Elad Gil, Lux Capital, Greenoaks, Neo, and Hanabi. | 高 | SI002, SI007, SI008 |
| CI025 | TechCrunch's July 2026 roundup of newly minted unicorns lists Applied Compute at a $1.3 billion valuation and $160 million raised to date, citing PitchBook, independently corroborating the company's own funding disclosures roughly three months after the round closed. | 中 | SI011 |
| CI026 | In January 2026, reporting attributed to The Information placed Applied Compute in talks to raise at a $1.3 billion valuation, up from roughly $500 million just months earlier -- a path that implies more than a 10x valuation increase in under a year even before the April 2026 round closed. | 中 | SI010 |
| CI027 | Applied Compute's own April 2026 fundraise announcement states the new financing will be used to grow the team, scale deployments, and bring to market 'the first generation of agent workforces built on specific models,' without a percentage breakdown of use of proceeds. | 中 | SI002 |
| CI028 | None of Applied Compute's public materials or its legal counsel's press release discloses any venture debt, project-finance facility, or credit line; every disclosed round to date is described as equity financing. | 高 | SI002, SI007 |
| CI029 | Applied Compute has not disclosed its cash balance, monthly cash burn, or runway following the April 2026 round in any source reviewed in this chapter. | 高 | SI002, SI011 |
| CI030 | Given the roughly 13x valuation escalation from an approximately $100 million seed to a $1.3 billion Series B within about a year, and no disclosed revenue base, Applied Compute's next financing looks more likely to be pulled forward by growth and investor demand than forced by an approaching cash shortfall, though this is an inference rather than a disclosed trigger. | 中 | SI010, SI011 |
| CI031 | A search of SEC EDGAR's company database for 'Applied Compute' returns no filings associated with the AI startup founded in 2025; the only similarly named registrant is an unrelated filer whose Exchange Act registration was revoked in 2006, confirming Applied Compute has no SEC-reportable public securities as of the run date. | 中 | SI024 |
| CI032 | Because Applied Compute is privately held with no SEC filings, none of the standard public-company disclosures (revenue, gross margin, cash position, headcount) available for a comparator like Palantir through its 10-Q are available for Applied Compute. | 高 | SI024, SI021 |
| CI033 | Applied Compute's public case studies name exactly three customers -- Cognition, DoorDash, and Mercor -- and no source reviewed in this chapter discloses a total paying-customer count, so the true customer base could be limited to these three plus undisclosed pilots or could be considerably larger. | 中 | SI004, SI005, SI006 |
| CI034 | Gartner forecasts that more than 40% of agentic AI projects will be canceled by the end of 2027 due to escalating costs, unclear business value, or inadequate risk controls, a category-wide demand risk directly relevant to a vendor like Applied Compute whose revenue depends on enterprises completing agentic AI deployments. | 中 | SI012 |
| CI035 | McKinsey's agentic-AI security playbook reports that 80% of surveyed organizations have already encountered risky agent behaviors such as improper data exposure, and recommends new governance, access-control, and contingency-planning investment before scaling agentic deployments -- incremental compliance costs that could raise the effective cost of Applied Compute's embedded delivery model without being reflected in any published price. | 中 | SI020 |
| CI036 | The IEA projects global data-center electricity consumption will nearly double to about 945 TWh by 2030, while Goldman Sachs Research forecasts data-center power demand will rise 165% by 2030 versus 2023 levels, a cost and supply backdrop that could raise the compute costs underlying any GPU-dependent vendor's gross margin, Applied Compute included. | 高 | SI013, SI014 |
| CI037 | a16z's analysis of AI-native application spending (based on Mercury banking data) found that some of the fastest-growing AI-native companies are built as end-to-end 'AI employee' substitutes rather than general productivity tools, a framing consistent with how Applied Compute's own case studies position its models as extensions of a customer's existing engineering or ML team. | 中 | SI025 |
| CI038 | Applied Compute's fundraise announcement and its launch essay both describe the company's mission in qualitative terms -- closing the gap between smart and useful, building Specific Intelligence -- without citing a single revenue, margin, or usage metric, consistent with a stage-appropriate but non-revenue-disclosing company. | 高 | SI002, SI003 |
| CI039 | Applied Compute's capital-adequacy profile combines a well-corroborated $160 million of cumulative funding and a $1.3 billion valuation with a complete absence of disclosed cash balance, burn rate, runway, or debt obligations, making cash-on-hand and burn the single largest open blockers to an outside runway assessment. | 高 | SI002, SI011 |
| CI040 | Bracketing each financing round against the hedged language used by the sources reporting it, Applied Compute's valuation moved within approximately $90-110 million at seed (mid-2025), $450-550 million at the September 2025 funding-talk stage, $650-750 million at the October 2025 stealth-exit round, and a confirmed $1.3 billion at the April 2026 Series B. | 中 | SI010, SI002, SI007 |
| CE001 | Applied Compute sells a Train/Serve/Improve platform for model training, inference, and continuous improvement rather than a standalone frontier base model. | 中 | SE001 |
| CE002 | Applied Compute says customers can train from the best available base models and later upgrade without changing the harness, data pipeline, or deployment stack. | 中 | SE001 |
| CE003 | Applied Compute says its Train layer supports post-training across text, images, code, and structured data using the customer’s own harnesses and graders. | 中 | SE001 |
| CE004 | Applied Compute says customers can serve the model in the same harness it was trained in, reducing train-to-deployment mismatch. | 中 | SE001 |
| CE005 | Applied Compute publicly claims support for dedicated single-tenant deployments in any region and operation either on its own cloud or fully inside the customer’s VPC. | 中 | SE001 |
| CE006 | Applied Compute publicly claims SOC 2 certification, role-based access control, audit logs on every dispatch, and data that never leaves the customer perimeter. | 中 | SE001 |
| CE007 | The Context Engine architecture is organized as Remember, Refine, and Retrieve: ingest enterprise resources and traces, refine them into a Contextbase, and expose retrieval APIs at runtime. | 中 | SE006 |
| CE008 | In Applied Compute’s APEX-Agents benchmark write-up, adding a Contextbase improved GPT-5.4 mean score@3 from 44.2% to 51.7% and GPT-5.4-mini from 33.4% to 38.7%. | 中 | SE006 |
| CE009 | Applied Compute’s internal ACL-Wiki memory system raised critical-memory retrieval from under 10% to around 20% over roughly two weeks as more traces and feedback accumulated. | 中 | SE007 |
| CE010 | Applied Compute’s “neural cheat-sheets” work explicitly tries to learn auditable context summaries that approach the utility of learned KV-cache-style memory without losing natural-language interpretability. | 中 | SE014 |
| CE011 | Modal describes Applied Compute’s core training loop as rollouts, evals, and inference operating continuously around custom customer tasks, with RL as the core training mechanism. | 中 | SE015 |
| CE012 | Modal says Applied Compute needed infrastructure that let rollouts, grading, and inference each run with different performance profiles while sharing state across the loop. | 中 | SE015 |
| CE013 | Modal says Applied Compute treats train-test mismatch as a consistent failure mode in deployed RL systems and therefore cares about high-fidelity mock environments and replayability. | 中 | SE015 |
| CE014 | DoorDash and Applied Compute turned internal QA labels into an automated grader and then used that grader as the reward function for reinforcement learning. | 中 | SE004 |
| CE015 | Applied Compute says its DoorDash menu-error-correction model reduced the share of low-quality menus by roughly 30% relative to baseline and was rolled out to all U.S. menu traffic, and ZenML repeats the same outcome in an independent LLMOps case summary. | 中 | SE004, SE031 |
| CE016 | Cognition says the jointly developed SWE-check bug-detection model matched frontier in-distribution performance much more closely than its starting point while running roughly 10x faster than Opus 4.6. | 中 | SE018 |
| CE017 | Cognition frames SWE-check as evidence that smaller, specialized models can rival frontier generalists on the tasks they are trained for while materially lowering cost and latency. | 中 | SE018 |
| CE018 | Mercor says fewer than 1,000 high-quality data points nearly doubled Pass@1 and mean score overall and tripled corporate-law Pass@1 in the cited long-horizon RL experiment. | 中 | SE019 |
| CE019 | Applied Compute and Mercor both present “Applied Compute: Small” as ranking #1 in corporate law and #4 overall on the APEX-Agents leaderboard at a fraction of the cost of larger frontier models. | 中 | SE005, SE019 |
| CE020 | Applied Compute’s routing research argues that model quality is not scalar and that different models excel at distinct task shapes, making routing a product capability rather than just an ops optimization. | 中 | SE010 |
| CE021 | In Applied Compute’s router experiment, oracle routing across Nemotron 3 Ultra, Claude Opus 4.7, and GPT-5.5 achieved a 0.890 average pass score versus 0.834 for the best single-model baseline. | 中 | SE010 |
| CE022 | Applied Compute’s inference benchmark says production multi-turn traces average about twenty tool turns with long tails into the hundreds, with assistant outputs in the low hundreds of tokens and prompts centered around roughly 10k tokens. | 中 | SE011 |
| CE023 | Applied Compute says its primary serving bottleneck on these agentic workloads is KV capacity and that cache management, scheduler pressure, and tool-latency tails are central system constraints. | 中 | SE011 |
| CE024 | Applied Compute’s staleness research says fully asynchronous RL improves utilization but creates off-policy lag whose magnitude depends on utilization, batch size, queue capacity, rollout concurrency, and response-length tailness. | 中 | SE013 |
| CE025 | Applied Compute’s high-leverage-samples research argues that when success probability is 10%, each successful rollout can carry roughly 81 times as much learning signal as a failed rollout. | 中 | SE008 |
| CE026 | Applied Compute’s entropy-preserving RL work says entropy collapse produces brittle, repetitive deployment behavior and worse continued training stability, while its REPO-R controller preserved entropy and kept multi-phase training improving. | 中 | SE012 |
| CE027 | Applied Compute’s RMSD research argues that self-distillation can teach out-of-distribution enterprise behaviors more effectively than standard SFT or sparse-reward RL while preserving existing capabilities. | 中 | SE009 |
| CE028 | Thinking Machines Lab independently argues that on-policy distillation combines the on-policy relevance of RL with dense token-level supervision and can be more compute-efficient than reward-only RL for later-stage training. | 中 | SE030 |
| CE029 | Anthropic recommends using the simplest architecture that works, reserving agent loops for cases where flexible model-directed action is necessary and explicitly warning that agents trade cost and latency for performance. | 中 | SE020 |
| CE030 | Cohere describes agentic workflows as automation of complex business processes with control, oversight, and security, emphasizing orchestration rather than pure chat UX. | 中 | SE021 |
| CE031 | OpenAI presents its Responses API as a foundation for production agents by bundling tool use, orchestration, and integrated tracing / observability into one surface. | 中 | SE022 |
| CE032 | Microsoft Copilot Studio positions itself as a platform for building and managing agents connected to business data, including autonomous capabilities and a shared control plane. | 中 | SE023 |
| CE033 | CISA says organizations adopting agentic AI need to align the systems with existing cybersecurity frameworks and strengthen oversight because agentic deployments introduce distinct security challenges and risks. | 中 | SE025 |
| CE034 | The Reward Hacking Benchmark reports exploit rates as high as 13.9% in tool-using RL-trained agents and says simple environmental hardening reduced exploit rates by 5.7 percentage points, or 87.7% relative. | 中 | SE026 |
| CE035 | Applied Compute’s GitHub organization shows one public repository and no public members, which limits how much external developers can independently verify about product breadth or community adoption. | 中 | SE016 |
| CE036 | The public `trie` repository is an Apache-2.0 lightweight harness for replaying inference traffic against an endpoint, with 21 stars, 3 forks, and a Jul. 2, 2026 update date visible in the fetched snapshot. | 中 | SE017 |
| CE037 | Applied Compute’s homepage attributes customer uses to more than coding: legal-agent training, support-ticket judgment, merchant menu onboarding, and biology benchmark work all appear in public testimonials. | 中 | SE001 |
| CE038 | Applied Compute says it can work with open-weight models from 1B to 1T+ parameters and swap models without rebuilding the stack. | 中 | SE001 |
| CE039 | DoorDash says its 2026 self-serve onboarding experience can help merchants launch more than 35% faster by using AI to pull photos, store hours, and menu items from an existing web presence. | 中 | SE024 |
| CE040 | Modal says Applied Compute evaluated nearly every sandbox and execution provider it could find before standardizing on Modal for rollouts, grading, and inference support. | 中 | SE015 |
| CU001 | Applied Compute’s homepage names Cognition, DoorDash, and Mercor as early customers and pairs each with a named operator quote. | 中 | SU001 |
| CU002 | Applied Compute’s homepage also presents Harvey, Bridge, and Latch Bio testimonials, indicating customer or design-partner breadth beyond the three flagship case studies. | 中 | SU001 |
| CU003 | Applied Compute says it is already building and validating customer models and agents in days instead of months, implying a high-touch deployment motion rather than pure self-serve software. | 中 | SU002 |
| CU004 | DoorDash uses Applied Compute for merchant menu onboarding and error correction, a workflow directly tied to merchant activation and order accuracy. | 中 | SU003 |
| CU005 | DoorDash had already built an AI system that converts photos, PDFs, and text menus into structured listings during merchant onboarding before Applied Compute improved the long tail of messy menus. | 中 | SU003 |
| CU006 | DoorDash and Applied Compute built an automated grader from QA labels and used it to RL-train a menu error correction model. | 中 | SU003 |
| CU007 | DoorDash’s case study says the share of low-quality menus fell by roughly 30% relative to baseline in testing. | 中 | SU003 |
| CU008 | DoorDash rolled the error-correction model to all menu traffic in the United States, which is strong evidence of production deployment rather than a limited lab pilot. | 高 | SU003, SU007 |
| CU009 | DoorDash says its 2026 AI-powered self-serve onboarding experience helps merchants launch more than 35% faster by pulling photos, store hours, and menu items from an existing web presence. | 中 | SU007 |
| CU010 | DoorDash says AI-powered websites built from existing merchant data are seeing average order conversion rates of nearly 10%, showing the company treats AI tooling as an ongoing merchant-growth motion rather than one-time setup. | 中 | SU007 |
| CU011 | Cognition’s Applied Compute case study frames the buyer as the product/engineering team behind Devin and Windsurf, with the end user being software engineers working inside an IDE. | 中 | SU004 |
| CU012 | Cognition tested viable open- and closed-source models for SWE-Check, but off-the-shelf options missed its combined quality, latency, and cost target for real-time IDE review. | 中 | SU004, SU008 |
| CU013 | Applied Compute and Cognition say SWE-Check delivers 10x faster bug detection than the frontier alternative and is in production powering Quick Review in Windsurf. | 高 | SU004, SU008 |
| CU014 | Cognition says SWE-Check was trained in a replica of the Windsurf environment and did not use customer code for training. | 中 | SU004, SU008 |
| CU015 | Cognition calibrated latency penalties to dogfooding-based user drop-off data, showing that the customer relationship includes product telemetry and iterative tuning rather than a static model handoff. | 中 | SU004 |
| CU016 | Cognition says Devin is deployed at some of the largest and most complex institutions in the world. | 中 | SU009 |
| CU017 | Cognition’s June 2026 Devin Desktop launch says millions of engineers use Windsurf and Devin. | 中 | SU011 |
| CU018 | Devin Desktop merges local and cloud agents into one command center, expanding the surface area where a specialized Applied Compute-trained model can remain useful after initial deployment. | 中 | SU010, SU011 |
| CU019 | Mercor describes itself as organizing human intelligence for the AI economy by sourcing and vetting domain experts to generate and evaluate frontier-model data. | 中 | SU005, SU013 |
| CU020 | Mercor’s APEX-Agents benchmark covers 480 tasks created by experts with 10+ years of experience across investment banking, management consulting, and corporate law. | 中 | SU005, SU012 |
| CU021 | Applied Compute’s custom-trained model ranked #1 on APEX-Agents for corporate law and 4th overall in the cited Mercor case study. | 中 | SU005 |
| CU022 | Mercor says it partnered with Applied Compute to post-train an open-source model using an expert-labeled dev set and saw substantial performance gains. | 中 | SU012 |
| CU023 | Applied Compute first deployed its long-horizon RL stack on Mercor’s expert-labeled dev set of 874 tasks across 50 worlds. | 中 | SU005 |
| CU024 | Mercor’s case study says Pass@1 and mean score nearly doubled overall, and corporate-law Pass@1 tripled from 4.4% to 16.3%. | 中 | SU005 |
| CU025 | Applied Compute’s Harvey case study says the engagement post-trained GLM-5.1 into the strongest available model on Harvey’s Legal Agent Benchmark by rubric pass rate. | 中 | SU006 |
| CU026 | Harvey says it helps law firms and legal teams work faster and smarter, and its LAB benchmark is designed for complex legal tasks. | 中 | SU006, SU014 |
| CU027 | Applied Compute says Harvey’s LAB contains more than 1,250 tasks across 24 legal practice areas and more than 75,000 binary criteria. | 中 | SU006 |
| CU028 | Harvey’s customer page says more than 142,000 lawyers across 1,500+ organizations in 60 countries rely on Harvey. | 中 | SU015 |
| CU029 | Harvey reports a 92% monthly adoption rate and 25+ hours saved per typical user per month, suggesting that a strong design partner could offer Applied Compute meaningful expansion volume if benchmark wins translate into production product surfaces. | 中 | SU015 |
| CU030 | Bridge’s testimonial on Applied Compute’s homepage says an agent now learns support-team judgment from SMEs in production, so expertise compounds with every ticket instead of living in a few people’s heads. | 中 | SU001 |
| CU031 | Bridge describes itself as stablecoin infrastructure that lets businesses receive, store, convert, issue, and spend stablecoins while Bridge handles regulatory, compliance, and technical complexity. | 中 | SU016 |
| CU032 | Applied Compute’s homepage includes a Latch Bio testimonial focused on benchmark quality and practical biology tasks, but it does not disclose deployment depth, production status, or commercial outcomes. | 中 | SU001 |
| CU033 | LatchBio’s scBench materials show the company is building benchmark infrastructure for biology agents, which supports Applied Compute’s claim that biology is an active target vertical even though public deployment evidence remains thin. | 中 | SU024, SU025 |
| CU034 | CISA says organizations adopting agentic AI need stronger oversight and should align deployments with existing cybersecurity frameworks, implying longer enterprise procurement cycles for customer-facing agent systems. | 中 | SU017 |
| CU035 | Gartner’s 2026 customer-service conference materials emphasize moving AI pilots beyond experimentation, redesigning operating models, and proving sustained ROI, which undercuts any assumption that customer logos automatically equal durable expansion. | 中 | SU018 |
| CU036 | Anthropic, Cohere, OpenAI, and Microsoft all frame enterprise agents as orchestration-heavy systems requiring control, tracing, and governance, which is consistent with Applied Compute’s specialization pitch but also means buyers can compare it against a widening field of platform alternatives. | 中 | SU019, SU020, SU021, SU022 |
| CU037 | Modal says Applied Compute embedded deeply with customers while running large-scale RL workloads, reinforcing that the company’s current customer motion likely remains services-intensive. | 中 | SU023 |
| CU038 | Applied Compute’s public materials do not disclose customer count, top-customer concentration, contract length, NRR, GRR, churn, or revenue by vertical. | 中 | SU001, SU002, SU026 |
| CU039 | Most quantified customer outcomes in the public record come from vendor-authored case studies or customer-partner blog posts rather than independent customer-authored ROI reports. | 中 | SU003, SU004, SU005, SU006, SU012 |
| CU040 | The current public customer set spans food delivery, developer tools, legal AI, fintech support workflows, expert-network evaluation, and biology benchmarking, suggesting broad applicability but also a bespoke, vertical-by-vertical deployment model. | 中 | SU001, SU003, SU004, SU005, SU006, SU016, SU024 |
| CR001 | Applied Compute’s public company narrative is unusually concentrated in three founders with frontier-lab pedigrees across Codex, RL reasoning, and RL infrastructure. | 高 | SR001, SR003 |
| CR002 | Applied Compute explicitly sells researchers working with customers on proprietary data and expertise, which raises key-person and scarce-talent dependency risk relative to a fully self-serve software company. | 中 | SR001, SR003 |
| CR003 | Applied Compute’s promise of swapping in stronger base models over time implies that the company does not own the foundational model layer and therefore faces some moat compression if foundation-model vendors improve vertical performance quickly. | 中 | SR001, SR010 |
| CR004 | DoorDash required onsite work with production QA labels and a custom grader, illustrating that some successful deployments may require forward-deployed customization rather than turnkey activation. | 中 | SR004 |
| CR005 | Cognition’s deployment required a replica of the Windsurf environment and iterative reward tuning against product telemetry, reinforcing integration complexity and execution risk on new accounts. | 中 | SR005 |
| CR006 | Mercor’s engagement depended on 874 expert-labeled tasks across 50 worlds, suggesting that customer-specific data creation can become a material delivery and margin bottleneck. | 中 | SR006 |
| CR007 | Harvey’s benchmark contains more than 1,250 tasks, 24 legal practice areas, and more than 75,000 binary criteria, implying that top-tier vertical wins may require unusually heavy evaluation and optimization effort. | 中 | SR007 |
| CR008 | Modal says Applied Compute embedded deeply with customers while standardizing on Modal after evaluating many execution providers, which creates both service-delivery dependence and third-party platform dependence. | 中 | SR008 |
| CR009 | Applied Compute’s April 2026 financing and the Latham announcement confirm fresh capital, but they also raise the performance bar attached to a $1.3 billion post-money valuation for a 2025-founded company. | 高 | SR002, SR009 |
| CR010 | CISA, NIST, and McKinsey all frame agentic AI deployment as a governance-intensive activity requiring oversight, secure design, and active risk management rather than simple model procurement. | 高 | SR013, SR014, SR019 |
| CR011 | McKinsey reports that 80% of organizations have encountered risky behaviors from AI agents, including improper data exposure and unauthorized system access. | 中 | SR019 |
| CR012 | CISA’s guidance explicitly warns that agentic AI services introduce security challenges and require organizations to strengthen oversight as adoption grows. | 中 | SR013 |
| CR013 | NIST’s AI RMF defines trustworthy AI to include security, resilience, accountability, transparency, privacy enhancement, and bias management, establishing a broad diligence surface for enterprise deployments. | 中 | SR014 |
| CR014 | The European Commission describes the AI Act as the first legal framework on AI and explicitly frames it around risk-based obligations. | 高 | SR015, SR016 |
| CR015 | The SEC EDGAR search result for Applied Compute reflects a thin public disclosure footprint, which is normal for a private startup but increases information asymmetry for diligence on governance, controls, and financing readiness. | 中 | SR017 |
| CR016 | Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of cost, unclear business value, or inadequate risk controls. | 中 | SR018 |
| CR017 | Gartner’s 2026 customer-service agenda emphasizes that enterprises still need operating-model redesign and hard ROI proof to scale AI beyond pilots. | 中 | SR031 |
| CR018 | The IEA and Goldman Sachs both point to rising power and data-center demand from AI, which threatens margin stability for any startup relying on compute-heavy post-training and inference loops. | 高 | SR020, SR021 |
| CR019 | Applied Compute’s product promise depends on repeatedly training and improving custom agents, so rising energy and GPU costs are strategically more important than for a lightweight SaaS wrapper. | 中 | SR001, SR020, SR021 |
| CR020 | Amazon Bedrock says it already powers generative AI for more than 100,000 organizations globally, giving AWS a distribution and infrastructure advantage that Applied Compute cannot match directly. | 中 | SR022 |
| CR021 | AWS AgentCore is marketed as a way to build, deploy, and operate AI agents securely and at scale without infrastructure management, tightening direct competition around the operating layer Applied Compute emphasizes. | 中 | SR023 |
| CR022 | OpenAI markets ChatGPT Enterprise as a secure and scalable enterprise offering, which increases the risk that buyers default to an incumbent generalist vendor instead of adopting a separate specialization layer. | 中 | SR024 |
| CR023 | Anthropic’s own guidance says the simplest workflow is often best and that full agent loops add cost and latency, which undermines the assumption that every customer problem needs Applied Compute-style heavy specialization. | 中 | SR025 |
| CR024 | Microsoft positions Copilot Studio as a way to build, manage, and govern enterprise agents, reinforcing the risk that large platform vendors converge on the same control-plane narrative Applied Compute uses. | 中 | SR026 |
| CR025 | Palantir markets AIP as a way to operationalize AI inside complex enterprises, making Palantir a particularly strong competitor for regulated, high-stakes deployments where ontology and workflow integration matter. | 中 | SR027 |
| CR026 | Google’s Gemini Enterprise page frames enterprise AI as something every employee can use securely, increasing competitive pressure from broad suite vendors with existing distribution. | 中 | SR028 |
| CR027 | Salesforce is positioning Agentforce as enterprise digital labor, extending competition into CRM-anchored workflows where Applied Compute would otherwise need to win greenfield trust. | 中 | SR029 |
| CR028 | Harvey’s customer page reports 142,000 lawyers and 1,500+ organizations, showing that specialized vertical incumbents can build large installed bases without relying on Applied Compute as a permanent platform layer. | 中 | SR012 |
| CR029 | Cognition says Devin is already deployed at some of the largest and most complex institutions in the world, which creates a realistic risk that sophisticated customers or partners may internalize more of the agent stack over time. | 中 | SR011 |
| CR030 | Applied Compute’s public model is explicitly embedded and managed, not self-serve, which can slow hiring leverage and keep gross margins below software-only expectations if not standardized quickly. | 高 | SR001, SR003, SR008 |
| CR031 | DoorDash, Cognition, Mercor, and Harvey each required customer-specific harnesses, graders, or benchmark environments, which raises onboarding complexity and implementation-time variance. | 中 | SR004, SR005, SR006, SR007 |
| CR032 | Applied Compute’s April 2026 raise brings total disclosed funding to $160 million, but the company has not publicly disclosed revenue, burn, or runway, making financial-model risk hard to bound from public data alone. | 高 | SR002, SR017 |
| CR033 | Kleiner Perkins’ framing of Applied Compute centers on closing the gap between frontier AI and real-world impact; if frontier models narrow that gap directly, the investor thesis weakens. | 中 | SR010, SR025 |
| CR034 | Palantir’s 2026 Q1 10-Q shows gross profit of $1.4 billion in a single quarter, underscoring the scale gap between Applied Compute and well-capitalized competitors that can subsidize enterprise AI platform adoption. | 中 | SR030 |
| CR035 | Applied Compute publicly claims SOC 2, VPC deployment, and auditability, but the public record still lacks a trust-center-style package that would let outside diligence verify scope and control maturity. | 中 | SR001, SR013, SR014 |
| CR036 | EU compliance obligations, NIST-style governance expectations, and CISA’s secure-adoption guidance together imply that enterprise sales cycles can lengthen materially in regulated or high-stakes workflows. | 高 | SR013, SR014, SR015, SR016 |
| CR037 | Because Applied Compute’s strongest references are also sophisticated product companies, loss or deterioration of a small number of flagship accounts would likely hurt both revenue concentration and go-to-market credibility. | 中 | SR004, SR005, SR006, SR007 |
| CR038 | The competitive field now spans AWS, OpenAI, Microsoft, Palantir, Google, Salesforce, and specialist incumbents, making sustained differentiation dependent on measured customer outcomes rather than frontier-AI branding alone. | 高 | SR022, SR024, SR026, SR027, SR028, SR029 |
| CR039 | OpenAI, Microsoft, AWS, and Salesforce each pair model access with broader enterprise distribution, increasing the risk that Applied Compute is judged as a premium professional-services layer rather than a must-have system of record. | 中 | SR022, SR024, SR026, SR029 |
| CR040 | The current public evidence supports strong technical capability but not yet a publishable proof set on implementation times, staffing ratios, or standardized launch playbooks. | 中 | SR001, SR003, SR008 |
| CR041 | In the absence of public revenue metrics, a plausible thesis-break trigger is benchmark or case-study success failing to translate into repeatable deployments with verifiable expansion economics. | 中 | SR002, SR018, SR031 |
| CR042 | A second plausible thesis-break trigger is if governance or compliance demands rise faster than Applied Compute can productize controls, turning each new regulated deployment into a bespoke security program. | 中 | SR013, SR014, SR015, SR016, SR019 |
| CV001 | Applied Compute’s April 2026 financing priced the company at a $1.3 billion post-money valuation and brought total disclosed funding to $160 million. | 高 | SV001, SV010 |
| CV002 | Applied Compute was founded in 2025, so the current public valuation has been reached within roughly a year of the company’s initial public emergence. | 高 | SV003, SV010 |
| CV003 | The public proof set behind that valuation is concentrated in four 2026 case studies—DoorDash, Cognition, Mercor, and Harvey—rather than a broad disclosed customer base with revenue metrics. | 中 | SV004, SV005, SV006, SV007 |
| CV004 | DoorDash provides the strongest production deployment proof, including rollout to all U.S. menu traffic, but no disclosed contract value, ROI, or retention metric. | 中 | SV005, SV016 |
| CV005 | Cognition provides a strong product outcome—10x faster bug detection in production—but still no disclosed revenue contribution or seat-scale economics for Applied Compute. | 中 | SV004, SV013 |
| CV006 | Mercor and Harvey are strategically valuable proof points, but both are benchmark-heavy references rather than clean public revenue or ARR disclosures for Applied Compute itself. | 中 | SV006, SV007, SV012 |
| CV007 | Latham’s financing announcement corroborates that the round was a formal financing event rather than only an informal press narrative. | 中 | SV008 |
| CV008 | Kleiner Perkins’ public thesis centers on closing the gap between frontier AI and real-world impact, which supports why investors would ascribe a premium narrative to Applied Compute. | 中 | SV009 |
| CV009 | TechCrunch’s 2026 unicorn tracker independently corroborates the $1.3 billion valuation and $160 million raised to date. | 中 | SV010 |
| CV010 | Harvey raised $200 million at an $11 billion valuation and had about $190 million ARR as of January 2026, implying a disclosed private-market ARR multiple of roughly 57.9x. | 中 | SV011 |
| CV011 | Harvey’s 142,000 lawyers across 1,500+ organizations indicate that the company’s valuation sits on top of a much more visibly scaled installed base than Applied Compute has publicly disclosed. | 中 | SV011, SV012 |
| CV012 | Cognition’s public materials show enterprise deployment scale for Devin, but Applied Compute has not publicly disclosed anything like the corresponding ARR or customer-count scale for its own platform. | 中 | SV013, SV014 |
| CV013 | The SEC company search result for Applied Compute highlights how little public operating disclosure is available on revenue, burn, or customer concentration. | 中 | SV014 |
| CV014 | Gartner’s warning that over 40% of agentic AI projects may be canceled by end-2027 is a direct adverse input to any valuation that assumes smooth commercialization of enterprise agents. | 中 | SV015 |
| CV015 | ZenML’s review of the DoorDash case study notes that total development cost, ongoing inference cost, and quantified business impact are not disclosed, which is unusually thin support for a $1.3 billion price if taken as the flagship public case. | 中 | SV016 |
| CV016 | The a16z AI application spending report supports the idea that application-layer AI value can be significant, but it does not provide company-specific evidence that Applied Compute has already captured that value at scale. | 中 | SV017 |
| CV017 | Palantir’s public filing showing $1.4 billion in quarterly gross profit underscores how far Applied Compute remains from public-company scale, even if both are positioned around high-stakes enterprise AI workflows. | 中 | SV018 |
| CV018 | Palantir IR, Salesforce IR, Microsoft Investor, and Alphabet Investor pages all show mature public-company disclosure environments that investors can triangulate, unlike Applied Compute’s still-opaque private-company file. | 高 | SV019, SV020, SV021, SV022 |
| CV019 | Microsoft publishes explicit Foundry pricing surfaces, which contrasts sharply with Applied Compute’s fully undisclosed platform pricing. | 高 | SV023, SV024 |
| CV020 | OpenAI’s pricing page publishes direct model and product prices, offering another transparent reference point that Applied Compute does not provide. | 中 | SV025 |
| CV021 | Together AI and Cerebras each publish infrastructure or inference pricing benchmarks, which helps anchor raw model economics but still does not reveal Applied Compute’s actual gross-margin structure. | 中 | SV026, SV027 |
| CV022 | Devin pricing is public, but Applied Compute’s pricing is not, reinforcing that even adjacent developer-agent products provide more economic transparency today. | 中 | SV028 |
| CV023 | OpenAI’s ChatGPT Enterprise offering and Palantir AIP reinforce that Applied Compute is pursuing a premium layer inside a market where giant vendors already bundle enterprise AI into broader platforms. | 中 | SV029, SV030 |
| CV024 | Because all publicly disclosed flagship deployments date from 2026, the current price is underwriting future repeatability more than a long measured operating history. | 中 | SV004, SV005, SV006, SV007, SV010 |
| CV025 | At a $1.3 billion enterprise-value proxy, a 25x ARR multiple would imply about $52 million of ARR. | 中 | SV001 |
| CV026 | At a $1.3 billion enterprise-value proxy, an 18x ARR multiple would imply about $72 million of ARR. | 中 | SV001 |
| CV027 | At a $1.3 billion enterprise-value proxy, a 12x ARR multiple would imply about $108 million of ARR. | 中 | SV001 |
| CV028 | At a $1.3 billion enterprise-value proxy, an 8x ARR multiple would imply about $163 million of ARR. | 中 | SV001 |
| CV029 | Those implied ARR hurdles are difficult to underwrite from public evidence because Applied Compute has not disclosed revenue, margins, retention, or contract sizes. | 高 | SV001, SV014, SV016 |
| CV030 | Harvey’s disclosed ~57.9x ARR multiple shows that the private market can pay extreme prices for enterprise AI, but Harvey paired that price with public ARR and much broader visible user scale than Applied Compute has. | 中 | SV011, SV012 |
| CV031 | A bullish case for Applied Compute requires believing that the current public case studies are the front edge of a repeatable high-value deployment engine rather than a small number of research-heavy flagship wins. | 中 | SV004, SV005, SV006, SV007, SV009 |
| CV032 | A bearish case starts with evidence opacity: public materials do not reveal ARR, gross margin, NRR, churn, or customer concentration, so the market may be overpaying for narrative before economics are proven. | 中 | SV014, SV015, SV016 |
| CV033 | The company’s valuation has public support as a financing fact, but not enough public support as a fair value conclusion. | 高 | SV001, SV010, SV014, SV016 |
| CV034 | Public-company comparator pages from Palantir, Salesforce, Microsoft, and Alphabet are useful mainly as disclosure and scale anchors; they are not clean like-for-like valuation comps for a 2025-founded AI specialist. | 中 | SV019, SV020, SV021, SV022 |
| CV035 | Published pricing from Microsoft, OpenAI, Together, Cerebras, and Devin suggests the market is increasingly transparent about AI-unit economics, which makes Applied Compute’s price harder to defend without private diligence access. | 中 | SV023, SV024, SV025, SV026, SV027, SV028 |
| CV036 | The most defendable recommendation from public evidence alone is not “buy” but “track / research more,” because the company quality signal is real while the price signal is under-supported. | 高 | SV001, SV014, SV015, SV016 |
| CV037 | The right valuation stance is “rich / underdetermined”: rich if Applied Compute is still materially below the ARR implied by premium private AI comps, underdetermined because that ARR is not public. | 中 | SV001, SV011, SV014 |
| CV038 | A base-case public fair-value range below the current mark is more reasonable than a point estimate at $1.3 billion unless private diligence can show stronger revenue and retention support. | 中 | SV001, SV014, SV015, SV016 |
| CV039 | Bull-case upside exists if customer specialization becomes repeatable and ARR is already well above what the public record suggests, but that is currently an evidence gap rather than a supported fact. | 中 | SV009, SV011, SV014 |
| CV040 | The key thesis-break trigger is not product failure; it is failure to translate marquee case studies into disclosed economic proof before competitors and transparent pricing make the narrative less scarce. | 中 | SV015, SV016, SV023, SV025, SV029 |
| CV041 | NVIDIA’s investor materials highlight how much AI equity value and public disclosure now sit at the infrastructure layer, which makes Applied Compute’s far smaller and less-disclosed position harder to price aggressively without private diligence. | 中 | SV031 |
| 编号 | 出版方 | 标题 | 引文 |
|---|---|---|---|
| SO001 | Applied Compute | Applied Compute | Build the AI that no one can buy |
| SO002 | Applied Compute | The Advantage You Own | Today, we're announcing $80 million in new financing at a $1.3 billion post-money valuation, led by Kleiner Perkins... This brings our total funding to $160 million. |
| SO003 | Applied Compute | It's time to get specific | Our founders all worked on different parts of this problem while they were researchers at OpenAI — Yash as a key member on the agentic software engineer effort (Codex), Rhythm as a core contributor to the first RL-trained reasoning model (o1), and Linden as a core contributor on ML systems and infrastructure for RL training. |
| SO004 | Latham & Watkins LLP | Latham & Watkins Represents Applied Compute in US$80 Million Fundraise | Latham & Watkins LLP represented Applied Compute in the funding round with an Emerging Companies & Growth team led by Bay Area partner Seth Gottlieb, with associates Kristine LaVeau, Camille N'Diaye-Muller, and Kavitha Babu. |
| SO005 | AIM Media House | How did Applied Compute raise $80M funding? | values the company at close to $700 million, according to people familiar with the deal. |
| SO006 | StartupHub.ai | Applied Compute's Agent Workforce Targets Niche AI with $80M | Inside sources close to the deal shared with StartupHub.ai that the round's valuation is close to $700 million. |
| SO007 | SiliconANGLE | Former OpenAI researchers launch Applied Compute with $80M in funding | The company didn't disclose its valuation. Last month, The Information reported that it was in the process of raising capital at a $500 million valuation. |
| SO008 | Pulse2 | Applied Compute Launches with $80 Million To Build Specific Intelligence For Enterprise AI Agents | |
| SO009 | KuCoin | Applied Compute Completes $80M Funding Round, Valued at $1.3B | bringing total raised capital to $160 million |
| SO010 | Fenado AI | Applied Compute's AI Thesis Drives Valuation to $1.3 Billion, Emphasizing Proprietary Data as New IP | securing $20 million in a seed round at a $100 million valuation in June 2025, followed by an $80 million raise that pushed its valuation to $700 million by October 2025, and subsequently to $1.3 billion by May 2026 |
| SO011 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | The result is a model, SWE-check, that enables 10x faster bug detection than the frontier alternative. |
| SO012 | Applied Compute | Automating Merchant Onboarding at DoorDash | By encoding our quality standards directly into model training, we scaled our internal expertise and raised the bar on menu accuracy on the platform. |
| SO013 | Applied Compute | Building State-of-the-Art Agents with Mercor | Applied Compute's custom-trained model, Applied Compute: Small, ranks #1 on the APEX-Agents leaderboard in corporate law, and 4th overall. |
| SO014 | Kleiner Perkins | Applied Compute: Closing the Gap Between Frontier AI and Real-World Impact | We're proud to partner with Yash, Rhythm, Linden, and the team for their Series B. |
| SO015 | Lux Capital | Applied Compute | including reinforcement learning infrastructure at OpenAI and data foundations at Scale AI, with additional experience from Together, Two Sigma, and Watershed. |
| SO016 | StartupsUnion | Applied Compute raised $80M - But Why ? | securing an initial $20 million at a $100 million valuation in June 2024. |
| SO017 | TechStartups | Applied Compute, AI startup founded by ex-OpenAI researchers, in talks to raise funds at $1.3B valuation just months after $500M round | Applied Compute was founded in May 2025 by Rhythm Garg, Linden Li, and Yash Patil, all alumni of OpenAI's technical teams. |
| SO018 | Upstarts Media | Scoop: Ex-OpenAI Staffers Raise $20M For New Startup Applied Compute | Applied Compute, the new company founded by Rhythm Garg, Linden Li and Yash Patil, has raised $20 million in a funding round led by Benchmark partner Victor Lazarte. |
| SO019 | Grokipedia | Applied Compute — Grokipedia | Applied Compute was founded in May 2025 by Rhythm Garg, Linden Li, and Yash Patil. |
| SO020 | The Information | Ex-OpenAI Trio in Funding Talks at $500 Million Valuation | |
| SO021 | Phemex | Applied Compute Raises $80M, Valuation Hits $1.3B | |
| SO022 | RocketReach | Applied Compute Information | Employees 29 (21 on RocketReach) |
| SO023 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | Applied Compute — $1.3 billion: The startup helps enterprises use their own data to train custom AI software and solutions. Founded in 2025, it last raised an $80 million round led by Kleiner Perkins... the company has raised $160 million in funding to date, according to Pitchbook. |
| SO024 | Gate.com | Applied Compute raises 80 million in funding and joins the unicorn club | |
| SO025 | Bizprofile.net (California Secretary of State registry) | Applied Compute, Inc. San Francisco, CA - filing information | Officially filed on October 10, 2025... document number B20250336266... Rhythm Garg... Chief Financial Officer; Rhythm Garg... Secretary; Yash Patil... Chief Executive Officer. |
| SO026 | Cognition | Introducing SWE-Check: 10x Faster Bug Detection | We've partnered with Applied Compute to put this to the test by collaborating to RL-train a bug detection model. |
| SO027 | Comcast NBCUniversal LIFT Labs | Applied Compute | Yash Patil, CEO, Rhythm Garg, CTO, Linden Li, Chief Architect |
| SO028 | ZenML | Doordash: Automating Merchant Onboarding with Reinforcement Learning - ZenML LLMOps Database | The case study is presented by Applied Compute, which naturally positions their tooling favorably, so readers should consider this promotional context when evaluating claims about tooling effectiveness. |
| SM001 | Applied Compute | Applied Compute | Build the AI that no one can buy |
| SM002 | Applied Compute | It's time to get specific | |
| SM003 | Applied Compute | Unlocking the AI Overhang | The overhang is the delta between a model's capability and its utility in a specific workflow, and forward deployment plays a critical role in bridging that gap. |
| SM004 | Applied Compute | The Advantage You Own | |
| SM005 | Applied Compute | Automating Merchant Onboarding at DoorDash | By encoding our quality standards directly into model training, we scaled our internal expertise and raised the bar on menu accuracy on the platform. |
| SM006 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | The result is a model, SWE-check, that enables 10x faster bug detection than the frontier alternative. |
| SM007 | Applied Compute | Building State-of-the-Art Agents with Mercor | Applied Compute's custom-trained model, Applied Compute: Small, ranks #1 on the APEX-Agents leaderboard in corporate law, and 4th overall. |
| SM008 | Applied Compute | Training a State-of-the-Art Legal Agent with Harvey | We collaborated with Harvey to post-train a frontier legal model on top of GLM-5.1. In Harvey's Legal Agent Benchmark (LAB), our trained model outperformed every available model on rubric pass rate. |
| SM009 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | Applied Compute — $1.3 billion: The startup helps enterprises use their own data to train custom AI software and solutions. |
| SM010 | SiliconANGLE | Former OpenAI researchers launch Applied Compute with $80M in funding | |
| SM011 | Gartner | Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026 | Forty percent of enterprise applications will be integrated with task-specific AI agents by the end of 2026, up from less than 5% today. |
| SM012 | Gartner | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. |
| SM013 | Gartner | Gartner Says Applying Uniform Governance Across AI Agents Will Lead to Enterprise AI Agent Failure | Gartner predicts that by 2027, 40% of enterprises will demote or decommission autonomous AI agents due to governance gaps identified only after production incidents occur. |
| SM014 | McKinsey & Company | The state of AI in 2025: Agents, innovation, and transformation | 88 percent report regular AI use in at least one business function, compared with 78 percent a year ago. |
| SM015 | Deloitte | The State of AI in the Enterprise - 2026 AI report | Only one in five companies has a mature model for governance of autonomous AI agents. |
| SM016 | Deloitte | Agentic AI enterprise adoption: Navigating key factors | |
| SM017 | Salesforce | New Agentic Enterprise Index Shows 119% Agent Growth in First Half of 2025 | Agent creation among first-mover companies surged 119% between January and June, and the average number of customer service conversations led by an agent grew 22 times in the first half of 2025. |
| SM018 | MarketsandMarkets | AI Agents Market Report 2025-2030 | The AI Agents market is projected to grow from USD 7.84 billion in 2025 to USD 52.62 billion by 2030, registering a CAGR of 46.3%. |
| SM019 | Grand View Research | AI Agents Market Size, Share And Trends Report, 2026-2033 | The global AI agents market size was valued at USD 7.6 billion in 2025 and is projected to grow from USD 10.9 billion in 2026 to USD 182.9 billion by 2033, at a CAGR of 49.6%. |
| SM020 | Precedence Research | AI Agents Market Size to Hit USD 294.66 Billion by 2035 | The global AI agents market size accounted for USD 7.92 billion in 2025 and is predicted to increase from USD 11.55 billion in 2026 to approximately USD 294.66 billion by 2035, expanding at a CAGR of 43.57%. |
| SM021 | International Energy Agency | Energy demand from AI | Our Base Case finds that global electricity consumption for data centres is projected to double to reach around 945 TWh by 2030. |
| SM022 | Goldman Sachs | How AI Is Transforming Data Centers and Ramping Up Power Demand | Global power demand from data centers, meanwhile, is forecast by Goldman Sachs Research to rise 165% by 2030 (from 2023 levels). |
| SM023 | Forbes | MIT Says 95% Of Enterprise AI Fail- Here’s What The 5% Are Doing Right | 95% of enterprise generative/agentic AI pilots fail to deliver measurable business value or significantly impact profit and loss. |
| SM024 | Amazon Web Services | AWS invests $1 billion to embed AI forward deployed engineers with customers | We are meeting that demand by creating a dedicated AWS Forward Deployed Engineering (FDE) organization. Backed by a $1 billion investment. |
| SM025 | Palantir | AI FDE - Overview - Palantir Foundry documentation | AI FDE, the AI-powered forward deployed engineer, is an interactive agent that operates Foundry for you through conversational commands. |
| SM026 | Mayfield | Insights from Mayfield’s CXO Network 2026 Survey on Agentic AI Adoption, Strategy, and Investment | Functional and Line-of-Business Leaders now have buying power. LOB leaders are now the largest decision-maker group at 46%, surpassing both CIOs (38%) and CTOs (38%). |
| SM027 | Software Strategies Blog | Roundup of agentic AI forecasts and market estimates, 2026 | Agentic AI spending is projected to reach $201.9 billion in 2026 (Gartner)... Four independent firms size the standalone market at $7-8 billion with 40%+ CAGRs... That 25x gap is not a contradiction. It is a measurement problem. |
| SP001 | Applied Compute | Applied Compute | |
| SP002 | Applied Compute | It's time to get specific | |
| SP003 | Applied Compute | Training a State-of-the-Art Legal Agent with Harvey | We collaborated with Harvey to post-train a frontier legal model on top of GLM-5.1. |
| SP004 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | |
| SP005 | Applied Compute | Automating Merchant Onboarding at DoorDash | |
| SP006 | Applied Compute | The Advantage You Own | |
| SP007 | Kleiner Perkins | Applied Compute: Closing the Gap Between Frontier AI and Real-World Impact | |
| SP008 | Harvey | Harvey | AI software for legal and professional services | |
| SP009 | Harvey | Harvey is built for high stakes: Who we work with | |
| SP010 | Harvey | Harvey Raises Growth Round at $11 Billion Valuation Co-led by GIC and Sequoia | Today we're announcing that we've raised $200M at an $11 billion valuation. |
| SP011 | Glean | AI Agents for Work: Build, Deploy & Orchestrate | Glean | |
| SP012 | Glean | Enterprise AI customer stories | Glean Work AI | |
| SP013 | Glean | Glean Raises $150M Series F at $7.2B Valuation to Accelerate Enterprise AI Agent Innovation Globally | |
| SP014 | Sierra | Product overview | Sierra | |
| SP015 | Sierra | Our customers: Sierra is trusted by industry leaders with millions of customers. | |
| SP016 | CNBC | Bret Taylor's Sierra raises nearly $1 billion months after last capital push | There's just a lot of competition. We are multiples larger than the next biggest and are trying to invest aggressively so that we can continue to expand our lead. |
| SP017 | Decagon | Decagon | The AI concierge for every customer | |
| SP018 | Decagon | Decagon raises series C at $1.5B valuation | |
| SP019 | Business Wire | Decagon's Valuation Triples to $4.5 Billion as it Ushers in the Age of AI Concierge | The round triples Decagon's valuation in just six months to $4.5 billion. |
| SP020 | Cognition | Cognition | |
| SP021 | Cognition | Devin | The AI Software Engineer | |
| SP022 | TechCrunch | Cognition AI defies turbulence with a $400M raise at $10.2B valuation | Last month, Cognition laid off 30 staffers and offered buyouts to the remaining 200 employees. |
| SP023 | OpenAI | New tools for building agents | |
| SP024 | OpenAI | The next phase of enterprise AI | |
| SP025 | OpenAI | Building agents | |
| SP026 | Salesforce | Agentforce: The AI Agent Platform | Salesforce | |
| SP027 | Salesforce | Salesforce Delivers Record Fourth Quarter Fiscal 2026 Results | |
| SP028 | Microsoft | Microsoft Copilot Studio | Create AI Agents | |
| SP029 | XtendedView | Microsoft Copilot Statistics 2026: Users, Growth, and ROI | |
| SP030 | StartupHub.ai | Applied Compute Alternatives & Competitors (2026) | |
| SP031 | eesel AI | Decagon vs Sierra: The 2026 guide to choosing your AI support agent | |
| SP032 | Gartner | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | |
| SP033 | Idlen | Devin, the AI Engineer: Review, Testing & Limitations in 2026 | Devin does not reliably identify or prevent security vulnerabilities. It may introduce SQL injection, XSS, or authentication bypass issues without awareness. |
| SP034 | TechCrunch | Almost 90 new unicorns have been minted so far this year -- here they are | |
| SI001 | Applied Compute | Applied Compute -- Homepage | |
| SI002 | Applied Compute | Applied Compute Raises $80M to Help Enterprises Advance from Generalized to Specific Intelligence | Today, we're announcing $80 million in new financing at a $1.3 billion post-money valuation, led by Kleiner Perkins... This brings our total funding to $160 million. |
| SI003 | Applied Compute | It's time to get specific | |
| SI004 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | |
| SI005 | Applied Compute | Automating Merchant Onboarding at DoorDash | |
| SI006 | Applied Compute | Building State-of-the-Art Agents with Mercor | |
| SI007 | Latham & Watkins LLP | Latham & Watkins Represents Applied Compute in US$80 Million Fundraise | |
| SI008 | Kleiner Perkins | Applied Compute: Closing the Gap Between Frontier AI and Real-World Impact | |
| SI009 | Lux Capital | Applied Compute (portfolio page) | |
| SI010 | TechStartups | Applied Compute, AI startup founded by ex-OpenAI researchers, in talks to raise funds at $1.3B valuation just months after $500M round | |
| SI011 | TechCrunch | Almost 90 new unicorns have been minted so far this year -- here they are | Applied Compute -- $1.3 billion: ... the company has raised $160 million in funding to date, according to Pitchbook. |
| SI012 | Gartner | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. |
| SI013 | International Energy Agency (IEA) | Energy demand from AI -- Energy and AI -- Analysis | |
| SI014 | Goldman Sachs | How AI Is Transforming Data Centers and Ramping Up Power Demand | |
| SI015 | Mercor | Expert data drives model performance | |
| SI016 | ZenML | Doordash: Automating Merchant Onboarding with Reinforcement Learning -- ZenML LLMOps Database | The case study doesn't discuss the total cost of development, ongoing inference costs, or quantified business impact... readers should view claims with appropriate skepticism. |
| SI017 | Microsoft | Microsoft Foundry - Pricing | Microsoft Azure | |
| SI018 | OpenAI | ChatGPT Pricing (Business) | |
| SI019 | Anthropic | Plans & Pricing | Claude by Anthropic | |
| SI020 | McKinsey & Company | Deploying agentic AI with safety and security: A playbook for technology leaders | 80 percent of organizations say they have encountered risky behaviors from AI agents, including improper data exposure and access to systems without authorization. |
| SI021 | Palantir Technologies Inc. | 2026 Q1 PLTR 10-Q | In the three months ended March 31, 2026 and 2025, our gross profit was $1.4 billion and $0.7 billion, respectively. |
| SI022 | Cognition (Devin) | Plans and Pricing | |
| SI023 | Together AI | Pricing | Together AI | |
| SI024 | U.S. Securities and Exchange Commission | EDGAR Company Search Results for "Applied Compute" | REVOKED -- Commission order revoking Exchange Act registration [Section 12(j)] ... 2006-08-28 |
| SI025 | Andreessen Horowitz (a16z) | The AI Application Spending Report: Where Startup Dollars Really Go | |
| SI026 | OpenAI | Reinforcement fine-tuning | OpenAI API | OpenAI is winding down the fine-tuning platform. The platform is no longer accessible to new users. |
| SI027 | Cerebras Systems | Inference - Cerebras | |
| SE001 | Applied Compute | Applied Compute | Applied Compute | Applied Compute is the cloud for model training, inference, and continuous improvement. |
| SE002 | Applied Compute | Blog | Applied Compute | Applied Compute | |
| SE003 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | |
| SE004 | Applied Compute | Automating Merchant Onboarding at DoorDash | |
| SE005 | Applied Compute | Building State-of-the-Art Agents with Mercor | |
| SE006 | Applied Compute | Remember, Refine, Retrieve: A Context Engine for Enterprise Agents | |
| SE007 | Applied Compute | Memory in the wild: how we use Context Engine on our own code | |
| SE008 | Applied Compute | Speeding up RL with high-leverage samples | |
| SE009 | Applied Compute | Bringing Capabilities in Distribution via Relevance-Masked Self-Distillation | |
| SE010 | Applied Compute | Training an Agentic Router for Optimal Cost-Performance on SWE Tasks | |
| SE011 | Applied Compute | Benchmarking Inference Engines on Agentic Workloads | |
| SE012 | Applied Compute | Continued Training with Entropy Preserving RL | |
| SE013 | Applied Compute | Predicting and Controlling Staleness in Fully Asynchronous RL Training | |
| SE014 | Applied Compute | Neural Cheat Sheets: Learning to Summarize with Reinforcement Learning | |
| SE015 | Modal | Scaling reinforcement learning at Applied Compute | |
| SE016 | GitHub | Applied Compute · GitHub | |
| SE017 | GitHub | GitHub - Applied-Compute/trie: Lightweight harness for replaying inference traffic against an endpoint | |
| SE018 | Cognition | Introducing SWE-Check: 10x Faster Bug Detection | |
| SE019 | Mercor | Expert data drives model performance | |
| SE020 | Anthropic | Building Effective AI Agents | |
| SE021 | Cohere | Agentic Workflows | Cohere | |
| SE022 | OpenAI | New tools for building agents | |
| SE023 | Microsoft | Microsoft Copilot Studio | Create AI Agents | |
| SE024 | DoorDash | DoorDash Introduces New AI-Powered Tools to Help Merchants Get Started Faster and Grow Across Channels | |
| SE025 | Cybersecurity and Infrastructure Security Agency (CISA) | Careful Adoption of Agentic AI Services | CISA | |
| SE026 | arXiv | Reward Hacking Benchmark: Measuring Exploits in LLM Agents with Tool Use | |
| SE027 | arXiv | Reinforcement Learning via Self-Distillation | |
| SE028 | arXiv | OpenClaw-RL: Train Any Agent Simply by Talking | |
| SE029 | arXiv | Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models | |
| SE030 | Thinking Machines Lab | On-Policy Distillation | |
| SE031 | ZenML | Automating Merchant Onboarding with Reinforcement Learning | |
| SU001 | Applied Compute | Applied Compute | Applied Compute | We are already seeing this with early customers including Cognition, DoorDash, and Mercor. |
| SU002 | Applied Compute | It's Time to Get Specific | Together we are building and validating models and agents in days instead of months, achieving state-of-the-art performance on customer evals and delivering measurable business value. |
| SU003 | Applied Compute | Automating Merchant Onboarding at DoorDash | In the test, the share of low-quality menus fell by roughly 30% relative to the baseline. |
| SU004 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | The result is a model, SWE-check, that enables 10x faster bug detection than the frontier alternative. SWE-check is in production, powering Quick Review in Windsurf. |
| SU005 | Applied Compute | Building State-of-the-Art Agents with Mercor | Applied Compute first deployed its long-horizon RL stack on Mercor’s expert-labeled dev set of 874 tasks across 50 worlds. |
| SU006 | Applied Compute | Training a State-of-the-Art Legal Agent with Harvey | We collaborated with Harvey to post-train a frontier legal model on top of GLM-5.1. |
| SU007 | DoorDash | DoorDash Introduces New AI-Powered Tools to Help Merchants Get Started Faster and Grow Across Channels | Our new AI-powered self-serve onboarding experience can help merchants launch more than 35% faster. |
| SU008 | Cognition | Introducing SWE-Check: 10x Faster Bug Detection | The result is a model, SWE-check, that enables 10x faster bug detection than the frontier alternative. SWE-check is in production, powering Quick Review in Windsurf. |
| SU009 | Cognition | Cognition | Devin is deployed at some of the largest and most complex institutions in the world. |
| SU010 | Devin | Devin Desktop | Devin | Devin Desktop is the home for coding agents to do your best work. |
| SU011 | Devin | Windsurf is now Devin Desktop | Devin | Millions of engineers use Windsurf and Devin. |
| SU012 | Mercor | Expert data drives model performance | Mercor partnered with Applied Compute to post-train an open-source model using one of our expert-labeled dev sets. |
| SU013 | Mercor | Mercor | Organizing human intelligence to power the AI economy | Mercor is organizing human intelligence to power the AI economy. |
| SU014 | Harvey | Harvey | AI software for legal and professional services | Today’s top law firms and in-house legal teams trust Harvey to elevate their craft and navigate complexity. |
| SU015 | Harvey | Harvey – Customers | More than 142,000 lawyers across 1,500+ organizations in 60 countries rely on Harvey to advance legal expertise faster. |
| SU016 | Bridge | Bridge | Stablecoin Infrastructure and APIs for Developers | Bridge handles all of the regulatory, compliance and technical complexities. |
| SU017 | CISA | Careful Adoption of Agentic AI Services | CISA | This guide outlines key security challenges and risks associated with agentic AI, and provides actionable steps for designing, deploying, and operating these systems safely. |
| SU018 | Gartner | Gartner Customer Service & Support Conference 2026 in Denver, CO | Discover practical frameworks to move AI pilots beyond experimentation and achieve sustained ROI. |
| SU019 | Anthropic | Building Effective AI Agents | The simplest solution is often the best: many workflows can be handled by compositional patterns rather than fully autonomous agents. |
| SU020 | Cohere | Agentic Workflows | Cohere | Agentic workflows automate complex business processes while preserving control, oversight, and security. |
| SU021 | OpenAI | New tools for building agents | The Responses API combines tools, orchestration, and tracing for building agents. |
| SU022 | Microsoft | Microsoft Copilot Studio | Microsoft positions Copilot Studio as a way to build, manage, and govern enterprise agents. |
| SU023 | Modal | How Applied Compute runs reinforcement learning on Modal | Applied Compute embedded deeply with customers while running large-scale RL workloads on its platform stack. |
| SU024 | GitHub | GitHub - latchbio/scbench: Benchmark for agentic single cell data analysis | scBench is a benchmark for agentic single-cell data analysis maintained by LatchBio. |
| SU025 | arXiv | scBench: Evaluating AI Agents on Single-Cell RNA-seq Analysis | scBench provides verifiable single-cell RNA-seq tasks for evaluating AI agents. |
| SU026 | Applied Compute | Blog | Applied Compute | Applied Compute | Applied Compute continued publishing customer and research updates throughout 2026. |
| SR001 | Applied Compute | Applied Compute | Applied Compute | Our researchers work with you to transform your data and expertise into frontier intelligence no one can buy. |
| SR002 | Applied Compute | Applied Compute Raises $80M to Help Enterprises Advance from Generalized to Specific Intelligence | Today, we’re announcing $80 million in new financing at a $1.3 billion post-money valuation. |
| SR003 | Applied Compute | It's time to get specific | Our founders all worked on different parts of this problem while they were researchers at OpenAI. |
| SR004 | Applied Compute | Automating Merchant Onboarding at DoorDash | Working onsite at DoorDash’s Sunnyvale office to translate production QA labels into an automated grader. |
| SR005 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | The model was trained inside a replica of the Windsurf environment. |
| SR006 | Applied Compute | Building State-of-the-Art Agents with Mercor | Applied Compute first deployed its long-horizon RL stack on Mercor’s expert-labeled dev set of 874 tasks across 50 worlds. |
| SR007 | Applied Compute | Training a State-of-the-Art Legal Agent with Harvey | LAB contains more than 1,250 tasks across 24 legal practice areas and more than 75,000 binary criteria. |
| SR008 | Modal | How Applied Compute runs reinforcement learning on Modal | Applied Compute embedded deeply with customers while evaluating sandbox and execution providers. |
| SR009 | Latham & Watkins LLP | Latham & Watkins Represents Applied Compute in US$80 Million Fundraise | |
| SR010 | Kleiner Perkins | Applied Compute: Closing the Gap Between Frontier AI and Real-World Impact | |
| SR011 | Cognition | Cognition | Devin is deployed at some of the largest and most complex institutions in the world. |
| SR012 | Harvey | Harvey – Customers | More than 142,000 lawyers across 1,500+ organizations in 60 countries rely on Harvey. |
| SR013 | CISA | Careful Adoption of Agentic AI Services | CISA | This guide outlines key security challenges and risks associated with agentic AI. |
| SR014 | NIST | AI Risk Management Framework | NIST | Trustworthy AI is valid, reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair. |
| SR015 | European Commission | AI Act - Shaping Europe’s digital future | The AI Act is the first-ever legal framework on AI, which addresses the risks of AI and positions Europe to play a leading role globally. |
| SR016 | EUR-Lex | Regulation (EU) 2024/1689 - Artificial Intelligence Act | |
| SR017 | U.S. Securities and Exchange Commission | EDGAR Company Search Results for Applied Compute | |
| SR018 | Gartner | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | Over 40% of agentic AI projects will be canceled by the end of 2027, due to escalating costs, unclear business value or inadequate risk controls. |
| SR019 | McKinsey & Company | Deploying agentic AI with safety and security: A playbook for technology leaders | 80 percent of organizations say they have encountered risky behaviors from AI agents. |
| SR020 | International Energy Agency (IEA) | Energy demand from AI -- Energy and AI -- Analysis | |
| SR021 | Goldman Sachs | How AI Is Transforming Data Centers and Ramping Up Power Demand | |
| SR022 | Amazon Web Services | Amazon Bedrock – Build genAI applications and agents at production scale – AWS | Amazon Bedrock powers generative AI for more than 100,000 organizations globally. |
| SR023 | Amazon Web Services | Amazon Bedrock AgentCore | Build, deploy, and operate AI agents securely and at scale without managing infrastructure. |
| SR024 | OpenAI | ChatGPT Enterprise | OpenAI markets ChatGPT Enterprise as secure, scalable, and enterprise-ready. |
| SR025 | Anthropic | Building Effective AI Agents | The simplest solution is often the best; many workflows do not require full agentic loops. |
| SR026 | Microsoft | Microsoft Copilot Studio | Microsoft positions Copilot Studio as a way to build, manage, and govern enterprise agents. |
| SR027 | Palantir | Artificial Intelligence Platform (AIP) | Palantir | Palantir markets AIP as a way to operationalize AI inside complex enterprises. |
| SR028 | Google Cloud | Gemini Enterprise app: Best of Google AI for Business | Gemini Enterprise app brings the best of Google AI to every employee. |
| SR029 | Salesforce | Agentforce | Salesforce | Salesforce is positioning Agentforce as a platform for enterprise digital labor. |
| SR030 | Palantir Technologies Inc. | 2026 Q1 PLTR 10-Q | In the three months ended March 31, 2026 and 2025, our gross profit was $1.4 billion and $0.7 billion, respectively. |
| SR031 | Gartner | Gartner Customer Service & Support Conference 2026 in Denver, CO | Discover practical frameworks to move AI pilots beyond experimentation and achieve sustained ROI. |
| SV001 | Applied Compute | Applied Compute Raises $80M to Help Enterprises Advance from Generalized to Specific Intelligence | Today, we’re announcing $80 million in new financing at a $1.3 billion post-money valuation. |
| SV002 | Applied Compute | Applied Compute | Applied Compute | Applied Compute is the cloud for model training, inference, and continuous improvement. |
| SV003 | Applied Compute | It's time to get specific | We are building Specific Intelligence for specific work at specific companies. |
| SV004 | Applied Compute | Unlocking Real-Time Bug Detection at Cognition | SWE-check enables 10x faster bug detection than the frontier alternative. |
| SV005 | Applied Compute | Automating Merchant Onboarding at DoorDash | DoorDash rolled out the error correction model to all menu traffic in the USA. |
| SV006 | Applied Compute | Building State-of-the-Art Agents with Mercor | Applied Compute: Small ranks #1 on the APEX-Agents leaderboard in corporate law, and 4th overall. |
| SV007 | Applied Compute | Training a State-of-the-Art Legal Agent with Harvey | Applied Compute post-trained GLM-5.1 into the strongest available model on Harvey’s Legal Agent Benchmark through full-stack optimization. |
| SV008 | Latham & Watkins LLP | Latham & Watkins Represents Applied Compute in US$80 Million Fundraise | |
| SV009 | Kleiner Perkins | Applied Compute: Closing the Gap Between Frontier AI and Real-World Impact | Closing the gap between frontier AI and real-world impact is the core thesis investors are backing. |
| SV010 | TechCrunch | Almost 90 new unicorns have been minted so far this year — here they are | TechCrunch lists Applied Compute at a $1.3 billion valuation and $160 million raised to date per PitchBook. |
| SV011 | CNBC | Legal AI startup Harvey raises $200 million at $11 billion valuation | Harvey raised $200 million at an $11 billion valuation and had about $190 million in ARR as of January 2026. |
| SV012 | Harvey | Harvey – Customers | More than 142,000 lawyers across 1,500+ organizations in 60 countries rely on Harvey. |
| SV013 | Cognition | Cognition | Devin is deployed at some of the largest and most complex institutions in the world. |
| SV014 | U.S. Securities and Exchange Commission | EDGAR Company Search Results for Applied Compute | |
| SV015 | Gartner | Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027 | Over 40% of agentic AI projects will be canceled by the end of 2027. |
| SV016 | ZenML | Doordash: Automating Merchant Onboarding with Reinforcement Learning -- ZenML LLMOps Database | The case study does not discuss the total cost of development, ongoing inference costs, or quantified business impact; readers should view claims with appropriate skepticism. |
| SV017 | Andreessen Horowitz (a16z) | The AI Application Spending Report: Where Startup Dollars Really Go | |
| SV018 | Palantir Technologies Inc. | 2026 Q1 PLTR 10-Q | In the three months ended March 31, 2026 and 2025, our gross profit was $1.4 billion and $0.7 billion, respectively. |
| SV019 | Palantir | Palantir IR | Quarterly Results SEC Filings |
| SV020 | Salesforce | Salesforce.com, Inc. - Salesforce Investor Relations | Salesforce is the #1 AI CRM, where humans with agents drive customer success together. |
| SV021 | Microsoft | Home page | Microsoft Corp (MSFT). |
| SV022 | Alphabet | Alphabet Investor Relations - Investors | Results & Financials. Earnings. SEC Filings. |
| SV023 | Microsoft Azure | Microsoft Foundry - Pricing | Microsoft Azure | Request a pricing quote. |
| SV024 | Microsoft Azure | Foundry Models Pricing | Microsoft Azure | |
| SV025 | OpenAI | OpenAI Pricing | |
| SV026 | Together AI | Pricing | Together AI | |
| SV027 | Cerebras Systems | Inference - Cerebras | |
| SV028 | Cognition (Devin) | Plans and Pricing | |
| SV029 | OpenAI | ChatGPT Enterprise | |
| SV030 | Palantir | Artificial Intelligence Platform (AIP) | Palantir | |
| SV031 | NVIDIA | NVIDIA Corporation - Home | NVIDIA is the pioneer of GPU-accelerated computing. |